public inbox for [email protected]
help / color / mirror / Atom feed[PATCH v18 08/18] tableam: finish_bulk_insert().
103+ messages / 21 participants
[nested] [flat]
* [PATCH v18 08/18] tableam: finish_bulk_insert().
@ 2019-03-08 00:31 Andres Freund <[email protected]>
0 siblings, 0 replies; 103+ messages in thread
From: Andres Freund @ 2019-03-08 00:31 UTC (permalink / raw)
Author:
Reviewed-By:
Discussion: https://postgr.es/m/
Backpatch:
---
src/backend/access/heap/heapam_handler.c | 12 ++++++++++++
src/backend/commands/copy.c | 7 +------
src/backend/commands/createas.c | 5 ++---
src/backend/commands/matview.c | 5 ++---
src/backend/commands/tablecmds.c | 4 +---
src/include/access/tableam.h | 18 ++++++++++++++++++
6 files changed, 36 insertions(+), 15 deletions(-)
diff --git a/src/backend/access/heap/heapam_handler.c b/src/backend/access/heap/heapam_handler.c
index ea8e3ee9ce5..3098cb96b60 100644
--- a/src/backend/access/heap/heapam_handler.c
+++ b/src/backend/access/heap/heapam_handler.c
@@ -541,6 +541,17 @@ retry:
return result;
}
+static void
+heapam_finish_bulk_insert(Relation relation, int options)
+{
+ /*
+ * If we skipped writing WAL, then we need to sync the heap (but not
+ * indexes since those use WAL anyway)
+ */
+ if (options & HEAP_INSERT_SKIP_WAL)
+ heap_sync(relation);
+}
+
/* ------------------------------------------------------------------------
* Definition of the heap table access method.
@@ -573,6 +584,7 @@ static const TableAmRoutine heapam_methods = {
.tuple_update = heapam_heap_update,
.multi_insert = heap_multi_insert,
.tuple_lock = heapam_lock_tuple,
+ .finish_bulk_insert = heapam_finish_bulk_insert,
.tuple_fetch_row_version = heapam_fetch_row_version,
.tuple_get_latest_tid = heap_get_latest_tid,
diff --git a/src/backend/commands/copy.c b/src/backend/commands/copy.c
index 312fd3bed31..1e7a06a72fb 100644
--- a/src/backend/commands/copy.c
+++ b/src/backend/commands/copy.c
@@ -3098,12 +3098,7 @@ CopyFrom(CopyState cstate)
FreeExecutorState(estate);
- /*
- * If we skipped writing WAL, then we need to sync the heap (but not
- * indexes since those use WAL anyway)
- */
- if (hi_options & HEAP_INSERT_SKIP_WAL)
- heap_sync(cstate->rel);
+ table_finish_bulk_insert(cstate->rel, hi_options);
return processed;
}
diff --git a/src/backend/commands/createas.c b/src/backend/commands/createas.c
index 0ac295cea3f..55f61854614 100644
--- a/src/backend/commands/createas.c
+++ b/src/backend/commands/createas.c
@@ -28,6 +28,7 @@
#include "access/reloptions.h"
#include "access/htup_details.h"
#include "access/sysattr.h"
+#include "access/tableam.h"
#include "access/xact.h"
#include "access/xlog.h"
#include "catalog/namespace.h"
@@ -601,9 +602,7 @@ intorel_shutdown(DestReceiver *self)
FreeBulkInsertState(myState->bistate);
- /* If we skipped using WAL, must heap_sync before commit */
- if (myState->hi_options & HEAP_INSERT_SKIP_WAL)
- heap_sync(myState->rel);
+ table_finish_bulk_insert(myState->rel, myState->hi_options);
/* close rel, but keep lock until commit */
table_close(myState->rel, NoLock);
diff --git a/src/backend/commands/matview.c b/src/backend/commands/matview.c
index 5a47be4b33c..62b76cfd358 100644
--- a/src/backend/commands/matview.c
+++ b/src/backend/commands/matview.c
@@ -18,6 +18,7 @@
#include "access/heapam.h"
#include "access/htup_details.h"
#include "access/multixact.h"
+#include "access/tableam.h"
#include "access/xact.h"
#include "access/xlog.h"
#include "catalog/catalog.h"
@@ -509,9 +510,7 @@ transientrel_shutdown(DestReceiver *self)
FreeBulkInsertState(myState->bistate);
- /* If we skipped using WAL, must heap_sync before commit */
- if (myState->hi_options & HEAP_INSERT_SKIP_WAL)
- heap_sync(myState->transientrel);
+ table_finish_bulk_insert(myState->transientrel, myState->hi_options);
/* close transientrel, but keep lock until commit */
table_close(myState->transientrel, NoLock);
diff --git a/src/backend/commands/tablecmds.c b/src/backend/commands/tablecmds.c
index 40839e14dbe..1f5a7e93155 100644
--- a/src/backend/commands/tablecmds.c
+++ b/src/backend/commands/tablecmds.c
@@ -4951,9 +4951,7 @@ ATRewriteTable(AlteredTableInfo *tab, Oid OIDNewHeap, LOCKMODE lockmode)
{
FreeBulkInsertState(bistate);
- /* If we skipped writing WAL, then we need to sync the heap. */
- if (hi_options & HEAP_INSERT_SKIP_WAL)
- heap_sync(newrel);
+ table_finish_bulk_insert(newrel, hi_options);
table_close(newrel, NoLock);
}
diff --git a/src/include/access/tableam.h b/src/include/access/tableam.h
index 7213d425e12..2c2d388dda6 100644
--- a/src/include/access/tableam.h
+++ b/src/include/access/tableam.h
@@ -284,6 +284,16 @@ typedef struct TableAmRoutine
uint8 flags,
HeapUpdateFailureData *hufd);
+ /*
+ * Perform operations necessary to complete insertions made via
+ * tuple_insert and multi_insert with a BulkInsertState specified. This
+ * e.g. may e.g. used to flush the relation when inserting with skipping
+ * WAL.
+ *
+ * May be NULL.
+ */
+ void (*finish_bulk_insert) (Relation rel, int options);
+
} TableAmRoutine;
@@ -687,6 +697,14 @@ table_lock_tuple(Relation rel, ItemPointer tid, Snapshot snapshot,
flags, hufd);
}
+static inline void
+table_finish_bulk_insert(Relation rel, int options)
+{
+ /* optional */
+ if (rel->rd_tableam && rel->rd_tableam->finish_bulk_insert)
+ rel->rd_tableam->finish_bulk_insert(rel, options);
+}
+
/* ----------------------------------------------------------------------------
* Functions to make modifications a bit simpler.
--
2.21.0.dirty
--yvn3crbc4qf4vymf
Content-Type: text/x-diff; charset=us-ascii
Content-Disposition: attachment;
filename="v18-0009-tableam-slotify-CREATE-TABLE-AS-and-CREATE-MATER.patch"
^ permalink raw reply [nested|flat] 103+ messages in thread
* MultiXact\SLRU buffers configuration
@ 2020-05-08 16:36 Andrey M. Borodin <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Andrey M. Borodin @ 2020-05-08 16:36 UTC (permalink / raw)
To: pgsql-hackers
Hi, hackers!
*** The problem ***
I'm investigating some cases of reduced database performance due to MultiXactOffsetLock contention (80% MultiXactOffsetLock, 20% IO DataFileRead).
The problem manifested itself during index repack and constraint validation. Both being effectively full table scans.
The database workload contains a lot of select for share\select for update queries. I've tried to construct synthetic world generator and could not achieve similar lock configuration: I see a lot of different locks in wait events, particularly a lot more MultiXactMemberLocks. But from my experiments with synthetic workload, contention of MultiXactOffsetLock can be reduced by increasing NUM_MXACTOFFSET_BUFFERS=8 to bigger numbers.
*** Question 1 ***
Is it safe to increase number of buffers of MultiXact\All SLRUs, recompile and run database as usual?
I cannot experiment much with production. But I'm mostly sure that bigger buffers will solve the problem.
*** Question 2 ***
Probably, we could do GUCs for SLRU sizes? Are there any reasons not to do them configurable? I think multis, clog, subtransactions and others will benefit from bigger buffer. But, probably, too much of knobs can be confusing.
*** Question 3 ***
MultiXact offset lock is always taken as exclusive lock. It turns MultiXact Offset subsystem to single threaded. If someone have good idea how to make it more concurrency-friendly, I'm willing to put some efforts into this.
Probably, I could just add LWlocks for each offset buffer page. Is it something worth doing? Or are there any hidden cavers and difficulties?
Thanks!
Best regards, Andrey Borodin.
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2020-05-11 11:17 Andrey M. Borodin <[email protected]>
parent: Andrey M. Borodin <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Andrey M. Borodin @ 2020-05-11 11:17 UTC (permalink / raw)
To: pgsql-hackers
> 8 мая 2020 г., в 21:36, Andrey M. Borodin <[email protected]> написал(а):
>
> *** The problem ***
> I'm investigating some cases of reduced database performance due to MultiXactOffsetLock contention (80% MultiXactOffsetLock, 20% IO DataFileRead).
> The problem manifested itself during index repack and constraint validation. Both being effectively full table scans.
> The database workload contains a lot of select for share\select for update queries. I've tried to construct synthetic world generator and could not achieve similar lock configuration: I see a lot of different locks in wait events, particularly a lot more MultiXactMemberLocks. But from my experiments with synthetic workload, contention of MultiXactOffsetLock can be reduced by increasing NUM_MXACTOFFSET_BUFFERS=8 to bigger numbers.
>
> *** Question 1 ***
> Is it safe to increase number of buffers of MultiXact\All SLRUs, recompile and run database as usual?
> I cannot experiment much with production. But I'm mostly sure that bigger buffers will solve the problem.
>
> *** Question 2 ***
> Probably, we could do GUCs for SLRU sizes? Are there any reasons not to do them configurable? I think multis, clog, subtransactions and others will benefit from bigger buffer. But, probably, too much of knobs can be confusing.
>
> *** Question 3 ***
> MultiXact offset lock is always taken as exclusive lock. It turns MultiXact Offset subsystem to single threaded. If someone have good idea how to make it more concurrency-friendly, I'm willing to put some efforts into this.
> Probably, I could just add LWlocks for each offset buffer page. Is it something worth doing? Or are there any hidden cavers and difficulties?
I've created benchmark[0] imitating MultiXact pressure on my laptop: 7 clients are concurrently running select "select * from table where primary_key = ANY ($1) for share" where $1 is array of identifiers so that each tuple in a table is locked by different set of XIDs. During this benchmark I observe contention of MultiXactControlLock in pg_stat_activity
пятница, 8 мая 2020 г. 15:08:37 (every 1s)
pid | wait_event | wait_event_type | state | query
-------+----------------------------+-----------------+--------+----------------------------------------------------
41344 | ClientRead | Client | idle | insert into t1 select generate_series(1,1000000,1)
41375 | MultiXactOffsetControlLock | LWLock | active | select * from t1 where i = ANY ($1) for share
41377 | MultiXactOffsetControlLock | LWLock | active | select * from t1 where i = ANY ($1) for share
41378 | | | active | select * from t1 where i = ANY ($1) for share
41379 | MultiXactOffsetControlLock | LWLock | active | select * from t1 where i = ANY ($1) for share
41381 | | | active | select * from t1 where i = ANY ($1) for share
41383 | MultiXactOffsetControlLock | LWLock | active | select * from t1 where i = ANY ($1) for share
41385 | MultiXactOffsetControlLock | LWLock | active | select * from t1 where i = ANY ($1) for share
(8 rows)
Finally, the benchmark is measuring time to execute select for update 42 times.
I've went ahead and created 3 patches:
1. Configurable SLRU buffer sizes for MultiXacOffsets and MultiXactMembers
2. Reduce locking level to shared on read of MultiXactId members
3. Configurable cache size
I've found out that:
1. When MultiXact working set does not fit into buffers - benchmark results grow very high. Yet, very big buffers slow down benchmark too. For this benchmark optimal SLRU size id 32 pages for offsets and 64 pages for members (defaults are 8 and 16 respectively).
2. Lock optimisation increases performance by 5% on default SLRU sizes. Actually, benchmark does not explicitly read MultiXactId members, but when it replaces one with another - it have to read previous set. I understand that we can construct benchmark to demonstrate dominance of any algorithm and 5% of synthetic workload is not a very big number. But it just make sense to try to take shared lock for reading.
3. Manipulations with cache size do not affect benchmark anyhow. It's somewhat expected: benchmark is designed to defeat cache, either way OffsetControlLock would not be stressed.
For our workload, I think we will just increase numbers of SLRU sizes. But patchset may be useful for tuning and as a performance optimisation of MultiXact.
Also MultiXacts seems to be not very good fit into SLRU design. I think it would be better to use B-tree as a container. Or at least make MultiXact members extendable in-place (reserve some size when multixact is created).
When we want to extend number of locks for a tuple currently we will:
1. Iterate through all SLRU buffers for offsets to read current offset (with exclusive lock for offsets)
2. Iterate through all buffers for members to find current members (with exclusive lock for members)
3. Create new members array with +1 xid
4. Iterate through all cache members to find out maybe there are any such cache item as what we are going to create
5. iterate over 1 again for write
6. Iterate over 2 again for write
Obviously this does not scale well - we cannot increase SLRU sizes for too long.
Thanks! I'd be happy to hear any feedback.
Best regards, Andrey Borodin.
[0] https://github.com/x4m/multixact_stress
Attachments:
[application/octet-stream] v1-0001-Add-GUCs-to-tune-MultiXact-SLRUs.patch (6.2K, ../../[email protected]/2-v1-0001-Add-GUCs-to-tune-MultiXact-SLRUs.patch)
download | inline diff:
From 17ba50c29aafd2de81462e0aca70c8629076e52c Mon Sep 17 00:00:00 2001
From: Andrey Borodin <[email protected]>
Date: Sat, 9 May 2020 16:42:07 +0500
Subject: [PATCH v1 1/3] Add GUCs to tune MultiXact SLRUs
---
doc/src/sgml/config.sgml | 31 ++++++++++++++++++++++++++
src/backend/access/transam/multixact.c | 8 +++----
src/backend/utils/init/globals.c | 3 +++
src/backend/utils/misc/guc.c | 22 ++++++++++++++++++
src/include/access/multixact.h | 4 ----
src/include/miscadmin.h | 2 ++
6 files changed, 62 insertions(+), 8 deletions(-)
diff --git a/doc/src/sgml/config.sgml b/doc/src/sgml/config.sgml
index 3aea1763b4..71aa5c4900 100644
--- a/doc/src/sgml/config.sgml
+++ b/doc/src/sgml/config.sgml
@@ -1746,6 +1746,37 @@ include_dir 'conf.d'
</listitem>
</varlistentry>
+ <varlistentry id="guc-multixact-offsets-slru-buffers" xreflabel="multixact_offsets_slru_buffers">
+ <term><varname>multixact_offsets_slru_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>multixact_offsets_slru_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of memory to be used for MultiXact offsets. MultiXact offsets
+ are used to store informaion about offsets of multiple row lockers (caused by SELECT FOR UPDATE and others).
+ It defaults to 64 kilobytes (<literal>64KB</literal>).
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-multixact-members-slru-buffers" xreflabel="multixact_members_slru_buffers">
+ <term><varname>multixact_members_slru_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>multixact_members_slru_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of memory to be used for MultiXact members. MultiXact members
+ are used to store informaion about XIDs of multiple row lockers. Tipically <varname>multixact_members_slru_buffers</varname>
+ is twice more than <varname>multixact_offsets_slru_buffers</varname>.
+ It defaults to 128 kilobytes (<literal>128KB</literal>).
+ </para>
+ </listitem>
+ </varlistentry>
+
<varlistentry id="guc-max-stack-depth" xreflabel="max_stack_depth">
<term><varname>max_stack_depth</varname> (<type>integer</type>)
<indexterm>
diff --git a/src/backend/access/transam/multixact.c b/src/backend/access/transam/multixact.c
index e2aa5c9ce4..d2ac029b98 100644
--- a/src/backend/access/transam/multixact.c
+++ b/src/backend/access/transam/multixact.c
@@ -1812,8 +1812,8 @@ MultiXactShmemSize(void)
mul_size(sizeof(MultiXactId) * 2, MaxOldestSlot))
size = SHARED_MULTIXACT_STATE_SIZE;
- size = add_size(size, SimpleLruShmemSize(NUM_MXACTOFFSET_BUFFERS, 0));
- size = add_size(size, SimpleLruShmemSize(NUM_MXACTMEMBER_BUFFERS, 0));
+ size = add_size(size, SimpleLruShmemSize(multixact_offsets_slru_buffers, 0));
+ size = add_size(size, SimpleLruShmemSize(multixact_members_slru_buffers, 0));
return size;
}
@@ -1829,11 +1829,11 @@ MultiXactShmemInit(void)
MultiXactMemberCtl->PagePrecedes = MultiXactMemberPagePrecedes;
SimpleLruInit(MultiXactOffsetCtl,
- "multixact_offset", NUM_MXACTOFFSET_BUFFERS, 0,
+ "multixact_offset", multixact_offsets_slru_buffers, 0,
MultiXactOffsetControlLock, "pg_multixact/offsets",
LWTRANCHE_MXACTOFFSET_BUFFERS);
SimpleLruInit(MultiXactMemberCtl,
- "multixact_member", NUM_MXACTMEMBER_BUFFERS, 0,
+ "multixact_member", multixact_members_slru_buffers, 0,
MultiXactMemberControlLock, "pg_multixact/members",
LWTRANCHE_MXACTMEMBER_BUFFERS);
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
index eb19644419..aebfbdb67c 100644
--- a/src/backend/utils/init/globals.c
+++ b/src/backend/utils/init/globals.c
@@ -148,3 +148,6 @@ int VacuumCostBalance = 0; /* working state for vacuum */
bool VacuumCostActive = false;
double vacuum_cleanup_index_scale_factor;
+
+int multixact_offsets_slru_buffers = 8;
+int multixact_members_slru_buffers = 16;
diff --git a/src/backend/utils/misc/guc.c b/src/backend/utils/misc/guc.c
index 5bdc02fce2..b139e387f3 100644
--- a/src/backend/utils/misc/guc.c
+++ b/src/backend/utils/misc/guc.c
@@ -2270,6 +2270,28 @@ static struct config_int ConfigureNamesInt[] =
NULL, NULL, NULL
},
+ {
+ {"multixact_offsets_slru_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for MultiXact offsets SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &multixact_offsets_slru_buffers,
+ 8, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"multixact_members_slru_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for MultiXact members SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &multixact_members_slru_buffers,
+ 16, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
{
{"temp_buffers", PGC_USERSET, RESOURCES_MEM,
gettext_noop("Sets the maximum number of temporary buffers used by each session."),
diff --git a/src/include/access/multixact.h b/src/include/access/multixact.h
index af4aac08bd..848398bf1e 100644
--- a/src/include/access/multixact.h
+++ b/src/include/access/multixact.h
@@ -28,10 +28,6 @@
#define MaxMultiXactOffset ((MultiXactOffset) 0xFFFFFFFF)
-/* Number of SLRU buffers to use for multixact */
-#define NUM_MXACTOFFSET_BUFFERS 8
-#define NUM_MXACTMEMBER_BUFFERS 16
-
/*
* Possible multixact lock modes ("status"). The first four modes are for
* tuple locks (FOR KEY SHARE, FOR SHARE, FOR NO KEY UPDATE, FOR UPDATE); the
diff --git a/src/include/miscadmin.h b/src/include/miscadmin.h
index 14fa127ab1..95f4633a88 100644
--- a/src/include/miscadmin.h
+++ b/src/include/miscadmin.h
@@ -161,6 +161,8 @@ extern PGDLLIMPORT int MaxBackends;
extern PGDLLIMPORT int MaxConnections;
extern PGDLLIMPORT int max_worker_processes;
extern PGDLLIMPORT int max_parallel_workers;
+extern PGDLLIMPORT int multixact_offsets_slru_buffers;
+extern PGDLLIMPORT int multixact_members_slru_buffers;
extern PGDLLIMPORT int MyProcPid;
extern PGDLLIMPORT pg_time_t MyStartTime;
--
2.24.2 (Apple Git-127)
[application/octet-stream] v1-0002-Use-shared-lock-in-GetMultiXactIdMembers-for-offs.patch (7.6K, ../../[email protected]/3-v1-0002-Use-shared-lock-in-GetMultiXactIdMembers-for-offs.patch)
download | inline diff:
From 0a0ebc6391c33ca77260656d886224c71fb8ad96 Mon Sep 17 00:00:00 2001
From: Andrey Borodin <[email protected]>
Date: Sat, 9 May 2020 21:19:18 +0500
Subject: [PATCH v1 2/3] Use shared lock in GetMultiXactIdMembers for offsets
and members
Previously read of multixact required exclusive control locks in
offstes and members SLRUs. This could lead to contention.
In this commit we take advantge of SimpleLruReadPage_ReadOnly
similar to CLOG usage.
---
src/backend/access/transam/clog.c | 2 +-
src/backend/access/transam/commit_ts.c | 2 +-
src/backend/access/transam/multixact.c | 12 ++++++------
src/backend/access/transam/slru.c | 8 +++++---
src/backend/access/transam/subtrans.c | 2 +-
src/backend/commands/async.c | 2 +-
src/backend/storage/lmgr/predicate.c | 2 +-
src/include/access/slru.h | 2 +-
8 files changed, 17 insertions(+), 15 deletions(-)
diff --git a/src/backend/access/transam/clog.c b/src/backend/access/transam/clog.c
index f8e7670f8d..2ae024608e 100644
--- a/src/backend/access/transam/clog.c
+++ b/src/backend/access/transam/clog.c
@@ -643,7 +643,7 @@ TransactionIdGetStatus(TransactionId xid, XLogRecPtr *lsn)
/* lock is acquired by SimpleLruReadPage_ReadOnly */
- slotno = SimpleLruReadPage_ReadOnly(ClogCtl, pageno, xid);
+ slotno = SimpleLruReadPage_ReadOnly(ClogCtl, pageno, xid, false);
byteptr = ClogCtl->shared->page_buffer[slotno] + byteno;
status = (*byteptr >> bshift) & CLOG_XACT_BITMASK;
diff --git a/src/backend/access/transam/commit_ts.c b/src/backend/access/transam/commit_ts.c
index 630df672cc..1efa269339 100644
--- a/src/backend/access/transam/commit_ts.c
+++ b/src/backend/access/transam/commit_ts.c
@@ -342,7 +342,7 @@ TransactionIdGetCommitTsData(TransactionId xid, TimestampTz *ts,
}
/* lock is acquired by SimpleLruReadPage_ReadOnly */
- slotno = SimpleLruReadPage_ReadOnly(CommitTsCtl, pageno, xid);
+ slotno = SimpleLruReadPage_ReadOnly(CommitTsCtl, pageno, xid, false);
memcpy(&entry,
CommitTsCtl->shared->page_buffer[slotno] +
SizeOfCommitTimestampEntry * entryno,
diff --git a/src/backend/access/transam/multixact.c b/src/backend/access/transam/multixact.c
index d2ac029b98..5387a3602d 100644
--- a/src/backend/access/transam/multixact.c
+++ b/src/backend/access/transam/multixact.c
@@ -1321,12 +1321,12 @@ GetMultiXactIdMembers(MultiXactId multi, MultiXactMember **members,
* time on every multixact creation.
*/
retry:
- LWLockAcquire(MultiXactOffsetControlLock, LW_EXCLUSIVE);
pageno = MultiXactIdToOffsetPage(multi);
entryno = MultiXactIdToOffsetEntry(multi);
- slotno = SimpleLruReadPage(MultiXactOffsetCtl, pageno, true, multi);
+ /* lock is acquired by SimpleLruReadPage_ReadOnly */
+ slotno = SimpleLruReadPage_ReadOnly(MultiXactOffsetCtl, pageno, multi, false);
offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
offptr += entryno;
offset = *offptr;
@@ -1358,7 +1358,7 @@ retry:
entryno = MultiXactIdToOffsetEntry(tmpMXact);
if (pageno != prev_pageno)
- slotno = SimpleLruReadPage(MultiXactOffsetCtl, pageno, true, tmpMXact);
+ slotno = SimpleLruReadPage_ReadOnly(MultiXactOffsetCtl, pageno, tmpMXact, true);
offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
offptr += entryno;
@@ -1382,7 +1382,7 @@ retry:
*members = ptr;
/* Now get the members themselves. */
- LWLockAcquire(MultiXactMemberControlLock, LW_EXCLUSIVE);
+ LWLockAcquire(MultiXactMemberControlLock, LW_SHARED);
truelength = 0;
prev_pageno = -1;
@@ -1399,7 +1399,7 @@ retry:
if (pageno != prev_pageno)
{
- slotno = SimpleLruReadPage(MultiXactMemberCtl, pageno, true, multi);
+ slotno = SimpleLruReadPage_ReadOnly(MultiXactMemberCtl, pageno, multi, true);
prev_pageno = pageno;
}
@@ -2745,7 +2745,7 @@ find_multixact_start(MultiXactId multi, MultiXactOffset *result)
return false;
/* lock is acquired by SimpleLruReadPage_ReadOnly */
- slotno = SimpleLruReadPage_ReadOnly(MultiXactOffsetCtl, pageno, multi);
+ slotno = SimpleLruReadPage_ReadOnly(MultiXactOffsetCtl, pageno, multi, false);
offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
offptr += entryno;
offset = *offptr;
diff --git a/src/backend/access/transam/slru.c b/src/backend/access/transam/slru.c
index b2316af779..ae63ad4e6c 100644
--- a/src/backend/access/transam/slru.c
+++ b/src/backend/access/transam/slru.c
@@ -470,17 +470,19 @@ SimpleLruReadPage(SlruCtl ctl, int pageno, bool write_ok,
* Return value is the shared-buffer slot number now holding the page.
* The buffer's LRU access info is updated.
*
- * Control lock must NOT be held at entry, but will be held at exit.
+ * Control lock will be held at exit. If lock_held is true at least
+ * shared control lock must be held.
* It is unspecified whether the lock will be shared or exclusive.
*/
int
-SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno, TransactionId xid)
+SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno, TransactionId xid, bool lock_held)
{
SlruShared shared = ctl->shared;
int slotno;
/* Try to find the page while holding only shared lock */
- LWLockAcquire(shared->ControlLock, LW_SHARED);
+ if (!lock_held)
+ LWLockAcquire(shared->ControlLock, LW_SHARED);
/* See if page is already in a buffer */
for (slotno = 0; slotno < shared->num_slots; slotno++)
diff --git a/src/backend/access/transam/subtrans.c b/src/backend/access/transam/subtrans.c
index 25d7d739cf..98700f26a7 100644
--- a/src/backend/access/transam/subtrans.c
+++ b/src/backend/access/transam/subtrans.c
@@ -123,7 +123,7 @@ SubTransGetParent(TransactionId xid)
/* lock is acquired by SimpleLruReadPage_ReadOnly */
- slotno = SimpleLruReadPage_ReadOnly(SubTransCtl, pageno, xid);
+ slotno = SimpleLruReadPage_ReadOnly(SubTransCtl, pageno, xid, false);
ptr = (TransactionId *) SubTransCtl->shared->page_buffer[slotno];
ptr += entryno;
diff --git a/src/backend/commands/async.c b/src/backend/commands/async.c
index 0c9d20ebfc..a12447ecb5 100644
--- a/src/backend/commands/async.c
+++ b/src/backend/commands/async.c
@@ -2005,7 +2005,7 @@ asyncQueueReadAllNotifications(void)
* of the page we will actually inspect.
*/
slotno = SimpleLruReadPage_ReadOnly(AsyncCtl, curpage,
- InvalidTransactionId);
+ InvalidTransactionId, false);
if (curpage == QUEUE_POS_PAGE(head))
{
/* we only want to read as far as head */
diff --git a/src/backend/storage/lmgr/predicate.c b/src/backend/storage/lmgr/predicate.c
index 654584b77a..2969992a90 100644
--- a/src/backend/storage/lmgr/predicate.c
+++ b/src/backend/storage/lmgr/predicate.c
@@ -948,7 +948,7 @@ OldSerXidGetMinConflictCommitSeqNo(TransactionId xid)
* but will return with that lock held, which must then be released.
*/
slotno = SimpleLruReadPage_ReadOnly(OldSerXidSlruCtl,
- OldSerXidPage(xid), xid);
+ OldSerXidPage(xid), xid, false);
val = OldSerXidValue(slotno, xid);
LWLockRelease(OldSerXidLock);
return val;
diff --git a/src/include/access/slru.h b/src/include/access/slru.h
index 00dbd803e1..916dd8203e 100644
--- a/src/include/access/slru.h
+++ b/src/include/access/slru.h
@@ -144,7 +144,7 @@ extern int SimpleLruZeroPage(SlruCtl ctl, int pageno);
extern int SimpleLruReadPage(SlruCtl ctl, int pageno, bool write_ok,
TransactionId xid);
extern int SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno,
- TransactionId xid);
+ TransactionId xid, bool lock_held);
extern void SimpleLruWritePage(SlruCtl ctl, int slotno);
extern void SimpleLruFlush(SlruCtl ctl, bool allow_redirtied);
extern void SimpleLruTruncate(SlruCtl ctl, int cutoffPage);
--
2.24.2 (Apple Git-127)
[application/octet-stream] v1-0003-Make-MultiXact-local-cache-size-configurable.patch (3.7K, ../../[email protected]/4-v1-0003-Make-MultiXact-local-cache-size-configurable.patch)
download | inline diff:
From 944551228e4c2b4ec015359821122d9676fcf7e0 Mon Sep 17 00:00:00 2001
From: Andrey Borodin <[email protected]>
Date: Mon, 11 May 2020 16:17:02 +0500
Subject: [PATCH v1 3/3] Make MultiXact local cache size configurable
---
doc/src/sgml/config.sgml | 16 ++++++++++++++++
src/backend/access/transam/multixact.c | 2 +-
src/backend/utils/init/globals.c | 1 +
src/backend/utils/misc/guc.c | 10 ++++++++++
src/include/miscadmin.h | 1 +
5 files changed, 29 insertions(+), 1 deletion(-)
diff --git a/doc/src/sgml/config.sgml b/doc/src/sgml/config.sgml
index 71aa5c4900..6dbc3ce5f1 100644
--- a/doc/src/sgml/config.sgml
+++ b/doc/src/sgml/config.sgml
@@ -1777,6 +1777,22 @@ include_dir 'conf.d'
</listitem>
</varlistentry>
+ <varlistentry id="guc-multixact-local-cache-entries" xreflabel="multixact_local-cache-entries">
+ <term><varname>multixact_local_cache_entries</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>multixact_local_cache_entries</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the number cached MultiXact by backend. Any SLRU lookup is preceeded
+ by cache lookup. Higher numbers of cache size help to deduplicate lock sets, while
+ lower values make cache lookup faster.
+ It defaults to 256.
+ </para>
+ </listitem>
+ </varlistentry>
+
<varlistentry id="guc-max-stack-depth" xreflabel="max_stack_depth">
<term><varname>max_stack_depth</varname> (<type>integer</type>)
<indexterm>
diff --git a/src/backend/access/transam/multixact.c b/src/backend/access/transam/multixact.c
index 5387a3602d..62f9bf73eb 100644
--- a/src/backend/access/transam/multixact.c
+++ b/src/backend/access/transam/multixact.c
@@ -1589,7 +1589,7 @@ mXactCachePut(MultiXactId multi, int nmembers, MultiXactMember *members)
qsort(entry->members, nmembers, sizeof(MultiXactMember), mxactMemberComparator);
dlist_push_head(&MXactCache, &entry->node);
- if (MXactCacheMembers++ >= MAX_CACHE_ENTRIES)
+ if (MXactCacheMembers++ >= multixact_local_cache_entries)
{
dlist_node *node;
mXactCacheEnt *entry;
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
index aebfbdb67c..b17382293c 100644
--- a/src/backend/utils/init/globals.c
+++ b/src/backend/utils/init/globals.c
@@ -151,3 +151,4 @@ double vacuum_cleanup_index_scale_factor;
int multixact_offsets_slru_buffers = 8;
int multixact_members_slru_buffers = 16;
+int multixact_local_cache_entries = 256;
diff --git a/src/backend/utils/misc/guc.c b/src/backend/utils/misc/guc.c
index b139e387f3..a74f8a8f20 100644
--- a/src/backend/utils/misc/guc.c
+++ b/src/backend/utils/misc/guc.c
@@ -2292,6 +2292,16 @@ static struct config_int ConfigureNamesInt[] =
NULL, NULL, NULL
},
+ {
+ {"multixact_local_cache_entries", PGC_SUSET, RESOURCES_MEM,
+ gettext_noop("Sets the number cached MultiXact by backend."),
+ NULL
+ },
+ &multixact_local_cache_entries,
+ 256, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
{
{"temp_buffers", PGC_USERSET, RESOURCES_MEM,
gettext_noop("Sets the maximum number of temporary buffers used by each session."),
diff --git a/src/include/miscadmin.h b/src/include/miscadmin.h
index 95f4633a88..27d5170ff1 100644
--- a/src/include/miscadmin.h
+++ b/src/include/miscadmin.h
@@ -163,6 +163,7 @@ extern PGDLLIMPORT int max_worker_processes;
extern PGDLLIMPORT int max_parallel_workers;
extern PGDLLIMPORT int multixact_offsets_slru_buffers;
extern PGDLLIMPORT int multixact_members_slru_buffers;
+extern PGDLLIMPORT int multixact_local_cache_entries;
extern PGDLLIMPORT int MyProcPid;
extern PGDLLIMPORT pg_time_t MyStartTime;
--
2.24.2 (Apple Git-127)
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2020-05-13 18:08 Andrey M. Borodin <[email protected]>
parent: Andrey M. Borodin <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Andrey M. Borodin @ 2020-05-13 18:08 UTC (permalink / raw)
To: pgsql-hackers
> 11 мая 2020 г., в 16:17, Andrey M. Borodin <[email protected]> написал(а):
>
> I've went ahead and created 3 patches:
> 1. Configurable SLRU buffer sizes for MultiXacOffsets and MultiXactMembers
> 2. Reduce locking level to shared on read of MultiXactId members
> 3. Configurable cache size
I'm looking more at MultiXact and it seems to me that we have a race condition there.
When we create a new MultiXact we do:
1. Generate new MultiXactId under MultiXactGenLock
2. Record new mxid with members and offset to WAL
3. Write offset to SLRU under MultiXactOffsetControlLock
4. Write members to SLRU under MultiXactMemberControlLock
When we read MultiXact we do:
1. Retrieve offset by mxid from SLRU under MultiXactOffsetControlLock
2. If offset is 0 - it's not filled in at step 4 of previous algorithm, we sleep and goto 1
3. Retrieve members from SLRU under MultiXactMemberControlLock
4. ..... what we do if there are just zeroes because step 4 is not executed yet? Nothing, return empty members list.
What am I missing?
Best regards, Andrey Borodin.
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2020-05-14 01:25 Kyotaro Horiguchi <[email protected]>
parent: Andrey M. Borodin <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Kyotaro Horiguchi @ 2020-05-14 01:25 UTC (permalink / raw)
To: [email protected]; +Cc: pgsql-hackers
At Wed, 13 May 2020 23:08:37 +0500, "Andrey M. Borodin" <[email protected]> wrote in
>
>
> > 11 мая 2020 г., в 16:17, Andrey M. Borodin <[email protected]> написал(а):
> >
> > I've went ahead and created 3 patches:
> > 1. Configurable SLRU buffer sizes for MultiXacOffsets and MultiXactMembers
> > 2. Reduce locking level to shared on read of MultiXactId members
> > 3. Configurable cache size
>
> I'm looking more at MultiXact and it seems to me that we have a race condition there.
>
> When we create a new MultiXact we do:
> 1. Generate new MultiXactId under MultiXactGenLock
> 2. Record new mxid with members and offset to WAL
> 3. Write offset to SLRU under MultiXactOffsetControlLock
> 4. Write members to SLRU under MultiXactMemberControlLock
But, don't we hold exclusive lock on the buffer through all the steps
above?
> When we read MultiXact we do:
> 1. Retrieve offset by mxid from SLRU under MultiXactOffsetControlLock
> 2. If offset is 0 - it's not filled in at step 4 of previous algorithm, we sleep and goto 1
> 3. Retrieve members from SLRU under MultiXactMemberControlLock
> 4. ..... what we do if there are just zeroes because step 4 is not executed yet? Nothing, return empty members list.
So transactions never see such incomplete mxids, I believe.
> What am I missing?
regards.
--
Kyotaro Horiguchi
NTT Open Source Software Center
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2020-05-14 05:19 Andrey M. Borodin <[email protected]>
parent: Kyotaro Horiguchi <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Andrey M. Borodin @ 2020-05-14 05:19 UTC (permalink / raw)
To: Kyotaro Horiguchi <[email protected]>; +Cc: pgsql-hackers
> 14 мая 2020 г., в 06:25, Kyotaro Horiguchi <[email protected]> написал(а):
>
> At Wed, 13 May 2020 23:08:37 +0500, "Andrey M. Borodin" <[email protected]> wrote in
>>
>>
>>> 11 мая 2020 г., в 16:17, Andrey M. Borodin <[email protected]> написал(а):
>>>
>>> I've went ahead and created 3 patches:
>>> 1. Configurable SLRU buffer sizes for MultiXacOffsets and MultiXactMembers
>>> 2. Reduce locking level to shared on read of MultiXactId members
>>> 3. Configurable cache size
>>
>> I'm looking more at MultiXact and it seems to me that we have a race condition there.
>>
>> When we create a new MultiXact we do:
>> 1. Generate new MultiXactId under MultiXactGenLock
>> 2. Record new mxid with members and offset to WAL
>> 3. Write offset to SLRU under MultiXactOffsetControlLock
>> 4. Write members to SLRU under MultiXactMemberControlLock
>
> But, don't we hold exclusive lock on the buffer through all the steps
> above?
Yes...Unless MultiXact is observed on StandBy. This could lead to observing inconsistent snapshot: one of lockers committed tuple delete, but standby sees it as alive.
>> When we read MultiXact we do:
>> 1. Retrieve offset by mxid from SLRU under MultiXactOffsetControlLock
>> 2. If offset is 0 - it's not filled in at step 4 of previous algorithm, we sleep and goto 1
>> 3. Retrieve members from SLRU under MultiXactMemberControlLock
>> 4. ..... what we do if there are just zeroes because step 4 is not executed yet? Nothing, return empty members list.
>
> So transactions never see such incomplete mxids, I believe.
I've observed sleep in step 2. I believe it's possible to observe special effects of step 4 too.
Maybe we could add lock on standby to dismiss this 1000us wait? Sometimes it hits hard on Standbys: if someone is locking whole table on primary - all seq scans on standbys follow him with MultiXactOffsetControlLock contention.
It looks like this:
0x00007fcd56896ff7 in __GI___select (nfds=nfds@entry=0, readfds=readfds@entry=0x0, writefds=writefds@entry=0x0, exceptfds=exceptfds@entry=0x0, timeout=timeout@entry=0x7ffd83376fe0) at ../sysdeps/unix/sysv/linux/select.c:41
#0 0x00007fcd56896ff7 in __GI___select (nfds=nfds@entry=0, readfds=readfds@entry=0x0, writefds=writefds@entry=0x0, exceptfds=exceptfds@entry=0x0, timeout=timeout@entry=0x7ffd83376fe0) at ../sysdeps/unix/sysv/linux/select.c:41
#1 0x000056186e0d54bd in pg_usleep (microsec=microsec@entry=1000) at ./build/../src/port/pgsleep.c:56
#2 0x000056186dd5edf2 in GetMultiXactIdMembers (from_pgupgrade=0 '\000', onlyLock=<optimized out>, members=0x7ffd83377080, multi=3106214809) at ./build/../src/backend/access/transam/multixact.c:1370
#3 GetMultiXactIdMembers () at ./build/../src/backend/access/transam/multixact.c:1202
#4 0x000056186dd2d2d9 in MultiXactIdGetUpdateXid (xmax=<optimized out>, t_infomask=<optimized out>) at ./build/../src/backend/access/heap/heapam.c:7039
#5 0x000056186dd35098 in HeapTupleGetUpdateXid (tuple=tuple@entry=0x7fcba3b63d58) at ./build/../src/backend/access/heap/heapam.c:7080
#6 0x000056186e0cd0f8 in HeapTupleSatisfiesMVCC (htup=<optimized out>, snapshot=0x56186f44a058, buffer=230684) at ./build/../src/backend/utils/time/tqual.c:1091
#7 0x000056186dd2d922 in heapgetpage (scan=scan@entry=0x56186f4c8e78, page=page@entry=3620) at ./build/../src/backend/access/heap/heapam.c:439
#8 0x000056186dd2ea7c in heapgettup_pagemode (key=0x0, nkeys=0, dir=ForwardScanDirection, scan=0x56186f4c8e78) at ./build/../src/backend/access/heap/heapam.c:1034
#9 heap_getnext (scan=scan@entry=0x56186f4c8e78, direction=direction@entry=ForwardScanDirection) at ./build/../src/backend/access/heap/heapam.c:1801
#10 0x000056186de84f51 in SeqNext (node=node@entry=0x56186f4a4f78) at ./build/../src/backend/executor/nodeSeqscan.c:81
#11 0x000056186de6a3f1 in ExecScanFetch (recheckMtd=0x56186de84ef0 <SeqRecheck>, accessMtd=0x56186de84f20 <SeqNext>, node=0x56186f4a4f78) at ./build/../src/backend/executor/execScan.c:97
#12 ExecScan (node=0x56186f4a4f78, accessMtd=0x56186de84f20 <SeqNext>, recheckMtd=0x56186de84ef0 <SeqRecheck>) at ./build/../src/backend/executor/execScan.c:164
Best regards, Andrey Borodin.
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2020-05-14 06:16 Kyotaro Horiguchi <[email protected]>
parent: Andrey M. Borodin <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Kyotaro Horiguchi @ 2020-05-14 06:16 UTC (permalink / raw)
To: [email protected]; +Cc: pgsql-hackers
At Thu, 14 May 2020 10:19:42 +0500, "Andrey M. Borodin" <[email protected]> wrote in
> >> I'm looking more at MultiXact and it seems to me that we have a race condition there.
> >>
> >> When we create a new MultiXact we do:
> >> 1. Generate new MultiXactId under MultiXactGenLock
> >> 2. Record new mxid with members and offset to WAL
> >> 3. Write offset to SLRU under MultiXactOffsetControlLock
> >> 4. Write members to SLRU under MultiXactMemberControlLock
> >
> > But, don't we hold exclusive lock on the buffer through all the steps
> > above?
> Yes...Unless MultiXact is observed on StandBy. This could lead to observing inconsistent snapshot: one of lockers committed tuple delete, but standby sees it as alive.
Ah, right. I looked from GetNewMultiXactId. Actually
XLOG_MULTIXACT_CREATE_ID is not protected from concurrent reference to
the creating mxact id. And GetMultiXactIdMembers is considering that
case.
> >> When we read MultiXact we do:
> >> 1. Retrieve offset by mxid from SLRU under MultiXactOffsetControlLock
> >> 2. If offset is 0 - it's not filled in at step 4 of previous algorithm, we sleep and goto 1
> >> 3. Retrieve members from SLRU under MultiXactMemberControlLock
> >> 4. ..... what we do if there are just zeroes because step 4 is not executed yet? Nothing, return empty members list.
> >
> > So transactions never see such incomplete mxids, I believe.
> I've observed sleep in step 2. I believe it's possible to observe special effects of step 4 too.
> Maybe we could add lock on standby to dismiss this 1000us wait? Sometimes it hits hard on Standbys: if someone is locking whole table on primary - all seq scans on standbys follow him with MultiXactOffsetControlLock contention.
GetMultiXactIdMembers believes that 4 is successfully done if 2
returned valid offset, but actually that is not obvious.
If we add a single giant lock just to isolate ,say,
GetMultiXactIdMember and RecordNewMultiXact, it reduces concurrency
unnecessarily. Perhaps we need finer-grained locking-key for standby
that works similary to buffer lock on primary, that doesn't cause
confilicts between irrelevant mxids.
regards.
--
Kyotaro Horiguchi
NTT Open Source Software Center
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2020-05-14 06:44 Andrey M. Borodin <[email protected]>
parent: Kyotaro Horiguchi <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Andrey M. Borodin @ 2020-05-14 06:44 UTC (permalink / raw)
To: Kyotaro Horiguchi <[email protected]>; +Cc: pgsql-hackers
> 14 мая 2020 г., в 11:16, Kyotaro Horiguchi <[email protected]> написал(а):
>
> At Thu, 14 May 2020 10:19:42 +0500, "Andrey M. Borodin" <[email protected]> wrote in
>>>> I'm looking more at MultiXact and it seems to me that we have a race condition there.
>>>>
>>>> When we create a new MultiXact we do:
>>>> 1. Generate new MultiXactId under MultiXactGenLock
>>>> 2. Record new mxid with members and offset to WAL
>>>> 3. Write offset to SLRU under MultiXactOffsetControlLock
>>>> 4. Write members to SLRU under MultiXactMemberControlLock
>>>
>>> But, don't we hold exclusive lock on the buffer through all the steps
>>> above?
>> Yes...Unless MultiXact is observed on StandBy. This could lead to observing inconsistent snapshot: one of lockers committed tuple delete, but standby sees it as alive.
>
> Ah, right. I looked from GetNewMultiXactId. Actually
> XLOG_MULTIXACT_CREATE_ID is not protected from concurrent reference to
> the creating mxact id. And GetMultiXactIdMembers is considering that
> case.
>
>>>> When we read MultiXact we do:
>>>> 1. Retrieve offset by mxid from SLRU under MultiXactOffsetControlLock
>>>> 2. If offset is 0 - it's not filled in at step 4 of previous algorithm, we sleep and goto 1
>>>> 3. Retrieve members from SLRU under MultiXactMemberControlLock
>>>> 4. ..... what we do if there are just zeroes because step 4 is not executed yet? Nothing, return empty members list.
>>>
>>> So transactions never see such incomplete mxids, I believe.
>> I've observed sleep in step 2. I believe it's possible to observe special effects of step 4 too.
>> Maybe we could add lock on standby to dismiss this 1000us wait? Sometimes it hits hard on Standbys: if someone is locking whole table on primary - all seq scans on standbys follow him with MultiXactOffsetControlLock contention.
>
> GetMultiXactIdMembers believes that 4 is successfully done if 2
> returned valid offset, but actually that is not obvious.
>
> If we add a single giant lock just to isolate ,say,
> GetMultiXactIdMember and RecordNewMultiXact, it reduces concurrency
> unnecessarily. Perhaps we need finer-grained locking-key for standby
> that works similary to buffer lock on primary, that doesn't cause
> confilicts between irrelevant mxids.
>
We can just replay members before offsets. If offset is already there - members are there too.
But I'd be happy if we could mitigate those 1000us too - with a hint about last maixd state in a shared MX state, for example.
Actually, if we read empty mxid array instead of something that is replayed just yet - it's not a problem of inconsistency, because transaction in this mxid could not commit before we started. ISTM.
So instead of fix, we, probably, can just add a comment. If this reasoning is correct.
Best regards, Andrey Borodin.
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2020-05-15 00:03 Kyotaro Horiguchi <[email protected]>
parent: Andrey M. Borodin <[email protected]>
0 siblings, 2 replies; 103+ messages in thread
From: Kyotaro Horiguchi @ 2020-05-15 00:03 UTC (permalink / raw)
To: [email protected]; +Cc: pgsql-hackers
At Thu, 14 May 2020 11:44:01 +0500, "Andrey M. Borodin" <[email protected]> wrote in
> > GetMultiXactIdMembers believes that 4 is successfully done if 2
> > returned valid offset, but actually that is not obvious.
> >
> > If we add a single giant lock just to isolate ,say,
> > GetMultiXactIdMember and RecordNewMultiXact, it reduces concurrency
> > unnecessarily. Perhaps we need finer-grained locking-key for standby
> > that works similary to buffer lock on primary, that doesn't cause
> > confilicts between irrelevant mxids.
> >
> We can just replay members before offsets. If offset is already there - members are there too.
> But I'd be happy if we could mitigate those 1000us too - with a hint about last maixd state in a shared MX state, for example.
Generally in such cases, condition variables would work. In the
attached PoC, the reader side gets no penalty in the "likely" code
path. The writer side always calls ConditionVariableBroadcast but the
waiter list is empty in almost all cases. But I couldn't cause the
situation where the sleep 1000u is reached.
> Actually, if we read empty mxid array instead of something that is replayed just yet - it's not a problem of inconsistency, because transaction in this mxid could not commit before we started. ISTM.
> So instead of fix, we, probably, can just add a comment. If this reasoning is correct.
The step 4 of the reader side reads the members of the target mxid. It
is already written if the offset of the *next* mxid is filled-in.
regards.
--
Kyotaro Horiguchi
NTT Open Source Software Center
Attachments:
[text/x-patch] mxid_wait_instead_of_sleep.patch (3.0K, ../../[email protected]/2-mxid_wait_instead_of_sleep.patch)
download | inline diff:
diff --git a/src/backend/access/transam/multixact.c b/src/backend/access/transam/multixact.c
index e2aa5c9ce4..9db8f6cddd 100644
--- a/src/backend/access/transam/multixact.c
+++ b/src/backend/access/transam/multixact.c
@@ -82,6 +82,7 @@
#include "lib/ilist.h"
#include "miscadmin.h"
#include "pg_trace.h"
+#include "pgstat.h"
#include "postmaster/autovacuum.h"
#include "storage/lmgr.h"
#include "storage/pmsignal.h"
@@ -233,6 +234,7 @@ typedef struct MultiXactStateData
/* support for members anti-wraparound measures */
MultiXactOffset offsetStopLimit; /* known if oldestOffsetKnown */
+ ConditionVariable nextoff_cv;
/*
* Per-backend data starts here. We have two arrays stored in the area
* immediately following the MultiXactStateData struct. Each is indexed by
@@ -873,6 +875,14 @@ RecordNewMultiXact(MultiXactId multi, MultiXactOffset offset,
/* Exchange our lock */
LWLockRelease(MultiXactOffsetControlLock);
+ /*
+ * Let everybody know the offset of this mxid is recorded now. The waiters
+ * are waiting for the offset of the mxid next of the target to know the
+ * number of members of the target mxid, so we don't need to wait for
+ * members of this mxid are recorded.
+ */
+ ConditionVariableBroadcast(&MultiXactState->nextoff_cv);
+
LWLockAcquire(MultiXactMemberControlLock, LW_EXCLUSIVE);
prev_pageno = -1;
@@ -1367,9 +1377,19 @@ retry:
if (nextMXOffset == 0)
{
/* Corner case 2: next multixact is still being filled in */
+
+ /*
+ * The recorder of the next mxid is just before writing the offset.
+ * Wait for the offset to be written.
+ */
+ ConditionVariablePrepareToSleep(&MultiXactState->nextoff_cv);
+
LWLockRelease(MultiXactOffsetControlLock);
CHECK_FOR_INTERRUPTS();
- pg_usleep(1000L);
+
+ ConditionVariableSleep(&MultiXactState->nextoff_cv,
+ WAIT_EVENT_WAIT_NEXT_MXMEMBERS);
+ ConditionVariableCancelSleep();
goto retry;
}
@@ -1847,6 +1867,7 @@ MultiXactShmemInit(void)
/* Make sure we zero out the per-backend state */
MemSet(MultiXactState, 0, SHARED_MULTIXACT_STATE_SIZE);
+ ConditionVariableInit(&MultiXactState->nextoff_cv);
}
else
Assert(found);
diff --git a/src/backend/postmaster/pgstat.c b/src/backend/postmaster/pgstat.c
index 0ecd29a1d9..1ac6b37188 100644
--- a/src/backend/postmaster/pgstat.c
+++ b/src/backend/postmaster/pgstat.c
@@ -3879,6 +3879,9 @@ pgstat_get_wait_ipc(WaitEventIPC w)
case WAIT_EVENT_SYNC_REP:
event_name = "SyncRep";
break;
+ case WAIT_EVENT_WAIT_NEXT_MXMEMBERS:
+ event_name = "Mact/WaitNextXactMembers";
+ break;
/* no default case, so that compiler will warn */
}
diff --git a/src/include/pgstat.h b/src/include/pgstat.h
index ae9a39573c..e79bba0bef 100644
--- a/src/include/pgstat.h
+++ b/src/include/pgstat.h
@@ -886,7 +886,8 @@ typedef enum
WAIT_EVENT_REPLICATION_ORIGIN_DROP,
WAIT_EVENT_REPLICATION_SLOT_DROP,
WAIT_EVENT_SAFE_SNAPSHOT,
- WAIT_EVENT_SYNC_REP
+ WAIT_EVENT_SYNC_REP,
+ WAIT_EVENT_WAIT_NEXT_MXMEMBERS
} WaitEventIPC;
/* ----------
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2020-05-15 09:01 Andrey M. Borodin <[email protected]>
parent: Kyotaro Horiguchi <[email protected]>
1 sibling, 1 reply; 103+ messages in thread
From: Andrey M. Borodin @ 2020-05-15 09:01 UTC (permalink / raw)
To: Kyotaro Horiguchi <[email protected]>; +Cc: pgsql-hackers
> 15 мая 2020 г., в 05:03, Kyotaro Horiguchi <[email protected]> написал(а):
>
> At Thu, 14 May 2020 11:44:01 +0500, "Andrey M. Borodin" <[email protected]> wrote in
>>> GetMultiXactIdMembers believes that 4 is successfully done if 2
>>> returned valid offset, but actually that is not obvious.
>>>
>>> If we add a single giant lock just to isolate ,say,
>>> GetMultiXactIdMember and RecordNewMultiXact, it reduces concurrency
>>> unnecessarily. Perhaps we need finer-grained locking-key for standby
>>> that works similary to buffer lock on primary, that doesn't cause
>>> confilicts between irrelevant mxids.
>>>
>> We can just replay members before offsets. If offset is already there - members are there too.
>> But I'd be happy if we could mitigate those 1000us too - with a hint about last maixd state in a shared MX state, for example.
>
> Generally in such cases, condition variables would work. In the
> attached PoC, the reader side gets no penalty in the "likely" code
> path. The writer side always calls ConditionVariableBroadcast but the
> waiter list is empty in almost all cases. But I couldn't cause the
> situation where the sleep 1000u is reached.
Thanks! That really looks like a good solution without magic timeouts. Beautiful!
I think I can create temporary extension which calls MultiXact API and tests edge-cases like this 1000us wait.
This extension will also be also useful for me to assess impact of bigger buffers, reduced read locking (as in my 2nd patch) and other tweaks.
>> Actually, if we read empty mxid array instead of something that is replayed just yet - it's not a problem of inconsistency, because transaction in this mxid could not commit before we started. ISTM.
>> So instead of fix, we, probably, can just add a comment. If this reasoning is correct.
>
> The step 4 of the reader side reads the members of the target mxid. It
> is already written if the offset of the *next* mxid is filled-in.
Most often - yes, but members are not guaranteed to be filled in order. Those who win MXMemberControlLock will write first.
But nobody can read members of MXID before it is returned. And its members will be written before returning MXID.
Best regards, Andrey Borodin.
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2020-05-20 04:54 Kyotaro Horiguchi <[email protected]>
parent: Andrey M. Borodin <[email protected]>
0 siblings, 0 replies; 103+ messages in thread
From: Kyotaro Horiguchi @ 2020-05-20 04:54 UTC (permalink / raw)
To: [email protected]; +Cc: pgsql-hackers
At Fri, 15 May 2020 14:01:46 +0500, "Andrey M. Borodin" <[email protected]> wrote in
>
>
> > 15 мая 2020 г., в 05:03, Kyotaro Horiguchi <[email protected]> написал(а):
> >
> > At Thu, 14 May 2020 11:44:01 +0500, "Andrey M. Borodin" <[email protected]> wrote in
> >>> GetMultiXactIdMembers believes that 4 is successfully done if 2
> >>> returned valid offset, but actually that is not obvious.
> >>>
> >>> If we add a single giant lock just to isolate ,say,
> >>> GetMultiXactIdMember and RecordNewMultiXact, it reduces concurrency
> >>> unnecessarily. Perhaps we need finer-grained locking-key for standby
> >>> that works similary to buffer lock on primary, that doesn't cause
> >>> confilicts between irrelevant mxids.
> >>>
> >> We can just replay members before offsets. If offset is already there - members are there too.
> >> But I'd be happy if we could mitigate those 1000us too - with a hint about last maixd state in a shared MX state, for example.
> >
> > Generally in such cases, condition variables would work. In the
> > attached PoC, the reader side gets no penalty in the "likely" code
> > path. The writer side always calls ConditionVariableBroadcast but the
> > waiter list is empty in almost all cases. But I couldn't cause the
> > situation where the sleep 1000u is reached.
> Thanks! That really looks like a good solution without magic timeouts. Beautiful!
> I think I can create temporary extension which calls MultiXact API and tests edge-cases like this 1000us wait.
> This extension will also be also useful for me to assess impact of bigger buffers, reduced read locking (as in my 2nd patch) and other tweaks.
Happy to hear that, It would need to use timeout just in case, though.
> >> Actually, if we read empty mxid array instead of something that is replayed just yet - it's not a problem of inconsistency, because transaction in this mxid could not commit before we started. ISTM.
> >> So instead of fix, we, probably, can just add a comment. If this reasoning is correct.
> >
> > The step 4 of the reader side reads the members of the target mxid. It
> > is already written if the offset of the *next* mxid is filled-in.
> Most often - yes, but members are not guaranteed to be filled in order. Those who win MXMemberControlLock will write first.
> But nobody can read members of MXID before it is returned. And its members will be written before returning MXID.
Yeah, right. Otherwise assertion failure happens.
regards.
--
Kyotaro Horiguchi
NTT Open Source Software Center
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2020-07-02 12:02 Daniel Gustafsson <[email protected]>
parent: Kyotaro Horiguchi <[email protected]>
1 sibling, 1 reply; 103+ messages in thread
From: Daniel Gustafsson @ 2020-07-02 12:02 UTC (permalink / raw)
To: Kyotaro Horiguchi <[email protected]>; +Cc: [email protected]; pgsql-hackers
> On 15 May 2020, at 02:03, Kyotaro Horiguchi <[email protected]> wrote:
> Generally in such cases, condition variables would work. In the
> attached PoC, the reader side gets no penalty in the "likely" code
> path. The writer side always calls ConditionVariableBroadcast but the
> waiter list is empty in almost all cases. But I couldn't cause the
> situation where the sleep 1000u is reached.
The submitted patch no longer applies, can you please submit an updated
version? I'm marking the patch Waiting on Author in the meantime.
cheers ./daniel
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2020-07-08 07:03 Andrey M. Borodin <[email protected]>
parent: Daniel Gustafsson <[email protected]>
0 siblings, 2 replies; 103+ messages in thread
From: Andrey M. Borodin @ 2020-07-08 07:03 UTC (permalink / raw)
To: Daniel Gustafsson <[email protected]>; +Cc: Kyotaro Horiguchi <[email protected]>; pgsql-hackers
> 2 июля 2020 г., в 17:02, Daniel Gustafsson <[email protected]> написал(а):
>
>> On 15 May 2020, at 02:03, Kyotaro Horiguchi <[email protected]> wrote:
>
>> Generally in such cases, condition variables would work. In the
>> attached PoC, the reader side gets no penalty in the "likely" code
>> path. The writer side always calls ConditionVariableBroadcast but the
>> waiter list is empty in almost all cases. But I couldn't cause the
>> situation where the sleep 1000u is reached.
>
> The submitted patch no longer applies, can you please submit an updated
> version? I'm marking the patch Waiting on Author in the meantime.
Thanks, Daniel! PFA V2
Best regards, Andrey Borodin.
Attachments:
[application/octet-stream] v2-0001-Use-shared-lock-in-GetMultiXactIdMembers-for-offs.patch (7.6K, ../../[email protected]/2-v2-0001-Use-shared-lock-in-GetMultiXactIdMembers-for-offs.patch)
download | inline diff:
From 264b475fbf6ee519dc3b4eb2cca25145d2aaa569 Mon Sep 17 00:00:00 2001
From: Andrey Borodin <[email protected]>
Date: Sat, 9 May 2020 21:19:18 +0500
Subject: [PATCH v2 1/4] Use shared lock in GetMultiXactIdMembers for offsets
and members
Previously read of multixact required exclusive control locks in
offstes and members SLRUs. This could lead to contention.
In this commit we take advantge of SimpleLruReadPage_ReadOnly
similar to CLOG usage.
---
src/backend/access/transam/clog.c | 2 +-
src/backend/access/transam/commit_ts.c | 2 +-
src/backend/access/transam/multixact.c | 12 ++++++------
src/backend/access/transam/slru.c | 8 +++++---
src/backend/access/transam/subtrans.c | 2 +-
src/backend/commands/async.c | 2 +-
src/backend/storage/lmgr/predicate.c | 2 +-
src/include/access/slru.h | 2 +-
8 files changed, 17 insertions(+), 15 deletions(-)
diff --git a/src/backend/access/transam/clog.c b/src/backend/access/transam/clog.c
index f3da40ae01..3614d0c774 100644
--- a/src/backend/access/transam/clog.c
+++ b/src/backend/access/transam/clog.c
@@ -643,7 +643,7 @@ TransactionIdGetStatus(TransactionId xid, XLogRecPtr *lsn)
/* lock is acquired by SimpleLruReadPage_ReadOnly */
- slotno = SimpleLruReadPage_ReadOnly(XactCtl, pageno, xid);
+ slotno = SimpleLruReadPage_ReadOnly(XactCtl, pageno, xid, false);
byteptr = XactCtl->shared->page_buffer[slotno] + byteno;
status = (*byteptr >> bshift) & CLOG_XACT_BITMASK;
diff --git a/src/backend/access/transam/commit_ts.c b/src/backend/access/transam/commit_ts.c
index 9cdb136435..5fbbf446e3 100644
--- a/src/backend/access/transam/commit_ts.c
+++ b/src/backend/access/transam/commit_ts.c
@@ -342,7 +342,7 @@ TransactionIdGetCommitTsData(TransactionId xid, TimestampTz *ts,
}
/* lock is acquired by SimpleLruReadPage_ReadOnly */
- slotno = SimpleLruReadPage_ReadOnly(CommitTsCtl, pageno, xid);
+ slotno = SimpleLruReadPage_ReadOnly(CommitTsCtl, pageno, xid, false);
memcpy(&entry,
CommitTsCtl->shared->page_buffer[slotno] +
SizeOfCommitTimestampEntry * entryno,
diff --git a/src/backend/access/transam/multixact.c b/src/backend/access/transam/multixact.c
index 79ec21c75a..241d9dd78d 100644
--- a/src/backend/access/transam/multixact.c
+++ b/src/backend/access/transam/multixact.c
@@ -1321,12 +1321,12 @@ GetMultiXactIdMembers(MultiXactId multi, MultiXactMember **members,
* time on every multixact creation.
*/
retry:
- LWLockAcquire(MultiXactOffsetSLRULock, LW_EXCLUSIVE);
pageno = MultiXactIdToOffsetPage(multi);
entryno = MultiXactIdToOffsetEntry(multi);
- slotno = SimpleLruReadPage(MultiXactOffsetCtl, pageno, true, multi);
+ /* lock is acquired by SimpleLruReadPage_ReadOnly */
+ slotno = SimpleLruReadPage_ReadOnly(MultiXactOffsetCtl, pageno, multi, false);
offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
offptr += entryno;
offset = *offptr;
@@ -1358,7 +1358,7 @@ retry:
entryno = MultiXactIdToOffsetEntry(tmpMXact);
if (pageno != prev_pageno)
- slotno = SimpleLruReadPage(MultiXactOffsetCtl, pageno, true, tmpMXact);
+ slotno = SimpleLruReadPage_ReadOnly(MultiXactOffsetCtl, pageno, tmpMXact, true);
offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
offptr += entryno;
@@ -1382,7 +1382,7 @@ retry:
*members = ptr;
/* Now get the members themselves. */
- LWLockAcquire(MultiXactMemberSLRULock, LW_EXCLUSIVE);
+ LWLockAcquire(MultiXactMemberSLRULock, LW_SHARED);
truelength = 0;
prev_pageno = -1;
@@ -1399,7 +1399,7 @@ retry:
if (pageno != prev_pageno)
{
- slotno = SimpleLruReadPage(MultiXactMemberCtl, pageno, true, multi);
+ slotno = SimpleLruReadPage_ReadOnly(MultiXactMemberCtl, pageno, multi, true);
prev_pageno = pageno;
}
@@ -2745,7 +2745,7 @@ find_multixact_start(MultiXactId multi, MultiXactOffset *result)
return false;
/* lock is acquired by SimpleLruReadPage_ReadOnly */
- slotno = SimpleLruReadPage_ReadOnly(MultiXactOffsetCtl, pageno, multi);
+ slotno = SimpleLruReadPage_ReadOnly(MultiXactOffsetCtl, pageno, multi, false);
offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
offptr += entryno;
offset = *offptr;
diff --git a/src/backend/access/transam/slru.c b/src/backend/access/transam/slru.c
index 61249f4a12..96d0def5c2 100644
--- a/src/backend/access/transam/slru.c
+++ b/src/backend/access/transam/slru.c
@@ -475,17 +475,19 @@ SimpleLruReadPage(SlruCtl ctl, int pageno, bool write_ok,
* Return value is the shared-buffer slot number now holding the page.
* The buffer's LRU access info is updated.
*
- * Control lock must NOT be held at entry, but will be held at exit.
+ * Control lock will be held at exit. If lock_held is true at least
+ * shared control lock must be held.
* It is unspecified whether the lock will be shared or exclusive.
*/
int
-SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno, TransactionId xid)
+SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno, TransactionId xid, bool lock_held)
{
SlruShared shared = ctl->shared;
int slotno;
/* Try to find the page while holding only shared lock */
- LWLockAcquire(shared->ControlLock, LW_SHARED);
+ if (!lock_held)
+ LWLockAcquire(shared->ControlLock, LW_SHARED);
/* See if page is already in a buffer */
for (slotno = 0; slotno < shared->num_slots; slotno++)
diff --git a/src/backend/access/transam/subtrans.c b/src/backend/access/transam/subtrans.c
index f33ae407a6..8d8e0d9ec7 100644
--- a/src/backend/access/transam/subtrans.c
+++ b/src/backend/access/transam/subtrans.c
@@ -123,7 +123,7 @@ SubTransGetParent(TransactionId xid)
/* lock is acquired by SimpleLruReadPage_ReadOnly */
- slotno = SimpleLruReadPage_ReadOnly(SubTransCtl, pageno, xid);
+ slotno = SimpleLruReadPage_ReadOnly(SubTransCtl, pageno, xid, false);
ptr = (TransactionId *) SubTransCtl->shared->page_buffer[slotno];
ptr += entryno;
diff --git a/src/backend/commands/async.c b/src/backend/commands/async.c
index 71b7577afc..70085fc037 100644
--- a/src/backend/commands/async.c
+++ b/src/backend/commands/async.c
@@ -2012,7 +2012,7 @@ asyncQueueReadAllNotifications(void)
* part of the page we will actually inspect.
*/
slotno = SimpleLruReadPage_ReadOnly(NotifyCtl, curpage,
- InvalidTransactionId);
+ InvalidTransactionId, false);
if (curpage == QUEUE_POS_PAGE(head))
{
/* we only want to read as far as head */
diff --git a/src/backend/storage/lmgr/predicate.c b/src/backend/storage/lmgr/predicate.c
index d24919f76b..e2a7d00d37 100644
--- a/src/backend/storage/lmgr/predicate.c
+++ b/src/backend/storage/lmgr/predicate.c
@@ -948,7 +948,7 @@ SerialGetMinConflictCommitSeqNo(TransactionId xid)
* but will return with that lock held, which must then be released.
*/
slotno = SimpleLruReadPage_ReadOnly(SerialSlruCtl,
- SerialPage(xid), xid);
+ SerialPage(xid), xid, false);
val = SerialValue(slotno, xid);
LWLockRelease(SerialSLRULock);
return val;
diff --git a/src/include/access/slru.h b/src/include/access/slru.h
index 61fbc80ef0..471c692021 100644
--- a/src/include/access/slru.h
+++ b/src/include/access/slru.h
@@ -140,7 +140,7 @@ extern int SimpleLruZeroPage(SlruCtl ctl, int pageno);
extern int SimpleLruReadPage(SlruCtl ctl, int pageno, bool write_ok,
TransactionId xid);
extern int SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno,
- TransactionId xid);
+ TransactionId xid, bool lock_held);
extern void SimpleLruWritePage(SlruCtl ctl, int slotno);
extern void SimpleLruFlush(SlruCtl ctl, bool allow_redirtied);
extern void SimpleLruTruncate(SlruCtl ctl, int cutoffPage);
--
2.24.2 (Apple Git-127)
[application/octet-stream] v2-0002-Make-MultiXact-local-cache-size-configurable.patch (3.7K, ../../[email protected]/3-v2-0002-Make-MultiXact-local-cache-size-configurable.patch)
download | inline diff:
From 520686d6cc22a599bc06b91391171ca218c3db75 Mon Sep 17 00:00:00 2001
From: Andrey Borodin <[email protected]>
Date: Mon, 11 May 2020 16:17:02 +0500
Subject: [PATCH v2 2/4] Make MultiXact local cache size configurable
---
doc/src/sgml/config.sgml | 16 ++++++++++++++++
src/backend/access/transam/multixact.c | 2 +-
src/backend/utils/init/globals.c | 1 +
src/backend/utils/misc/guc.c | 10 ++++++++++
src/include/miscadmin.h | 1 +
5 files changed, 29 insertions(+), 1 deletion(-)
diff --git a/doc/src/sgml/config.sgml b/doc/src/sgml/config.sgml
index 94d99373a7..31d7c1f4b3 100644
--- a/doc/src/sgml/config.sgml
+++ b/doc/src/sgml/config.sgml
@@ -1777,6 +1777,22 @@ include_dir 'conf.d'
</listitem>
</varlistentry>
+ <varlistentry id="guc-multixact-local-cache-entries" xreflabel="multixact_local-cache-entries">
+ <term><varname>multixact_local_cache_entries</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>multixact_local_cache_entries</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the number cached MultiXact by backend. Any SLRU lookup is preceeded
+ by cache lookup. Higher numbers of cache size help to deduplicate lock sets, while
+ lower values make cache lookup faster.
+ It defaults to 256.
+ </para>
+ </listitem>
+ </varlistentry>
+
<varlistentry id="guc-max-stack-depth" xreflabel="max_stack_depth">
<term><varname>max_stack_depth</varname> (<type>integer</type>)
<indexterm>
diff --git a/src/backend/access/transam/multixact.c b/src/backend/access/transam/multixact.c
index 241d9dd78d..445b6258a4 100644
--- a/src/backend/access/transam/multixact.c
+++ b/src/backend/access/transam/multixact.c
@@ -1589,7 +1589,7 @@ mXactCachePut(MultiXactId multi, int nmembers, MultiXactMember *members)
qsort(entry->members, nmembers, sizeof(MultiXactMember), mxactMemberComparator);
dlist_push_head(&MXactCache, &entry->node);
- if (MXactCacheMembers++ >= MAX_CACHE_ENTRIES)
+ if (MXactCacheMembers++ >= multixact_local_cache_entries)
{
dlist_node *node;
mXactCacheEnt *entry;
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
index 01e739da40..f404167ad0 100644
--- a/src/backend/utils/init/globals.c
+++ b/src/backend/utils/init/globals.c
@@ -151,3 +151,4 @@ double vacuum_cleanup_index_scale_factor;
int multixact_offsets_slru_buffers = 8;
int multixact_members_slru_buffers = 16;
+int multixact_local_cache_entries = 256;
diff --git a/src/backend/utils/misc/guc.c b/src/backend/utils/misc/guc.c
index d1fbb75cd5..6443f595cc 100644
--- a/src/backend/utils/misc/guc.c
+++ b/src/backend/utils/misc/guc.c
@@ -2273,6 +2273,16 @@ static struct config_int ConfigureNamesInt[] =
NULL, NULL, NULL
},
+ {
+ {"multixact_local_cache_entries", PGC_SUSET, RESOURCES_MEM,
+ gettext_noop("Sets the number cached MultiXact by backend."),
+ NULL
+ },
+ &multixact_local_cache_entries,
+ 256, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
{
{"temp_buffers", PGC_USERSET, RESOURCES_MEM,
gettext_noop("Sets the maximum number of temporary buffers used by each session."),
diff --git a/src/include/miscadmin.h b/src/include/miscadmin.h
index bb1475ad2b..9ce2adea63 100644
--- a/src/include/miscadmin.h
+++ b/src/include/miscadmin.h
@@ -163,6 +163,7 @@ extern PGDLLIMPORT int max_worker_processes;
extern PGDLLIMPORT int max_parallel_workers;
extern PGDLLIMPORT int multixact_offsets_slru_buffers;
extern PGDLLIMPORT int multixact_members_slru_buffers;
+extern PGDLLIMPORT int multixact_local_cache_entries;
extern PGDLLIMPORT int MyProcPid;
extern PGDLLIMPORT pg_time_t MyStartTime;
--
2.24.2 (Apple Git-127)
[application/octet-stream] v2-0003-Add-conditional-variable-to-wait-for-next-MultXac.patch (3.4K, ../../[email protected]/4-v2-0003-Add-conditional-variable-to-wait-for-next-MultXac.patch)
download | inline diff:
From ddde253b10b87d060d34f8d7250be27a27c18c0a Mon Sep 17 00:00:00 2001
From: Andrey Borodin <[email protected]>
Date: Fri, 22 May 2020 14:13:33 +0500
Subject: [PATCH v2 3/4] Add conditional variable to wait for next MultXact
offset in edge case
---
src/backend/access/transam/multixact.c | 23 ++++++++++++++++++++++-
src/backend/postmaster/pgstat.c | 2 ++
src/include/pgstat.h | 1 +
3 files changed, 25 insertions(+), 1 deletion(-)
diff --git a/src/backend/access/transam/multixact.c b/src/backend/access/transam/multixact.c
index 445b6258a4..02de9a7c13 100644
--- a/src/backend/access/transam/multixact.c
+++ b/src/backend/access/transam/multixact.c
@@ -82,6 +82,7 @@
#include "lib/ilist.h"
#include "miscadmin.h"
#include "pg_trace.h"
+#include "pgstat.h"
#include "postmaster/autovacuum.h"
#include "storage/lmgr.h"
#include "storage/pmsignal.h"
@@ -233,6 +234,7 @@ typedef struct MultiXactStateData
/* support for members anti-wraparound measures */
MultiXactOffset offsetStopLimit; /* known if oldestOffsetKnown */
+ ConditionVariable nextoff_cv;
/*
* Per-backend data starts here. We have two arrays stored in the area
* immediately following the MultiXactStateData struct. Each is indexed by
@@ -873,6 +875,14 @@ RecordNewMultiXact(MultiXactId multi, MultiXactOffset offset,
/* Exchange our lock */
LWLockRelease(MultiXactOffsetSLRULock);
+ /*
+ * Let everybody know the offset of this mxid is recorded now. The waiters
+ * are waiting for the offset of the mxid next of the target to know the
+ * number of members of the target mxid, so we don't need to wait for
+ * members of this mxid are recorded.
+ */
+ ConditionVariableBroadcast(&MultiXactState->nextoff_cv);
+
LWLockAcquire(MultiXactMemberSLRULock, LW_EXCLUSIVE);
prev_pageno = -1;
@@ -1367,9 +1377,19 @@ retry:
if (nextMXOffset == 0)
{
/* Corner case 2: next multixact is still being filled in */
+
+ /*
+ * The recorder of the next mxid is just before writing the offset.
+ * Wait for the offset to be written.
+ */
+ ConditionVariablePrepareToSleep(&MultiXactState->nextoff_cv);
+
LWLockRelease(MultiXactOffsetSLRULock);
CHECK_FOR_INTERRUPTS();
- pg_usleep(1000L);
+
+ ConditionVariableSleep(&MultiXactState->nextoff_cv,
+ WAIT_EVENT_WAIT_NEXT_MXMEMBERS);
+ ConditionVariableCancelSleep();
goto retry;
}
@@ -1847,6 +1867,7 @@ MultiXactShmemInit(void)
/* Make sure we zero out the per-backend state */
MemSet(MultiXactState, 0, SHARED_MULTIXACT_STATE_SIZE);
+ ConditionVariableInit(&MultiXactState->nextoff_cv);
}
else
Assert(found);
diff --git a/src/backend/postmaster/pgstat.c b/src/backend/postmaster/pgstat.c
index edfa774ee4..4bce534c7a 100644
--- a/src/backend/postmaster/pgstat.c
+++ b/src/backend/postmaster/pgstat.c
@@ -3877,6 +3877,8 @@ pgstat_get_wait_ipc(WaitEventIPC w)
break;
case WAIT_EVENT_XACT_GROUP_UPDATE:
event_name = "XactGroupUpdate";
+ case WAIT_EVENT_WAIT_NEXT_MXMEMBERS:
+ event_name = "MultiXact/WaitNextXactMembers";
break;
/* no default case, so that compiler will warn */
}
diff --git a/src/include/pgstat.h b/src/include/pgstat.h
index 1387201382..26f828597b 100644
--- a/src/include/pgstat.h
+++ b/src/include/pgstat.h
@@ -887,6 +887,7 @@ typedef enum
WAIT_EVENT_REPLICATION_SLOT_DROP,
WAIT_EVENT_SAFE_SNAPSHOT,
WAIT_EVENT_SYNC_REP,
+ WAIT_EVENT_WAIT_NEXT_MXMEMBERS,
WAIT_EVENT_XACT_GROUP_UPDATE
} WaitEventIPC;
--
2.24.2 (Apple Git-127)
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2020-08-02 21:30 Daniel Gustafsson <[email protected]>
parent: Andrey M. Borodin <[email protected]>
1 sibling, 0 replies; 103+ messages in thread
From: Daniel Gustafsson @ 2020-08-02 21:30 UTC (permalink / raw)
To: Andrey M. Borodin <[email protected]>; +Cc: Kyotaro Horiguchi <[email protected]>; pgsql-hackers
> On 8 Jul 2020, at 09:03, Andrey M. Borodin <[email protected]> wrote:
>
>
>
>> 2 июля 2020 г., в 17:02, Daniel Gustafsson <[email protected]> написал(а):
>>
>>> On 15 May 2020, at 02:03, Kyotaro Horiguchi <[email protected]> wrote:
>>
>>> Generally in such cases, condition variables would work. In the
>>> attached PoC, the reader side gets no penalty in the "likely" code
>>> path. The writer side always calls ConditionVariableBroadcast but the
>>> waiter list is empty in almost all cases. But I couldn't cause the
>>> situation where the sleep 1000u is reached.
>>
>> The submitted patch no longer applies, can you please submit an updated
>> version? I'm marking the patch Waiting on Author in the meantime.
> Thanks, Daniel! PFA V2
This version too has stopped applying according to the CFbot. I've moved it to
the next commitfest as we're out of time in this one and it was only pointed
out now, but kept it Waiting on Author.
cheers ./daniel
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2020-08-28 18:08 Anastasia Lubennikova <[email protected]>
parent: Andrey M. Borodin <[email protected]>
1 sibling, 1 reply; 103+ messages in thread
From: Anastasia Lubennikova @ 2020-08-28 18:08 UTC (permalink / raw)
To: Andrey M. Borodin <[email protected]>; Daniel Gustafsson <[email protected]>; +Cc: Kyotaro Horiguchi <[email protected]>; pgsql-hackers
On 08.07.2020 10:03, Andrey M. Borodin wrote:
>
>> 2 июля 2020 г., в 17:02, Daniel Gustafsson <[email protected]> написал(а):
>>
>>> On 15 May 2020, at 02:03, Kyotaro Horiguchi <[email protected]> wrote:
>>> Generally in such cases, condition variables would work. In the
>>> attached PoC, the reader side gets no penalty in the "likely" code
>>> path. The writer side always calls ConditionVariableBroadcast but the
>>> waiter list is empty in almost all cases. But I couldn't cause the
>>> situation where the sleep 1000u is reached.
>> The submitted patch no longer applies, can you please submit an updated
>> version? I'm marking the patch Waiting on Author in the meantime.
> Thanks, Daniel! PFA V2
>
> Best regards, Andrey Borodin.
1) The first patch is sensible and harmless, so I think it is ready for
committer. I haven't tested the performance impact, though.
2) I like the initial proposal to make various SLRU buffers
configurable, however, I doubt if it really solves the problem, or just
moves it to another place?
The previous patch you sent was based on some version that contained
changes for other slru buffers numbers: 'multixact_offsets_slru_buffers'
and 'multixact_members_slru_buffers'. Have you just forgot to attach
them? The patch message "[PATCH v2 2/4]" hints that you had 4 patches)
Meanwhile, I attach the rebased patch to calm down the CFbot. The
changes are trivial.
2.1) I think that both min and max values for this parameter are too
extreme. Have you tested them?
+ &multixact_local_cache_entries,
+ 256, 2, INT_MAX / 2,
2.2) MAX_CACHE_ENTRIES is not used anymore, so it can be deleted.
3) No changes for third patch. I just renamed it for consistency.
--
Anastasia Lubennikova
Postgres Professional: http://www.postgrespro.com
The Russian Postgres Company
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2020-09-28 14:41 Andrey M. Borodin <[email protected]>
parent: Anastasia Lubennikova <[email protected]>
0 siblings, 2 replies; 103+ messages in thread
From: Andrey M. Borodin @ 2020-09-28 14:41 UTC (permalink / raw)
To: Anastasia Lubennikova <[email protected]>; +Cc: Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
Hi, Anastasia!
> 28 авг. 2020 г., в 23:08, Anastasia Lubennikova <[email protected]> написал(а):
>
> 1) The first patch is sensible and harmless, so I think it is ready for committer. I haven't tested the performance impact, though.
>
> 2) I like the initial proposal to make various SLRU buffers configurable, however, I doubt if it really solves the problem, or just moves it to another place?
>
> The previous patch you sent was based on some version that contained changes for other slru buffers numbers: 'multixact_offsets_slru_buffers' and 'multixact_members_slru_buffers'. Have you just forgot to attach them? The patch message "[PATCH v2 2/4]" hints that you had 4 patches)
> Meanwhile, I attach the rebased patch to calm down the CFbot. The changes are trivial.
>
> 2.1) I think that both min and max values for this parameter are too extreme. Have you tested them?
>
> + &multixact_local_cache_entries,
> + 256, 2, INT_MAX / 2,
>
> 2.2) MAX_CACHE_ENTRIES is not used anymore, so it can be deleted.
>
> 3) No changes for third patch. I just renamed it for consistency.
Thank you for your review.
Indeed, I had 4th patch with tests, but these tests didn't work well: I still did not manage to stress SLRUs to reproduce problem from production...
You are absolutely correct in point 2: I did only tests with sane values. And observed extreme performance degradation with values ~ 64 megabytes. I do not know which highest values should we pick? 1Gb? Or highest possible functioning value?
I greatly appreciate your review, sorry for so long delay. Thanks!
Best regards, Andrey Borodin.
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2020-10-07 14:23 Anastasia Lubennikova <[email protected]>
parent: Andrey M. Borodin <[email protected]>
1 sibling, 0 replies; 103+ messages in thread
From: Anastasia Lubennikova @ 2020-10-07 14:23 UTC (permalink / raw)
To: Andrey M. Borodin <[email protected]>; +Cc: Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
On 28.09.2020 17:41, Andrey M. Borodin wrote:
> Hi, Anastasia!
>
>> 28 авг. 2020 г., в 23:08, Anastasia Lubennikova <[email protected]> написал(а):
>>
>> 1) The first patch is sensible and harmless, so I think it is ready for committer. I haven't tested the performance impact, though.
>>
>> 2) I like the initial proposal to make various SLRU buffers configurable, however, I doubt if it really solves the problem, or just moves it to another place?
>>
>> The previous patch you sent was based on some version that contained changes for other slru buffers numbers: 'multixact_offsets_slru_buffers' and 'multixact_members_slru_buffers'. Have you just forgot to attach them? The patch message "[PATCH v2 2/4]" hints that you had 4 patches)
>> Meanwhile, I attach the rebased patch to calm down the CFbot. The changes are trivial.
>>
>> 2.1) I think that both min and max values for this parameter are too extreme. Have you tested them?
>>
>> + &multixact_local_cache_entries,
>> + 256, 2, INT_MAX / 2,
>>
>> 2.2) MAX_CACHE_ENTRIES is not used anymore, so it can be deleted.
>>
>> 3) No changes for third patch. I just renamed it for consistency.
> Thank you for your review.
>
> Indeed, I had 4th patch with tests, but these tests didn't work well: I still did not manage to stress SLRUs to reproduce problem from production...
>
> You are absolutely correct in point 2: I did only tests with sane values. And observed extreme performance degradation with values ~ 64 megabytes. I do not know which highest values should we pick? 1Gb? Or highest possible functioning value?
I would go with the values that we consider adequate for this setting.
As I see, there is no strict rule about it in guc.c and many variables
have large border values. Anyway, we need to explain it at least in the
documentation and code comments.
It seems that the default was conservative enough, so it can be also a
minimal value too. As for maximum, can you provide any benchmark
results? If we have a peak and a noticeable performance degradation
after that, we can use it to calculate the preferable maxvalue.
--
Anastasia Lubennikova
Postgres Professional: http://www.postgrespro.com
The Russian Postgres Company
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2020-10-26 01:05 Alexander Korotkov <[email protected]>
parent: Andrey M. Borodin <[email protected]>
1 sibling, 1 reply; 103+ messages in thread
From: Alexander Korotkov @ 2020-10-26 01:05 UTC (permalink / raw)
To: Andrey M. Borodin <[email protected]>; +Cc: Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
Hi!
On Mon, Sep 28, 2020 at 5:41 PM Andrey M. Borodin <[email protected]> wrote:
> > 28 авг. 2020 г., в 23:08, Anastasia Lubennikova <[email protected]> написал(а):
> >
> > 1) The first patch is sensible and harmless, so I think it is ready for committer. I haven't tested the performance impact, though.
> >
> > 2) I like the initial proposal to make various SLRU buffers configurable, however, I doubt if it really solves the problem, or just moves it to another place?
> >
> > The previous patch you sent was based on some version that contained changes for other slru buffers numbers: 'multixact_offsets_slru_buffers' and 'multixact_members_slru_buffers'. Have you just forgot to attach them? The patch message "[PATCH v2 2/4]" hints that you had 4 patches)
> > Meanwhile, I attach the rebased patch to calm down the CFbot. The changes are trivial.
> >
> > 2.1) I think that both min and max values for this parameter are too extreme. Have you tested them?
> >
> > + &multixact_local_cache_entries,
> > + 256, 2, INT_MAX / 2,
> >
> > 2.2) MAX_CACHE_ENTRIES is not used anymore, so it can be deleted.
> >
> > 3) No changes for third patch. I just renamed it for consistency.
>
> Thank you for your review.
>
> Indeed, I had 4th patch with tests, but these tests didn't work well: I still did not manage to stress SLRUs to reproduce problem from production...
>
> You are absolutely correct in point 2: I did only tests with sane values. And observed extreme performance degradation with values ~ 64 megabytes. I do not know which highest values should we pick? 1Gb? Or highest possible functioning value?
>
> I greatly appreciate your review, sorry for so long delay. Thanks!
I took a look at this patchset.
The 1st and 3rd patches look good to me. I made just minor improvements.
1) There is still a case when SimpleLruReadPage_ReadOnly() relocks the
SLRU lock, which is already taken in exclusive mode. I've evaded this
by passing the lock mode as a parameter to
SimpleLruReadPage_ReadOnly().
3) CHECK_FOR_INTERRUPTS() is not needed anymore, because it's called
inside ConditionVariableSleep() if needed. Also, no current wait
events use slashes, and I don't think we should introduce slashes
here. Even if we should, then we should also rename existing wait
events to be consistent with a new one. So, I've renamed the new wait
event to remove the slash.
Regarding the patch 2. I see the current documentation in the patch
doesn't explain to the user how to set the new parameter. I think we
should give users an idea what workloads need high values of
multixact_local_cache_entries parameter and what doesn't. Also, we
should explain the negative aspects of high values
multixact_local_cache_entries. Ideally, we should get the advantage
without overloading users with new nontrivial parameters, but I don't
have a particular idea here.
I'd like to propose committing 1 and 3, but leave 2 for further review.
------
Regards,
Alexander Korotkov
Attachments:
[application/octet-stream] v4-0001-Use-shared-lock-in-GetMultiXactIdMembers-for-offsets.patch (12.2K, ../../CAPpHfduekT=BrM3rLqvo4ajzKYmDk+6aCoYJrNHhK0Je+v2z=Q@mail.gmail.com/2-v4-0001-Use-shared-lock-in-GetMultiXactIdMembers-for-offsets.patch)
download | inline diff:
From 12bbf5dbd372a431d106e7e8c3ed50ffa5707a19 Mon Sep 17 00:00:00 2001
From: Alexander Korotkov <[email protected]>
Date: Mon, 26 Oct 2020 01:50:08 +0300
Subject: [PATCH 1/3] Use shared lock in GetMultiXactIdMembers for offsets and
members
Previously the read of multixact required exclusive control locks for both
offsets and members SLRUs. This could lead to the excessive lock contention.
This commit we makes multixacts SLRU take advantage of SimpleLruReadPage_ReadOnly
similar to clog, commit_ts and subtrans.
In order to evade extra reacquiring of CLRU lock, we teach
SimpleLruReadPage_ReadOnly() to take into account the current lock mode and
report resulting lock mode back.
Discussion: https://postgr.es/m/a7f1c4e1-1015-92a4-2bd4-6736bd13d03e%40postgrespro.ru#c496c4e75fc0605094a0e1f763e6a6ec
Author: Andrey Borodin
Reviewed-by: Kyotaro Horiguchi, Daniel Gustafsson
Reviewed-by: Anastasia Lubennikova, Alexander Korotkov
---
src/backend/access/transam/clog.c | 4 +++-
src/backend/access/transam/commit_ts.c | 4 +++-
src/backend/access/transam/multixact.c | 22 +++++++++++++++-------
src/backend/access/transam/slru.c | 23 +++++++++++++++++------
src/backend/access/transam/subtrans.c | 4 +++-
src/backend/commands/async.c | 4 +++-
src/backend/storage/lmgr/predicate.c | 4 +++-
src/include/access/slru.h | 2 +-
src/include/storage/lwlock.h | 2 ++
9 files changed, 50 insertions(+), 19 deletions(-)
diff --git a/src/backend/access/transam/clog.c b/src/backend/access/transam/clog.c
index 034349aa7b9..4c372065dee 100644
--- a/src/backend/access/transam/clog.c
+++ b/src/backend/access/transam/clog.c
@@ -640,10 +640,11 @@ TransactionIdGetStatus(TransactionId xid, XLogRecPtr *lsn)
int lsnindex;
char *byteptr;
XidStatus status;
+ LWLockMode lockmode = LW_NONE;
/* lock is acquired by SimpleLruReadPage_ReadOnly */
- slotno = SimpleLruReadPage_ReadOnly(XactCtl, pageno, xid);
+ slotno = SimpleLruReadPage_ReadOnly(XactCtl, pageno, xid, &lockmode);
byteptr = XactCtl->shared->page_buffer[slotno] + byteno;
status = (*byteptr >> bshift) & CLOG_XACT_BITMASK;
@@ -651,6 +652,7 @@ TransactionIdGetStatus(TransactionId xid, XLogRecPtr *lsn)
lsnindex = GetLSNIndex(slotno, xid);
*lsn = XactCtl->shared->group_lsn[lsnindex];
+ Assert(lockmode != LW_NONE);
LWLockRelease(XactSLRULock);
return status;
diff --git a/src/backend/access/transam/commit_ts.c b/src/backend/access/transam/commit_ts.c
index cb8a9688018..c19b0b5399e 100644
--- a/src/backend/access/transam/commit_ts.c
+++ b/src/backend/access/transam/commit_ts.c
@@ -288,6 +288,7 @@ TransactionIdGetCommitTsData(TransactionId xid, TimestampTz *ts,
CommitTimestampEntry entry;
TransactionId oldestCommitTsXid;
TransactionId newestCommitTsXid;
+ LWLockMode lockmode = LW_NONE;
if (!TransactionIdIsValid(xid))
ereport(ERROR,
@@ -342,7 +343,7 @@ TransactionIdGetCommitTsData(TransactionId xid, TimestampTz *ts,
}
/* lock is acquired by SimpleLruReadPage_ReadOnly */
- slotno = SimpleLruReadPage_ReadOnly(CommitTsCtl, pageno, xid);
+ slotno = SimpleLruReadPage_ReadOnly(CommitTsCtl, pageno, xid, &lockmode);
memcpy(&entry,
CommitTsCtl->shared->page_buffer[slotno] +
SizeOfCommitTimestampEntry * entryno,
@@ -352,6 +353,7 @@ TransactionIdGetCommitTsData(TransactionId xid, TimestampTz *ts,
if (nodeid)
*nodeid = entry.nodeid;
+ Assert(lockmode != LW_NONE);
LWLockRelease(CommitTsSLRULock);
return *ts != 0;
}
diff --git a/src/backend/access/transam/multixact.c b/src/backend/access/transam/multixact.c
index 43653fe5721..ccbce90f0ea 100644
--- a/src/backend/access/transam/multixact.c
+++ b/src/backend/access/transam/multixact.c
@@ -1237,6 +1237,7 @@ GetMultiXactIdMembers(MultiXactId multi, MultiXactMember **members,
MultiXactId tmpMXact;
MultiXactOffset nextOffset;
MultiXactMember *ptr;
+ LWLockMode lockmode = LW_NONE;
debug_elog3(DEBUG2, "GetMembers: asked for %u", multi);
@@ -1340,12 +1341,13 @@ GetMultiXactIdMembers(MultiXactId multi, MultiXactMember **members,
* time on every multixact creation.
*/
retry:
- LWLockAcquire(MultiXactOffsetSLRULock, LW_EXCLUSIVE);
pageno = MultiXactIdToOffsetPage(multi);
entryno = MultiXactIdToOffsetEntry(multi);
- slotno = SimpleLruReadPage(MultiXactOffsetCtl, pageno, true, multi);
+ /* lock is acquired by SimpleLruReadPage_ReadOnly */
+ slotno = SimpleLruReadPage_ReadOnly(MultiXactOffsetCtl, pageno,
+ multi, &lockmode);
offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
offptr += entryno;
offset = *offptr;
@@ -1377,7 +1379,8 @@ retry:
entryno = MultiXactIdToOffsetEntry(tmpMXact);
if (pageno != prev_pageno)
- slotno = SimpleLruReadPage(MultiXactOffsetCtl, pageno, true, tmpMXact);
+ slotno = SimpleLruReadPage_ReadOnly(MultiXactOffsetCtl, pageno,
+ tmpMXact, &lockmode);
offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
offptr += entryno;
@@ -1395,14 +1398,14 @@ retry:
length = nextMXOffset - offset;
}
+ Assert(lockmode != LW_NONE);
LWLockRelease(MultiXactOffsetSLRULock);
ptr = (MultiXactMember *) palloc(length * sizeof(MultiXactMember));
*members = ptr;
/* Now get the members themselves. */
- LWLockAcquire(MultiXactMemberSLRULock, LW_EXCLUSIVE);
-
+ lockmode = LW_NONE;
truelength = 0;
prev_pageno = -1;
for (i = 0; i < length; i++, offset++)
@@ -1418,7 +1421,8 @@ retry:
if (pageno != prev_pageno)
{
- slotno = SimpleLruReadPage(MultiXactMemberCtl, pageno, true, multi);
+ slotno = SimpleLruReadPage_ReadOnly(MultiXactMemberCtl, pageno,
+ multi, &lockmode);
prev_pageno = pageno;
}
@@ -1441,6 +1445,7 @@ retry:
truelength++;
}
+ Assert(lockmode != LW_NONE);
LWLockRelease(MultiXactMemberSLRULock);
/*
@@ -2733,6 +2738,7 @@ find_multixact_start(MultiXactId multi, MultiXactOffset *result)
int entryno;
int slotno;
MultiXactOffset *offptr;
+ LWLockMode lockmode = LW_NONE;
Assert(MultiXactState->finishedStartup);
@@ -2749,10 +2755,12 @@ find_multixact_start(MultiXactId multi, MultiXactOffset *result)
return false;
/* lock is acquired by SimpleLruReadPage_ReadOnly */
- slotno = SimpleLruReadPage_ReadOnly(MultiXactOffsetCtl, pageno, multi);
+ slotno = SimpleLruReadPage_ReadOnly(MultiXactOffsetCtl, pageno,
+ multi, &lockmode);
offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
offptr += entryno;
offset = *offptr;
+ Assert(lockmode != LW_NONE);
LWLockRelease(MultiXactOffsetSLRULock);
*result = offset;
diff --git a/src/backend/access/transam/slru.c b/src/backend/access/transam/slru.c
index 16a78986971..ca0a23ab875 100644
--- a/src/backend/access/transam/slru.c
+++ b/src/backend/access/transam/slru.c
@@ -487,17 +487,23 @@ SimpleLruReadPage(SlruCtl ctl, int pageno, bool write_ok,
* Return value is the shared-buffer slot number now holding the page.
* The buffer's LRU access info is updated.
*
- * Control lock must NOT be held at entry, but will be held at exit.
- * It is unspecified whether the lock will be shared or exclusive.
+ * Control lock must be held in *lock_mode mode, which may be LW_NONE. Control
+ * lock will be held at exit in at least shared mode. Resulting control lock
+ * mode is set to *lock_mode.
*/
int
-SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno, TransactionId xid)
+SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno, TransactionId xid,
+ LWLockMode *lock_mode)
{
SlruShared shared = ctl->shared;
int slotno;
/* Try to find the page while holding only shared lock */
- LWLockAcquire(shared->ControlLock, LW_SHARED);
+ if (*lock_mode == LW_NONE)
+ {
+ LWLockAcquire(shared->ControlLock, LW_SHARED);
+ *lock_mode = LW_SHARED;
+ }
/* See if page is already in a buffer */
for (slotno = 0; slotno < shared->num_slots; slotno++)
@@ -517,8 +523,13 @@ SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno, TransactionId xid)
}
/* No luck, so switch to normal exclusive lock and do regular read */
- LWLockRelease(shared->ControlLock);
- LWLockAcquire(shared->ControlLock, LW_EXCLUSIVE);
+ if (*lock_mode != LW_EXCLUSIVE)
+ {
+ Assert(*lock_mode == LW_NONE);
+ LWLockRelease(shared->ControlLock);
+ LWLockAcquire(shared->ControlLock, LW_EXCLUSIVE);
+ *lock_mode = LW_EXCLUSIVE;
+ }
return SimpleLruReadPage(ctl, pageno, true, xid);
}
diff --git a/src/backend/access/transam/subtrans.c b/src/backend/access/transam/subtrans.c
index 0111e867c79..7bb15431893 100644
--- a/src/backend/access/transam/subtrans.c
+++ b/src/backend/access/transam/subtrans.c
@@ -113,6 +113,7 @@ SubTransGetParent(TransactionId xid)
int slotno;
TransactionId *ptr;
TransactionId parent;
+ LWLockMode lockmode = LW_NONE;
/* Can't ask about stuff that might not be around anymore */
Assert(TransactionIdFollowsOrEquals(xid, TransactionXmin));
@@ -123,12 +124,13 @@ SubTransGetParent(TransactionId xid)
/* lock is acquired by SimpleLruReadPage_ReadOnly */
- slotno = SimpleLruReadPage_ReadOnly(SubTransCtl, pageno, xid);
+ slotno = SimpleLruReadPage_ReadOnly(SubTransCtl, pageno, xid, &lockmode);
ptr = (TransactionId *) SubTransCtl->shared->page_buffer[slotno];
ptr += entryno;
parent = *ptr;
+ Assert(lockmode != LW_NONE);
LWLockRelease(SubtransSLRULock);
return parent;
diff --git a/src/backend/commands/async.c b/src/backend/commands/async.c
index 8dbcace3f93..8b419575590 100644
--- a/src/backend/commands/async.c
+++ b/src/backend/commands/async.c
@@ -2002,6 +2002,7 @@ asyncQueueReadAllNotifications(void)
int curoffset = QUEUE_POS_OFFSET(pos);
int slotno;
int copysize;
+ LWLockMode lockmode = LW_NONE;
/*
* We copy the data from SLRU into a local buffer, so as to avoid
@@ -2010,7 +2011,7 @@ asyncQueueReadAllNotifications(void)
* part of the page we will actually inspect.
*/
slotno = SimpleLruReadPage_ReadOnly(NotifyCtl, curpage,
- InvalidTransactionId);
+ InvalidTransactionId, &lockmode);
if (curpage == QUEUE_POS_PAGE(head))
{
/* we only want to read as far as head */
@@ -2027,6 +2028,7 @@ asyncQueueReadAllNotifications(void)
NotifyCtl->shared->page_buffer[slotno] + curoffset,
copysize);
/* Release lock that we got from SimpleLruReadPage_ReadOnly() */
+ Assert(lockmode != LW_NONE);
LWLockRelease(NotifySLRULock);
/*
diff --git a/src/backend/storage/lmgr/predicate.c b/src/backend/storage/lmgr/predicate.c
index 8a365b400c6..a4df90a8aeb 100644
--- a/src/backend/storage/lmgr/predicate.c
+++ b/src/backend/storage/lmgr/predicate.c
@@ -924,6 +924,7 @@ SerialGetMinConflictCommitSeqNo(TransactionId xid)
TransactionId tailXid;
SerCommitSeqNo val;
int slotno;
+ LWLockMode lockmode = LW_NONE;
Assert(TransactionIdIsValid(xid));
@@ -946,8 +947,9 @@ SerialGetMinConflictCommitSeqNo(TransactionId xid)
* but will return with that lock held, which must then be released.
*/
slotno = SimpleLruReadPage_ReadOnly(SerialSlruCtl,
- SerialPage(xid), xid);
+ SerialPage(xid), xid, &lockmode);
val = SerialValue(slotno, xid);
+ Assert(lockmode != LW_NONE);
LWLockRelease(SerialSLRULock);
return val;
}
diff --git a/src/include/access/slru.h b/src/include/access/slru.h
index b39b43504d8..4b66d3b592b 100644
--- a/src/include/access/slru.h
+++ b/src/include/access/slru.h
@@ -142,7 +142,7 @@ extern int SimpleLruZeroPage(SlruCtl ctl, int pageno);
extern int SimpleLruReadPage(SlruCtl ctl, int pageno, bool write_ok,
TransactionId xid);
extern int SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno,
- TransactionId xid);
+ TransactionId xid, LWLockMode *lock_mode);
extern void SimpleLruWritePage(SlruCtl ctl, int slotno);
extern void SimpleLruWriteAll(SlruCtl ctl, bool allow_redirtied);
extern void SimpleLruTruncate(SlruCtl ctl, int cutoffPage);
diff --git a/src/include/storage/lwlock.h b/src/include/storage/lwlock.h
index af9b41795d2..2684c8e51c9 100644
--- a/src/include/storage/lwlock.h
+++ b/src/include/storage/lwlock.h
@@ -129,6 +129,8 @@ extern PGDLLIMPORT int NamedLWLockTrancheRequests;
typedef enum LWLockMode
{
+ LW_NONE, /* Not a lock mode. Indicates that there is
+ * no lock. */
LW_EXCLUSIVE,
LW_SHARED,
LW_WAIT_UNTIL_FREE /* A special mode used in PGPROC->lwWaitMode,
--
2.14.3
[application/octet-stream] v4-0002-Make-MultiXact-local-cache-size-configurable.patch (3.7K, ../../CAPpHfduekT=BrM3rLqvo4ajzKYmDk+6aCoYJrNHhK0Je+v2z=Q@mail.gmail.com/3-v4-0002-Make-MultiXact-local-cache-size-configurable.patch)
download | inline diff:
From b1af887f4c4e5402dce1b6ea226ac8b81b92443f Mon Sep 17 00:00:00 2001
From: Alexander Korotkov <[email protected]>
Date: Mon, 26 Oct 2020 03:44:50 +0300
Subject: [PATCH 2/3] Make MultiXact local cache size configurable
---
doc/src/sgml/config.sgml | 16 ++++++++++++++++
src/backend/access/transam/multixact.c | 2 +-
src/backend/utils/init/globals.c | 2 ++
src/backend/utils/misc/guc.c | 10 ++++++++++
src/include/miscadmin.h | 2 ++
5 files changed, 31 insertions(+), 1 deletion(-)
diff --git a/doc/src/sgml/config.sgml b/doc/src/sgml/config.sgml
index f043433e318..fd4ca293476 100644
--- a/doc/src/sgml/config.sgml
+++ b/doc/src/sgml/config.sgml
@@ -1823,6 +1823,22 @@ include_dir 'conf.d'
</listitem>
</varlistentry>
+ <varlistentry id="guc-multixact-local-cache-entries" xreflabel="multixact_local-cache-entries">
+ <term><varname>multixact_local_cache_entries</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>multixact_local_cache_entries</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the number cached MultiXact by backend. Any SLRU lookup is preceeded
+ by cache lookup. Higher numbers of cache size help to deduplicate lock sets, while
+ lower values make cache lookup faster.
+ It defaults to 256.
+ </para>
+ </listitem>
+ </varlistentry>
+
<varlistentry id="guc-max-stack-depth" xreflabel="max_stack_depth">
<term><varname>max_stack_depth</varname> (<type>integer</type>)
<indexterm>
diff --git a/src/backend/access/transam/multixact.c b/src/backend/access/transam/multixact.c
index ccbce90f0ea..57be24c0cc1 100644
--- a/src/backend/access/transam/multixact.c
+++ b/src/backend/access/transam/multixact.c
@@ -1613,7 +1613,7 @@ mXactCachePut(MultiXactId multi, int nmembers, MultiXactMember *members)
qsort(entry->members, nmembers, sizeof(MultiXactMember), mxactMemberComparator);
dlist_push_head(&MXactCache, &entry->node);
- if (MXactCacheMembers++ >= MAX_CACHE_ENTRIES)
+ if (MXactCacheMembers++ >= multixact_local_cache_entries)
{
dlist_node *node;
mXactCacheEnt *entry;
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
index 6ab82168398..9ca71933dcc 100644
--- a/src/backend/utils/init/globals.c
+++ b/src/backend/utils/init/globals.c
@@ -149,3 +149,5 @@ int VacuumCostBalance = 0; /* working state for vacuum */
bool VacuumCostActive = false;
double vacuum_cleanup_index_scale_factor;
+
+int multixact_local_cache_entries = 256;
diff --git a/src/backend/utils/misc/guc.c b/src/backend/utils/misc/guc.c
index a62d64eaa47..d667578cc33 100644
--- a/src/backend/utils/misc/guc.c
+++ b/src/backend/utils/misc/guc.c
@@ -2257,6 +2257,16 @@ static struct config_int ConfigureNamesInt[] =
NULL, NULL, NULL
},
+ {
+ {"multixact_local_cache_entries", PGC_SUSET, RESOURCES_MEM,
+ gettext_noop("Sets the number of cached MultiXact by backend."),
+ NULL
+ },
+ &multixact_local_cache_entries,
+ 256, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
{
{"temp_buffers", PGC_USERSET, RESOURCES_MEM,
gettext_noop("Sets the maximum number of temporary buffers used by each session."),
diff --git a/src/include/miscadmin.h b/src/include/miscadmin.h
index 72e33523984..01af61c963a 100644
--- a/src/include/miscadmin.h
+++ b/src/include/miscadmin.h
@@ -162,6 +162,8 @@ extern PGDLLIMPORT int MaxConnections;
extern PGDLLIMPORT int max_worker_processes;
extern PGDLLIMPORT int max_parallel_workers;
+extern PGDLLIMPORT int multixact_local_cache_entries;
+
extern PGDLLIMPORT int MyProcPid;
extern PGDLLIMPORT pg_time_t MyStartTime;
extern PGDLLIMPORT TimestampTz MyStartTimestamp;
--
2.14.3
[application/octet-stream] v4-0003-Add-conditional-variable-to-wait-for-next-MultXact-o.patch (4.2K, ../../CAPpHfduekT=BrM3rLqvo4ajzKYmDk+6aCoYJrNHhK0Je+v2z=Q@mail.gmail.com/4-v4-0003-Add-conditional-variable-to-wait-for-next-MultXact-o.patch)
download | inline diff:
From b878ffe74a10d34d312a7be66df7a53b423aaaec Mon Sep 17 00:00:00 2001
From: Alexander Korotkov <[email protected]>
Date: Mon, 26 Oct 2020 01:55:31 +0300
Subject: [PATCH 3/3] Add conditional variable to wait for next MultXact offset
in corner case
GetMultiXactIdMembers() has a corner case, when the next multixact offset is
not yet set. In this case GetMultiXactIdMembers() has to sleep till this offset
is set. Currently the sleeping is implemented in naive way using pg_sleep()
and retry. This commit implements sleeping with conditional variable, which
provides more efficient way for waiting till the event.
Discussion: https://postgr.es/m/a7f1c4e1-1015-92a4-2bd4-6736bd13d03e%40postgrespro.ru#c496c4e75fc0605094a0e1f763e6a6ec
Author: Andrey Borodin
Reviewed-by: Kyotaro Horiguchi, Daniel Gustafsson
Reviewed-by: Anastasia Lubennikova, Alexander Korotkov
---
src/backend/access/transam/multixact.c | 29 +++++++++++++++++++++++++++--
src/backend/postmaster/pgstat.c | 2 ++
src/include/pgstat.h | 1 +
3 files changed, 30 insertions(+), 2 deletions(-)
diff --git a/src/backend/access/transam/multixact.c b/src/backend/access/transam/multixact.c
index 57be24c0cc1..09df91899ec 100644
--- a/src/backend/access/transam/multixact.c
+++ b/src/backend/access/transam/multixact.c
@@ -82,6 +82,7 @@
#include "lib/ilist.h"
#include "miscadmin.h"
#include "pg_trace.h"
+#include "pgstat.h"
#include "postmaster/autovacuum.h"
#include "storage/lmgr.h"
#include "storage/pmsignal.h"
@@ -233,6 +234,7 @@ typedef struct MultiXactStateData
/* support for members anti-wraparound measures */
MultiXactOffset offsetStopLimit; /* known if oldestOffsetKnown */
+ ConditionVariable nextoff_cv;
/*
* Per-backend data starts here. We have two arrays stored in the area
* immediately following the MultiXactStateData struct. Each is indexed by
@@ -892,6 +894,14 @@ RecordNewMultiXact(MultiXactId multi, MultiXactOffset offset,
/* Exchange our lock */
LWLockRelease(MultiXactOffsetSLRULock);
+ /*
+ * Let everybody know the offset of this mxid is recorded now. The waiters
+ * are waiting for the offset of the mxid next of the target to know the
+ * number of members of the target mxid, so we don't need to wait for
+ * members of this mxid are recorded.
+ */
+ ConditionVariableBroadcast(&MultiXactState->nextoff_cv);
+
LWLockAcquire(MultiXactMemberSLRULock, LW_EXCLUSIVE);
prev_pageno = -1;
@@ -1389,9 +1399,23 @@ retry:
if (nextMXOffset == 0)
{
/* Corner case 2: next multixact is still being filled in */
+
+ /*
+ * The recorder of the next mxid is just before writing the offset.
+ * Wait for the offset to be written.
+ */
+ ConditionVariablePrepareToSleep(&MultiXactState->nextoff_cv);
+
+ /*
+ * We don't have to recheck if multixact was filled in during
+ * ConditionVariablePrepareToSleep(), because we were holding
+ * MultiXactOffsetSLRULock.
+ */
LWLockRelease(MultiXactOffsetSLRULock);
- CHECK_FOR_INTERRUPTS();
- pg_usleep(1000L);
+
+ ConditionVariableSleep(&MultiXactState->nextoff_cv,
+ WAIT_EVENT_WAIT_NEXT_MXMEMBERS);
+ ConditionVariableCancelSleep();
goto retry;
}
@@ -1873,6 +1897,7 @@ MultiXactShmemInit(void)
/* Make sure we zero out the per-backend state */
MemSet(MultiXactState, 0, SHARED_MULTIXACT_STATE_SIZE);
+ ConditionVariableInit(&MultiXactState->nextoff_cv);
}
else
Assert(found);
diff --git a/src/backend/postmaster/pgstat.c b/src/backend/postmaster/pgstat.c
index 822f0ebc628..b99398a97e9 100644
--- a/src/backend/postmaster/pgstat.c
+++ b/src/backend/postmaster/pgstat.c
@@ -4020,6 +4020,8 @@ pgstat_get_wait_ipc(WaitEventIPC w)
break;
case WAIT_EVENT_XACT_GROUP_UPDATE:
event_name = "XactGroupUpdate";
+ case WAIT_EVENT_WAIT_NEXT_MXMEMBERS:
+ event_name = "MultiXactWaitNextMembers";
break;
/* no default case, so that compiler will warn */
}
diff --git a/src/include/pgstat.h b/src/include/pgstat.h
index a821ff4f158..75ede141ab7 100644
--- a/src/include/pgstat.h
+++ b/src/include/pgstat.h
@@ -952,6 +952,7 @@ typedef enum
WAIT_EVENT_REPLICATION_SLOT_DROP,
WAIT_EVENT_SAFE_SNAPSHOT,
WAIT_EVENT_SYNC_REP,
+ WAIT_EVENT_WAIT_NEXT_MXMEMBERS,
WAIT_EVENT_XACT_GROUP_UPDATE
} WaitEventIPC;
--
2.14.3
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2020-10-26 15:45 Andrey Borodin <[email protected]>
parent: Alexander Korotkov <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Andrey Borodin @ 2020-10-26 15:45 UTC (permalink / raw)
To: Alexander Korotkov <[email protected]>; +Cc: Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
> 26 окт. 2020 г., в 06:05, Alexander Korotkov <[email protected]> написал(а):
>
> Hi!
>
> On Mon, Sep 28, 2020 at 5:41 PM Andrey M. Borodin <[email protected]> wrote:
>>> 28 авг. 2020 г., в 23:08, Anastasia Lubennikova <[email protected]> написал(а):
>>>
>>> 1) The first patch is sensible and harmless, so I think it is ready for committer. I haven't tested the performance impact, though.
>>>
>>> 2) I like the initial proposal to make various SLRU buffers configurable, however, I doubt if it really solves the problem, or just moves it to another place?
>>>
>>> The previous patch you sent was based on some version that contained changes for other slru buffers numbers: 'multixact_offsets_slru_buffers' and 'multixact_members_slru_buffers'. Have you just forgot to attach them? The patch message "[PATCH v2 2/4]" hints that you had 4 patches)
>>> Meanwhile, I attach the rebased patch to calm down the CFbot. The changes are trivial.
>>>
>>> 2.1) I think that both min and max values for this parameter are too extreme. Have you tested them?
>>>
>>> + &multixact_local_cache_entries,
>>> + 256, 2, INT_MAX / 2,
>>>
>>> 2.2) MAX_CACHE_ENTRIES is not used anymore, so it can be deleted.
>>>
>>> 3) No changes for third patch. I just renamed it for consistency.
>>
>> Thank you for your review.
>>
>> Indeed, I had 4th patch with tests, but these tests didn't work well: I still did not manage to stress SLRUs to reproduce problem from production...
>>
>> You are absolutely correct in point 2: I did only tests with sane values. And observed extreme performance degradation with values ~ 64 megabytes. I do not know which highest values should we pick? 1Gb? Or highest possible functioning value?
>>
>> I greatly appreciate your review, sorry for so long delay. Thanks!
>
> I took a look at this patchset.
>
> The 1st and 3rd patches look good to me. I made just minor improvements.
> 1) There is still a case when SimpleLruReadPage_ReadOnly() relocks the
> SLRU lock, which is already taken in exclusive mode. I've evaded this
> by passing the lock mode as a parameter to
> SimpleLruReadPage_ReadOnly().
> 3) CHECK_FOR_INTERRUPTS() is not needed anymore, because it's called
> inside ConditionVariableSleep() if needed. Also, no current wait
> events use slashes, and I don't think we should introduce slashes
> here. Even if we should, then we should also rename existing wait
> events to be consistent with a new one. So, I've renamed the new wait
> event to remove the slash.
>
> Regarding the patch 2. I see the current documentation in the patch
> doesn't explain to the user how to set the new parameter. I think we
> should give users an idea what workloads need high values of
> multixact_local_cache_entries parameter and what doesn't. Also, we
> should explain the negative aspects of high values
> multixact_local_cache_entries. Ideally, we should get the advantage
> without overloading users with new nontrivial parameters, but I don't
> have a particular idea here.
>
> I'd like to propose committing 1 and 3, but leave 2 for further review.
Thanks for your review, Alexander!
+1 for avoiding double locking in SimpleLruReadPage_ReadOnly().
Other changes seem correct to me too.
I've tried to find optimal value for cache size and it seems to me that it affects multixact scalability much less than sizes of offsets\members buffers. I concur that patch 2 of the patchset does not seem documented enough.
Best regards, Andrey Borodin.
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2020-10-27 17:02 Alexander Korotkov <[email protected]>
parent: Andrey Borodin <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Alexander Korotkov @ 2020-10-27 17:02 UTC (permalink / raw)
To: Andrey Borodin <[email protected]>; +Cc: Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
On Mon, Oct 26, 2020 at 6:45 PM Andrey Borodin <[email protected]> wrote:
> Thanks for your review, Alexander!
> +1 for avoiding double locking in SimpleLruReadPage_ReadOnly().
> Other changes seem correct to me too.
>
>
> I've tried to find optimal value for cache size and it seems to me that it affects multixact scalability much less than sizes of offsets\members buffers. I concur that patch 2 of the patchset does not seem documented enough.
Thank you. I've made a few more minor adjustments to the patchset.
I'm going to push 0001 and 0003 if there are no objections.
------
Regards,
Alexander Korotkov
Attachments:
[application/octet-stream] v5-0001-Use-shared-lock-in-GetMultiXactIdMembers-for-offs.patch (12.9K, ../../CAPpHfduT0Kb+8fMt4t0EF5G_3hig+wYgeRadQnNq20F4VxP7LA@mail.gmail.com/2-v5-0001-Use-shared-lock-in-GetMultiXactIdMembers-for-offs.patch)
download | inline diff:
From e81bccfe1dbf0403163b8e1e836211e0dab1d7a4 Mon Sep 17 00:00:00 2001
From: Alexander Korotkov <[email protected]>
Date: Tue, 27 Oct 2020 19:29:56 +0300
Subject: [PATCH 1/3] Use shared lock in GetMultiXactIdMembers for offsets and
members
Previously the read of multixact required exclusive control locks for both
offsets and members SLRUs. This could lead to the excessive lock contention.
This commit we makes multixacts SLRU take advantage of SimpleLruReadPage_ReadOnly
similar to clog, commit_ts and subtrans.
In order to evade extra reacquiring of CLRU lock, we teach
SimpleLruReadPage_ReadOnly() to take into account the current lock mode and
report resulting lock mode back.
Discussion: https://postgr.es/m/a7f1c4e1-1015-92a4-2bd4-6736bd13d03e%40postgrespro.ru#c496c4e75fc0605094a0e1f763e6a6ec
Author: Andrey Borodin
Reviewed-by: Kyotaro Horiguchi, Daniel Gustafsson
Reviewed-by: Anastasia Lubennikova, Alexander Korotkov
---
src/backend/access/transam/clog.c | 4 +++-
src/backend/access/transam/commit_ts.c | 4 +++-
src/backend/access/transam/multixact.c | 22 +++++++++++++++-------
src/backend/access/transam/slru.c | 23 +++++++++++++++++------
src/backend/access/transam/subtrans.c | 4 +++-
src/backend/commands/async.c | 4 +++-
src/backend/storage/lmgr/lwlock.c | 3 +++
src/backend/storage/lmgr/predicate.c | 4 +++-
src/include/access/slru.h | 2 +-
src/include/storage/lwlock.h | 2 ++
10 files changed, 53 insertions(+), 19 deletions(-)
diff --git a/src/backend/access/transam/clog.c b/src/backend/access/transam/clog.c
index 034349aa7b9..4c372065dee 100644
--- a/src/backend/access/transam/clog.c
+++ b/src/backend/access/transam/clog.c
@@ -640,10 +640,11 @@ TransactionIdGetStatus(TransactionId xid, XLogRecPtr *lsn)
int lsnindex;
char *byteptr;
XidStatus status;
+ LWLockMode lockmode = LW_NONE;
/* lock is acquired by SimpleLruReadPage_ReadOnly */
- slotno = SimpleLruReadPage_ReadOnly(XactCtl, pageno, xid);
+ slotno = SimpleLruReadPage_ReadOnly(XactCtl, pageno, xid, &lockmode);
byteptr = XactCtl->shared->page_buffer[slotno] + byteno;
status = (*byteptr >> bshift) & CLOG_XACT_BITMASK;
@@ -651,6 +652,7 @@ TransactionIdGetStatus(TransactionId xid, XLogRecPtr *lsn)
lsnindex = GetLSNIndex(slotno, xid);
*lsn = XactCtl->shared->group_lsn[lsnindex];
+ Assert(lockmode != LW_NONE);
LWLockRelease(XactSLRULock);
return status;
diff --git a/src/backend/access/transam/commit_ts.c b/src/backend/access/transam/commit_ts.c
index cb8a9688018..c19b0b5399e 100644
--- a/src/backend/access/transam/commit_ts.c
+++ b/src/backend/access/transam/commit_ts.c
@@ -288,6 +288,7 @@ TransactionIdGetCommitTsData(TransactionId xid, TimestampTz *ts,
CommitTimestampEntry entry;
TransactionId oldestCommitTsXid;
TransactionId newestCommitTsXid;
+ LWLockMode lockmode = LW_NONE;
if (!TransactionIdIsValid(xid))
ereport(ERROR,
@@ -342,7 +343,7 @@ TransactionIdGetCommitTsData(TransactionId xid, TimestampTz *ts,
}
/* lock is acquired by SimpleLruReadPage_ReadOnly */
- slotno = SimpleLruReadPage_ReadOnly(CommitTsCtl, pageno, xid);
+ slotno = SimpleLruReadPage_ReadOnly(CommitTsCtl, pageno, xid, &lockmode);
memcpy(&entry,
CommitTsCtl->shared->page_buffer[slotno] +
SizeOfCommitTimestampEntry * entryno,
@@ -352,6 +353,7 @@ TransactionIdGetCommitTsData(TransactionId xid, TimestampTz *ts,
if (nodeid)
*nodeid = entry.nodeid;
+ Assert(lockmode != LW_NONE);
LWLockRelease(CommitTsSLRULock);
return *ts != 0;
}
diff --git a/src/backend/access/transam/multixact.c b/src/backend/access/transam/multixact.c
index 43653fe5721..ccbce90f0ea 100644
--- a/src/backend/access/transam/multixact.c
+++ b/src/backend/access/transam/multixact.c
@@ -1237,6 +1237,7 @@ GetMultiXactIdMembers(MultiXactId multi, MultiXactMember **members,
MultiXactId tmpMXact;
MultiXactOffset nextOffset;
MultiXactMember *ptr;
+ LWLockMode lockmode = LW_NONE;
debug_elog3(DEBUG2, "GetMembers: asked for %u", multi);
@@ -1340,12 +1341,13 @@ GetMultiXactIdMembers(MultiXactId multi, MultiXactMember **members,
* time on every multixact creation.
*/
retry:
- LWLockAcquire(MultiXactOffsetSLRULock, LW_EXCLUSIVE);
pageno = MultiXactIdToOffsetPage(multi);
entryno = MultiXactIdToOffsetEntry(multi);
- slotno = SimpleLruReadPage(MultiXactOffsetCtl, pageno, true, multi);
+ /* lock is acquired by SimpleLruReadPage_ReadOnly */
+ slotno = SimpleLruReadPage_ReadOnly(MultiXactOffsetCtl, pageno,
+ multi, &lockmode);
offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
offptr += entryno;
offset = *offptr;
@@ -1377,7 +1379,8 @@ retry:
entryno = MultiXactIdToOffsetEntry(tmpMXact);
if (pageno != prev_pageno)
- slotno = SimpleLruReadPage(MultiXactOffsetCtl, pageno, true, tmpMXact);
+ slotno = SimpleLruReadPage_ReadOnly(MultiXactOffsetCtl, pageno,
+ tmpMXact, &lockmode);
offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
offptr += entryno;
@@ -1395,14 +1398,14 @@ retry:
length = nextMXOffset - offset;
}
+ Assert(lockmode != LW_NONE);
LWLockRelease(MultiXactOffsetSLRULock);
ptr = (MultiXactMember *) palloc(length * sizeof(MultiXactMember));
*members = ptr;
/* Now get the members themselves. */
- LWLockAcquire(MultiXactMemberSLRULock, LW_EXCLUSIVE);
-
+ lockmode = LW_NONE;
truelength = 0;
prev_pageno = -1;
for (i = 0; i < length; i++, offset++)
@@ -1418,7 +1421,8 @@ retry:
if (pageno != prev_pageno)
{
- slotno = SimpleLruReadPage(MultiXactMemberCtl, pageno, true, multi);
+ slotno = SimpleLruReadPage_ReadOnly(MultiXactMemberCtl, pageno,
+ multi, &lockmode);
prev_pageno = pageno;
}
@@ -1441,6 +1445,7 @@ retry:
truelength++;
}
+ Assert(lockmode != LW_NONE);
LWLockRelease(MultiXactMemberSLRULock);
/*
@@ -2733,6 +2738,7 @@ find_multixact_start(MultiXactId multi, MultiXactOffset *result)
int entryno;
int slotno;
MultiXactOffset *offptr;
+ LWLockMode lockmode = LW_NONE;
Assert(MultiXactState->finishedStartup);
@@ -2749,10 +2755,12 @@ find_multixact_start(MultiXactId multi, MultiXactOffset *result)
return false;
/* lock is acquired by SimpleLruReadPage_ReadOnly */
- slotno = SimpleLruReadPage_ReadOnly(MultiXactOffsetCtl, pageno, multi);
+ slotno = SimpleLruReadPage_ReadOnly(MultiXactOffsetCtl, pageno,
+ multi, &lockmode);
offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
offptr += entryno;
offset = *offptr;
+ Assert(lockmode != LW_NONE);
LWLockRelease(MultiXactOffsetSLRULock);
*result = offset;
diff --git a/src/backend/access/transam/slru.c b/src/backend/access/transam/slru.c
index 16a78986971..ca0a23ab875 100644
--- a/src/backend/access/transam/slru.c
+++ b/src/backend/access/transam/slru.c
@@ -487,17 +487,23 @@ SimpleLruReadPage(SlruCtl ctl, int pageno, bool write_ok,
* Return value is the shared-buffer slot number now holding the page.
* The buffer's LRU access info is updated.
*
- * Control lock must NOT be held at entry, but will be held at exit.
- * It is unspecified whether the lock will be shared or exclusive.
+ * Control lock must be held in *lock_mode mode, which may be LW_NONE. Control
+ * lock will be held at exit in at least shared mode. Resulting control lock
+ * mode is set to *lock_mode.
*/
int
-SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno, TransactionId xid)
+SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno, TransactionId xid,
+ LWLockMode *lock_mode)
{
SlruShared shared = ctl->shared;
int slotno;
/* Try to find the page while holding only shared lock */
- LWLockAcquire(shared->ControlLock, LW_SHARED);
+ if (*lock_mode == LW_NONE)
+ {
+ LWLockAcquire(shared->ControlLock, LW_SHARED);
+ *lock_mode = LW_SHARED;
+ }
/* See if page is already in a buffer */
for (slotno = 0; slotno < shared->num_slots; slotno++)
@@ -517,8 +523,13 @@ SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno, TransactionId xid)
}
/* No luck, so switch to normal exclusive lock and do regular read */
- LWLockRelease(shared->ControlLock);
- LWLockAcquire(shared->ControlLock, LW_EXCLUSIVE);
+ if (*lock_mode != LW_EXCLUSIVE)
+ {
+ Assert(*lock_mode == LW_NONE);
+ LWLockRelease(shared->ControlLock);
+ LWLockAcquire(shared->ControlLock, LW_EXCLUSIVE);
+ *lock_mode = LW_EXCLUSIVE;
+ }
return SimpleLruReadPage(ctl, pageno, true, xid);
}
diff --git a/src/backend/access/transam/subtrans.c b/src/backend/access/transam/subtrans.c
index 0111e867c79..7bb15431893 100644
--- a/src/backend/access/transam/subtrans.c
+++ b/src/backend/access/transam/subtrans.c
@@ -113,6 +113,7 @@ SubTransGetParent(TransactionId xid)
int slotno;
TransactionId *ptr;
TransactionId parent;
+ LWLockMode lockmode = LW_NONE;
/* Can't ask about stuff that might not be around anymore */
Assert(TransactionIdFollowsOrEquals(xid, TransactionXmin));
@@ -123,12 +124,13 @@ SubTransGetParent(TransactionId xid)
/* lock is acquired by SimpleLruReadPage_ReadOnly */
- slotno = SimpleLruReadPage_ReadOnly(SubTransCtl, pageno, xid);
+ slotno = SimpleLruReadPage_ReadOnly(SubTransCtl, pageno, xid, &lockmode);
ptr = (TransactionId *) SubTransCtl->shared->page_buffer[slotno];
ptr += entryno;
parent = *ptr;
+ Assert(lockmode != LW_NONE);
LWLockRelease(SubtransSLRULock);
return parent;
diff --git a/src/backend/commands/async.c b/src/backend/commands/async.c
index 8dbcace3f93..8b419575590 100644
--- a/src/backend/commands/async.c
+++ b/src/backend/commands/async.c
@@ -2002,6 +2002,7 @@ asyncQueueReadAllNotifications(void)
int curoffset = QUEUE_POS_OFFSET(pos);
int slotno;
int copysize;
+ LWLockMode lockmode = LW_NONE;
/*
* We copy the data from SLRU into a local buffer, so as to avoid
@@ -2010,7 +2011,7 @@ asyncQueueReadAllNotifications(void)
* part of the page we will actually inspect.
*/
slotno = SimpleLruReadPage_ReadOnly(NotifyCtl, curpage,
- InvalidTransactionId);
+ InvalidTransactionId, &lockmode);
if (curpage == QUEUE_POS_PAGE(head))
{
/* we only want to read as far as head */
@@ -2027,6 +2028,7 @@ asyncQueueReadAllNotifications(void)
NotifyCtl->shared->page_buffer[slotno] + curoffset,
copysize);
/* Release lock that we got from SimpleLruReadPage_ReadOnly() */
+ Assert(lockmode != LW_NONE);
LWLockRelease(NotifySLRULock);
/*
diff --git a/src/backend/storage/lmgr/lwlock.c b/src/backend/storage/lmgr/lwlock.c
index 2fa90cc0954..8a7726f11c0 100644
--- a/src/backend/storage/lmgr/lwlock.c
+++ b/src/backend/storage/lmgr/lwlock.c
@@ -1068,6 +1068,8 @@ LWLockWakeup(LWLock *lock)
static void
LWLockQueueSelf(LWLock *lock, LWLockMode mode)
{
+ Assert(mode != LW_NONE);
+
/*
* If we don't have a PGPROC structure, there's no way to wait. This
* should never occur, since MyProc should only be null during shared
@@ -1828,6 +1830,7 @@ LWLockRelease(LWLock *lock)
elog(ERROR, "lock %s is not held", T_NAME(lock));
mode = held_lwlocks[i].mode;
+ Assert(mode == LW_EXCLUSIVE || mode == LW_SHARED);
num_held_lwlocks--;
for (; i < num_held_lwlocks; i++)
diff --git a/src/backend/storage/lmgr/predicate.c b/src/backend/storage/lmgr/predicate.c
index 8a365b400c6..a4df90a8aeb 100644
--- a/src/backend/storage/lmgr/predicate.c
+++ b/src/backend/storage/lmgr/predicate.c
@@ -924,6 +924,7 @@ SerialGetMinConflictCommitSeqNo(TransactionId xid)
TransactionId tailXid;
SerCommitSeqNo val;
int slotno;
+ LWLockMode lockmode = LW_NONE;
Assert(TransactionIdIsValid(xid));
@@ -946,8 +947,9 @@ SerialGetMinConflictCommitSeqNo(TransactionId xid)
* but will return with that lock held, which must then be released.
*/
slotno = SimpleLruReadPage_ReadOnly(SerialSlruCtl,
- SerialPage(xid), xid);
+ SerialPage(xid), xid, &lockmode);
val = SerialValue(slotno, xid);
+ Assert(lockmode != LW_NONE);
LWLockRelease(SerialSLRULock);
return val;
}
diff --git a/src/include/access/slru.h b/src/include/access/slru.h
index b39b43504d8..4b66d3b592b 100644
--- a/src/include/access/slru.h
+++ b/src/include/access/slru.h
@@ -142,7 +142,7 @@ extern int SimpleLruZeroPage(SlruCtl ctl, int pageno);
extern int SimpleLruReadPage(SlruCtl ctl, int pageno, bool write_ok,
TransactionId xid);
extern int SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno,
- TransactionId xid);
+ TransactionId xid, LWLockMode *lock_mode);
extern void SimpleLruWritePage(SlruCtl ctl, int slotno);
extern void SimpleLruWriteAll(SlruCtl ctl, bool allow_redirtied);
extern void SimpleLruTruncate(SlruCtl ctl, int cutoffPage);
diff --git a/src/include/storage/lwlock.h b/src/include/storage/lwlock.h
index af9b41795d2..e680c6397aa 100644
--- a/src/include/storage/lwlock.h
+++ b/src/include/storage/lwlock.h
@@ -129,6 +129,8 @@ extern PGDLLIMPORT int NamedLWLockTrancheRequests;
typedef enum LWLockMode
{
+ LW_NONE, /* Not a lock mode. Indicates that there is no
+ * lock. */
LW_EXCLUSIVE,
LW_SHARED,
LW_WAIT_UNTIL_FREE /* A special mode used in PGPROC->lwWaitMode,
--
2.14.3
[application/octet-stream] v5-0002-Make-MultiXact-local-cache-size-configurable.patch (3.7K, ../../CAPpHfduT0Kb+8fMt4t0EF5G_3hig+wYgeRadQnNq20F4VxP7LA@mail.gmail.com/3-v5-0002-Make-MultiXact-local-cache-size-configurable.patch)
download | inline diff:
From 53ef71094688048002ffd1a62dbbb1a4a90d4691 Mon Sep 17 00:00:00 2001
From: Alexander Korotkov <[email protected]>
Date: Mon, 26 Oct 2020 03:44:50 +0300
Subject: [PATCH 2/3] Make MultiXact local cache size configurable
---
doc/src/sgml/config.sgml | 16 ++++++++++++++++
src/backend/access/transam/multixact.c | 2 +-
src/backend/utils/init/globals.c | 2 ++
src/backend/utils/misc/guc.c | 10 ++++++++++
src/include/miscadmin.h | 2 ++
5 files changed, 31 insertions(+), 1 deletion(-)
diff --git a/doc/src/sgml/config.sgml b/doc/src/sgml/config.sgml
index f043433e318..fd4ca293476 100644
--- a/doc/src/sgml/config.sgml
+++ b/doc/src/sgml/config.sgml
@@ -1823,6 +1823,22 @@ include_dir 'conf.d'
</listitem>
</varlistentry>
+ <varlistentry id="guc-multixact-local-cache-entries" xreflabel="multixact_local-cache-entries">
+ <term><varname>multixact_local_cache_entries</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>multixact_local_cache_entries</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the number cached MultiXact by backend. Any SLRU lookup is preceeded
+ by cache lookup. Higher numbers of cache size help to deduplicate lock sets, while
+ lower values make cache lookup faster.
+ It defaults to 256.
+ </para>
+ </listitem>
+ </varlistentry>
+
<varlistentry id="guc-max-stack-depth" xreflabel="max_stack_depth">
<term><varname>max_stack_depth</varname> (<type>integer</type>)
<indexterm>
diff --git a/src/backend/access/transam/multixact.c b/src/backend/access/transam/multixact.c
index ccbce90f0ea..57be24c0cc1 100644
--- a/src/backend/access/transam/multixact.c
+++ b/src/backend/access/transam/multixact.c
@@ -1613,7 +1613,7 @@ mXactCachePut(MultiXactId multi, int nmembers, MultiXactMember *members)
qsort(entry->members, nmembers, sizeof(MultiXactMember), mxactMemberComparator);
dlist_push_head(&MXactCache, &entry->node);
- if (MXactCacheMembers++ >= MAX_CACHE_ENTRIES)
+ if (MXactCacheMembers++ >= multixact_local_cache_entries)
{
dlist_node *node;
mXactCacheEnt *entry;
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
index 6ab82168398..9ca71933dcc 100644
--- a/src/backend/utils/init/globals.c
+++ b/src/backend/utils/init/globals.c
@@ -149,3 +149,5 @@ int VacuumCostBalance = 0; /* working state for vacuum */
bool VacuumCostActive = false;
double vacuum_cleanup_index_scale_factor;
+
+int multixact_local_cache_entries = 256;
diff --git a/src/backend/utils/misc/guc.c b/src/backend/utils/misc/guc.c
index a62d64eaa47..d667578cc33 100644
--- a/src/backend/utils/misc/guc.c
+++ b/src/backend/utils/misc/guc.c
@@ -2257,6 +2257,16 @@ static struct config_int ConfigureNamesInt[] =
NULL, NULL, NULL
},
+ {
+ {"multixact_local_cache_entries", PGC_SUSET, RESOURCES_MEM,
+ gettext_noop("Sets the number of cached MultiXact by backend."),
+ NULL
+ },
+ &multixact_local_cache_entries,
+ 256, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
{
{"temp_buffers", PGC_USERSET, RESOURCES_MEM,
gettext_noop("Sets the maximum number of temporary buffers used by each session."),
diff --git a/src/include/miscadmin.h b/src/include/miscadmin.h
index 72e33523984..01af61c963a 100644
--- a/src/include/miscadmin.h
+++ b/src/include/miscadmin.h
@@ -162,6 +162,8 @@ extern PGDLLIMPORT int MaxConnections;
extern PGDLLIMPORT int max_worker_processes;
extern PGDLLIMPORT int max_parallel_workers;
+extern PGDLLIMPORT int multixact_local_cache_entries;
+
extern PGDLLIMPORT int MyProcPid;
extern PGDLLIMPORT pg_time_t MyStartTime;
extern PGDLLIMPORT TimestampTz MyStartTimestamp;
--
2.14.3
[application/octet-stream] v5-0003-Add-conditional-variable-to-wait-for-next-MultXact.patch (4.4K, ../../CAPpHfduT0Kb+8fMt4t0EF5G_3hig+wYgeRadQnNq20F4VxP7LA@mail.gmail.com/4-v5-0003-Add-conditional-variable-to-wait-for-next-MultXact.patch)
download | inline diff:
From 8876966a703f35abf98b575978c9d38ff4297473 Mon Sep 17 00:00:00 2001
From: Alexander Korotkov <[email protected]>
Date: Mon, 26 Oct 2020 01:55:31 +0300
Subject: [PATCH 3/3] Add conditional variable to wait for next MultXact offset
in corner case
GetMultiXactIdMembers() has a corner case, when the next multixact offset is
not yet set. In this case GetMultiXactIdMembers() has to sleep till this offset
is set. Currently the sleeping is implemented in naive way using pg_sleep()
and retry. This commit implements sleeping with conditional variable, which
provides more efficient way for waiting till the event.
Discussion: https://postgr.es/m/a7f1c4e1-1015-92a4-2bd4-6736bd13d03e%40postgrespro.ru#c496c4e75fc0605094a0e1f763e6a6ec
Author: Andrey Borodin
Reviewed-by: Kyotaro Horiguchi, Daniel Gustafsson
Reviewed-by: Anastasia Lubennikova, Alexander Korotkov
---
src/backend/access/transam/multixact.c | 35 ++++++++++++++++++++++++++++++++--
src/backend/postmaster/pgstat.c | 2 ++
src/include/pgstat.h | 1 +
3 files changed, 36 insertions(+), 2 deletions(-)
diff --git a/src/backend/access/transam/multixact.c b/src/backend/access/transam/multixact.c
index 57be24c0cc1..70d977cba33 100644
--- a/src/backend/access/transam/multixact.c
+++ b/src/backend/access/transam/multixact.c
@@ -82,6 +82,7 @@
#include "lib/ilist.h"
#include "miscadmin.h"
#include "pg_trace.h"
+#include "pgstat.h"
#include "postmaster/autovacuum.h"
#include "storage/lmgr.h"
#include "storage/pmsignal.h"
@@ -233,6 +234,13 @@ typedef struct MultiXactStateData
/* support for members anti-wraparound measures */
MultiXactOffset offsetStopLimit; /* known if oldestOffsetKnown */
+ /*
+ * Conditional variable for waiting till the filling of the next multixact
+ * will be finished. See GetMultiXactIdMembers() and RecordNewMultiXact()
+ * for details.
+ */
+ ConditionVariable nextoffCV;
+
/*
* Per-backend data starts here. We have two arrays stored in the area
* immediately following the MultiXactStateData struct. Each is indexed by
@@ -892,6 +900,14 @@ RecordNewMultiXact(MultiXactId multi, MultiXactOffset offset,
/* Exchange our lock */
LWLockRelease(MultiXactOffsetSLRULock);
+ /*
+ * Let everybody know the offset of this mxid is recorded now. The
+ * waiters are waiting for the offset of the mxid next of the target to
+ * know the number of members of the target mxid, so we don't need to wait
+ * for members of this mxid are recorded.
+ */
+ ConditionVariableBroadcast(&MultiXactState->nextoffCV);
+
LWLockAcquire(MultiXactMemberSLRULock, LW_EXCLUSIVE);
prev_pageno = -1;
@@ -1389,9 +1405,23 @@ retry:
if (nextMXOffset == 0)
{
/* Corner case 2: next multixact is still being filled in */
+
+ /*
+ * The recorder of the next mxid is just before writing the
+ * offset. Wait for the offset to be written.
+ */
+ ConditionVariablePrepareToSleep(&MultiXactState->nextoffCV);
+
+ /*
+ * We don't have to recheck if multixact was filled in during
+ * ConditionVariablePrepareToSleep(), because we were holding
+ * MultiXactOffsetSLRULock.
+ */
LWLockRelease(MultiXactOffsetSLRULock);
- CHECK_FOR_INTERRUPTS();
- pg_usleep(1000L);
+
+ ConditionVariableSleep(&MultiXactState->nextoffCV,
+ WAIT_EVENT_WAIT_NEXT_MXMEMBERS);
+ ConditionVariableCancelSleep();
goto retry;
}
@@ -1873,6 +1903,7 @@ MultiXactShmemInit(void)
/* Make sure we zero out the per-backend state */
MemSet(MultiXactState, 0, SHARED_MULTIXACT_STATE_SIZE);
+ ConditionVariableInit(&MultiXactState->nextoffCV);
}
else
Assert(found);
diff --git a/src/backend/postmaster/pgstat.c b/src/backend/postmaster/pgstat.c
index 822f0ebc628..b99398a97e9 100644
--- a/src/backend/postmaster/pgstat.c
+++ b/src/backend/postmaster/pgstat.c
@@ -4020,6 +4020,8 @@ pgstat_get_wait_ipc(WaitEventIPC w)
break;
case WAIT_EVENT_XACT_GROUP_UPDATE:
event_name = "XactGroupUpdate";
+ case WAIT_EVENT_WAIT_NEXT_MXMEMBERS:
+ event_name = "MultiXactWaitNextMembers";
break;
/* no default case, so that compiler will warn */
}
diff --git a/src/include/pgstat.h b/src/include/pgstat.h
index a821ff4f158..75ede141ab7 100644
--- a/src/include/pgstat.h
+++ b/src/include/pgstat.h
@@ -952,6 +952,7 @@ typedef enum
WAIT_EVENT_REPLICATION_SLOT_DROP,
WAIT_EVENT_SAFE_SNAPSHOT,
WAIT_EVENT_SYNC_REP,
+ WAIT_EVENT_WAIT_NEXT_MXMEMBERS,
WAIT_EVENT_XACT_GROUP_UPDATE
} WaitEventIPC;
--
2.14.3
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2020-10-27 17:23 Alexander Korotkov <[email protected]>
parent: Alexander Korotkov <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Alexander Korotkov @ 2020-10-27 17:23 UTC (permalink / raw)
To: Andrey Borodin <[email protected]>; +Cc: Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
On Tue, Oct 27, 2020 at 8:02 PM Alexander Korotkov <[email protected]> wrote:
> On Mon, Oct 26, 2020 at 6:45 PM Andrey Borodin <[email protected]> wrote:
> > Thanks for your review, Alexander!
> > +1 for avoiding double locking in SimpleLruReadPage_ReadOnly().
> > Other changes seem correct to me too.
> >
> >
> > I've tried to find optimal value for cache size and it seems to me that it affects multixact scalability much less than sizes of offsets\members buffers. I concur that patch 2 of the patchset does not seem documented enough.
>
> Thank you. I've made a few more minor adjustments to the patchset.
> I'm going to push 0001 and 0003 if there are no objections.
I get that patchset v5 doesn't pass the tests due to typo in assert.
The fixes version is attached.
------
Regards,
Alexander Korotkov
Attachments:
[application/octet-stream] v6-0001-Use-shared-lock-in-GetMultiXactIdMembers-for-offs.patch (12.9K, ../../CAPpHfdsUtSGxRrKn65F_wrE66sNztNY4TpDGDHh2kvAZZDOMqA@mail.gmail.com/2-v6-0001-Use-shared-lock-in-GetMultiXactIdMembers-for-offs.patch)
download | inline diff:
From 9307192bbb9d1ee48aba5c335ae54cff7b34ffb7 Mon Sep 17 00:00:00 2001
From: Alexander Korotkov <[email protected]>
Date: Tue, 27 Oct 2020 19:29:56 +0300
Subject: [PATCH 1/3] Use shared lock in GetMultiXactIdMembers for offsets and
members
Previously the read of multixact required exclusive control locks for both
offsets and members SLRUs. This could lead to the excessive lock contention.
This commit we makes multixacts SLRU take advantage of SimpleLruReadPage_ReadOnly
similar to clog, commit_ts and subtrans.
In order to evade extra reacquiring of CLRU lock, we teach
SimpleLruReadPage_ReadOnly() to take into account the current lock mode and
report resulting lock mode back.
Discussion: https://postgr.es/m/a7f1c4e1-1015-92a4-2bd4-6736bd13d03e%40postgrespro.ru#c496c4e75fc0605094a0e1f763e6a6ec
Author: Andrey Borodin
Reviewed-by: Kyotaro Horiguchi, Daniel Gustafsson
Reviewed-by: Anastasia Lubennikova, Alexander Korotkov
---
src/backend/access/transam/clog.c | 4 +++-
src/backend/access/transam/commit_ts.c | 4 +++-
src/backend/access/transam/multixact.c | 22 +++++++++++++++-------
src/backend/access/transam/slru.c | 23 +++++++++++++++++------
src/backend/access/transam/subtrans.c | 4 +++-
src/backend/commands/async.c | 4 +++-
src/backend/storage/lmgr/lwlock.c | 3 +++
src/backend/storage/lmgr/predicate.c | 4 +++-
src/include/access/slru.h | 2 +-
src/include/storage/lwlock.h | 2 ++
10 files changed, 53 insertions(+), 19 deletions(-)
diff --git a/src/backend/access/transam/clog.c b/src/backend/access/transam/clog.c
index 034349aa7b9..4c372065dee 100644
--- a/src/backend/access/transam/clog.c
+++ b/src/backend/access/transam/clog.c
@@ -640,10 +640,11 @@ TransactionIdGetStatus(TransactionId xid, XLogRecPtr *lsn)
int lsnindex;
char *byteptr;
XidStatus status;
+ LWLockMode lockmode = LW_NONE;
/* lock is acquired by SimpleLruReadPage_ReadOnly */
- slotno = SimpleLruReadPage_ReadOnly(XactCtl, pageno, xid);
+ slotno = SimpleLruReadPage_ReadOnly(XactCtl, pageno, xid, &lockmode);
byteptr = XactCtl->shared->page_buffer[slotno] + byteno;
status = (*byteptr >> bshift) & CLOG_XACT_BITMASK;
@@ -651,6 +652,7 @@ TransactionIdGetStatus(TransactionId xid, XLogRecPtr *lsn)
lsnindex = GetLSNIndex(slotno, xid);
*lsn = XactCtl->shared->group_lsn[lsnindex];
+ Assert(lockmode != LW_NONE);
LWLockRelease(XactSLRULock);
return status;
diff --git a/src/backend/access/transam/commit_ts.c b/src/backend/access/transam/commit_ts.c
index cb8a9688018..c19b0b5399e 100644
--- a/src/backend/access/transam/commit_ts.c
+++ b/src/backend/access/transam/commit_ts.c
@@ -288,6 +288,7 @@ TransactionIdGetCommitTsData(TransactionId xid, TimestampTz *ts,
CommitTimestampEntry entry;
TransactionId oldestCommitTsXid;
TransactionId newestCommitTsXid;
+ LWLockMode lockmode = LW_NONE;
if (!TransactionIdIsValid(xid))
ereport(ERROR,
@@ -342,7 +343,7 @@ TransactionIdGetCommitTsData(TransactionId xid, TimestampTz *ts,
}
/* lock is acquired by SimpleLruReadPage_ReadOnly */
- slotno = SimpleLruReadPage_ReadOnly(CommitTsCtl, pageno, xid);
+ slotno = SimpleLruReadPage_ReadOnly(CommitTsCtl, pageno, xid, &lockmode);
memcpy(&entry,
CommitTsCtl->shared->page_buffer[slotno] +
SizeOfCommitTimestampEntry * entryno,
@@ -352,6 +353,7 @@ TransactionIdGetCommitTsData(TransactionId xid, TimestampTz *ts,
if (nodeid)
*nodeid = entry.nodeid;
+ Assert(lockmode != LW_NONE);
LWLockRelease(CommitTsSLRULock);
return *ts != 0;
}
diff --git a/src/backend/access/transam/multixact.c b/src/backend/access/transam/multixact.c
index 43653fe5721..ccbce90f0ea 100644
--- a/src/backend/access/transam/multixact.c
+++ b/src/backend/access/transam/multixact.c
@@ -1237,6 +1237,7 @@ GetMultiXactIdMembers(MultiXactId multi, MultiXactMember **members,
MultiXactId tmpMXact;
MultiXactOffset nextOffset;
MultiXactMember *ptr;
+ LWLockMode lockmode = LW_NONE;
debug_elog3(DEBUG2, "GetMembers: asked for %u", multi);
@@ -1340,12 +1341,13 @@ GetMultiXactIdMembers(MultiXactId multi, MultiXactMember **members,
* time on every multixact creation.
*/
retry:
- LWLockAcquire(MultiXactOffsetSLRULock, LW_EXCLUSIVE);
pageno = MultiXactIdToOffsetPage(multi);
entryno = MultiXactIdToOffsetEntry(multi);
- slotno = SimpleLruReadPage(MultiXactOffsetCtl, pageno, true, multi);
+ /* lock is acquired by SimpleLruReadPage_ReadOnly */
+ slotno = SimpleLruReadPage_ReadOnly(MultiXactOffsetCtl, pageno,
+ multi, &lockmode);
offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
offptr += entryno;
offset = *offptr;
@@ -1377,7 +1379,8 @@ retry:
entryno = MultiXactIdToOffsetEntry(tmpMXact);
if (pageno != prev_pageno)
- slotno = SimpleLruReadPage(MultiXactOffsetCtl, pageno, true, tmpMXact);
+ slotno = SimpleLruReadPage_ReadOnly(MultiXactOffsetCtl, pageno,
+ tmpMXact, &lockmode);
offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
offptr += entryno;
@@ -1395,14 +1398,14 @@ retry:
length = nextMXOffset - offset;
}
+ Assert(lockmode != LW_NONE);
LWLockRelease(MultiXactOffsetSLRULock);
ptr = (MultiXactMember *) palloc(length * sizeof(MultiXactMember));
*members = ptr;
/* Now get the members themselves. */
- LWLockAcquire(MultiXactMemberSLRULock, LW_EXCLUSIVE);
-
+ lockmode = LW_NONE;
truelength = 0;
prev_pageno = -1;
for (i = 0; i < length; i++, offset++)
@@ -1418,7 +1421,8 @@ retry:
if (pageno != prev_pageno)
{
- slotno = SimpleLruReadPage(MultiXactMemberCtl, pageno, true, multi);
+ slotno = SimpleLruReadPage_ReadOnly(MultiXactMemberCtl, pageno,
+ multi, &lockmode);
prev_pageno = pageno;
}
@@ -1441,6 +1445,7 @@ retry:
truelength++;
}
+ Assert(lockmode != LW_NONE);
LWLockRelease(MultiXactMemberSLRULock);
/*
@@ -2733,6 +2738,7 @@ find_multixact_start(MultiXactId multi, MultiXactOffset *result)
int entryno;
int slotno;
MultiXactOffset *offptr;
+ LWLockMode lockmode = LW_NONE;
Assert(MultiXactState->finishedStartup);
@@ -2749,10 +2755,12 @@ find_multixact_start(MultiXactId multi, MultiXactOffset *result)
return false;
/* lock is acquired by SimpleLruReadPage_ReadOnly */
- slotno = SimpleLruReadPage_ReadOnly(MultiXactOffsetCtl, pageno, multi);
+ slotno = SimpleLruReadPage_ReadOnly(MultiXactOffsetCtl, pageno,
+ multi, &lockmode);
offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
offptr += entryno;
offset = *offptr;
+ Assert(lockmode != LW_NONE);
LWLockRelease(MultiXactOffsetSLRULock);
*result = offset;
diff --git a/src/backend/access/transam/slru.c b/src/backend/access/transam/slru.c
index 16a78986971..cf89006e131 100644
--- a/src/backend/access/transam/slru.c
+++ b/src/backend/access/transam/slru.c
@@ -487,17 +487,23 @@ SimpleLruReadPage(SlruCtl ctl, int pageno, bool write_ok,
* Return value is the shared-buffer slot number now holding the page.
* The buffer's LRU access info is updated.
*
- * Control lock must NOT be held at entry, but will be held at exit.
- * It is unspecified whether the lock will be shared or exclusive.
+ * Control lock must be held in *lock_mode mode, which may be LW_NONE. Control
+ * lock will be held at exit in at least shared mode. Resulting control lock
+ * mode is set to *lock_mode.
*/
int
-SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno, TransactionId xid)
+SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno, TransactionId xid,
+ LWLockMode *lock_mode)
{
SlruShared shared = ctl->shared;
int slotno;
/* Try to find the page while holding only shared lock */
- LWLockAcquire(shared->ControlLock, LW_SHARED);
+ if (*lock_mode == LW_NONE)
+ {
+ LWLockAcquire(shared->ControlLock, LW_SHARED);
+ *lock_mode = LW_SHARED;
+ }
/* See if page is already in a buffer */
for (slotno = 0; slotno < shared->num_slots; slotno++)
@@ -517,8 +523,13 @@ SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno, TransactionId xid)
}
/* No luck, so switch to normal exclusive lock and do regular read */
- LWLockRelease(shared->ControlLock);
- LWLockAcquire(shared->ControlLock, LW_EXCLUSIVE);
+ if (*lock_mode != LW_EXCLUSIVE)
+ {
+ Assert(*lock_mode != LW_NONE);
+ LWLockRelease(shared->ControlLock);
+ LWLockAcquire(shared->ControlLock, LW_EXCLUSIVE);
+ *lock_mode = LW_EXCLUSIVE;
+ }
return SimpleLruReadPage(ctl, pageno, true, xid);
}
diff --git a/src/backend/access/transam/subtrans.c b/src/backend/access/transam/subtrans.c
index 0111e867c79..7bb15431893 100644
--- a/src/backend/access/transam/subtrans.c
+++ b/src/backend/access/transam/subtrans.c
@@ -113,6 +113,7 @@ SubTransGetParent(TransactionId xid)
int slotno;
TransactionId *ptr;
TransactionId parent;
+ LWLockMode lockmode = LW_NONE;
/* Can't ask about stuff that might not be around anymore */
Assert(TransactionIdFollowsOrEquals(xid, TransactionXmin));
@@ -123,12 +124,13 @@ SubTransGetParent(TransactionId xid)
/* lock is acquired by SimpleLruReadPage_ReadOnly */
- slotno = SimpleLruReadPage_ReadOnly(SubTransCtl, pageno, xid);
+ slotno = SimpleLruReadPage_ReadOnly(SubTransCtl, pageno, xid, &lockmode);
ptr = (TransactionId *) SubTransCtl->shared->page_buffer[slotno];
ptr += entryno;
parent = *ptr;
+ Assert(lockmode != LW_NONE);
LWLockRelease(SubtransSLRULock);
return parent;
diff --git a/src/backend/commands/async.c b/src/backend/commands/async.c
index 8dbcace3f93..8b419575590 100644
--- a/src/backend/commands/async.c
+++ b/src/backend/commands/async.c
@@ -2002,6 +2002,7 @@ asyncQueueReadAllNotifications(void)
int curoffset = QUEUE_POS_OFFSET(pos);
int slotno;
int copysize;
+ LWLockMode lockmode = LW_NONE;
/*
* We copy the data from SLRU into a local buffer, so as to avoid
@@ -2010,7 +2011,7 @@ asyncQueueReadAllNotifications(void)
* part of the page we will actually inspect.
*/
slotno = SimpleLruReadPage_ReadOnly(NotifyCtl, curpage,
- InvalidTransactionId);
+ InvalidTransactionId, &lockmode);
if (curpage == QUEUE_POS_PAGE(head))
{
/* we only want to read as far as head */
@@ -2027,6 +2028,7 @@ asyncQueueReadAllNotifications(void)
NotifyCtl->shared->page_buffer[slotno] + curoffset,
copysize);
/* Release lock that we got from SimpleLruReadPage_ReadOnly() */
+ Assert(lockmode != LW_NONE);
LWLockRelease(NotifySLRULock);
/*
diff --git a/src/backend/storage/lmgr/lwlock.c b/src/backend/storage/lmgr/lwlock.c
index 2fa90cc0954..8a7726f11c0 100644
--- a/src/backend/storage/lmgr/lwlock.c
+++ b/src/backend/storage/lmgr/lwlock.c
@@ -1068,6 +1068,8 @@ LWLockWakeup(LWLock *lock)
static void
LWLockQueueSelf(LWLock *lock, LWLockMode mode)
{
+ Assert(mode != LW_NONE);
+
/*
* If we don't have a PGPROC structure, there's no way to wait. This
* should never occur, since MyProc should only be null during shared
@@ -1828,6 +1830,7 @@ LWLockRelease(LWLock *lock)
elog(ERROR, "lock %s is not held", T_NAME(lock));
mode = held_lwlocks[i].mode;
+ Assert(mode == LW_EXCLUSIVE || mode == LW_SHARED);
num_held_lwlocks--;
for (; i < num_held_lwlocks; i++)
diff --git a/src/backend/storage/lmgr/predicate.c b/src/backend/storage/lmgr/predicate.c
index 8a365b400c6..a4df90a8aeb 100644
--- a/src/backend/storage/lmgr/predicate.c
+++ b/src/backend/storage/lmgr/predicate.c
@@ -924,6 +924,7 @@ SerialGetMinConflictCommitSeqNo(TransactionId xid)
TransactionId tailXid;
SerCommitSeqNo val;
int slotno;
+ LWLockMode lockmode = LW_NONE;
Assert(TransactionIdIsValid(xid));
@@ -946,8 +947,9 @@ SerialGetMinConflictCommitSeqNo(TransactionId xid)
* but will return with that lock held, which must then be released.
*/
slotno = SimpleLruReadPage_ReadOnly(SerialSlruCtl,
- SerialPage(xid), xid);
+ SerialPage(xid), xid, &lockmode);
val = SerialValue(slotno, xid);
+ Assert(lockmode != LW_NONE);
LWLockRelease(SerialSLRULock);
return val;
}
diff --git a/src/include/access/slru.h b/src/include/access/slru.h
index b39b43504d8..4b66d3b592b 100644
--- a/src/include/access/slru.h
+++ b/src/include/access/slru.h
@@ -142,7 +142,7 @@ extern int SimpleLruZeroPage(SlruCtl ctl, int pageno);
extern int SimpleLruReadPage(SlruCtl ctl, int pageno, bool write_ok,
TransactionId xid);
extern int SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno,
- TransactionId xid);
+ TransactionId xid, LWLockMode *lock_mode);
extern void SimpleLruWritePage(SlruCtl ctl, int slotno);
extern void SimpleLruWriteAll(SlruCtl ctl, bool allow_redirtied);
extern void SimpleLruTruncate(SlruCtl ctl, int cutoffPage);
diff --git a/src/include/storage/lwlock.h b/src/include/storage/lwlock.h
index af9b41795d2..e680c6397aa 100644
--- a/src/include/storage/lwlock.h
+++ b/src/include/storage/lwlock.h
@@ -129,6 +129,8 @@ extern PGDLLIMPORT int NamedLWLockTrancheRequests;
typedef enum LWLockMode
{
+ LW_NONE, /* Not a lock mode. Indicates that there is no
+ * lock. */
LW_EXCLUSIVE,
LW_SHARED,
LW_WAIT_UNTIL_FREE /* A special mode used in PGPROC->lwWaitMode,
--
2.14.3
[application/octet-stream] v6-0002-Make-MultiXact-local-cache-size-configurable.patch (3.7K, ../../CAPpHfdsUtSGxRrKn65F_wrE66sNztNY4TpDGDHh2kvAZZDOMqA@mail.gmail.com/3-v6-0002-Make-MultiXact-local-cache-size-configurable.patch)
download | inline diff:
From 2c40a5878da97b18e7e64a775f5caa0641e2f39c Mon Sep 17 00:00:00 2001
From: Alexander Korotkov <[email protected]>
Date: Mon, 26 Oct 2020 03:44:50 +0300
Subject: [PATCH 2/3] Make MultiXact local cache size configurable
---
doc/src/sgml/config.sgml | 16 ++++++++++++++++
src/backend/access/transam/multixact.c | 2 +-
src/backend/utils/init/globals.c | 2 ++
src/backend/utils/misc/guc.c | 10 ++++++++++
src/include/miscadmin.h | 2 ++
5 files changed, 31 insertions(+), 1 deletion(-)
diff --git a/doc/src/sgml/config.sgml b/doc/src/sgml/config.sgml
index f043433e318..fd4ca293476 100644
--- a/doc/src/sgml/config.sgml
+++ b/doc/src/sgml/config.sgml
@@ -1823,6 +1823,22 @@ include_dir 'conf.d'
</listitem>
</varlistentry>
+ <varlistentry id="guc-multixact-local-cache-entries" xreflabel="multixact_local-cache-entries">
+ <term><varname>multixact_local_cache_entries</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>multixact_local_cache_entries</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the number cached MultiXact by backend. Any SLRU lookup is preceeded
+ by cache lookup. Higher numbers of cache size help to deduplicate lock sets, while
+ lower values make cache lookup faster.
+ It defaults to 256.
+ </para>
+ </listitem>
+ </varlistentry>
+
<varlistentry id="guc-max-stack-depth" xreflabel="max_stack_depth">
<term><varname>max_stack_depth</varname> (<type>integer</type>)
<indexterm>
diff --git a/src/backend/access/transam/multixact.c b/src/backend/access/transam/multixact.c
index ccbce90f0ea..57be24c0cc1 100644
--- a/src/backend/access/transam/multixact.c
+++ b/src/backend/access/transam/multixact.c
@@ -1613,7 +1613,7 @@ mXactCachePut(MultiXactId multi, int nmembers, MultiXactMember *members)
qsort(entry->members, nmembers, sizeof(MultiXactMember), mxactMemberComparator);
dlist_push_head(&MXactCache, &entry->node);
- if (MXactCacheMembers++ >= MAX_CACHE_ENTRIES)
+ if (MXactCacheMembers++ >= multixact_local_cache_entries)
{
dlist_node *node;
mXactCacheEnt *entry;
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
index 6ab82168398..9ca71933dcc 100644
--- a/src/backend/utils/init/globals.c
+++ b/src/backend/utils/init/globals.c
@@ -149,3 +149,5 @@ int VacuumCostBalance = 0; /* working state for vacuum */
bool VacuumCostActive = false;
double vacuum_cleanup_index_scale_factor;
+
+int multixact_local_cache_entries = 256;
diff --git a/src/backend/utils/misc/guc.c b/src/backend/utils/misc/guc.c
index a62d64eaa47..d667578cc33 100644
--- a/src/backend/utils/misc/guc.c
+++ b/src/backend/utils/misc/guc.c
@@ -2257,6 +2257,16 @@ static struct config_int ConfigureNamesInt[] =
NULL, NULL, NULL
},
+ {
+ {"multixact_local_cache_entries", PGC_SUSET, RESOURCES_MEM,
+ gettext_noop("Sets the number of cached MultiXact by backend."),
+ NULL
+ },
+ &multixact_local_cache_entries,
+ 256, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
{
{"temp_buffers", PGC_USERSET, RESOURCES_MEM,
gettext_noop("Sets the maximum number of temporary buffers used by each session."),
diff --git a/src/include/miscadmin.h b/src/include/miscadmin.h
index 72e33523984..01af61c963a 100644
--- a/src/include/miscadmin.h
+++ b/src/include/miscadmin.h
@@ -162,6 +162,8 @@ extern PGDLLIMPORT int MaxConnections;
extern PGDLLIMPORT int max_worker_processes;
extern PGDLLIMPORT int max_parallel_workers;
+extern PGDLLIMPORT int multixact_local_cache_entries;
+
extern PGDLLIMPORT int MyProcPid;
extern PGDLLIMPORT pg_time_t MyStartTime;
extern PGDLLIMPORT TimestampTz MyStartTimestamp;
--
2.14.3
[application/octet-stream] v6-0003-Add-conditional-variable-to-wait-for-next-MultXact.patch (4.4K, ../../CAPpHfdsUtSGxRrKn65F_wrE66sNztNY4TpDGDHh2kvAZZDOMqA@mail.gmail.com/4-v6-0003-Add-conditional-variable-to-wait-for-next-MultXact.patch)
download | inline diff:
From da9cc973a4e7d568466b1c4de15479f707ad04f6 Mon Sep 17 00:00:00 2001
From: Alexander Korotkov <[email protected]>
Date: Mon, 26 Oct 2020 01:55:31 +0300
Subject: [PATCH 3/3] Add conditional variable to wait for next MultXact offset
in corner case
GetMultiXactIdMembers() has a corner case, when the next multixact offset is
not yet set. In this case GetMultiXactIdMembers() has to sleep till this offset
is set. Currently the sleeping is implemented in naive way using pg_sleep()
and retry. This commit implements sleeping with conditional variable, which
provides more efficient way for waiting till the event.
Discussion: https://postgr.es/m/a7f1c4e1-1015-92a4-2bd4-6736bd13d03e%40postgrespro.ru#c496c4e75fc0605094a0e1f763e6a6ec
Author: Andrey Borodin
Reviewed-by: Kyotaro Horiguchi, Daniel Gustafsson
Reviewed-by: Anastasia Lubennikova, Alexander Korotkov
---
src/backend/access/transam/multixact.c | 35 ++++++++++++++++++++++++++++++++--
src/backend/postmaster/pgstat.c | 2 ++
src/include/pgstat.h | 1 +
3 files changed, 36 insertions(+), 2 deletions(-)
diff --git a/src/backend/access/transam/multixact.c b/src/backend/access/transam/multixact.c
index 57be24c0cc1..70d977cba33 100644
--- a/src/backend/access/transam/multixact.c
+++ b/src/backend/access/transam/multixact.c
@@ -82,6 +82,7 @@
#include "lib/ilist.h"
#include "miscadmin.h"
#include "pg_trace.h"
+#include "pgstat.h"
#include "postmaster/autovacuum.h"
#include "storage/lmgr.h"
#include "storage/pmsignal.h"
@@ -233,6 +234,13 @@ typedef struct MultiXactStateData
/* support for members anti-wraparound measures */
MultiXactOffset offsetStopLimit; /* known if oldestOffsetKnown */
+ /*
+ * Conditional variable for waiting till the filling of the next multixact
+ * will be finished. See GetMultiXactIdMembers() and RecordNewMultiXact()
+ * for details.
+ */
+ ConditionVariable nextoffCV;
+
/*
* Per-backend data starts here. We have two arrays stored in the area
* immediately following the MultiXactStateData struct. Each is indexed by
@@ -892,6 +900,14 @@ RecordNewMultiXact(MultiXactId multi, MultiXactOffset offset,
/* Exchange our lock */
LWLockRelease(MultiXactOffsetSLRULock);
+ /*
+ * Let everybody know the offset of this mxid is recorded now. The
+ * waiters are waiting for the offset of the mxid next of the target to
+ * know the number of members of the target mxid, so we don't need to wait
+ * for members of this mxid are recorded.
+ */
+ ConditionVariableBroadcast(&MultiXactState->nextoffCV);
+
LWLockAcquire(MultiXactMemberSLRULock, LW_EXCLUSIVE);
prev_pageno = -1;
@@ -1389,9 +1405,23 @@ retry:
if (nextMXOffset == 0)
{
/* Corner case 2: next multixact is still being filled in */
+
+ /*
+ * The recorder of the next mxid is just before writing the
+ * offset. Wait for the offset to be written.
+ */
+ ConditionVariablePrepareToSleep(&MultiXactState->nextoffCV);
+
+ /*
+ * We don't have to recheck if multixact was filled in during
+ * ConditionVariablePrepareToSleep(), because we were holding
+ * MultiXactOffsetSLRULock.
+ */
LWLockRelease(MultiXactOffsetSLRULock);
- CHECK_FOR_INTERRUPTS();
- pg_usleep(1000L);
+
+ ConditionVariableSleep(&MultiXactState->nextoffCV,
+ WAIT_EVENT_WAIT_NEXT_MXMEMBERS);
+ ConditionVariableCancelSleep();
goto retry;
}
@@ -1873,6 +1903,7 @@ MultiXactShmemInit(void)
/* Make sure we zero out the per-backend state */
MemSet(MultiXactState, 0, SHARED_MULTIXACT_STATE_SIZE);
+ ConditionVariableInit(&MultiXactState->nextoffCV);
}
else
Assert(found);
diff --git a/src/backend/postmaster/pgstat.c b/src/backend/postmaster/pgstat.c
index 822f0ebc628..b99398a97e9 100644
--- a/src/backend/postmaster/pgstat.c
+++ b/src/backend/postmaster/pgstat.c
@@ -4020,6 +4020,8 @@ pgstat_get_wait_ipc(WaitEventIPC w)
break;
case WAIT_EVENT_XACT_GROUP_UPDATE:
event_name = "XactGroupUpdate";
+ case WAIT_EVENT_WAIT_NEXT_MXMEMBERS:
+ event_name = "MultiXactWaitNextMembers";
break;
/* no default case, so that compiler will warn */
}
diff --git a/src/include/pgstat.h b/src/include/pgstat.h
index a821ff4f158..75ede141ab7 100644
--- a/src/include/pgstat.h
+++ b/src/include/pgstat.h
@@ -952,6 +952,7 @@ typedef enum
WAIT_EVENT_REPLICATION_SLOT_DROP,
WAIT_EVENT_SAFE_SNAPSHOT,
WAIT_EVENT_SYNC_REP,
+ WAIT_EVENT_WAIT_NEXT_MXMEMBERS,
WAIT_EVENT_XACT_GROUP_UPDATE
} WaitEventIPC;
--
2.14.3
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2020-10-28 01:36 Tomas Vondra <[email protected]>
parent: Alexander Korotkov <[email protected]>
0 siblings, 2 replies; 103+ messages in thread
From: Tomas Vondra @ 2020-10-28 01:36 UTC (permalink / raw)
To: Alexander Korotkov <[email protected]>; +Cc: Andrey Borodin <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
On Tue, Oct 27, 2020 at 08:23:26PM +0300, Alexander Korotkov wrote:
>On Tue, Oct 27, 2020 at 8:02 PM Alexander Korotkov <[email protected]> wrote:
>> On Mon, Oct 26, 2020 at 6:45 PM Andrey Borodin <[email protected]> wrote:
>> > Thanks for your review, Alexander!
>> > +1 for avoiding double locking in SimpleLruReadPage_ReadOnly().
>> > Other changes seem correct to me too.
>> >
>> >
>> > I've tried to find optimal value for cache size and it seems to me that it affects multixact scalability much less than sizes of offsets\members buffers. I concur that patch 2 of the patchset does not seem documented enough.
>>
>> Thank you. I've made a few more minor adjustments to the patchset.
>> I'm going to push 0001 and 0003 if there are no objections.
>
>I get that patchset v5 doesn't pass the tests due to typo in assert.
>The fixes version is attached.
>
I did a quick review on this patch series. A couple comments:
0001
----
This looks quite suspicious to me - SimpleLruReadPage_ReadOnly is
changed to return information about what lock was used, merely to allow
the callers to do an Assert() that the value is not LW_NONE.
IMO we could achieve exactly the same thing by passing a simple flag
that would say 'make sure we got a lock' or something like that. In
fact, aren't all callers doing the assert? That'd mean we can just do
the check always, without the flag. (I see GetMultiXactIdMembers does
two calls and only checks the second result, but I wonder if that's
intended or omission.)
In any case, it'd make the lwlock.c changes unnecessary, I think.
0002
----
Specifies the number cached MultiXact by backend. Any SLRU lookup ...
should be 'number of cached ...'
0003
----
* Conditional variable for waiting till the filling of the next multixact
* will be finished. See GetMultiXactIdMembers() and RecordNewMultiXact()
* for details.
Perhaps 'till the next multixact is filled' or 'gets full' would be
better. Not sure.
This thread started with a discussion about making the SLRU sizes
configurable, but this patch version only adds a local cache. Does this
achieve the same goal, or would we still gain something by having GUCs
for the SLRUs?
If we're claiming this improves performance, it'd be good to have some
workload demonstrating that and measurements. I don't see anything like
that in this thread, so it's a bit hand-wavy. Can someone share details
of such workload (even synthetic one) and some basic measurements?
regards
--
Tomas Vondra http://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2020-10-28 07:34 Andrey Borodin <[email protected]>
parent: Tomas Vondra <[email protected]>
1 sibling, 1 reply; 103+ messages in thread
From: Andrey Borodin @ 2020-10-28 07:34 UTC (permalink / raw)
To: Tomas Vondra <[email protected]>; +Cc: Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
Tomas, thanks for looking into this!
> 28 окт. 2020 г., в 06:36, Tomas Vondra <[email protected]> написал(а):
>
>
> This thread started with a discussion about making the SLRU sizes
> configurable, but this patch version only adds a local cache. Does this
> achieve the same goal, or would we still gain something by having GUCs
> for the SLRUs?
>
> If we're claiming this improves performance, it'd be good to have some
> workload demonstrating that and measurements. I don't see anything like
> that in this thread, so it's a bit hand-wavy. Can someone share details
> of such workload (even synthetic one) and some basic measurements?
All patches in this thread aim at the same goal: improve performance in presence of MultiXact locks contention.
I could not build synthetical reproduction of the problem, however I did some MultiXact stressing here [0]. It's a clumsy test program, because it still is not clear to me which parameters of workload trigger MultiXact locks contention. In generic case I was encountering other locks like *GenLock: XidGenLock, MultixactGenLock etc. Yet our production system encounters this problem approximately once in a month through this year.
Test program locks for share different set of tuples in presence of concurrent full scans.
To produce a set of locks we choose one of 14 bits. If a row number has this bit set to 0 we add lock it.
I've been measuring time to lock all rows 3 time for each of 14 bits, observing total time to set all locks.
During the test I was observing locks in pg_stat_activity, if they did not contain enough MultiXact locks I was tuning parameters further (number of concurrent clients, number of bits, select queries etc).
Why is it so complicated? It seems that other reproductions of a problem were encountering other locks.
Lets describe patches in this thread from the POV of these test.
*** Configurable SLRU buffers for MultiXact members and offsets.
From tests it is clear that high and low values for these buffers affect the test time. Here are time for a one test run with different offsets and members sizes [1]
Our production currently runs with (numbers are pages of buffers)
+#define NUM_MXACTOFFSET_BUFFERS 32
+#define NUM_MXACTMEMBER_BUFFERS 64
And, looking back to incidents in summer and fall 2020, seems like it helped mostly.
But it's hard to give some tuning advises based on test results. Values (32,64) produce 10% better result than current hardcoded values (8,16). In generic case this is not what someone should tune first.
*** Configurable caches of MultiXacts.
Tests were specifically designed to beat caches. So, according to test the bigger cache is - the more time it takes to accomplish the test [2].
Anyway cache is local for backend and it's purpose is deduplication of written MultiXacts, not enhancing reads.
*** Using advantage of SimpleLruReadPage_ReadOnly() in MultiXacts.
This simply aligns MultiXact with Subtransactions and CLOG. Other SLRUs already take advantage of reading SLRU with shared lock.
On synthetical tests without background selects this patch adds another ~4.7% of performance [3] against [4]. This improvement seems consistent between different parameter values, yet within measurements deviation (see difference between warmup run [5] and closing run [6]).
All in all, these attempts to measure impact are hand-wavy too. But it makes sense to use consistent approach among similar subsystems (MultiXacts, Subtrans, CLOG etc).
*** Reduce sleep in GetMultiXactIdMembers() on standby.
The problem with pg_usleep(1000L) within GetMultiXactIdMembers() manifests on standbys during contention of MultiXactOffsetControlLock. It's even harder to reproduce.
Yet it seems obvious that reducing sleep to shorter time frame will make count of sleeping backend smaller.
For consistency I've returned patch with SLRU buffer configs to patchset (other patches are intact). But I'm mostly concerned about patches 1 and 3.
Thanks!
Best regards, Andrey Borodin.
[0] https://github.com/x4m/multixact_stress
[1] https://github.com/x4m/multixact_stress/blob/master/testresults.txt#L22-L39
[2] https://github.com/x4m/multixact_stress/blob/master/testresults.txt#L83-L99
[3] https://github.com/x4m/multixact_stress/blob/master/testresults.txt#L9
[4] https://github.com/x4m/multixact_stress/blob/master/testresults.txt#L29
[5] https://github.com/x4m/multixact_stress/blob/master/testresults.txt#L3
[6] https://github.com/x4m/multixact_stress/blob/master/testresults.txt#L19
Attachments:
[application/octet-stream] v6-0001-Use-shared-lock-in-GetMultiXactIdMembers-for-offs.patch (12.9K, ../../[email protected]/2-v6-0001-Use-shared-lock-in-GetMultiXactIdMembers-for-offs.patch)
download | inline diff:
From bcbd4f6d88d7b6dd9cf779b62e0e0261edadc71b Mon Sep 17 00:00:00 2001
From: Alexander Korotkov <[email protected]>
Date: Tue, 27 Oct 2020 19:29:56 +0300
Subject: [PATCH v6 1/4] Use shared lock in GetMultiXactIdMembers for offsets
and members
Previously the read of multixact required exclusive control locks for both
offsets and members SLRUs. This could lead to the excessive lock contention.
This commit we makes multixacts SLRU take advantage of SimpleLruReadPage_ReadOnly
similar to clog, commit_ts and subtrans.
In order to evade extra reacquiring of CLRU lock, we teach
SimpleLruReadPage_ReadOnly() to take into account the current lock mode and
report resulting lock mode back.
Discussion: https://postgr.es/m/a7f1c4e1-1015-92a4-2bd4-6736bd13d03e%40postgrespro.ru#c496c4e75fc0605094a0e1f763e6a6ec
Author: Andrey Borodin
Reviewed-by: Kyotaro Horiguchi, Daniel Gustafsson
Reviewed-by: Anastasia Lubennikova, Alexander Korotkov
---
src/backend/access/transam/clog.c | 4 +++-
src/backend/access/transam/commit_ts.c | 4 +++-
src/backend/access/transam/multixact.c | 22 +++++++++++++++-------
src/backend/access/transam/slru.c | 23 +++++++++++++++++------
src/backend/access/transam/subtrans.c | 4 +++-
src/backend/commands/async.c | 4 +++-
src/backend/storage/lmgr/lwlock.c | 3 +++
src/backend/storage/lmgr/predicate.c | 4 +++-
src/include/access/slru.h | 2 +-
src/include/storage/lwlock.h | 2 ++
10 files changed, 53 insertions(+), 19 deletions(-)
diff --git a/src/backend/access/transam/clog.c b/src/backend/access/transam/clog.c
index 034349aa7b..4c372065de 100644
--- a/src/backend/access/transam/clog.c
+++ b/src/backend/access/transam/clog.c
@@ -640,10 +640,11 @@ TransactionIdGetStatus(TransactionId xid, XLogRecPtr *lsn)
int lsnindex;
char *byteptr;
XidStatus status;
+ LWLockMode lockmode = LW_NONE;
/* lock is acquired by SimpleLruReadPage_ReadOnly */
- slotno = SimpleLruReadPage_ReadOnly(XactCtl, pageno, xid);
+ slotno = SimpleLruReadPage_ReadOnly(XactCtl, pageno, xid, &lockmode);
byteptr = XactCtl->shared->page_buffer[slotno] + byteno;
status = (*byteptr >> bshift) & CLOG_XACT_BITMASK;
@@ -651,6 +652,7 @@ TransactionIdGetStatus(TransactionId xid, XLogRecPtr *lsn)
lsnindex = GetLSNIndex(slotno, xid);
*lsn = XactCtl->shared->group_lsn[lsnindex];
+ Assert(lockmode != LW_NONE);
LWLockRelease(XactSLRULock);
return status;
diff --git a/src/backend/access/transam/commit_ts.c b/src/backend/access/transam/commit_ts.c
index cb8a968801..c19b0b5399 100644
--- a/src/backend/access/transam/commit_ts.c
+++ b/src/backend/access/transam/commit_ts.c
@@ -288,6 +288,7 @@ TransactionIdGetCommitTsData(TransactionId xid, TimestampTz *ts,
CommitTimestampEntry entry;
TransactionId oldestCommitTsXid;
TransactionId newestCommitTsXid;
+ LWLockMode lockmode = LW_NONE;
if (!TransactionIdIsValid(xid))
ereport(ERROR,
@@ -342,7 +343,7 @@ TransactionIdGetCommitTsData(TransactionId xid, TimestampTz *ts,
}
/* lock is acquired by SimpleLruReadPage_ReadOnly */
- slotno = SimpleLruReadPage_ReadOnly(CommitTsCtl, pageno, xid);
+ slotno = SimpleLruReadPage_ReadOnly(CommitTsCtl, pageno, xid, &lockmode);
memcpy(&entry,
CommitTsCtl->shared->page_buffer[slotno] +
SizeOfCommitTimestampEntry * entryno,
@@ -352,6 +353,7 @@ TransactionIdGetCommitTsData(TransactionId xid, TimestampTz *ts,
if (nodeid)
*nodeid = entry.nodeid;
+ Assert(lockmode != LW_NONE);
LWLockRelease(CommitTsSLRULock);
return *ts != 0;
}
diff --git a/src/backend/access/transam/multixact.c b/src/backend/access/transam/multixact.c
index 43653fe572..ccbce90f0e 100644
--- a/src/backend/access/transam/multixact.c
+++ b/src/backend/access/transam/multixact.c
@@ -1237,6 +1237,7 @@ GetMultiXactIdMembers(MultiXactId multi, MultiXactMember **members,
MultiXactId tmpMXact;
MultiXactOffset nextOffset;
MultiXactMember *ptr;
+ LWLockMode lockmode = LW_NONE;
debug_elog3(DEBUG2, "GetMembers: asked for %u", multi);
@@ -1340,12 +1341,13 @@ GetMultiXactIdMembers(MultiXactId multi, MultiXactMember **members,
* time on every multixact creation.
*/
retry:
- LWLockAcquire(MultiXactOffsetSLRULock, LW_EXCLUSIVE);
pageno = MultiXactIdToOffsetPage(multi);
entryno = MultiXactIdToOffsetEntry(multi);
- slotno = SimpleLruReadPage(MultiXactOffsetCtl, pageno, true, multi);
+ /* lock is acquired by SimpleLruReadPage_ReadOnly */
+ slotno = SimpleLruReadPage_ReadOnly(MultiXactOffsetCtl, pageno,
+ multi, &lockmode);
offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
offptr += entryno;
offset = *offptr;
@@ -1377,7 +1379,8 @@ retry:
entryno = MultiXactIdToOffsetEntry(tmpMXact);
if (pageno != prev_pageno)
- slotno = SimpleLruReadPage(MultiXactOffsetCtl, pageno, true, tmpMXact);
+ slotno = SimpleLruReadPage_ReadOnly(MultiXactOffsetCtl, pageno,
+ tmpMXact, &lockmode);
offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
offptr += entryno;
@@ -1395,14 +1398,14 @@ retry:
length = nextMXOffset - offset;
}
+ Assert(lockmode != LW_NONE);
LWLockRelease(MultiXactOffsetSLRULock);
ptr = (MultiXactMember *) palloc(length * sizeof(MultiXactMember));
*members = ptr;
/* Now get the members themselves. */
- LWLockAcquire(MultiXactMemberSLRULock, LW_EXCLUSIVE);
-
+ lockmode = LW_NONE;
truelength = 0;
prev_pageno = -1;
for (i = 0; i < length; i++, offset++)
@@ -1418,7 +1421,8 @@ retry:
if (pageno != prev_pageno)
{
- slotno = SimpleLruReadPage(MultiXactMemberCtl, pageno, true, multi);
+ slotno = SimpleLruReadPage_ReadOnly(MultiXactMemberCtl, pageno,
+ multi, &lockmode);
prev_pageno = pageno;
}
@@ -1441,6 +1445,7 @@ retry:
truelength++;
}
+ Assert(lockmode != LW_NONE);
LWLockRelease(MultiXactMemberSLRULock);
/*
@@ -2733,6 +2738,7 @@ find_multixact_start(MultiXactId multi, MultiXactOffset *result)
int entryno;
int slotno;
MultiXactOffset *offptr;
+ LWLockMode lockmode = LW_NONE;
Assert(MultiXactState->finishedStartup);
@@ -2749,10 +2755,12 @@ find_multixact_start(MultiXactId multi, MultiXactOffset *result)
return false;
/* lock is acquired by SimpleLruReadPage_ReadOnly */
- slotno = SimpleLruReadPage_ReadOnly(MultiXactOffsetCtl, pageno, multi);
+ slotno = SimpleLruReadPage_ReadOnly(MultiXactOffsetCtl, pageno,
+ multi, &lockmode);
offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
offptr += entryno;
offset = *offptr;
+ Assert(lockmode != LW_NONE);
LWLockRelease(MultiXactOffsetSLRULock);
*result = offset;
diff --git a/src/backend/access/transam/slru.c b/src/backend/access/transam/slru.c
index 16a7898697..ca0a23ab87 100644
--- a/src/backend/access/transam/slru.c
+++ b/src/backend/access/transam/slru.c
@@ -487,17 +487,23 @@ SimpleLruReadPage(SlruCtl ctl, int pageno, bool write_ok,
* Return value is the shared-buffer slot number now holding the page.
* The buffer's LRU access info is updated.
*
- * Control lock must NOT be held at entry, but will be held at exit.
- * It is unspecified whether the lock will be shared or exclusive.
+ * Control lock must be held in *lock_mode mode, which may be LW_NONE. Control
+ * lock will be held at exit in at least shared mode. Resulting control lock
+ * mode is set to *lock_mode.
*/
int
-SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno, TransactionId xid)
+SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno, TransactionId xid,
+ LWLockMode *lock_mode)
{
SlruShared shared = ctl->shared;
int slotno;
/* Try to find the page while holding only shared lock */
- LWLockAcquire(shared->ControlLock, LW_SHARED);
+ if (*lock_mode == LW_NONE)
+ {
+ LWLockAcquire(shared->ControlLock, LW_SHARED);
+ *lock_mode = LW_SHARED;
+ }
/* See if page is already in a buffer */
for (slotno = 0; slotno < shared->num_slots; slotno++)
@@ -517,8 +523,13 @@ SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno, TransactionId xid)
}
/* No luck, so switch to normal exclusive lock and do regular read */
- LWLockRelease(shared->ControlLock);
- LWLockAcquire(shared->ControlLock, LW_EXCLUSIVE);
+ if (*lock_mode != LW_EXCLUSIVE)
+ {
+ Assert(*lock_mode == LW_NONE);
+ LWLockRelease(shared->ControlLock);
+ LWLockAcquire(shared->ControlLock, LW_EXCLUSIVE);
+ *lock_mode = LW_EXCLUSIVE;
+ }
return SimpleLruReadPage(ctl, pageno, true, xid);
}
diff --git a/src/backend/access/transam/subtrans.c b/src/backend/access/transam/subtrans.c
index 0111e867c7..7bb1543189 100644
--- a/src/backend/access/transam/subtrans.c
+++ b/src/backend/access/transam/subtrans.c
@@ -113,6 +113,7 @@ SubTransGetParent(TransactionId xid)
int slotno;
TransactionId *ptr;
TransactionId parent;
+ LWLockMode lockmode = LW_NONE;
/* Can't ask about stuff that might not be around anymore */
Assert(TransactionIdFollowsOrEquals(xid, TransactionXmin));
@@ -123,12 +124,13 @@ SubTransGetParent(TransactionId xid)
/* lock is acquired by SimpleLruReadPage_ReadOnly */
- slotno = SimpleLruReadPage_ReadOnly(SubTransCtl, pageno, xid);
+ slotno = SimpleLruReadPage_ReadOnly(SubTransCtl, pageno, xid, &lockmode);
ptr = (TransactionId *) SubTransCtl->shared->page_buffer[slotno];
ptr += entryno;
parent = *ptr;
+ Assert(lockmode != LW_NONE);
LWLockRelease(SubtransSLRULock);
return parent;
diff --git a/src/backend/commands/async.c b/src/backend/commands/async.c
index 8dbcace3f9..8b41957559 100644
--- a/src/backend/commands/async.c
+++ b/src/backend/commands/async.c
@@ -2002,6 +2002,7 @@ asyncQueueReadAllNotifications(void)
int curoffset = QUEUE_POS_OFFSET(pos);
int slotno;
int copysize;
+ LWLockMode lockmode = LW_NONE;
/*
* We copy the data from SLRU into a local buffer, so as to avoid
@@ -2010,7 +2011,7 @@ asyncQueueReadAllNotifications(void)
* part of the page we will actually inspect.
*/
slotno = SimpleLruReadPage_ReadOnly(NotifyCtl, curpage,
- InvalidTransactionId);
+ InvalidTransactionId, &lockmode);
if (curpage == QUEUE_POS_PAGE(head))
{
/* we only want to read as far as head */
@@ -2027,6 +2028,7 @@ asyncQueueReadAllNotifications(void)
NotifyCtl->shared->page_buffer[slotno] + curoffset,
copysize);
/* Release lock that we got from SimpleLruReadPage_ReadOnly() */
+ Assert(lockmode != LW_NONE);
LWLockRelease(NotifySLRULock);
/*
diff --git a/src/backend/storage/lmgr/lwlock.c b/src/backend/storage/lmgr/lwlock.c
index 2fa90cc095..8a7726f11c 100644
--- a/src/backend/storage/lmgr/lwlock.c
+++ b/src/backend/storage/lmgr/lwlock.c
@@ -1068,6 +1068,8 @@ LWLockWakeup(LWLock *lock)
static void
LWLockQueueSelf(LWLock *lock, LWLockMode mode)
{
+ Assert(mode != LW_NONE);
+
/*
* If we don't have a PGPROC structure, there's no way to wait. This
* should never occur, since MyProc should only be null during shared
@@ -1828,6 +1830,7 @@ LWLockRelease(LWLock *lock)
elog(ERROR, "lock %s is not held", T_NAME(lock));
mode = held_lwlocks[i].mode;
+ Assert(mode == LW_EXCLUSIVE || mode == LW_SHARED);
num_held_lwlocks--;
for (; i < num_held_lwlocks; i++)
diff --git a/src/backend/storage/lmgr/predicate.c b/src/backend/storage/lmgr/predicate.c
index 8a365b400c..a4df90a8ae 100644
--- a/src/backend/storage/lmgr/predicate.c
+++ b/src/backend/storage/lmgr/predicate.c
@@ -924,6 +924,7 @@ SerialGetMinConflictCommitSeqNo(TransactionId xid)
TransactionId tailXid;
SerCommitSeqNo val;
int slotno;
+ LWLockMode lockmode = LW_NONE;
Assert(TransactionIdIsValid(xid));
@@ -946,8 +947,9 @@ SerialGetMinConflictCommitSeqNo(TransactionId xid)
* but will return with that lock held, which must then be released.
*/
slotno = SimpleLruReadPage_ReadOnly(SerialSlruCtl,
- SerialPage(xid), xid);
+ SerialPage(xid), xid, &lockmode);
val = SerialValue(slotno, xid);
+ Assert(lockmode != LW_NONE);
LWLockRelease(SerialSLRULock);
return val;
}
diff --git a/src/include/access/slru.h b/src/include/access/slru.h
index b39b43504d..4b66d3b592 100644
--- a/src/include/access/slru.h
+++ b/src/include/access/slru.h
@@ -142,7 +142,7 @@ extern int SimpleLruZeroPage(SlruCtl ctl, int pageno);
extern int SimpleLruReadPage(SlruCtl ctl, int pageno, bool write_ok,
TransactionId xid);
extern int SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno,
- TransactionId xid);
+ TransactionId xid, LWLockMode *lock_mode);
extern void SimpleLruWritePage(SlruCtl ctl, int slotno);
extern void SimpleLruWriteAll(SlruCtl ctl, bool allow_redirtied);
extern void SimpleLruTruncate(SlruCtl ctl, int cutoffPage);
diff --git a/src/include/storage/lwlock.h b/src/include/storage/lwlock.h
index af9b41795d..e680c6397a 100644
--- a/src/include/storage/lwlock.h
+++ b/src/include/storage/lwlock.h
@@ -129,6 +129,8 @@ extern PGDLLIMPORT int NamedLWLockTrancheRequests;
typedef enum LWLockMode
{
+ LW_NONE, /* Not a lock mode. Indicates that there is no
+ * lock. */
LW_EXCLUSIVE,
LW_SHARED,
LW_WAIT_UNTIL_FREE /* A special mode used in PGPROC->lwWaitMode,
--
2.24.3 (Apple Git-128)
[application/octet-stream] v6-0004-Add-GUCs-to-tune-MultiXact-SLRUs.patch (6.3K, ../../[email protected]/3-v6-0004-Add-GUCs-to-tune-MultiXact-SLRUs.patch)
download | inline diff:
From a5e15ae42bd6c99d39bac7a9988f586c48ceba13 Mon Sep 17 00:00:00 2001
From: Andrey Borodin <[email protected]>
Date: Sat, 9 May 2020 16:42:07 +0500
Subject: [PATCH v6 4/4] Add GUCs to tune MultiXact SLRUs
---
doc/src/sgml/config.sgml | 31 ++++++++++++++++++++++++++
src/backend/access/transam/multixact.c | 8 +++----
src/backend/utils/init/globals.c | 2 ++
src/backend/utils/misc/guc.c | 22 ++++++++++++++++++
src/include/access/multixact.h | 4 ----
src/include/miscadmin.h | 2 ++
6 files changed, 61 insertions(+), 8 deletions(-)
diff --git a/doc/src/sgml/config.sgml b/doc/src/sgml/config.sgml
index fd4ca29347..d2ca4934de 100644
--- a/doc/src/sgml/config.sgml
+++ b/doc/src/sgml/config.sgml
@@ -1838,6 +1838,37 @@ include_dir 'conf.d'
</para>
</listitem>
</varlistentry>
+
+ <varlistentry id="guc-multixact-offsets-slru-buffers" xreflabel="multixact_offsets_slru_buffers">
+ <term><varname>multixact_offsets_slru_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>multixact_offsets_slru_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of memory to be used for MultiXact offsets. MultiXact offsets
+ are used to store informaion about offsets of multiple row lockers (caused by SELECT FOR UPDATE and others).
+ It defaults to 64 kilobytes (<literal>64KB</literal>).
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-multixact-members-slru-buffers" xreflabel="multixact_members_slru_buffers">
+ <term><varname>multixact_members_slru_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>multixact_members_slru_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of memory to be used for MultiXact members. MultiXact members
+ are used to store informaion about XIDs of multiple row lockers. Tipically <varname>multixact_members_slru_buffers</varname>
+ is twice more than <varname>multixact_offsets_slru_buffers</varname>.
+ It defaults to 128 kilobytes (<literal>128KB</literal>).
+ </para>
+ </listitem>
+ </varlistentry>
<varlistentry id="guc-max-stack-depth" xreflabel="max_stack_depth">
<term><varname>max_stack_depth</varname> (<type>integer</type>)
diff --git a/src/backend/access/transam/multixact.c b/src/backend/access/transam/multixact.c
index 70d977cba3..1c177242c5 100644
--- a/src/backend/access/transam/multixact.c
+++ b/src/backend/access/transam/multixact.c
@@ -1866,8 +1866,8 @@ MultiXactShmemSize(void)
mul_size(sizeof(MultiXactId) * 2, MaxOldestSlot))
size = SHARED_MULTIXACT_STATE_SIZE;
- size = add_size(size, SimpleLruShmemSize(NUM_MULTIXACTOFFSET_BUFFERS, 0));
- size = add_size(size, SimpleLruShmemSize(NUM_MULTIXACTMEMBER_BUFFERS, 0));
+ size = add_size(size, SimpleLruShmemSize(multixact_offsets_slru_buffers, 0));
+ size = add_size(size, SimpleLruShmemSize(multixact_members_slru_buffers, 0));
return size;
}
@@ -1883,12 +1883,12 @@ MultiXactShmemInit(void)
MultiXactMemberCtl->PagePrecedes = MultiXactMemberPagePrecedes;
SimpleLruInit(MultiXactOffsetCtl,
- "MultiXactOffset", NUM_MULTIXACTOFFSET_BUFFERS, 0,
+ "MultiXactOffset", multixact_offsets_slru_buffers, 0,
MultiXactOffsetSLRULock, "pg_multixact/offsets",
LWTRANCHE_MULTIXACTOFFSET_BUFFER,
SYNC_HANDLER_MULTIXACT_OFFSET);
SimpleLruInit(MultiXactMemberCtl,
- "MultiXactMember", NUM_MULTIXACTMEMBER_BUFFERS, 0,
+ "MultiXactMember", multixact_members_slru_buffers, 0,
MultiXactMemberSLRULock, "pg_multixact/members",
LWTRANCHE_MULTIXACTMEMBER_BUFFER,
SYNC_HANDLER_MULTIXACT_MEMBER);
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
index 9ca71933dc..a5ec7bfe88 100644
--- a/src/backend/utils/init/globals.c
+++ b/src/backend/utils/init/globals.c
@@ -151,3 +151,5 @@ bool VacuumCostActive = false;
double vacuum_cleanup_index_scale_factor;
int multixact_local_cache_entries = 256;
+int multixact_offsets_slru_buffers = 8;
+int multixact_members_slru_buffers = 16;
diff --git a/src/backend/utils/misc/guc.c b/src/backend/utils/misc/guc.c
index d667578cc3..7682c8be1e 100644
--- a/src/backend/utils/misc/guc.c
+++ b/src/backend/utils/misc/guc.c
@@ -2267,6 +2267,28 @@ static struct config_int ConfigureNamesInt[] =
NULL, NULL, NULL
},
+ {
+ {"multixact_offsets_slru_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for MultiXact offsets SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &multixact_offsets_slru_buffers,
+ 8, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"multixact_members_slru_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for MultiXact members SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &multixact_members_slru_buffers,
+ 16, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
{
{"temp_buffers", PGC_USERSET, RESOURCES_MEM,
gettext_noop("Sets the maximum number of temporary buffers used by each session."),
diff --git a/src/include/access/multixact.h b/src/include/access/multixact.h
index 9a30380901..630ceaea4d 100644
--- a/src/include/access/multixact.h
+++ b/src/include/access/multixact.h
@@ -29,10 +29,6 @@
#define MaxMultiXactOffset ((MultiXactOffset) 0xFFFFFFFF)
-/* Number of SLRU buffers to use for multixact */
-#define NUM_MULTIXACTOFFSET_BUFFERS 8
-#define NUM_MULTIXACTMEMBER_BUFFERS 16
-
/*
* Possible multixact lock modes ("status"). The first four modes are for
* tuple locks (FOR KEY SHARE, FOR SHARE, FOR NO KEY UPDATE, FOR UPDATE); the
diff --git a/src/include/miscadmin.h b/src/include/miscadmin.h
index 01af61c963..ef8abea84d 100644
--- a/src/include/miscadmin.h
+++ b/src/include/miscadmin.h
@@ -161,6 +161,8 @@ extern PGDLLIMPORT int MaxBackends;
extern PGDLLIMPORT int MaxConnections;
extern PGDLLIMPORT int max_worker_processes;
extern PGDLLIMPORT int max_parallel_workers;
+extern PGDLLIMPORT int multixact_offsets_slru_buffers;
+extern PGDLLIMPORT int multixact_members_slru_buffers;
extern PGDLLIMPORT int multixact_local_cache_entries;
--
2.24.3 (Apple Git-128)
[application/octet-stream] v6-0003-Add-conditional-variable-to-wait-for-next-MultXac.patch (4.4K, ../../[email protected]/4-v6-0003-Add-conditional-variable-to-wait-for-next-MultXac.patch)
download | inline diff:
From d25e7d68a7979fc391c5d988b256a49cdd916bfb Mon Sep 17 00:00:00 2001
From: Alexander Korotkov <[email protected]>
Date: Mon, 26 Oct 2020 01:55:31 +0300
Subject: [PATCH v6 3/4] Add conditional variable to wait for next MultXact
offset in corner case
GetMultiXactIdMembers() has a corner case, when the next multixact offset is
not yet set. In this case GetMultiXactIdMembers() has to sleep till this offset
is set. Currently the sleeping is implemented in naive way using pg_sleep()
and retry. This commit implements sleeping with conditional variable, which
provides more efficient way for waiting till the event.
Discussion: https://postgr.es/m/a7f1c4e1-1015-92a4-2bd4-6736bd13d03e%40postgrespro.ru#c496c4e75fc0605094a0e1f763e6a6ec
Author: Andrey Borodin
Reviewed-by: Kyotaro Horiguchi, Daniel Gustafsson
Reviewed-by: Anastasia Lubennikova, Alexander Korotkov
---
src/backend/access/transam/multixact.c | 35 ++++++++++++++++++++++++--
src/backend/postmaster/pgstat.c | 2 ++
src/include/pgstat.h | 1 +
3 files changed, 36 insertions(+), 2 deletions(-)
diff --git a/src/backend/access/transam/multixact.c b/src/backend/access/transam/multixact.c
index 57be24c0cc..70d977cba3 100644
--- a/src/backend/access/transam/multixact.c
+++ b/src/backend/access/transam/multixact.c
@@ -82,6 +82,7 @@
#include "lib/ilist.h"
#include "miscadmin.h"
#include "pg_trace.h"
+#include "pgstat.h"
#include "postmaster/autovacuum.h"
#include "storage/lmgr.h"
#include "storage/pmsignal.h"
@@ -233,6 +234,13 @@ typedef struct MultiXactStateData
/* support for members anti-wraparound measures */
MultiXactOffset offsetStopLimit; /* known if oldestOffsetKnown */
+ /*
+ * Conditional variable for waiting till the filling of the next multixact
+ * will be finished. See GetMultiXactIdMembers() and RecordNewMultiXact()
+ * for details.
+ */
+ ConditionVariable nextoffCV;
+
/*
* Per-backend data starts here. We have two arrays stored in the area
* immediately following the MultiXactStateData struct. Each is indexed by
@@ -892,6 +900,14 @@ RecordNewMultiXact(MultiXactId multi, MultiXactOffset offset,
/* Exchange our lock */
LWLockRelease(MultiXactOffsetSLRULock);
+ /*
+ * Let everybody know the offset of this mxid is recorded now. The
+ * waiters are waiting for the offset of the mxid next of the target to
+ * know the number of members of the target mxid, so we don't need to wait
+ * for members of this mxid are recorded.
+ */
+ ConditionVariableBroadcast(&MultiXactState->nextoffCV);
+
LWLockAcquire(MultiXactMemberSLRULock, LW_EXCLUSIVE);
prev_pageno = -1;
@@ -1389,9 +1405,23 @@ retry:
if (nextMXOffset == 0)
{
/* Corner case 2: next multixact is still being filled in */
+
+ /*
+ * The recorder of the next mxid is just before writing the
+ * offset. Wait for the offset to be written.
+ */
+ ConditionVariablePrepareToSleep(&MultiXactState->nextoffCV);
+
+ /*
+ * We don't have to recheck if multixact was filled in during
+ * ConditionVariablePrepareToSleep(), because we were holding
+ * MultiXactOffsetSLRULock.
+ */
LWLockRelease(MultiXactOffsetSLRULock);
- CHECK_FOR_INTERRUPTS();
- pg_usleep(1000L);
+
+ ConditionVariableSleep(&MultiXactState->nextoffCV,
+ WAIT_EVENT_WAIT_NEXT_MXMEMBERS);
+ ConditionVariableCancelSleep();
goto retry;
}
@@ -1873,6 +1903,7 @@ MultiXactShmemInit(void)
/* Make sure we zero out the per-backend state */
MemSet(MultiXactState, 0, SHARED_MULTIXACT_STATE_SIZE);
+ ConditionVariableInit(&MultiXactState->nextoffCV);
}
else
Assert(found);
diff --git a/src/backend/postmaster/pgstat.c b/src/backend/postmaster/pgstat.c
index 822f0ebc62..b99398a97e 100644
--- a/src/backend/postmaster/pgstat.c
+++ b/src/backend/postmaster/pgstat.c
@@ -4020,6 +4020,8 @@ pgstat_get_wait_ipc(WaitEventIPC w)
break;
case WAIT_EVENT_XACT_GROUP_UPDATE:
event_name = "XactGroupUpdate";
+ case WAIT_EVENT_WAIT_NEXT_MXMEMBERS:
+ event_name = "MultiXactWaitNextMembers";
break;
/* no default case, so that compiler will warn */
}
diff --git a/src/include/pgstat.h b/src/include/pgstat.h
index a821ff4f15..75ede141ab 100644
--- a/src/include/pgstat.h
+++ b/src/include/pgstat.h
@@ -952,6 +952,7 @@ typedef enum
WAIT_EVENT_REPLICATION_SLOT_DROP,
WAIT_EVENT_SAFE_SNAPSHOT,
WAIT_EVENT_SYNC_REP,
+ WAIT_EVENT_WAIT_NEXT_MXMEMBERS,
WAIT_EVENT_XACT_GROUP_UPDATE
} WaitEventIPC;
--
2.24.3 (Apple Git-128)
[application/octet-stream] v6-0002-Make-MultiXact-local-cache-size-configurable.patch (3.7K, ../../[email protected]/5-v6-0002-Make-MultiXact-local-cache-size-configurable.patch)
download | inline diff:
From d0cd1ce556ad3f4e7bbebf0468dc6e00765d79ae Mon Sep 17 00:00:00 2001
From: Alexander Korotkov <[email protected]>
Date: Mon, 26 Oct 2020 03:44:50 +0300
Subject: [PATCH v6 2/4] Make MultiXact local cache size configurable
---
doc/src/sgml/config.sgml | 16 ++++++++++++++++
src/backend/access/transam/multixact.c | 2 +-
src/backend/utils/init/globals.c | 2 ++
src/backend/utils/misc/guc.c | 10 ++++++++++
src/include/miscadmin.h | 2 ++
5 files changed, 31 insertions(+), 1 deletion(-)
diff --git a/doc/src/sgml/config.sgml b/doc/src/sgml/config.sgml
index f043433e31..fd4ca29347 100644
--- a/doc/src/sgml/config.sgml
+++ b/doc/src/sgml/config.sgml
@@ -1823,6 +1823,22 @@ include_dir 'conf.d'
</listitem>
</varlistentry>
+ <varlistentry id="guc-multixact-local-cache-entries" xreflabel="multixact_local-cache-entries">
+ <term><varname>multixact_local_cache_entries</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>multixact_local_cache_entries</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the number cached MultiXact by backend. Any SLRU lookup is preceeded
+ by cache lookup. Higher numbers of cache size help to deduplicate lock sets, while
+ lower values make cache lookup faster.
+ It defaults to 256.
+ </para>
+ </listitem>
+ </varlistentry>
+
<varlistentry id="guc-max-stack-depth" xreflabel="max_stack_depth">
<term><varname>max_stack_depth</varname> (<type>integer</type>)
<indexterm>
diff --git a/src/backend/access/transam/multixact.c b/src/backend/access/transam/multixact.c
index ccbce90f0e..57be24c0cc 100644
--- a/src/backend/access/transam/multixact.c
+++ b/src/backend/access/transam/multixact.c
@@ -1613,7 +1613,7 @@ mXactCachePut(MultiXactId multi, int nmembers, MultiXactMember *members)
qsort(entry->members, nmembers, sizeof(MultiXactMember), mxactMemberComparator);
dlist_push_head(&MXactCache, &entry->node);
- if (MXactCacheMembers++ >= MAX_CACHE_ENTRIES)
+ if (MXactCacheMembers++ >= multixact_local_cache_entries)
{
dlist_node *node;
mXactCacheEnt *entry;
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
index 6ab8216839..9ca71933dc 100644
--- a/src/backend/utils/init/globals.c
+++ b/src/backend/utils/init/globals.c
@@ -149,3 +149,5 @@ int VacuumCostBalance = 0; /* working state for vacuum */
bool VacuumCostActive = false;
double vacuum_cleanup_index_scale_factor;
+
+int multixact_local_cache_entries = 256;
diff --git a/src/backend/utils/misc/guc.c b/src/backend/utils/misc/guc.c
index a62d64eaa4..d667578cc3 100644
--- a/src/backend/utils/misc/guc.c
+++ b/src/backend/utils/misc/guc.c
@@ -2257,6 +2257,16 @@ static struct config_int ConfigureNamesInt[] =
NULL, NULL, NULL
},
+ {
+ {"multixact_local_cache_entries", PGC_SUSET, RESOURCES_MEM,
+ gettext_noop("Sets the number of cached MultiXact by backend."),
+ NULL
+ },
+ &multixact_local_cache_entries,
+ 256, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
{
{"temp_buffers", PGC_USERSET, RESOURCES_MEM,
gettext_noop("Sets the maximum number of temporary buffers used by each session."),
diff --git a/src/include/miscadmin.h b/src/include/miscadmin.h
index 72e3352398..01af61c963 100644
--- a/src/include/miscadmin.h
+++ b/src/include/miscadmin.h
@@ -162,6 +162,8 @@ extern PGDLLIMPORT int MaxConnections;
extern PGDLLIMPORT int max_worker_processes;
extern PGDLLIMPORT int max_parallel_workers;
+extern PGDLLIMPORT int multixact_local_cache_entries;
+
extern PGDLLIMPORT int MyProcPid;
extern PGDLLIMPORT pg_time_t MyStartTime;
extern PGDLLIMPORT TimestampTz MyStartTimestamp;
--
2.24.3 (Apple Git-128)
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2020-10-28 19:36 Alexander Korotkov <[email protected]>
parent: Tomas Vondra <[email protected]>
1 sibling, 1 reply; 103+ messages in thread
From: Alexander Korotkov @ 2020-10-28 19:36 UTC (permalink / raw)
To: Tomas Vondra <[email protected]>; +Cc: Andrey Borodin <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
Hi, Tomas!
Thank you for your review.
On Wed, Oct 28, 2020 at 4:36 AM Tomas Vondra
<[email protected]> wrote:
> I did a quick review on this patch series. A couple comments:
>
>
> 0001
> ----
>
> This looks quite suspicious to me - SimpleLruReadPage_ReadOnly is
> changed to return information about what lock was used, merely to allow
> the callers to do an Assert() that the value is not LW_NONE.
Yes, but this is not merely to allow callers to do an Assert().
Sometimes in multixacts it could save us some relocks. So, we can
skip relocking lock to exclusive mode if it's in exclusive already.
Adding Assert() to every caller is probably overkill.
> IMO we could achieve exactly the same thing by passing a simple flag
> that would say 'make sure we got a lock' or something like that. In
> fact, aren't all callers doing the assert? That'd mean we can just do
> the check always, without the flag. (I see GetMultiXactIdMembers does
> two calls and only checks the second result, but I wonder if that's
> intended or omission.)
Having just the flag is exactly what the original version by Andrey
did. But if we have to read two multixact offsets pages or multiple
members page in one GetMultiXactIdMembers()), then it does relocks
from exclusive mode to exclusive mode. I decide that once we decide
to optimize this locks, this situation is nice to evade.
> In any case, it'd make the lwlock.c changes unnecessary, I think.
I agree that it would be better to not touch lwlock.c. But I didn't
find a way to evade relocking exclusive mode to exclusive mode without
touching lwlock.c or making code cumbersome in other places.
> 0002
> ----
>
> Specifies the number cached MultiXact by backend. Any SLRU lookup ...
>
> should be 'number of cached ...'
Sounds reasonable.
> 0003
> ----
>
> * Conditional variable for waiting till the filling of the next multixact
> * will be finished. See GetMultiXactIdMembers() and RecordNewMultiXact()
> * for details.
>
> Perhaps 'till the next multixact is filled' or 'gets full' would be
> better. Not sure.
Sounds reasonable as well.
------
Regards,
Alexander Korotkov
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2020-10-28 23:21 Tomas Vondra <[email protected]>
parent: Alexander Korotkov <[email protected]>
0 siblings, 0 replies; 103+ messages in thread
From: Tomas Vondra @ 2020-10-28 23:21 UTC (permalink / raw)
To: Alexander Korotkov <[email protected]>; +Cc: Andrey Borodin <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
On Wed, Oct 28, 2020 at 10:36:39PM +0300, Alexander Korotkov wrote:
>Hi, Tomas!
>
>Thank you for your review.
>
>On Wed, Oct 28, 2020 at 4:36 AM Tomas Vondra
><[email protected]> wrote:
>> I did a quick review on this patch series. A couple comments:
>>
>>
>> 0001
>> ----
>>
>> This looks quite suspicious to me - SimpleLruReadPage_ReadOnly is
>> changed to return information about what lock was used, merely to allow
>> the callers to do an Assert() that the value is not LW_NONE.
>
>Yes, but this is not merely to allow callers to do an Assert().
>Sometimes in multixacts it could save us some relocks. So, we can
>skip relocking lock to exclusive mode if it's in exclusive already.
>Adding Assert() to every caller is probably overkill.
>
Hmm, OK. That can only happen in GetMultiXactIdMembers, which is the
only place where we do retry, right? Do we actually know this makes any
measurable difference? It seems we're mostly imagining that it might
help, but we don't have any actual proof of that (e.g. a workload which
we might benchmark). Or am I missing something?
For me, the extra conditions make it way harder to reason about the
behavior of the code, and I can't convince myself it's worth it.
>> IMO we could achieve exactly the same thing by passing a simple flag
>> that would say 'make sure we got a lock' or something like that. In
>> fact, aren't all callers doing the assert? That'd mean we can just do
>> the check always, without the flag. (I see GetMultiXactIdMembers does
>> two calls and only checks the second result, but I wonder if that's
>> intended or omission.)
>
>Having just the flag is exactly what the original version by Andrey
>did. But if we have to read two multixact offsets pages or multiple
>members page in one GetMultiXactIdMembers()), then it does relocks
>from exclusive mode to exclusive mode. I decide that once we decide
>to optimize this locks, this situation is nice to evade.
>
>> In any case, it'd make the lwlock.c changes unnecessary, I think.
>
>I agree that it would be better to not touch lwlock.c. But I didn't
>find a way to evade relocking exclusive mode to exclusive mode without
>touching lwlock.c or making code cumbersome in other places.
>
Hmm. OK.
regards
--
Tomas Vondra http://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2020-10-28 23:32 Tomas Vondra <[email protected]>
parent: Andrey Borodin <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Tomas Vondra @ 2020-10-28 23:32 UTC (permalink / raw)
To: Andrey Borodin <[email protected]>; +Cc: Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
Hi,
On Wed, Oct 28, 2020 at 12:34:58PM +0500, Andrey Borodin wrote:
>Tomas, thanks for looking into this!
>
>> 28 окт. 2020 г., в 06:36, Tomas Vondra <[email protected]> написал(а):
>>
>>
>> This thread started with a discussion about making the SLRU sizes
>> configurable, but this patch version only adds a local cache. Does this
>> achieve the same goal, or would we still gain something by having GUCs
>> for the SLRUs?
>>
>> If we're claiming this improves performance, it'd be good to have some
>> workload demonstrating that and measurements. I don't see anything like
>> that in this thread, so it's a bit hand-wavy. Can someone share details
>> of such workload (even synthetic one) and some basic measurements?
>
>All patches in this thread aim at the same goal: improve performance in presence of MultiXact locks contention.
>I could not build synthetical reproduction of the problem, however I did some MultiXact stressing here [0]. It's a clumsy test program, because it still is not clear to me which parameters of workload trigger MultiXact locks contention. In generic case I was encountering other locks like *GenLock: XidGenLock, MultixactGenLock etc. Yet our production system encounters this problem approximately once in a month through this year.
>
>Test program locks for share different set of tuples in presence of concurrent full scans.
>To produce a set of locks we choose one of 14 bits. If a row number has this bit set to 0 we add lock it.
>I've been measuring time to lock all rows 3 time for each of 14 bits, observing total time to set all locks.
>During the test I was observing locks in pg_stat_activity, if they did not contain enough MultiXact locks I was tuning parameters further (number of concurrent clients, number of bits, select queries etc).
>
>Why is it so complicated? It seems that other reproductions of a problem were encountering other locks.
>
It's not my intention to be mean or anything like that, but to me this
means we don't really understand the problem we're trying to solve. Had
we understood it, we should be able to construct a workload reproducing
the issue ...
I understand what the individual patches are doing, and maybe those
changes are desirable in general. But without any benchmarks from a
plausible workload I find it hard to convince myself that:
(a) it actually will help with the issue you're observing on production
and
(b) it's actually worth the extra complexity (e.g. the lwlock changes)
I'm willing to invest some of my time into reviewing/testing this, but I
think we badly need better insight into the issue, so that we can build
a workload reproducing it. Perhaps collecting some perf profiles and a
sample of the queries might help, but I assume you already tried that.
regards
--
Tomas Vondra http://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2020-10-29 07:08 Andrey Borodin <[email protected]>
parent: Tomas Vondra <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Andrey Borodin @ 2020-10-29 07:08 UTC (permalink / raw)
To: Tomas Vondra <[email protected]>; +Cc: Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
> 29 окт. 2020 г., в 04:32, Tomas Vondra <[email protected]> написал(а):
>
> It's not my intention to be mean or anything like that, but to me this
> means we don't really understand the problem we're trying to solve. Had
> we understood it, we should be able to construct a workload reproducing
> the issue ...
>
> I understand what the individual patches are doing, and maybe those
> changes are desirable in general. But without any benchmarks from a
> plausible workload I find it hard to convince myself that:
>
> (a) it actually will help with the issue you're observing on production
>
> and
> (b) it's actually worth the extra complexity (e.g. the lwlock changes)
>
>
> I'm willing to invest some of my time into reviewing/testing this, but I
> think we badly need better insight into the issue, so that we can build
> a workload reproducing it. Perhaps collecting some perf profiles and a
> sample of the queries might help, but I assume you already tried that.
Thanks, Tomas! This totally makes sense.
Indeed, collecting queries did not help yet. We have loadtest environment equivalent to production (but with 10x less shards), copy of production workload queries. But the problem does not manifest there.
Why do I think the problem is in MultiXacts?
Here is a chart with number of wait events on each host
During the problem MultiXactOffsetControlLock and SLRURead dominate all other lock types. After primary switchover to another node SLRURead continued for a bit there, then disappeared.
Backtraces on standbys during the problem show that most of backends are sleeping in pg_sleep(1000L) and are not included into wait stats on these charts.
Currently I'm considering writing test that directly calls MultiXactIdExpand(), MultiXactIdCreate(), and GetMultiXactIdMembers() from an extension. How do you think, would benchmarks in such tests be meaningful?
Thanks!
Best regards, Andrey Borodin.
Attachments:
[image/png] Графика-1.png (422.0K, ../../[email protected]/3-%D0%93%D1%80%D0%B0%D1%84%D0%B8%D0%BA%D0%B0-1.png)
download | view image
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2020-10-29 13:49 Tomas Vondra <[email protected]>
parent: Andrey Borodin <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Tomas Vondra @ 2020-10-29 13:49 UTC (permalink / raw)
To: Andrey Borodin <[email protected]>; +Cc: Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
On Thu, Oct 29, 2020 at 12:08:21PM +0500, Andrey Borodin wrote:
>
>
>> 29 окт. 2020 г., в 04:32, Tomas Vondra <[email protected]> написал(а):
>>
>> It's not my intention to be mean or anything like that, but to me this
>> means we don't really understand the problem we're trying to solve. Had
>> we understood it, we should be able to construct a workload reproducing
>> the issue ...
>>
>> I understand what the individual patches are doing, and maybe those
>> changes are desirable in general. But without any benchmarks from a
>> plausible workload I find it hard to convince myself that:
>>
>> (a) it actually will help with the issue you're observing on production
>>
>> and
>> (b) it's actually worth the extra complexity (e.g. the lwlock changes)
>>
>>
>> I'm willing to invest some of my time into reviewing/testing this, but I
>> think we badly need better insight into the issue, so that we can build
>> a workload reproducing it. Perhaps collecting some perf profiles and a
>> sample of the queries might help, but I assume you already tried that.
>
>Thanks, Tomas! This totally makes sense.
>
>Indeed, collecting queries did not help yet. We have loadtest environment equivalent to production (but with 10x less shards), copy of production workload queries. But the problem does not manifest there.
>Why do I think the problem is in MultiXacts?
>Here is a chart with number of wait events on each host
>
>
>During the problem MultiXactOffsetControlLock and SLRURead dominate all other lock types. After primary switchover to another node SLRURead continued for a bit there, then disappeared.
OK, so most of this seems to be due to SLRURead and
MultiXactOffsetControlLock. Could it be that there were too many
multixact members, triggering autovacuum to prevent multixact
wraparound? That might generate a lot of IO on the SLRU. Are you
monitoring the size of the pg_multixact directory?
>Backtraces on standbys during the problem show that most of backends are sleeping in pg_sleep(1000L) and are not included into wait stats on these charts.
>
>Currently I'm considering writing test that directly calls MultiXactIdExpand(), MultiXactIdCreate(), and GetMultiXactIdMembers() from an extension. How do you think, would benchmarks in such tests be meaningful?
>
I don't know. I'd much rather have a SQL-level benchmark than an
extension doing this kind of stuff.
regards
--
Tomas Vondra http://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2020-11-02 12:45 Andrey Borodin <[email protected]>
parent: Tomas Vondra <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Andrey Borodin @ 2020-11-02 12:45 UTC (permalink / raw)
To: Tomas Vondra <[email protected]>; +Cc: Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
> 29 окт. 2020 г., в 18:49, Tomas Vondra <[email protected]> написал(а):
>
> On Thu, Oct 29, 2020 at 12:08:21PM +0500, Andrey Borodin wrote:
>>
>>
>>> 29 окт. 2020 г., в 04:32, Tomas Vondra <[email protected]> написал(а):
>>>
>>> It's not my intention to be mean or anything like that, but to me this
>>> means we don't really understand the problem we're trying to solve. Had
>>> we understood it, we should be able to construct a workload reproducing
>>> the issue ...
>>>
>>> I understand what the individual patches are doing, and maybe those
>>> changes are desirable in general. But without any benchmarks from a
>>> plausible workload I find it hard to convince myself that:
>>>
>>> (a) it actually will help with the issue you're observing on production
>>>
>>> and
>>> (b) it's actually worth the extra complexity (e.g. the lwlock changes)
>>>
>>>
>>> I'm willing to invest some of my time into reviewing/testing this, but I
>>> think we badly need better insight into the issue, so that we can build
>>> a workload reproducing it. Perhaps collecting some perf profiles and a
>>> sample of the queries might help, but I assume you already tried that.
>>
>> Thanks, Tomas! This totally makes sense.
>>
>> Indeed, collecting queries did not help yet. We have loadtest environment equivalent to production (but with 10x less shards), copy of production workload queries. But the problem does not manifest there.
>> Why do I think the problem is in MultiXacts?
>> Here is a chart with number of wait events on each host
>>
>>
>> During the problem MultiXactOffsetControlLock and SLRURead dominate all other lock types. After primary switchover to another node SLRURead continued for a bit there, then disappeared.
>
> OK, so most of this seems to be due to SLRURead and
> MultiXactOffsetControlLock. Could it be that there were too many
> multixact members, triggering autovacuum to prevent multixact
> wraparound? That might generate a lot of IO on the SLRU. Are you
> monitoring the size of the pg_multixact directory?
Yes, we had some problems with 'multixact "members" limit exceeded' long time ago.
We tuned autovacuum_multixact_freeze_max_age = 200000000 and vacuum_multixact_freeze_table_age = 75000000 (half of defaults) and since then did not ever encounter this problem (~5 months).
But the MultiXactOffsetControlLock problem persists. Partially the problem was solved by adding more shards. But when one of shards encounters a problem it's either MultiXacts or vacuum causing relation truncation (unrelated to this thread).
Thanks!
Best regards, Andrey Borodin.
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2020-11-10 00:13 Tomas Vondra <[email protected]>
parent: Andrey Borodin <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Tomas Vondra @ 2020-11-10 00:13 UTC (permalink / raw)
To: Andrey Borodin <[email protected]>; Tomas Vondra <[email protected]>; +Cc: Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
Hi,
After the issue reported in [1] got fixed, I've restarted the multi-xact
stress test, hoping to reproduce the issue. But so far no luck :-(
I've started slightly different tests on two machines - on one machine
I've done this:
a) init.sql
create table t (a int);
insert into t select i from generate_series(1,100000000) s(i);
alter table t add primary key (a);
b) select.sql
SELECT * FROM t
WHERE a = (1+mod(abs(hashint4(extract(epoch from now())::int)),
100000000)) FOR KEY SHARE;
c) pgbench -n -c 32 -j 8 -f select.sql -T $((24*3600)) test
The idea is to have large table and many clients hitting a small random
subset of the rows. A sample of wait events from ~24h run looks like this:
e_type | e_name | sum
----------+----------------------+----------
LWLock | BufferContent | 13913863
| | 7194679
LWLock | WALWrite | 1710507
Activity | LogicalLauncherMain | 726599
Activity | AutoVacuumMain | 726127
Activity | WalWriterMain | 725183
Activity | CheckpointerMain | 604694
Client | ClientRead | 599900
IO | WALSync | 502904
Activity | BgWriterMain | 378110
Activity | BgWriterHibernate | 348464
IO | WALWrite | 129412
LWLock | ProcArray | 6633
LWLock | WALInsert | 5714
IO | SLRUWrite | 2580
IPC | ProcArrayGroupUpdate | 2216
LWLock | XactSLRU | 2196
Timeout | VacuumDelay | 1078
IPC | XactGroupUpdate | 737
LWLock | LockManager | 503
LWLock | WALBufMapping | 295
LWLock | MultiXactMemberSLRU | 267
IO | DataFileWrite | 68
LWLock | BufferIO | 59
IO | DataFileRead | 27
IO | DataFileFlush | 14
LWLock | MultiXactGen | 7
LWLock | BufferMapping | 1
So, nothing particularly interesting - there certainly are not many wait
events related to SLRU.
On the other machine I did this:
a) init.sql
create table t (a int primary key);
insert into t select i from generate_series(1,1000) s(i);
b) select.sql
select * from t for key share;
c) pgbench -n -c 32 -j 8 -f select.sql -T $((24*3600)) test
and the wait events (24h run too) look like this:
e_type | e_name | sum
-----------+-----------------------+----------
LWLock | BufferContent | 20804925
| | 2575369
Activity | LogicalLauncherMain | 745780
Activity | AutoVacuumMain | 745292
Activity | WalWriterMain | 740507
Activity | CheckpointerMain | 737691
Activity | BgWriterHibernate | 731123
LWLock | WALWrite | 570107
IO | WALSync | 452603
Client | ClientRead | 151438
BufferPin | BufferPin | 23466
LWLock | WALInsert | 21631
IO | WALWrite | 19050
LWLock | ProcArray | 15082
Activity | BgWriterMain | 14655
IPC | ProcArrayGroupUpdate | 7772
LWLock | WALBufMapping | 3555
IO | SLRUWrite | 1951
LWLock | MultiXactGen | 1661
LWLock | MultiXactMemberSLRU | 359
LWLock | MultiXactOffsetSLRU | 242
LWLock | XactSLRU | 141
IPC | XactGroupUpdate | 104
LWLock | LockManager | 28
IO | DataFileRead | 4
IO | ControlFileSyncUpdate | 1
Timeout | VacuumDelay | 1
IO | WALInitWrite | 1
Also nothing particularly interesting - few SLRU wait events.
So unfortunately this does not really reproduce the SLRU locking issues
you're observing - clearly, there has to be something else triggering
it. Perhaps this workload is too simplistic, or maybe we need to run
different queries. Or maybe the hw needs to be somewhat different (more
CPUs? different storage?)
[1]
https://www.postgresql.org/message-id/20201104013205.icogbi773przyny5@development
regards
--
Tomas Vondra
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2020-11-10 06:16 Andrey Borodin <[email protected]>
parent: Tomas Vondra <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Andrey Borodin @ 2020-11-10 06:16 UTC (permalink / raw)
To: Tomas Vondra <[email protected]>; +Cc: Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
> 10 нояб. 2020 г., в 05:13, Tomas Vondra <[email protected]> написал(а):
> After the issue reported in [1] got fixed, I've restarted the multi-xact
> stress test, hoping to reproduce the issue. But so far no luck :-(
Tomas, many thanks for looking into this. I figured out that to make multixact sets bigger transactions must hang for a while and lock large set of tuples. But not continuous range to avoid locking on buffer_content.
I did not manage to implement this via pgbench, that's why I was trying to hack on separate go program. But, essentially, no luck either.
I was observing something resemblant though
пятница, 8 мая 2020 г. 15:08:37 (every 1s)
pid | wait_event | wait_event_type | state | query
-------+----------------------------+-----------------+--------+----------------------------------------------------
41344 | ClientRead | Client | idle | insert into t1 select generate_series(1,1000000,1)
41375 | MultiXactOffsetControlLock | LWLock | active | select * from t1 where i = ANY ($1) for share
41377 | MultiXactOffsetControlLock | LWLock | active | select * from t1 where i = ANY ($1) for share
41378 | | | active | select * from t1 where i = ANY ($1) for share
41379 | MultiXactOffsetControlLock | LWLock | active | select * from t1 where i = ANY ($1) for share
41381 | | | active | select * from t1 where i = ANY ($1) for share
41383 | MultiXactOffsetControlLock | LWLock | active | select * from t1 where i = ANY ($1) for share
41385 | MultiXactOffsetControlLock | LWLock | active | select * from t1 where i = ANY ($1) for share
(8 rows)
but this picture was not stable.
How do you collect wait events for aggregation? just insert into some table with cron?
Thanks!
Best regards, Andrey Borodin.
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2020-11-10 18:07 Tomas Vondra <[email protected]>
parent: Andrey Borodin <[email protected]>
0 siblings, 2 replies; 103+ messages in thread
From: Tomas Vondra @ 2020-11-10 18:07 UTC (permalink / raw)
To: Andrey Borodin <[email protected]>; +Cc: Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
On 11/10/20 7:16 AM, Andrey Borodin wrote:
>
>
>> 10 нояб. 2020 г., в 05:13, Tomas Vondra <[email protected]> написал(а):
>> After the issue reported in [1] got fixed, I've restarted the multi-xact
>> stress test, hoping to reproduce the issue. But so far no luck :-(
>
>
> Tomas, many thanks for looking into this. I figured out that to make multixact sets bigger transactions must hang for a while and lock large set of tuples. But not continuous range to avoid locking on buffer_content.
> I did not manage to implement this via pgbench, that's why I was trying to hack on separate go program. But, essentially, no luck either.
> I was observing something resemblant though
>
> пятница, 8 мая 2020 г. 15:08:37 (every 1s)
>
> pid | wait_event | wait_event_type | state | query
> -------+----------------------------+-----------------+--------+----------------------------------------------------
> 41344 | ClientRead | Client | idle | insert into t1 select generate_series(1,1000000,1)
> 41375 | MultiXactOffsetControlLock | LWLock | active | select * from t1 where i = ANY ($1) for share
> 41377 | MultiXactOffsetControlLock | LWLock | active | select * from t1 where i = ANY ($1) for share
> 41378 | | | active | select * from t1 where i = ANY ($1) for share
> 41379 | MultiXactOffsetControlLock | LWLock | active | select * from t1 where i = ANY ($1) for share
> 41381 | | | active | select * from t1 where i = ANY ($1) for share
> 41383 | MultiXactOffsetControlLock | LWLock | active | select * from t1 where i = ANY ($1) for share
> 41385 | MultiXactOffsetControlLock | LWLock | active | select * from t1 where i = ANY ($1) for share
> (8 rows)
>
> but this picture was not stable.
>
Seems we haven't made much progress in reproducing the issue :-( I guess
we'll need to know more about the machine where this happens. Is there
anything special about the hardware/config? Are you monitoring size of
the pg_multixact directory?
> How do you collect wait events for aggregation? just insert into some table with cron?
>
No, I have a simple shell script (attached) sampling data from
pg_stat_activity regularly. Then I load it into a table and aggregate to
get a summary.
regards
--
Tomas Vondra
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
Attachments:
[application/x-shellscript] collect-wait-events.sh (234B, ../../[email protected]/2-collect-wait-events.sh)
download
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2020-11-10 18:41 Thomas Munro <[email protected]>
parent: Tomas Vondra <[email protected]>
1 sibling, 0 replies; 103+ messages in thread
From: Thomas Munro @ 2020-11-10 18:41 UTC (permalink / raw)
To: Tomas Vondra <[email protected]>; +Cc: Andrey Borodin <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
On Wed, Nov 11, 2020 at 7:07 AM Tomas Vondra
<[email protected]> wrote:
> Seems we haven't made much progress in reproducing the issue :-( I guess
> we'll need to know more about the machine where this happens. Is there
> anything special about the hardware/config? Are you monitoring size of
> the pg_multixact directory?
Which release was the original problem seen on?
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2020-11-13 11:49 Andrey Borodin <[email protected]>
parent: Tomas Vondra <[email protected]>
1 sibling, 1 reply; 103+ messages in thread
From: Andrey Borodin @ 2020-11-13 11:49 UTC (permalink / raw)
To: Tomas Vondra <[email protected]>; +Cc: Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
> 10 нояб. 2020 г., в 23:07, Tomas Vondra <[email protected]> написал(а):
>
> On 11/10/20 7:16 AM, Andrey Borodin wrote:
>>
>>
>> but this picture was not stable.
>>
>
> Seems we haven't made much progress in reproducing the issue :-( I guess
> we'll need to know more about the machine where this happens. Is there
> anything special about the hardware/config? Are you monitoring size of
> the pg_multixact directory?
It's Ubuntu 18.04.4 LTS, Intel Xeon E5-2660 v4, 56 CPU cores with 256Gb of RAM.
PostgreSQL 10.14, compiled by gcc 7.5.0, 64-bit
No, unfortunately we do not have signals for SLRU sizes.
3.5Tb mdadm raid10 over 28 SSD drives, 82% full.
First incident triggering investigation was on 2020-04-19, at that time cluster was running on PG 10.11. But I think it was happening before.
I'd say nothing special...
>
>> How do you collect wait events for aggregation? just insert into some table with cron?
>>
>
> No, I have a simple shell script (attached) sampling data from
> pg_stat_activity regularly. Then I load it into a table and aggregate to
> get a summary.
Thanks!
Best regards, Andrey Borodin.
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2020-12-08 16:05 Gilles Darold <[email protected]>
parent: Andrey Borodin <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Gilles Darold @ 2020-12-08 16:05 UTC (permalink / raw)
To: Andrey Borodin <[email protected]>; Tomas Vondra <[email protected]>; +Cc: Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
Le 13/11/2020 à 12:49, Andrey Borodin a écrit :
>
>> 10 нояб. 2020 г., в 23:07, Tomas Vondra <[email protected]> написал(а):
>>
>> On 11/10/20 7:16 AM, Andrey Borodin wrote:
>>>
>>> but this picture was not stable.
>>>
>> Seems we haven't made much progress in reproducing the issue :-( I guess
>> we'll need to know more about the machine where this happens. Is there
>> anything special about the hardware/config? Are you monitoring size of
>> the pg_multixact directory?
> It's Ubuntu 18.04.4 LTS, Intel Xeon E5-2660 v4, 56 CPU cores with 256Gb of RAM.
> PostgreSQL 10.14, compiled by gcc 7.5.0, 64-bit
>
> No, unfortunately we do not have signals for SLRU sizes.
> 3.5Tb mdadm raid10 over 28 SSD drives, 82% full.
>
> First incident triggering investigation was on 2020-04-19, at that time cluster was running on PG 10.11. But I think it was happening before.
>
> I'd say nothing special...
>
>>> How do you collect wait events for aggregation? just insert into some table with cron?
>>>
>> No, I have a simple shell script (attached) sampling data from
>> pg_stat_activity regularly. Then I load it into a table and aggregate to
>> get a summary.
> Thanks!
>
> Best regards, Andrey Borodin.
Hi,
Some time ago I have encountered a contention on
MultiXactOffsetControlLock with a performances benchmark. Here are the
wait event monitoring result with a pooling each 10 seconds and a 30
minutes run for the benchmarl:
event_type | event | sum
------------+----------------------------+----------
Client | ClientRead | 44722952
LWLock | MultiXactOffsetControlLock | 30343060
LWLock | multixact_offset | 16735250
LWLock | MultiXactMemberControlLock | 1601470
LWLock | buffer_content | 991344
LWLock | multixact_member | 805624
Lock | transactionid | 204997
Activity | LogicalLauncherMain | 198834
Activity | CheckpointerMain | 198834
Activity | AutoVacuumMain | 198469
Activity | BgWriterMain | 184066
Activity | WalWriterMain | 171571
LWLock | WALWriteLock | 72428
IO | DataFileRead | 35708
Activity | BgWriterHibernate | 12741
IO | SLRURead | 9121
Lock | relation | 8858
LWLock | ProcArrayLock | 7309
LWLock | lock_manager | 6677
LWLock | pg_stat_statements | 4194
LWLock | buffer_mapping | 3222
After reading this thread I change the value of the buffer size to 32
and 64 and obtain the following results:
event_type | event | sum
------------+----------------------------+-----------
Client | ClientRead | 268297572
LWLock | MultiXactMemberControlLock | 65162906
LWLock | multixact_member | 33397714
LWLock | buffer_content | 4737065
Lock | transactionid | 2143750
LWLock | SubtransControlLock | 1318230
LWLock | WALWriteLock | 1038999
Activity | LogicalLauncherMain | 940598
Activity | AutoVacuumMain | 938566
Activity | CheckpointerMain | 799052
Activity | WalWriterMain | 749069
LWLock | subtrans | 710163
Activity | BgWriterHibernate | 536763
Lock | object | 514225
Activity | BgWriterMain | 394206
LWLock | lock_manager | 295616
IO | DataFileRead | 274236
LWLock | ProcArrayLock | 77099
Lock | tuple | 59043
IO | CopyFileWrite | 45611
Lock | relation | 42714
There was still contention on multixact but less than the first run. I
have increased the buffers to 128 and 512 and obtain the best results
for this bench:
event_type | event | sum
------------+----------------------------+-----------
Client | ClientRead | 160463037
LWLock | MultiXactMemberControlLock | 5334188
LWLock | buffer_content | 5228256
LWLock | buffer_mapping | 2368505
LWLock | SubtransControlLock | 2289977
IPC | ProcArrayGroupUpdate | 1560875
LWLock | ProcArrayLock | 1437750
Lock | transactionid | 825561
LWLock | subtrans | 772701
LWLock | WALWriteLock | 666138
Activity | LogicalLauncherMain | 492585
Activity | CheckpointerMain | 492458
Activity | AutoVacuumMain | 491548
LWLock | lock_manager | 426531
Lock | object | 403581
Activity | WalWriterMain | 394668
Activity | BgWriterHibernate | 293112
Activity | BgWriterMain | 195312
LWLock | MultiXactGenLock | 177820
LWLock | pg_stat_statements | 173864
IO | DataFileRead | 173009
I hope these metrics can have some interest to show the utility of this
patch but unfortunately I can not be more precise and provide reports
for the entire patch. The problem is that this benchmark is run on an
application that use PostgreSQL 11 and I can not back port the full
patch, there was too much changes since PG11. I have just increase the
size of NUM_MXACTOFFSET_BUFFERS and NUM_MXACTMEMBER_BUFFERS. This allow
us to triple the number of simultaneous connections between the first
and the last test.
I know that this report is not really helpful but at least I can give
more information on the benchmark that was used. This is the proprietary
zRef benchmark which compares the same Cobol programs (transactional and
batch) executed both on mainframes and on x86 servers. Instead of a DB2
z/os database we use PostgreSQL v11. This test has extensive use of
cursors (each select, even read only, is executed through a cursor) and
the contention was observed with update on tables with some foreign
keys. There is no explicit FOR SHARE on the queries, only some FOR
UPDATE clauses. I guess that the multixact contention is the result of
the for share locks produced for FK.
So in our case being able to tune the multixact buffers could help a lot
to improve the performances.
--
Gilles Darold
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2020-12-08 17:52 Andrey Borodin <[email protected]>
parent: Gilles Darold <[email protected]>
0 siblings, 2 replies; 103+ messages in thread
From: Andrey Borodin @ 2020-12-08 17:52 UTC (permalink / raw)
To: Gilles Darold <[email protected]>; +Cc: Tomas Vondra <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
Hi Gilles!
Many thanks for your message!
> 8 дек. 2020 г., в 21:05, Gilles Darold <[email protected]> написал(а):
>
> I know that this report is not really helpful
Quite contrary - this benchmarks prove that controllable reproduction exists. I've rebased patches for PG11. Can you please benchmark them (without extending SLRU)?
Best regards, Andrey Borodin.
Attachments:
[application/octet-stream] v1106-0001-Use-shared-lock-in-GetMultiXactIdMembers-for-o.patch (12.7K, ../../[email protected]/2-v1106-0001-Use-shared-lock-in-GetMultiXactIdMembers-for-o.patch)
download | inline diff:
From dee5bf87314606e020a574b85cf1e9cbe066588d Mon Sep 17 00:00:00 2001
From: Alexander Korotkov <[email protected]>
Date: Tue, 27 Oct 2020 19:29:56 +0300
Subject: [PATCH v1106 1/4] Use shared lock in GetMultiXactIdMembers for
offsets and members
Previously the read of multixact required exclusive control locks for both
offsets and members SLRUs. This could lead to the excessive lock contention.
This commit we makes multixacts SLRU take advantage of SimpleLruReadPage_ReadOnly
similar to clog, commit_ts and subtrans.
In order to evade extra reacquiring of CLRU lock, we teach
SimpleLruReadPage_ReadOnly() to take into account the current lock mode and
report resulting lock mode back.
Discussion: https://postgr.es/m/a7f1c4e1-1015-92a4-2bd4-6736bd13d03e%40postgrespro.ru#c496c4e75fc0605094a0e1f763e6a6ec
Author: Andrey Borodin
Reviewed-by: Kyotaro Horiguchi, Daniel Gustafsson
Reviewed-by: Anastasia Lubennikova, Alexander Korotkov
---
src/backend/access/transam/clog.c | 3 ++-
src/backend/access/transam/commit_ts.c | 3 ++-
src/backend/access/transam/multixact.c | 21 ++++++++++++++-------
src/backend/access/transam/slru.c | 23 +++++++++++++++++------
src/backend/access/transam/subtrans.c | 3 ++-
src/backend/commands/async.c | 6 ++++--
src/backend/storage/lmgr/lwlock.c | 3 +++
src/backend/storage/lmgr/predicate.c | 4 +++-
src/include/access/slru.h | 8 ++++----
src/include/storage/lwlock.h | 2 ++
10 files changed, 53 insertions(+), 23 deletions(-)
diff --git a/src/backend/access/transam/clog.c b/src/backend/access/transam/clog.c
index a8d1080f17..bdfbd10f96 100644
--- a/src/backend/access/transam/clog.c
+++ b/src/backend/access/transam/clog.c
@@ -640,10 +640,11 @@ TransactionIdGetStatus(TransactionId xid, XLogRecPtr *lsn)
int lsnindex;
char *byteptr;
XidStatus status;
+ LWLockMode lockmode = LW_NONE;
/* lock is acquired by SimpleLruReadPage_ReadOnly */
- slotno = SimpleLruReadPage_ReadOnly(ClogCtl, pageno, xid);
+ slotno = SimpleLruReadPage_ReadOnly(ClogCtl, pageno, xid, &lockmode);
byteptr = ClogCtl->shared->page_buffer[slotno] + byteno;
status = (*byteptr >> bshift) & CLOG_XACT_BITMASK;
diff --git a/src/backend/access/transam/commit_ts.c b/src/backend/access/transam/commit_ts.c
index 6c4911d9bc..d7b6f583e1 100644
--- a/src/backend/access/transam/commit_ts.c
+++ b/src/backend/access/transam/commit_ts.c
@@ -288,6 +288,7 @@ TransactionIdGetCommitTsData(TransactionId xid, TimestampTz *ts,
CommitTimestampEntry entry;
TransactionId oldestCommitTsXid;
TransactionId newestCommitTsXid;
+ LWLockMode lockmode = LW_NONE;
if (!TransactionIdIsValid(xid))
ereport(ERROR,
@@ -342,7 +343,7 @@ TransactionIdGetCommitTsData(TransactionId xid, TimestampTz *ts,
}
/* lock is acquired by SimpleLruReadPage_ReadOnly */
- slotno = SimpleLruReadPage_ReadOnly(CommitTsCtl, pageno, xid);
+ slotno = SimpleLruReadPage_ReadOnly(CommitTsCtl, pageno, xid, &lockmode);
memcpy(&entry,
CommitTsCtl->shared->page_buffer[slotno] +
SizeOfCommitTimestampEntry * entryno,
diff --git a/src/backend/access/transam/multixact.c b/src/backend/access/transam/multixact.c
index 365daf153a..5ca10a139a 100644
--- a/src/backend/access/transam/multixact.c
+++ b/src/backend/access/transam/multixact.c
@@ -1216,6 +1216,7 @@ GetMultiXactIdMembers(MultiXactId multi, MultiXactMember **members,
MultiXactId tmpMXact;
MultiXactOffset nextOffset;
MultiXactMember *ptr;
+ LWLockMode lockmode = LW_NONE;
debug_elog3(DEBUG2, "GetMembers: asked for %u", multi);
@@ -1319,12 +1320,13 @@ GetMultiXactIdMembers(MultiXactId multi, MultiXactMember **members,
* time on every multixact creation.
*/
retry:
- LWLockAcquire(MultiXactOffsetControlLock, LW_EXCLUSIVE);
pageno = MultiXactIdToOffsetPage(multi);
entryno = MultiXactIdToOffsetEntry(multi);
- slotno = SimpleLruReadPage(MultiXactOffsetCtl, pageno, true, multi);
+ /* lock is acquired by SimpleLruReadPage_ReadOnly */
+ slotno = SimpleLruReadPage_ReadOnly(MultiXactOffsetCtl, pageno,
+ multi, &lockmode);
offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
offptr += entryno;
offset = *offptr;
@@ -1356,7 +1358,8 @@ retry:
entryno = MultiXactIdToOffsetEntry(tmpMXact);
if (pageno != prev_pageno)
- slotno = SimpleLruReadPage(MultiXactOffsetCtl, pageno, true, tmpMXact);
+ slotno = SimpleLruReadPage_ReadOnly(MultiXactOffsetCtl, pageno,
+ tmpMXact, &lockmode);
offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
offptr += entryno;
@@ -1380,8 +1383,7 @@ retry:
*members = ptr;
/* Now get the members themselves. */
- LWLockAcquire(MultiXactMemberControlLock, LW_EXCLUSIVE);
-
+ lockmode = LW_NONE;
truelength = 0;
prev_pageno = -1;
for (i = 0; i < length; i++, offset++)
@@ -1397,7 +1399,8 @@ retry:
if (pageno != prev_pageno)
{
- slotno = SimpleLruReadPage(MultiXactMemberCtl, pageno, true, multi);
+ slotno = SimpleLruReadPage_ReadOnly(MultiXactMemberCtl, pageno,
+ multi, &lockmode);
prev_pageno = pageno;
}
@@ -1420,6 +1423,7 @@ retry:
truelength++;
}
+ Assert(lockmode != LW_NONE);
LWLockRelease(MultiXactMemberControlLock);
/*
@@ -2723,6 +2727,7 @@ find_multixact_start(MultiXactId multi, MultiXactOffset *result)
int entryno;
int slotno;
MultiXactOffset *offptr;
+ LWLockMode lockmode = LW_NONE;
Assert(MultiXactState->finishedStartup);
@@ -2743,10 +2748,12 @@ find_multixact_start(MultiXactId multi, MultiXactOffset *result)
return false;
/* lock is acquired by SimpleLruReadPage_ReadOnly */
- slotno = SimpleLruReadPage_ReadOnly(MultiXactOffsetCtl, pageno, multi);
+ slotno = SimpleLruReadPage_ReadOnly(MultiXactOffsetCtl, pageno,
+ multi, &lockmode);
offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
offptr += entryno;
offset = *offptr;
+ Assert(lockmode != LW_NONE);
LWLockRelease(MultiXactOffsetControlLock);
*result = offset;
diff --git a/src/backend/access/transam/slru.c b/src/backend/access/transam/slru.c
index cfe9513827..f86950eb51 100644
--- a/src/backend/access/transam/slru.c
+++ b/src/backend/access/transam/slru.c
@@ -460,17 +460,23 @@ SimpleLruReadPage(SlruCtl ctl, int pageno, bool write_ok,
* Return value is the shared-buffer slot number now holding the page.
* The buffer's LRU access info is updated.
*
- * Control lock must NOT be held at entry, but will be held at exit.
- * It is unspecified whether the lock will be shared or exclusive.
+ * Control lock must be held in *lock_mode mode, which may be LW_NONE. Control
+ * lock will be held at exit in at least shared mode. Resulting control lock
+ * mode is set to *lock_mode.
*/
int
-SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno, TransactionId xid)
+SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno, TransactionId xid,
+ LWLockMode *lock_mode)
{
SlruShared shared = ctl->shared;
int slotno;
/* Try to find the page while holding only shared lock */
- LWLockAcquire(shared->ControlLock, LW_SHARED);
+ if (*lock_mode == LW_NONE)
+ {
+ LWLockAcquire(shared->ControlLock, LW_SHARED);
+ *lock_mode = LW_SHARED;
+ }
/* See if page is already in a buffer */
for (slotno = 0; slotno < shared->num_slots; slotno++)
@@ -486,8 +492,13 @@ SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno, TransactionId xid)
}
/* No luck, so switch to normal exclusive lock and do regular read */
- LWLockRelease(shared->ControlLock);
- LWLockAcquire(shared->ControlLock, LW_EXCLUSIVE);
+ if (*lock_mode != LW_EXCLUSIVE)
+ {
+ Assert(*lock_mode == LW_NONE);
+ LWLockRelease(shared->ControlLock);
+ LWLockAcquire(shared->ControlLock, LW_EXCLUSIVE);
+ *lock_mode = LW_EXCLUSIVE;
+ }
return SimpleLruReadPage(ctl, pageno, true, xid);
}
diff --git a/src/backend/access/transam/subtrans.c b/src/backend/access/transam/subtrans.c
index ef63b6d98a..440a21415a 100644
--- a/src/backend/access/transam/subtrans.c
+++ b/src/backend/access/transam/subtrans.c
@@ -113,6 +113,7 @@ SubTransGetParent(TransactionId xid)
int slotno;
TransactionId *ptr;
TransactionId parent;
+ LWLockMode lockmode = LW_NONE;
/* Can't ask about stuff that might not be around anymore */
Assert(TransactionIdFollowsOrEquals(xid, TransactionXmin));
@@ -123,7 +124,7 @@ SubTransGetParent(TransactionId xid)
/* lock is acquired by SimpleLruReadPage_ReadOnly */
- slotno = SimpleLruReadPage_ReadOnly(SubTransCtl, pageno, xid);
+ slotno = SimpleLruReadPage_ReadOnly(SubTransCtl, pageno, xid, &lockmode);
ptr = (TransactionId *) SubTransCtl->shared->page_buffer[slotno];
ptr += entryno;
diff --git a/src/backend/commands/async.c b/src/backend/commands/async.c
index 7f4e11bd90..4ebc069a5d 100644
--- a/src/backend/commands/async.c
+++ b/src/backend/commands/async.c
@@ -1834,6 +1834,7 @@ asyncQueueReadAllNotifications(void)
int curoffset = QUEUE_POS_OFFSET(pos);
int slotno;
int copysize;
+ LWLockMode lockmode = LW_NONE;
/*
* We copy the data from SLRU into a local buffer, so as to avoid
@@ -1842,7 +1843,7 @@ asyncQueueReadAllNotifications(void)
* of the page we will actually inspect.
*/
slotno = SimpleLruReadPage_ReadOnly(AsyncCtl, curpage,
- InvalidTransactionId);
+ InvalidTransactionId, &lockmode);
if (curpage == QUEUE_POS_PAGE(head))
{
/* we only want to read as far as head */
@@ -1859,7 +1860,8 @@ asyncQueueReadAllNotifications(void)
AsyncCtl->shared->page_buffer[slotno] + curoffset,
copysize);
/* Release lock that we got from SimpleLruReadPage_ReadOnly() */
- LWLockRelease(AsyncCtlLock);
+ Assert(lockmode != LW_NONE);
+ LWLockRelease(AsyncCtl);
/*
* Process messages up to the stop position, end of page, or an
diff --git a/src/backend/storage/lmgr/lwlock.c b/src/backend/storage/lmgr/lwlock.c
index 5aae233995..066dbb8ccf 100644
--- a/src/backend/storage/lmgr/lwlock.c
+++ b/src/backend/storage/lmgr/lwlock.c
@@ -982,6 +982,8 @@ LWLockWakeup(LWLock *lock)
static void
LWLockQueueSelf(LWLock *lock, LWLockMode mode)
{
+ Assert(mode != LW_NONE);
+
/*
* If we don't have a PGPROC structure, there's no way to wait. This
* should never occur, since MyProc should only be null during shared
@@ -1742,6 +1744,7 @@ LWLockRelease(LWLock *lock)
elog(ERROR, "lock %s is not held", T_NAME(lock));
mode = held_lwlocks[i].mode;
+ Assert(mode == LW_EXCLUSIVE || mode == LW_SHARED);
num_held_lwlocks--;
for (; i < num_held_lwlocks; i++)
diff --git a/src/backend/storage/lmgr/predicate.c b/src/backend/storage/lmgr/predicate.c
index 002d040ba6..9513fd915b 100644
--- a/src/backend/storage/lmgr/predicate.c
+++ b/src/backend/storage/lmgr/predicate.c
@@ -908,6 +908,7 @@ OldSerXidGetMinConflictCommitSeqNo(TransactionId xid)
TransactionId tailXid;
SerCommitSeqNo val;
int slotno;
+ LWLockMode lockmode = LW_NONE;
Assert(TransactionIdIsValid(xid));
@@ -930,8 +931,9 @@ OldSerXidGetMinConflictCommitSeqNo(TransactionId xid)
* but will return with that lock held, which must then be released.
*/
slotno = SimpleLruReadPage_ReadOnly(OldSerXidSlruCtl,
- OldSerXidPage(xid), xid);
+ OldSerXidPage(xid), xid, &lockmode);
val = OldSerXidValue(slotno, xid);
+ Assert(lockmode != LW_NONE);
LWLockRelease(OldSerXidLock);
return val;
}
diff --git a/src/include/access/slru.h b/src/include/access/slru.h
index 0e89e48c97..51a3df62e0 100644
--- a/src/include/access/slru.h
+++ b/src/include/access/slru.h
@@ -141,10 +141,10 @@ extern Size SimpleLruShmemSize(int nslots, int nlsns);
extern void SimpleLruInit(SlruCtl ctl, const char *name, int nslots, int nlsns,
LWLock *ctllock, const char *subdir, int tranche_id);
extern int SimpleLruZeroPage(SlruCtl ctl, int pageno);
-extern int SimpleLruReadPage(SlruCtl ctl, int pageno, bool write_ok,
- TransactionId xid);
-extern int SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno,
- TransactionId xid);
+extern int SimpleLruReadPage(SlruCtl ctl, int pageno, bool write_ok,
+ TransactionId xid);
+extern int SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno,
+ TransactionId xid, LWLockMode *lock_mode);
extern void SimpleLruWritePage(SlruCtl ctl, int slotno);
extern void SimpleLruFlush(SlruCtl ctl, bool allow_redirtied);
extern void SimpleLruTruncate(SlruCtl ctl, int cutoffPage);
diff --git a/src/include/storage/lwlock.h b/src/include/storage/lwlock.h
index c21bfe2f66..3263c0fe62 100644
--- a/src/include/storage/lwlock.h
+++ b/src/include/storage/lwlock.h
@@ -131,6 +131,8 @@ extern PGDLLIMPORT int NamedLWLockTrancheRequests;
typedef enum LWLockMode
{
+ LW_NONE, /* Not a lock mode. Indicates that there is no
+ * lock. */
LW_EXCLUSIVE,
LW_SHARED,
LW_WAIT_UNTIL_FREE /* A special mode used in PGPROC->lwlockMode,
--
2.24.3 (Apple Git-128)
[application/octet-stream] v1106-0002-Make-MultiXact-local-cache-size-configurable.patch (4.8K, ../../[email protected]/3-v1106-0002-Make-MultiXact-local-cache-size-configurable.patch)
download | inline diff:
From 649c5c25349c32aecf02fad46197234af95f9860 Mon Sep 17 00:00:00 2001
From: Alexander Korotkov <[email protected]>
Date: Mon, 26 Oct 2020 03:44:50 +0300
Subject: [PATCH v1106 2/4] Make MultiXact local cache size configurable
---
doc/src/sgml/config.sgml | 37 ++++++++++++++++++++++++++
src/backend/access/transam/multixact.c | 2 +-
src/backend/utils/init/globals.c | 2 ++
src/backend/utils/misc/guc.c | 10 +++++++
src/include/miscadmin.h | 2 ++
5 files changed, 52 insertions(+), 1 deletion(-)
diff --git a/doc/src/sgml/config.sgml b/doc/src/sgml/config.sgml
index 3570b422be..15fb2f2bde 100644
--- a/doc/src/sgml/config.sgml
+++ b/doc/src/sgml/config.sgml
@@ -1637,6 +1637,43 @@ include_dir 'conf.d'
</listitem>
</varlistentry>
+ <varlistentry id="guc-logical-decoding-work-mem" xreflabel="logical_decoding_work_mem">
+ <term><varname>logical_decoding_work_mem</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>logical_decoding_work_mem</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the maximum amount of memory to be used by logical decoding,
+ before some of the decoded changes are written to local disk. This
+ limits the amount of memory used by logical streaming replication
+ connections. It defaults to 64 megabytes (<literal>64MB</literal>).
+ Since each replication connection only uses a single buffer of this size,
+ and an installation normally doesn't have many such connections
+ concurrently (as limited by <varname>max_wal_senders</varname>), it's
+ safe to set this value significantly higher than <varname>work_mem</varname>,
+ reducing the amount of decoded changes written to disk.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-multixact-local-cache-entries" xreflabel="multixact_local-cache-entries">
+ <term><varname>multixact_local_cache_entries</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>multixact_local_cache_entries</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the number cached MultiXact by backend. Any SLRU lookup is preceeded
+ by cache lookup. Higher numbers of cache size help to deduplicate lock sets, while
+ lower values make cache lookup faster.
+ It defaults to 256.
+ </para>
+ </listitem>
+ </varlistentry>
+
<varlistentry id="guc-max-stack-depth" xreflabel="max_stack_depth">
<term><varname>max_stack_depth</varname> (<type>integer</type>)
<indexterm>
diff --git a/src/backend/access/transam/multixact.c b/src/backend/access/transam/multixact.c
index 5ca10a139a..6203be0aa3 100644
--- a/src/backend/access/transam/multixact.c
+++ b/src/backend/access/transam/multixact.c
@@ -1591,7 +1591,7 @@ mXactCachePut(MultiXactId multi, int nmembers, MultiXactMember *members)
qsort(entry->members, nmembers, sizeof(MultiXactMember), mxactMemberComparator);
dlist_push_head(&MXactCache, &entry->node);
- if (MXactCacheMembers++ >= MAX_CACHE_ENTRIES)
+ if (MXactCacheMembers++ >= multixact_local_cache_entries)
{
dlist_node *node;
mXactCacheEnt *entry;
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
index f7d6617a13..22af834150 100644
--- a/src/backend/utils/init/globals.c
+++ b/src/backend/utils/init/globals.c
@@ -147,3 +147,5 @@ int VacuumCostBalance = 0; /* working state for vacuum */
bool VacuumCostActive = false;
double vacuum_cleanup_index_scale_factor;
+
+int multixact_local_cache_entries = 256;
diff --git a/src/backend/utils/misc/guc.c b/src/backend/utils/misc/guc.c
index 710344c72b..b54a063782 100644
--- a/src/backend/utils/misc/guc.c
+++ b/src/backend/utils/misc/guc.c
@@ -2026,6 +2026,16 @@ static struct config_int ConfigureNamesInt[] =
NULL, NULL, NULL
},
+ {
+ {"multixact_local_cache_entries", PGC_SUSET, RESOURCES_MEM,
+ gettext_noop("Sets the number of cached MultiXact by backend."),
+ NULL
+ },
+ &multixact_local_cache_entries,
+ 256, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
{
{"temp_buffers", PGC_USERSET, RESOURCES_MEM,
gettext_noop("Sets the maximum number of temporary buffers used by each session."),
diff --git a/src/include/miscadmin.h b/src/include/miscadmin.h
index 8024145535..11c3c8e6c5 100644
--- a/src/include/miscadmin.h
+++ b/src/include/miscadmin.h
@@ -161,6 +161,8 @@ extern PGDLLIMPORT int MaxConnections;
extern PGDLLIMPORT int max_worker_processes;
extern PGDLLIMPORT int max_parallel_workers;
+extern PGDLLIMPORT int multixact_local_cache_entries;
+
extern PGDLLIMPORT int MyProcPid;
extern PGDLLIMPORT pg_time_t MyStartTime;
extern PGDLLIMPORT struct Port *MyProcPort;
--
2.24.3 (Apple Git-128)
[application/octet-stream] v1106-0003-Add-conditional-variable-to-wait-for-next-Mult.patch (4.4K, ../../[email protected]/4-v1106-0003-Add-conditional-variable-to-wait-for-next-Mult.patch)
download | inline diff:
From 38dc550f363c1c8878fbf06d5616614a36112fac Mon Sep 17 00:00:00 2001
From: Alexander Korotkov <[email protected]>
Date: Mon, 26 Oct 2020 01:55:31 +0300
Subject: [PATCH v1106 3/4] Add conditional variable to wait for next MultXact
offset in corner case
GetMultiXactIdMembers() has a corner case, when the next multixact offset is
not yet set. In this case GetMultiXactIdMembers() has to sleep till this offset
is set. Currently the sleeping is implemented in naive way using pg_sleep()
and retry. This commit implements sleeping with conditional variable, which
provides more efficient way for waiting till the event.
Discussion: https://postgr.es/m/a7f1c4e1-1015-92a4-2bd4-6736bd13d03e%40postgrespro.ru#c496c4e75fc0605094a0e1f763e6a6ec
Author: Andrey Borodin
Reviewed-by: Kyotaro Horiguchi, Daniel Gustafsson
Reviewed-by: Anastasia Lubennikova, Alexander Korotkov
---
src/backend/access/transam/multixact.c | 35 ++++++++++++++++++++++++--
src/backend/postmaster/pgstat.c | 3 +++
src/include/pgstat.h | 3 ++-
3 files changed, 38 insertions(+), 3 deletions(-)
diff --git a/src/backend/access/transam/multixact.c b/src/backend/access/transam/multixact.c
index 6203be0aa3..5d2bbb1ca6 100644
--- a/src/backend/access/transam/multixact.c
+++ b/src/backend/access/transam/multixact.c
@@ -82,6 +82,7 @@
#include "lib/ilist.h"
#include "miscadmin.h"
#include "pg_trace.h"
+#include "pgstat.h"
#include "postmaster/autovacuum.h"
#include "storage/lmgr.h"
#include "storage/pmsignal.h"
@@ -233,6 +234,13 @@ typedef struct MultiXactStateData
/* support for members anti-wraparound measures */
MultiXactOffset offsetStopLimit; /* known if oldestOffsetKnown */
+ /*
+ * Conditional variable for waiting till the filling of the next multixact
+ * will be finished. See GetMultiXactIdMembers() and RecordNewMultiXact()
+ * for details.
+ */
+ ConditionVariable nextoffCV;
+
/*
* Per-backend data starts here. We have two arrays stored in the area
* immediately following the MultiXactStateData struct. Each is indexed by
@@ -871,6 +879,14 @@ RecordNewMultiXact(MultiXactId multi, MultiXactOffset offset,
/* Exchange our lock */
LWLockRelease(MultiXactOffsetControlLock);
+ /*
+ * Let everybody know the offset of this mxid is recorded now. The
+ * waiters are waiting for the offset of the mxid next of the target to
+ * know the number of members of the target mxid, so we don't need to wait
+ * for members of this mxid are recorded.
+ */
+ ConditionVariableBroadcast(&MultiXactState->nextoffCV);
+
LWLockAcquire(MultiXactMemberControlLock, LW_EXCLUSIVE);
prev_pageno = -1;
@@ -1368,9 +1384,23 @@ retry:
if (nextMXOffset == 0)
{
/* Corner case 2: next multixact is still being filled in */
+
+ /*
+ * The recorder of the next mxid is just before writing the
+ * offset. Wait for the offset to be written.
+ */
+ ConditionVariablePrepareToSleep(&MultiXactState->nextoffCV);
+
+ /*
+ * We don't have to recheck if multixact was filled in during
+ * ConditionVariablePrepareToSleep(), because we were holding
+ * MultiXactOffsetSLRULock.
+ */
LWLockRelease(MultiXactOffsetControlLock);
- CHECK_FOR_INTERRUPTS();
- pg_usleep(1000L);
+
+ ConditionVariableSleep(&MultiXactState->nextoffCV,
+ WAIT_EVENT_WAIT_NEXT_MXMEMBERS);
+ ConditionVariableCancelSleep();
goto retry;
}
@@ -1849,6 +1879,7 @@ MultiXactShmemInit(void)
/* Make sure we zero out the per-backend state */
MemSet(MultiXactState, 0, SHARED_MULTIXACT_STATE_SIZE);
+ ConditionVariableInit(&MultiXactState->nextoffCV);
}
else
Assert(found);
diff --git a/src/backend/postmaster/pgstat.c b/src/backend/postmaster/pgstat.c
index 4895354a28..8a8b9a4eef 100644
--- a/src/backend/postmaster/pgstat.c
+++ b/src/backend/postmaster/pgstat.c
@@ -3730,6 +3730,9 @@ pgstat_get_wait_ipc(WaitEventIPC w)
case WAIT_EVENT_SYNC_REP:
event_name = "SyncRep";
break;
+ case WAIT_EVENT_WAIT_NEXT_MXMEMBERS:
+ event_name = "MultiXactWaitNextMembers";
+ break;
/* no default case, so that compiler will warn */
}
diff --git a/src/include/pgstat.h b/src/include/pgstat.h
index 58e2e71c6f..e89d946669 100644
--- a/src/include/pgstat.h
+++ b/src/include/pgstat.h
@@ -832,7 +832,8 @@ typedef enum
WAIT_EVENT_REPLICATION_ORIGIN_DROP,
WAIT_EVENT_REPLICATION_SLOT_DROP,
WAIT_EVENT_SAFE_SNAPSHOT,
- WAIT_EVENT_SYNC_REP
+ WAIT_EVENT_SYNC_REP,
+ WAIT_EVENT_WAIT_NEXT_MXMEMBERS
} WaitEventIPC;
/* ----------
--
2.24.3 (Apple Git-128)
[application/octet-stream] v1106-0004-Add-GUCs-to-tune-MultiXact-SLRUs.patch (6.2K, ../../[email protected]/5-v1106-0004-Add-GUCs-to-tune-MultiXact-SLRUs.patch)
download | inline diff:
From e273a508592236dbf5c60417b5878961c2483bca Mon Sep 17 00:00:00 2001
From: Andrey Borodin <[email protected]>
Date: Sat, 9 May 2020 16:42:07 +0500
Subject: [PATCH v1106 4/4] Add GUCs to tune MultiXact SLRUs
---
doc/src/sgml/config.sgml | 31 ++++++++++++++++++++++++++
src/backend/access/transam/multixact.c | 8 +++----
src/backend/utils/init/globals.c | 2 ++
src/backend/utils/misc/guc.c | 22 ++++++++++++++++++
src/include/access/multixact.h | 4 ----
src/include/miscadmin.h | 2 ++
6 files changed, 61 insertions(+), 8 deletions(-)
diff --git a/doc/src/sgml/config.sgml b/doc/src/sgml/config.sgml
index 15fb2f2bde..44ebacc713 100644
--- a/doc/src/sgml/config.sgml
+++ b/doc/src/sgml/config.sgml
@@ -1673,6 +1673,37 @@ include_dir 'conf.d'
</para>
</listitem>
</varlistentry>
+
+ <varlistentry id="guc-multixact-offsets-slru-buffers" xreflabel="multixact_offsets_slru_buffers">
+ <term><varname>multixact_offsets_slru_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>multixact_offsets_slru_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of memory to be used for MultiXact offsets. MultiXact offsets
+ are used to store informaion about offsets of multiple row lockers (caused by SELECT FOR UPDATE and others).
+ It defaults to 64 kilobytes (<literal>64KB</literal>).
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-multixact-members-slru-buffers" xreflabel="multixact_members_slru_buffers">
+ <term><varname>multixact_members_slru_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>multixact_members_slru_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of memory to be used for MultiXact members. MultiXact members
+ are used to store informaion about XIDs of multiple row lockers. Tipically <varname>multixact_members_slru_buffers</varname>
+ is twice more than <varname>multixact_offsets_slru_buffers</varname>.
+ It defaults to 128 kilobytes (<literal>128KB</literal>).
+ </para>
+ </listitem>
+ </varlistentry>
<varlistentry id="guc-max-stack-depth" xreflabel="max_stack_depth">
<term><varname>max_stack_depth</varname> (<type>integer</type>)
diff --git a/src/backend/access/transam/multixact.c b/src/backend/access/transam/multixact.c
index 5d2bbb1ca6..31324714b8 100644
--- a/src/backend/access/transam/multixact.c
+++ b/src/backend/access/transam/multixact.c
@@ -1844,8 +1844,8 @@ MultiXactShmemSize(void)
mul_size(sizeof(MultiXactId) * 2, MaxOldestSlot))
size = SHARED_MULTIXACT_STATE_SIZE;
- size = add_size(size, SimpleLruShmemSize(NUM_MXACTOFFSET_BUFFERS, 0));
- size = add_size(size, SimpleLruShmemSize(NUM_MXACTMEMBER_BUFFERS, 0));
+ size = add_size(size, SimpleLruShmemSize(multixact_offsets_slru_buffers, 0));
+ size = add_size(size, SimpleLruShmemSize(multixact_members_slru_buffers, 0));
return size;
}
@@ -1861,11 +1861,11 @@ MultiXactShmemInit(void)
MultiXactMemberCtl->PagePrecedes = MultiXactMemberPagePrecedes;
SimpleLruInit(MultiXactOffsetCtl,
- "multixact_offset", NUM_MXACTOFFSET_BUFFERS, 0,
+ "multixact_offset", multixact_offsets_slru_buffers, 0,
MultiXactOffsetControlLock, "pg_multixact/offsets",
LWTRANCHE_MXACTOFFSET_BUFFERS);
SimpleLruInit(MultiXactMemberCtl,
- "multixact_member", NUM_MXACTMEMBER_BUFFERS, 0,
+ "multixact_member", multixact_members_slru_buffers, 0,
MultiXactMemberControlLock, "pg_multixact/members",
LWTRANCHE_MXACTMEMBER_BUFFERS);
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
index 22af834150..5a8517f8eb 100644
--- a/src/backend/utils/init/globals.c
+++ b/src/backend/utils/init/globals.c
@@ -149,3 +149,5 @@ bool VacuumCostActive = false;
double vacuum_cleanup_index_scale_factor;
int multixact_local_cache_entries = 256;
+int multixact_offsets_slru_buffers = 8;
+int multixact_members_slru_buffers = 16;
diff --git a/src/backend/utils/misc/guc.c b/src/backend/utils/misc/guc.c
index b54a063782..578452d757 100644
--- a/src/backend/utils/misc/guc.c
+++ b/src/backend/utils/misc/guc.c
@@ -2036,6 +2036,28 @@ static struct config_int ConfigureNamesInt[] =
NULL, NULL, NULL
},
+ {
+ {"multixact_offsets_slru_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for MultiXact offsets SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &multixact_offsets_slru_buffers,
+ 8, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"multixact_members_slru_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for MultiXact members SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &multixact_members_slru_buffers,
+ 16, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
{
{"temp_buffers", PGC_USERSET, RESOURCES_MEM,
gettext_noop("Sets the maximum number of temporary buffers used by each session."),
diff --git a/src/include/access/multixact.h b/src/include/access/multixact.h
index 18fe380c5f..6e14fb7b29 100644
--- a/src/include/access/multixact.h
+++ b/src/include/access/multixact.h
@@ -28,10 +28,6 @@
#define MaxMultiXactOffset ((MultiXactOffset) 0xFFFFFFFF)
-/* Number of SLRU buffers to use for multixact */
-#define NUM_MXACTOFFSET_BUFFERS 8
-#define NUM_MXACTMEMBER_BUFFERS 16
-
/*
* Possible multixact lock modes ("status"). The first four modes are for
* tuple locks (FOR KEY SHARE, FOR SHARE, FOR NO KEY UPDATE, FOR UPDATE); the
diff --git a/src/include/miscadmin.h b/src/include/miscadmin.h
index 11c3c8e6c5..e3fd0a264d 100644
--- a/src/include/miscadmin.h
+++ b/src/include/miscadmin.h
@@ -160,6 +160,8 @@ extern PGDLLIMPORT int MaxBackends;
extern PGDLLIMPORT int MaxConnections;
extern PGDLLIMPORT int max_worker_processes;
extern PGDLLIMPORT int max_parallel_workers;
+extern PGDLLIMPORT int multixact_offsets_slru_buffers;
+extern PGDLLIMPORT int multixact_members_slru_buffers;
extern PGDLLIMPORT int multixact_local_cache_entries;
--
2.24.3 (Apple Git-128)
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2020-12-09 10:51 Gilles Darold <[email protected]>
parent: Andrey Borodin <[email protected]>
1 sibling, 1 reply; 103+ messages in thread
From: Gilles Darold @ 2020-12-09 10:51 UTC (permalink / raw)
To: Andrey Borodin <[email protected]>; +Cc: Tomas Vondra <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
Hi Andrey,
Thanks for the backport. I have issue with the first patch "Use shared
lock in GetMultiXactIdMembers for offsets and members"
(v1106-0001-Use-shared-lock-in-GetMultiXactIdMembers-for-o.patch) the
applications are not working anymore when I'm applying it. Also PG
regression tests are failing too on several part.
test insert_conflict ... ok
test create_function_1 ... FAILED
test create_type ... FAILED
test create_table ... FAILED
test create_function_2 ... FAILED
test copy ... FAILED
test copyselect ... ok
test copydml ... ok
test create_misc ... FAILED
test create_operator ... FAILED
test create_procedure ... ok
test create_index ... FAILED
test index_including ... ok
test create_view ... FAILED
test create_aggregate ... ok
test create_function_3 ... ok
test create_cast ... ok
test constraints ... FAILED
test triggers ... FAILED
test inherit ...
^C
This is also where I left my last try to back port for PG11, I will try
to fix it again but it could take time to have it working.
Best regards,
--
Gilles Darold
Le 08/12/2020 à 18:52, Andrey Borodin a écrit :
> Hi Gilles!
>
> Many thanks for your message!
>
>> 8 дек. 2020 г., в 21:05, Gilles Darold <[email protected]> написал(а):
>>
>> I know that this report is not really helpful
> Quite contrary - this benchmarks prove that controllable reproduction exists. I've rebased patches for PG11. Can you please benchmark them (without extending SLRU)?
>
> Best regards, Andrey Borodin.
>
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2020-12-09 11:06 Gilles Darold <[email protected]>
parent: Gilles Darold <[email protected]>
0 siblings, 0 replies; 103+ messages in thread
From: Gilles Darold @ 2020-12-09 11:06 UTC (permalink / raw)
To: Andrey Borodin <[email protected]>; +Cc: Tomas Vondra <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
Le 09/12/2020 à 11:51, Gilles Darold a écrit :
> Also PG regression tests are failing too on several part.
Forget this, I have not run the regression tests in the right repository:
...
=======================
All 189 tests passed.
=======================
I'm looking why the application is failing.
--
Gilles Darold
http://www.darold.net/
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2020-12-10 14:45 Gilles Darold <[email protected]>
parent: Andrey Borodin <[email protected]>
1 sibling, 1 reply; 103+ messages in thread
From: Gilles Darold @ 2020-12-10 14:45 UTC (permalink / raw)
To: Andrey Borodin <[email protected]>; +Cc: Tomas Vondra <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
Le 08/12/2020 à 18:52, Andrey Borodin a écrit :
> Hi Gilles!
>
> Many thanks for your message!
>
>> 8 дек. 2020 г., в 21:05, Gilles Darold <[email protected]> написал(а):
>>
>> I know that this report is not really helpful
> Quite contrary - this benchmarks prove that controllable reproduction exists. I've rebased patches for PG11. Can you please benchmark them (without extending SLRU)?
>
> Best regards, Andrey Borodin.
>
Hi,
Running tests yesterday with the patches has reported log of failures
with error on INSERT and UPDATE statements:
ERROR: lock MultiXactOffsetControlLock is not held
After a patch review this morning I think I have found what's going
wrong. In patch
v6-0001-Use-shared-lock-in-GetMultiXactIdMembers-for-offs.patch I think
there is a missing reinitialisation of the lockmode variable to LW_NONE
inside the retry loop after the call to LWLockRelease() in
src/backend/access/transam/multixact.c:1392:GetMultiXactIdMembers().
I've attached a new version of the patch for master that include the fix
I'm using now with PG11 and with which everything works very well now.
I'm running more tests to see the impact on the performances to play
with multixact_offsets_slru_buffers, multixact_members_slru_buffers and
multixact_local_cache_entries. I will reports the results later today.
--
Gilles Darold
http://www.darold.net/
Attachments:
[text/x-patch] v7-0001-Use-shared-lock-in-GetMultiXactIdMembers-for-offs.patch (11.6K, ../../[email protected]/3-v7-0001-Use-shared-lock-in-GetMultiXactIdMembers-for-offs.patch)
download | inline diff:
diff --git a/src/backend/access/transam/clog.c b/src/backend/access/transam/clog.c
index 034349aa7b..4c372065de 100644
--- a/src/backend/access/transam/clog.c
+++ b/src/backend/access/transam/clog.c
@@ -640,10 +640,11 @@ TransactionIdGetStatus(TransactionId xid, XLogRecPtr *lsn)
int lsnindex;
char *byteptr;
XidStatus status;
+ LWLockMode lockmode = LW_NONE;
/* lock is acquired by SimpleLruReadPage_ReadOnly */
- slotno = SimpleLruReadPage_ReadOnly(XactCtl, pageno, xid);
+ slotno = SimpleLruReadPage_ReadOnly(XactCtl, pageno, xid, &lockmode);
byteptr = XactCtl->shared->page_buffer[slotno] + byteno;
status = (*byteptr >> bshift) & CLOG_XACT_BITMASK;
@@ -651,6 +652,7 @@ TransactionIdGetStatus(TransactionId xid, XLogRecPtr *lsn)
lsnindex = GetLSNIndex(slotno, xid);
*lsn = XactCtl->shared->group_lsn[lsnindex];
+ Assert(lockmode != LW_NONE);
LWLockRelease(XactSLRULock);
return status;
diff --git a/src/backend/access/transam/commit_ts.c b/src/backend/access/transam/commit_ts.c
index 2fe551f17e..2699de033d 100644
--- a/src/backend/access/transam/commit_ts.c
+++ b/src/backend/access/transam/commit_ts.c
@@ -288,6 +288,7 @@ TransactionIdGetCommitTsData(TransactionId xid, TimestampTz *ts,
CommitTimestampEntry entry;
TransactionId oldestCommitTsXid;
TransactionId newestCommitTsXid;
+ LWLockMode lockmode = LW_NONE;
if (!TransactionIdIsValid(xid))
ereport(ERROR,
@@ -342,7 +343,7 @@ TransactionIdGetCommitTsData(TransactionId xid, TimestampTz *ts,
}
/* lock is acquired by SimpleLruReadPage_ReadOnly */
- slotno = SimpleLruReadPage_ReadOnly(CommitTsCtl, pageno, xid);
+ slotno = SimpleLruReadPage_ReadOnly(CommitTsCtl, pageno, xid, &lockmode);
memcpy(&entry,
CommitTsCtl->shared->page_buffer[slotno] +
SizeOfCommitTimestampEntry * entryno,
@@ -352,6 +353,7 @@ TransactionIdGetCommitTsData(TransactionId xid, TimestampTz *ts,
if (nodeid)
*nodeid = entry.nodeid;
+ Assert(lockmode != LW_NONE);
LWLockRelease(CommitTsSLRULock);
return *ts != 0;
}
diff --git a/src/backend/access/transam/multixact.c b/src/backend/access/transam/multixact.c
index eb8de7cf32..56bdd04364 100644
--- a/src/backend/access/transam/multixact.c
+++ b/src/backend/access/transam/multixact.c
@@ -1237,6 +1237,7 @@ GetMultiXactIdMembers(MultiXactId multi, MultiXactMember **members,
MultiXactId tmpMXact;
MultiXactOffset nextOffset;
MultiXactMember *ptr;
+ LWLockMode lockmode = LW_NONE;
debug_elog3(DEBUG2, "GetMembers: asked for %u", multi);
@@ -1340,12 +1341,13 @@ GetMultiXactIdMembers(MultiXactId multi, MultiXactMember **members,
* time on every multixact creation.
*/
retry:
- LWLockAcquire(MultiXactOffsetSLRULock, LW_EXCLUSIVE);
pageno = MultiXactIdToOffsetPage(multi);
entryno = MultiXactIdToOffsetEntry(multi);
- slotno = SimpleLruReadPage(MultiXactOffsetCtl, pageno, true, multi);
+ /* lock is acquired by SimpleLruReadPage_ReadOnly */
+ slotno = SimpleLruReadPage_ReadOnly(MultiXactOffsetCtl, pageno,
+ multi, &lockmode);
offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
offptr += entryno;
offset = *offptr;
@@ -1377,7 +1379,8 @@ retry:
entryno = MultiXactIdToOffsetEntry(tmpMXact);
if (pageno != prev_pageno)
- slotno = SimpleLruReadPage(MultiXactOffsetCtl, pageno, true, tmpMXact);
+ slotno = SimpleLruReadPage_ReadOnly(MultiXactOffsetCtl, pageno,
+ tmpMXact, &lockmode);
offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
offptr += entryno;
@@ -1387,6 +1390,7 @@ retry:
{
/* Corner case 2: next multixact is still being filled in */
LWLockRelease(MultiXactOffsetSLRULock);
+ lockmode = LW_NONE;
CHECK_FOR_INTERRUPTS();
pg_usleep(1000L);
goto retry;
@@ -1395,14 +1399,14 @@ retry:
length = nextMXOffset - offset;
}
+ Assert(lockmode != LW_NONE);
LWLockRelease(MultiXactOffsetSLRULock);
ptr = (MultiXactMember *) palloc(length * sizeof(MultiXactMember));
*members = ptr;
/* Now get the members themselves. */
- LWLockAcquire(MultiXactMemberSLRULock, LW_EXCLUSIVE);
-
+ lockmode = LW_NONE;
truelength = 0;
prev_pageno = -1;
for (i = 0; i < length; i++, offset++)
@@ -1418,7 +1422,8 @@ retry:
if (pageno != prev_pageno)
{
- slotno = SimpleLruReadPage(MultiXactMemberCtl, pageno, true, multi);
+ slotno = SimpleLruReadPage_ReadOnly(MultiXactMemberCtl, pageno,
+ multi, &lockmode);
prev_pageno = pageno;
}
@@ -1441,6 +1446,7 @@ retry:
truelength++;
}
+ Assert(lockmode != LW_NONE);
LWLockRelease(MultiXactMemberSLRULock);
/*
@@ -2733,6 +2739,7 @@ find_multixact_start(MultiXactId multi, MultiXactOffset *result)
int entryno;
int slotno;
MultiXactOffset *offptr;
+ LWLockMode lockmode = LW_NONE;
Assert(MultiXactState->finishedStartup);
@@ -2749,10 +2756,12 @@ find_multixact_start(MultiXactId multi, MultiXactOffset *result)
return false;
/* lock is acquired by SimpleLruReadPage_ReadOnly */
- slotno = SimpleLruReadPage_ReadOnly(MultiXactOffsetCtl, pageno, multi);
+ slotno = SimpleLruReadPage_ReadOnly(MultiXactOffsetCtl, pageno,
+ multi, &lockmode);
offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
offptr += entryno;
offset = *offptr;
+ Assert(lockmode != LW_NONE);
LWLockRelease(MultiXactOffsetSLRULock);
*result = offset;
diff --git a/src/backend/access/transam/slru.c b/src/backend/access/transam/slru.c
index cec17cb2ae..e271b37d7b 100644
--- a/src/backend/access/transam/slru.c
+++ b/src/backend/access/transam/slru.c
@@ -487,17 +487,23 @@ SimpleLruReadPage(SlruCtl ctl, int pageno, bool write_ok,
* Return value is the shared-buffer slot number now holding the page.
* The buffer's LRU access info is updated.
*
- * Control lock must NOT be held at entry, but will be held at exit.
- * It is unspecified whether the lock will be shared or exclusive.
+ * Control lock must be held in *lock_mode mode, which may be LW_NONE. Control
+ * lock will be held at exit in at least shared mode. Resulting control lock
+ * mode is set to *lock_mode.
*/
int
-SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno, TransactionId xid)
+SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno, TransactionId xid,
+ LWLockMode *lock_mode)
{
SlruShared shared = ctl->shared;
int slotno;
/* Try to find the page while holding only shared lock */
- LWLockAcquire(shared->ControlLock, LW_SHARED);
+ if (*lock_mode == LW_NONE)
+ {
+ LWLockAcquire(shared->ControlLock, LW_SHARED);
+ *lock_mode = LW_SHARED;
+ }
/* See if page is already in a buffer */
for (slotno = 0; slotno < shared->num_slots; slotno++)
@@ -517,8 +523,13 @@ SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno, TransactionId xid)
}
/* No luck, so switch to normal exclusive lock and do regular read */
- LWLockRelease(shared->ControlLock);
- LWLockAcquire(shared->ControlLock, LW_EXCLUSIVE);
+ if (*lock_mode != LW_EXCLUSIVE)
+ {
+ Assert(*lock_mode == LW_NONE);
+ LWLockRelease(shared->ControlLock);
+ LWLockAcquire(shared->ControlLock, LW_EXCLUSIVE);
+ *lock_mode = LW_EXCLUSIVE;
+ }
return SimpleLruReadPage(ctl, pageno, true, xid);
}
diff --git a/src/backend/access/transam/subtrans.c b/src/backend/access/transam/subtrans.c
index 0111e867c7..7bb1543189 100644
--- a/src/backend/access/transam/subtrans.c
+++ b/src/backend/access/transam/subtrans.c
@@ -113,6 +113,7 @@ SubTransGetParent(TransactionId xid)
int slotno;
TransactionId *ptr;
TransactionId parent;
+ LWLockMode lockmode = LW_NONE;
/* Can't ask about stuff that might not be around anymore */
Assert(TransactionIdFollowsOrEquals(xid, TransactionXmin));
@@ -123,12 +124,13 @@ SubTransGetParent(TransactionId xid)
/* lock is acquired by SimpleLruReadPage_ReadOnly */
- slotno = SimpleLruReadPage_ReadOnly(SubTransCtl, pageno, xid);
+ slotno = SimpleLruReadPage_ReadOnly(SubTransCtl, pageno, xid, &lockmode);
ptr = (TransactionId *) SubTransCtl->shared->page_buffer[slotno];
ptr += entryno;
parent = *ptr;
+ Assert(lockmode != LW_NONE);
LWLockRelease(SubtransSLRULock);
return parent;
diff --git a/src/backend/commands/async.c b/src/backend/commands/async.c
index c0763c63e2..ef1dbbe53d 100644
--- a/src/backend/commands/async.c
+++ b/src/backend/commands/async.c
@@ -2011,6 +2011,7 @@ asyncQueueReadAllNotifications(void)
int curoffset = QUEUE_POS_OFFSET(pos);
int slotno;
int copysize;
+ LWLockMode lockmode = LW_NONE;
/*
* We copy the data from SLRU into a local buffer, so as to avoid
@@ -2019,7 +2020,7 @@ asyncQueueReadAllNotifications(void)
* part of the page we will actually inspect.
*/
slotno = SimpleLruReadPage_ReadOnly(NotifyCtl, curpage,
- InvalidTransactionId);
+ InvalidTransactionId, &lockmode);
if (curpage == QUEUE_POS_PAGE(head))
{
/* we only want to read as far as head */
@@ -2036,6 +2037,7 @@ asyncQueueReadAllNotifications(void)
NotifyCtl->shared->page_buffer[slotno] + curoffset,
copysize);
/* Release lock that we got from SimpleLruReadPage_ReadOnly() */
+ Assert(lockmode != LW_NONE);
LWLockRelease(NotifySLRULock);
/*
diff --git a/src/backend/storage/lmgr/lwlock.c b/src/backend/storage/lmgr/lwlock.c
index 108e652179..ad1dcc7140 100644
--- a/src/backend/storage/lmgr/lwlock.c
+++ b/src/backend/storage/lmgr/lwlock.c
@@ -1067,6 +1067,8 @@ LWLockWakeup(LWLock *lock)
static void
LWLockQueueSelf(LWLock *lock, LWLockMode mode)
{
+ Assert(mode != LW_NONE);
+
/*
* If we don't have a PGPROC structure, there's no way to wait. This
* should never occur, since MyProc should only be null during shared
@@ -1827,6 +1829,7 @@ LWLockRelease(LWLock *lock)
elog(ERROR, "lock %s is not held", T_NAME(lock));
mode = held_lwlocks[i].mode;
+ Assert(mode == LW_EXCLUSIVE || mode == LW_SHARED);
num_held_lwlocks--;
for (; i < num_held_lwlocks; i++)
diff --git a/src/backend/storage/lmgr/predicate.c b/src/backend/storage/lmgr/predicate.c
index 8a365b400c..a4df90a8ae 100644
--- a/src/backend/storage/lmgr/predicate.c
+++ b/src/backend/storage/lmgr/predicate.c
@@ -924,6 +924,7 @@ SerialGetMinConflictCommitSeqNo(TransactionId xid)
TransactionId tailXid;
SerCommitSeqNo val;
int slotno;
+ LWLockMode lockmode = LW_NONE;
Assert(TransactionIdIsValid(xid));
@@ -946,8 +947,9 @@ SerialGetMinConflictCommitSeqNo(TransactionId xid)
* but will return with that lock held, which must then be released.
*/
slotno = SimpleLruReadPage_ReadOnly(SerialSlruCtl,
- SerialPage(xid), xid);
+ SerialPage(xid), xid, &lockmode);
val = SerialValue(slotno, xid);
+ Assert(lockmode != LW_NONE);
LWLockRelease(SerialSLRULock);
return val;
}
diff --git a/src/include/access/slru.h b/src/include/access/slru.h
index b39b43504d..4b66d3b592 100644
--- a/src/include/access/slru.h
+++ b/src/include/access/slru.h
@@ -142,7 +142,7 @@ extern int SimpleLruZeroPage(SlruCtl ctl, int pageno);
extern int SimpleLruReadPage(SlruCtl ctl, int pageno, bool write_ok,
TransactionId xid);
extern int SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno,
- TransactionId xid);
+ TransactionId xid, LWLockMode *lock_mode);
extern void SimpleLruWritePage(SlruCtl ctl, int slotno);
extern void SimpleLruWriteAll(SlruCtl ctl, bool allow_redirtied);
extern void SimpleLruTruncate(SlruCtl ctl, int cutoffPage);
diff --git a/src/include/storage/lwlock.h b/src/include/storage/lwlock.h
index af9b41795d..e680c6397a 100644
--- a/src/include/storage/lwlock.h
+++ b/src/include/storage/lwlock.h
@@ -129,6 +129,8 @@ extern PGDLLIMPORT int NamedLWLockTrancheRequests;
typedef enum LWLockMode
{
+ LW_NONE, /* Not a lock mode. Indicates that there is no
+ * lock. */
LW_EXCLUSIVE,
LW_SHARED,
LW_WAIT_UNTIL_FREE /* A special mode used in PGPROC->lwWaitMode,
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2020-12-11 17:50 Gilles Darold <[email protected]>
parent: Gilles Darold <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Gilles Darold @ 2020-12-11 17:50 UTC (permalink / raw)
To: Andrey Borodin <[email protected]>; +Cc: Tomas Vondra <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
Le 10/12/2020 à 15:45, Gilles Darold a écrit :
> Le 08/12/2020 à 18:52, Andrey Borodin a écrit :
>> Hi Gilles!
>>
>> Many thanks for your message!
>>
>>> 8 дек. 2020 г., в 21:05, Gilles Darold<[email protected]> написал(а):
>>>
>>> I know that this report is not really helpful
>> Quite contrary - this benchmarks prove that controllable reproduction exists. I've rebased patches for PG11. Can you please benchmark them (without extending SLRU)?
>>
>> Best regards, Andrey Borodin.
>>
> Hi,
>
>
> Running tests yesterday with the patches has reported log of failures
> with error on INSERT and UPDATE statements:
>
>
> ERROR: lock MultiXactOffsetControlLock is not held
>
>
> After a patch review this morning I think I have found what's going
> wrong. In patch
> v6-0001-Use-shared-lock-in-GetMultiXactIdMembers-for-offs.patch I
> think there is a missing reinitialisation of the lockmode variable to
> LW_NONE inside the retry loop after the call to LWLockRelease() in
> src/backend/access/transam/multixact.c:1392:GetMultiXactIdMembers().
> I've attached a new version of the patch for master that include the
> fix I'm using now with PG11 and with which everything works very well now.
>
>
> I'm running more tests to see the impact on the performances to play
> with multixact_offsets_slru_buffers, multixact_members_slru_buffers
> and multixact_local_cache_entries. I will reports the results later today.
>
Hi,
Sorry for the delay, I have done some further tests to try to reach the
limit without bottlenecks on multixact or shared buffers. The tests was
done on a Microsoft Asure machine with 2TB of RAM and 4 sockets Intel
Xeon Platinum 8280M (128 cpu). PG configuration:
max_connections = 4096
shared_buffers = 64GB
max_prepared_transactions = 2048
work_mem = 256MB
maintenance_work_mem = 2GB
wal_level = minimal
synchronous_commit = off
commit_delay = 1000
commit_siblings = 10
checkpoint_timeout = 1h
max_wal_size = 32GB
checkpoint_completion_target = 0.9
I have tested with several values for the different buffer's variables
starting from:
multixact_offsets_slru_buffers = 64
multixact_members_slru_buffers = 128
multixact_local_cache_entries = 256
to the values with the best performances we achieve with this test to
avoid MultiXactOffsetControlLock or MultiXactMemberControlLock:
multixact_offsets_slru_buffers = 128
multixact_members_slru_buffers = 512
multixact_local_cache_entries = 1024
Also shared_buffers have been increased up to 256GB to avoid
buffer_mapping contention.
Our last best test reports the following wait events:
event_type | event | sum
------------+----------------------------+-----------
Client | ClientRead | 321690211
LWLock | buffer_content | 2970016
IPC | ProcArrayGroupUpdate | 2317388
LWLock | ProcArrayLock | 1445828
LWLock | WALWriteLock | 1187606
LWLock | SubtransControlLock | 972889
Lock | transactionid | 840560
Lock | relation | 587600
Activity | LogicalLauncherMain | 529599
Activity | AutoVacuumMain | 528097
At this stage I don't think we can have better performances by tuning
these buffers at least with PG11.
About performances gain related to the patch for shared lock in
GetMultiXactIdMembers unfortunately I can not see a difference with or
without this patch, it could be related to our particular benchmark. But
clearly the patch on multixact buffers should be committed as this is
really helpfull to be able to tuned PG when multixact bottlenecks are found.
Best regards,
--
Gilles Darold
LzLabs GmbH
https://www.lzlabs.com/
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2020-12-13 09:17 Gilles Darold <[email protected]>
parent: Gilles Darold <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Gilles Darold @ 2020-12-13 09:17 UTC (permalink / raw)
To: Andrey Borodin <[email protected]>; +Cc: Tomas Vondra <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
Le 11/12/2020 à 18:50, Gilles Darold a écrit :
> Le 10/12/2020 à 15:45, Gilles Darold a écrit :
>> Le 08/12/2020 à 18:52, Andrey Borodin a écrit :
>>> Hi Gilles!
>>>
>>> Many thanks for your message!
>>>
>>>> 8 дек. 2020 г., в 21:05, Gilles Darold <[email protected]> написал(а):
>>>>
>>>> I know that this report is not really helpful
>>> Quite contrary - this benchmarks prove that controllable reproduction exists. I've rebased patches for PG11. Can you please benchmark them (without extending SLRU)?
>>>
>>> Best regards, Andrey Borodin.
>>>
>> Hi,
>>
>>
>> Running tests yesterday with the patches has reported log of failures
>> with error on INSERT and UPDATE statements:
>>
>>
>> ERROR: lock MultiXactOffsetControlLock is not held
>>
>>
>> After a patch review this morning I think I have found what's going
>> wrong. In patch
>> v6-0001-Use-shared-lock-in-GetMultiXactIdMembers-for-offs.patch I
>> think there is a missing reinitialisation of the lockmode variable to
>> LW_NONE inside the retry loop after the call to LWLockRelease() in
>> src/backend/access/transam/multixact.c:1392:GetMultiXactIdMembers().
>> I've attached a new version of the patch for master that include the
>> fix I'm using now with PG11 and with which everything works very well
>> now.
>>
>>
>> I'm running more tests to see the impact on the performances to play
>> with multixact_offsets_slru_buffers, multixact_members_slru_buffers
>> and multixact_local_cache_entries. I will reports the results later
>> today.
>>
>
> Hi,
>
> Sorry for the delay, I have done some further tests to try to reach
> the limit without bottlenecks on multixact or shared buffers. The
> tests was done on a Microsoft Asure machine with 2TB of RAM and 4
> sockets Intel Xeon Platinum 8280M (128 cpu). PG configuration:
>
> max_connections = 4096
> shared_buffers = 64GB
> max_prepared_transactions = 2048
> work_mem = 256MB
> maintenance_work_mem = 2GB
> wal_level = minimal
> synchronous_commit = off
> commit_delay = 1000
> commit_siblings = 10
> checkpoint_timeout = 1h
> max_wal_size = 32GB
> checkpoint_completion_target = 0.9
>
> I have tested with several values for the different buffer's variables
> starting from:
>
> multixact_offsets_slru_buffers = 64
> multixact_members_slru_buffers = 128
> multixact_local_cache_entries = 256
>
> to the values with the best performances we achieve with this test to
> avoid MultiXactOffsetControlLock or MultiXactMemberControlLock:
>
> multixact_offsets_slru_buffers = 128
> multixact_members_slru_buffers = 512
> multixact_local_cache_entries = 1024
>
> Also shared_buffers have been increased up to 256GB to avoid
> buffer_mapping contention.
>
> Our last best test reports the following wait events:
>
> event_type | event | sum
> ------------+----------------------------+-----------
> Client | ClientRead | 321690211
> LWLock | buffer_content | 2970016
> IPC | ProcArrayGroupUpdate | 2317388
> LWLock | ProcArrayLock | 1445828
> LWLock | WALWriteLock | 1187606
> LWLock | SubtransControlLock | 972889
> Lock | transactionid | 840560
> Lock | relation | 587600
> Activity | LogicalLauncherMain | 529599
> Activity | AutoVacuumMain | 528097
>
> At this stage I don't think we can have better performances by tuning
> these buffers at least with PG11.
>
> About performances gain related to the patch for shared lock in
> GetMultiXactIdMembers unfortunately I can not see a difference with or
> without this patch, it could be related to our particular benchmark.
> But clearly the patch on multixact buffers should be committed as this
> is really helpfull to be able to tuned PG when multixact bottlenecks
> are found.
I've done more review on these patches.
1) as reported in my previous message patch 0001 looks useless as it
doesn't allow measurable performances gain.
2) In patch 0004 there is two typo: s/informaion/information/ will fix them
3) the GUC are missing in the postgresql.conf.sample file, see patch in
attachment for a proposal.
Best regards,
--
Gilles Darold
LzLabs GmbH
https://www.lzlabs.com/
Attachments:
[text/x-patch] postgresql_conf_multixact_buffers_GUCs.patch (819B, ../../[email protected]/3-postgresql_conf_multixact_buffers_GUCs.patch)
download | inline diff:
diff --git a/src/backend/utils/misc/postgresql.conf.sample b/src/backend/utils/misc/postgresql.conf.sample
index b7fb2ec1fe..1fb33809ee 100644
--- a/src/backend/utils/misc/postgresql.conf.sample
+++ b/src/backend/utils/misc/postgresql.conf.sample
@@ -189,6 +189,14 @@
# (change requires restart)
#backend_flush_after = 0 # measured in pages, 0 disables
+# - MultiXact Buffers -
+
+#multixact_offsets_slru_buffers = 8 # memory used for MultiXact offsets
+ # (change requires restart)
+#multixact_members_slru_buffers = 16 # memory used for MultiXact members
+ # (change requires restart)
+#multixact_local_cache_entries = 256 # number of cached MultiXact by backend
+ # (change requires restart)
#------------------------------------------------------------------------------
# WRITE-AHEAD LOG
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2020-12-13 17:24 Andrey Borodin <[email protected]>
parent: Gilles Darold <[email protected]>
0 siblings, 2 replies; 103+ messages in thread
From: Andrey Borodin @ 2020-12-13 17:24 UTC (permalink / raw)
To: Gilles Darold <[email protected]>; +Cc: Tomas Vondra <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
> 13 дек. 2020 г., в 14:17, Gilles Darold <[email protected]> написал(а):
>
> I've done more review on these patches.
Thanks, Gilles! I'll incorporate all your fixes to patchset.
Can you also benchmark conditional variable sleep? The patch "Add conditional variable to wait for next MultXact offset in corner case"?
The problem manifests on Standby when Primary is heavily loaded with MultiXactOffsetControlLock.
Thanks!
Best regards, Andrey Borodin.
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2020-12-14 06:31 Andrey Borodin <[email protected]>
parent: Andrey Borodin <[email protected]>
1 sibling, 0 replies; 103+ messages in thread
From: Andrey Borodin @ 2020-12-14 06:31 UTC (permalink / raw)
To: Gilles Darold <[email protected]>; +Cc: Tomas Vondra <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
> 13 дек. 2020 г., в 22:24, Andrey Borodin <[email protected]> написал(а):
>
>
>
>> 13 дек. 2020 г., в 14:17, Gilles Darold <[email protected]> написал(а):
>>
>> I've done more review on these patches.
>
> Thanks, Gilles! I'll incorporate all your fixes to patchset.
PFA patches.
Also, I've noted that patch "Add conditional variable to wait for next MultXact offset in corner case" removes CHECK_FOR_INTERRUPTS();, I'm not sure it's correct.
Thanks!
Best regards, Andrey Borodin.
Attachments:
[application/octet-stream] v8-0004-Add-GUCs-to-tune-MultiXact-SLRUs.patch (7.2K, ../../[email protected]/2-v8-0004-Add-GUCs-to-tune-MultiXact-SLRUs.patch)
download | inline diff:
From 122fcbf0a0e38d6ba7880b22ec093aa8bb8011e4 Mon Sep 17 00:00:00 2001
From: Andrey Borodin <[email protected]>
Date: Sat, 9 May 2020 16:42:07 +0500
Subject: [PATCH v8 4/4] Add GUCs to tune MultiXact SLRUs
---
doc/src/sgml/config.sgml | 31 +++++++++++++++++++
src/backend/access/transam/multixact.c | 8 ++---
src/backend/utils/init/globals.c | 2 ++
src/backend/utils/misc/guc.c | 22 +++++++++++++
src/backend/utils/misc/postgresql.conf.sample | 8 +++++
src/include/access/multixact.h | 4 ---
src/include/miscadmin.h | 2 ++
7 files changed, 69 insertions(+), 8 deletions(-)
diff --git a/doc/src/sgml/config.sgml b/doc/src/sgml/config.sgml
index d9c27daa49..fab9715e83 100644
--- a/doc/src/sgml/config.sgml
+++ b/doc/src/sgml/config.sgml
@@ -1860,6 +1860,37 @@ include_dir 'conf.d'
</para>
</listitem>
</varlistentry>
+
+ <varlistentry id="guc-multixact-offsets-slru-buffers" xreflabel="multixact_offsets_slru_buffers">
+ <term><varname>multixact_offsets_slru_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>multixact_offsets_slru_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of memory to be used for MultiXact offsets. MultiXact offsets
+ are used to store information about offsets of multiple row lockers (caused by SELECT FOR UPDATE and others).
+ It defaults to 64 kilobytes (<literal>64KB</literal>).
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-multixact-members-slru-buffers" xreflabel="multixact_members_slru_buffers">
+ <term><varname>multixact_members_slru_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>multixact_members_slru_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of memory to be used for MultiXact members. MultiXact members
+ are used to store information about XIDs of multiple row lockers. Tipically <varname>multixact_members_slru_buffers</varname>
+ is twice more than <varname>multixact_offsets_slru_buffers</varname>.
+ It defaults to 128 kilobytes (<literal>128KB</literal>).
+ </para>
+ </listitem>
+ </varlistentry>
<varlistentry id="guc-max-stack-depth" xreflabel="max_stack_depth">
<term><varname>max_stack_depth</varname> (<type>integer</type>)
diff --git a/src/backend/access/transam/multixact.c b/src/backend/access/transam/multixact.c
index ffd35c2aea..76b4ad75bc 100644
--- a/src/backend/access/transam/multixact.c
+++ b/src/backend/access/transam/multixact.c
@@ -1867,8 +1867,8 @@ MultiXactShmemSize(void)
mul_size(sizeof(MultiXactId) * 2, MaxOldestSlot))
size = SHARED_MULTIXACT_STATE_SIZE;
- size = add_size(size, SimpleLruShmemSize(NUM_MULTIXACTOFFSET_BUFFERS, 0));
- size = add_size(size, SimpleLruShmemSize(NUM_MULTIXACTMEMBER_BUFFERS, 0));
+ size = add_size(size, SimpleLruShmemSize(multixact_offsets_slru_buffers, 0));
+ size = add_size(size, SimpleLruShmemSize(multixact_members_slru_buffers, 0));
return size;
}
@@ -1884,12 +1884,12 @@ MultiXactShmemInit(void)
MultiXactMemberCtl->PagePrecedes = MultiXactMemberPagePrecedes;
SimpleLruInit(MultiXactOffsetCtl,
- "MultiXactOffset", NUM_MULTIXACTOFFSET_BUFFERS, 0,
+ "MultiXactOffset", multixact_offsets_slru_buffers, 0,
MultiXactOffsetSLRULock, "pg_multixact/offsets",
LWTRANCHE_MULTIXACTOFFSET_BUFFER,
SYNC_HANDLER_MULTIXACT_OFFSET);
SimpleLruInit(MultiXactMemberCtl,
- "MultiXactMember", NUM_MULTIXACTMEMBER_BUFFERS, 0,
+ "MultiXactMember", multixact_members_slru_buffers, 0,
MultiXactMemberSLRULock, "pg_multixact/members",
LWTRANCHE_MULTIXACTMEMBER_BUFFER,
SYNC_HANDLER_MULTIXACT_MEMBER);
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
index 9ca71933dc..a5ec7bfe88 100644
--- a/src/backend/utils/init/globals.c
+++ b/src/backend/utils/init/globals.c
@@ -151,3 +151,5 @@ bool VacuumCostActive = false;
double vacuum_cleanup_index_scale_factor;
int multixact_local_cache_entries = 256;
+int multixact_offsets_slru_buffers = 8;
+int multixact_members_slru_buffers = 16;
diff --git a/src/backend/utils/misc/guc.c b/src/backend/utils/misc/guc.c
index ae2a3e1304..d86d34b4a5 100644
--- a/src/backend/utils/misc/guc.c
+++ b/src/backend/utils/misc/guc.c
@@ -2257,6 +2257,28 @@ static struct config_int ConfigureNamesInt[] =
NULL, NULL, NULL
},
+ {
+ {"multixact_offsets_slru_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for MultiXact offsets SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &multixact_offsets_slru_buffers,
+ 8, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"multixact_members_slru_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for MultiXact members SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &multixact_members_slru_buffers,
+ 16, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
{
{"temp_buffers", PGC_USERSET, RESOURCES_MEM,
gettext_noop("Sets the maximum number of temporary buffers used by each session."),
diff --git a/src/backend/utils/misc/postgresql.conf.sample b/src/backend/utils/misc/postgresql.conf.sample
index b7fb2ec1fe..1fb33809ee 100644
--- a/src/backend/utils/misc/postgresql.conf.sample
+++ b/src/backend/utils/misc/postgresql.conf.sample
@@ -189,6 +189,14 @@
# (change requires restart)
#backend_flush_after = 0 # measured in pages, 0 disables
+# - MultiXact Buffers -
+
+#multixact_offsets_slru_buffers = 8 # memory used for MultiXact offsets
+ # (change requires restart)
+#multixact_members_slru_buffers = 16 # memory used for MultiXact members
+ # (change requires restart)
+#multixact_local_cache_entries = 256 # number of cached MultiXact by backend
+ # (change requires restart)
#------------------------------------------------------------------------------
# WRITE-AHEAD LOG
diff --git a/src/include/access/multixact.h b/src/include/access/multixact.h
index 9a30380901..630ceaea4d 100644
--- a/src/include/access/multixact.h
+++ b/src/include/access/multixact.h
@@ -29,10 +29,6 @@
#define MaxMultiXactOffset ((MultiXactOffset) 0xFFFFFFFF)
-/* Number of SLRU buffers to use for multixact */
-#define NUM_MULTIXACTOFFSET_BUFFERS 8
-#define NUM_MULTIXACTMEMBER_BUFFERS 16
-
/*
* Possible multixact lock modes ("status"). The first four modes are for
* tuple locks (FOR KEY SHARE, FOR SHARE, FOR NO KEY UPDATE, FOR UPDATE); the
diff --git a/src/include/miscadmin.h b/src/include/miscadmin.h
index 01af61c963..ef8abea84d 100644
--- a/src/include/miscadmin.h
+++ b/src/include/miscadmin.h
@@ -161,6 +161,8 @@ extern PGDLLIMPORT int MaxBackends;
extern PGDLLIMPORT int MaxConnections;
extern PGDLLIMPORT int max_worker_processes;
extern PGDLLIMPORT int max_parallel_workers;
+extern PGDLLIMPORT int multixact_offsets_slru_buffers;
+extern PGDLLIMPORT int multixact_members_slru_buffers;
extern PGDLLIMPORT int multixact_local_cache_entries;
--
2.24.3 (Apple Git-128)
[application/octet-stream] v8-0003-Add-conditional-variable-to-wait-for-next-MultXac.patch (4.4K, ../../[email protected]/3-v8-0003-Add-conditional-variable-to-wait-for-next-MultXac.patch)
download | inline diff:
From 2706550cd4fc9ef53b8fa09d7414e88a31ea6e0e Mon Sep 17 00:00:00 2001
From: Alexander Korotkov <[email protected]>
Date: Mon, 26 Oct 2020 01:55:31 +0300
Subject: [PATCH v8 3/4] Add conditional variable to wait for next MultXact
offset in corner case
GetMultiXactIdMembers() has a corner case, when the next multixact offset is
not yet set. In this case GetMultiXactIdMembers() has to sleep till this offset
is set. Currently the sleeping is implemented in naive way using pg_sleep()
and retry. This commit implements sleeping with conditional variable, which
provides more efficient way for waiting till the event.
Discussion: https://postgr.es/m/a7f1c4e1-1015-92a4-2bd4-6736bd13d03e%40postgrespro.ru#c496c4e75fc0605094a0e1f763e6a6ec
Author: Andrey Borodin
Reviewed-by: Kyotaro Horiguchi, Daniel Gustafsson
Reviewed-by: Anastasia Lubennikova, Alexander Korotkov
---
src/backend/access/transam/multixact.c | 35 ++++++++++++++++++++++++--
src/backend/postmaster/pgstat.c | 2 ++
src/include/pgstat.h | 1 +
3 files changed, 36 insertions(+), 2 deletions(-)
diff --git a/src/backend/access/transam/multixact.c b/src/backend/access/transam/multixact.c
index 171e2ab681..ffd35c2aea 100644
--- a/src/backend/access/transam/multixact.c
+++ b/src/backend/access/transam/multixact.c
@@ -82,6 +82,7 @@
#include "lib/ilist.h"
#include "miscadmin.h"
#include "pg_trace.h"
+#include "pgstat.h"
#include "postmaster/autovacuum.h"
#include "storage/lmgr.h"
#include "storage/pmsignal.h"
@@ -233,6 +234,13 @@ typedef struct MultiXactStateData
/* support for members anti-wraparound measures */
MultiXactOffset offsetStopLimit; /* known if oldestOffsetKnown */
+ /*
+ * Conditional variable for waiting till the filling of the next multixact
+ * will be finished. See GetMultiXactIdMembers() and RecordNewMultiXact()
+ * for details.
+ */
+ ConditionVariable nextoffCV;
+
/*
* Per-backend data starts here. We have two arrays stored in the area
* immediately following the MultiXactStateData struct. Each is indexed by
@@ -892,6 +900,14 @@ RecordNewMultiXact(MultiXactId multi, MultiXactOffset offset,
/* Exchange our lock */
LWLockRelease(MultiXactOffsetSLRULock);
+ /*
+ * Let everybody know the offset of this mxid is recorded now. The
+ * waiters are waiting for the offset of the mxid next of the target to
+ * know the number of members of the target mxid, so we don't need to wait
+ * for members of this mxid are recorded.
+ */
+ ConditionVariableBroadcast(&MultiXactState->nextoffCV);
+
LWLockAcquire(MultiXactMemberSLRULock, LW_EXCLUSIVE);
prev_pageno = -1;
@@ -1389,10 +1405,24 @@ retry:
if (nextMXOffset == 0)
{
/* Corner case 2: next multixact is still being filled in */
+
+ /*
+ * The recorder of the next mxid is just before writing the
+ * offset. Wait for the offset to be written.
+ */
+ ConditionVariablePrepareToSleep(&MultiXactState->nextoffCV);
+
+ /*
+ * We don't have to recheck if multixact was filled in during
+ * ConditionVariablePrepareToSleep(), because we were holding
+ * MultiXactOffsetSLRULock.
+ */
LWLockRelease(MultiXactOffsetSLRULock);
lockmode = LW_NONE;
- CHECK_FOR_INTERRUPTS();
- pg_usleep(1000L);
+
+ ConditionVariableSleep(&MultiXactState->nextoffCV,
+ WAIT_EVENT_WAIT_NEXT_MXMEMBERS);
+ ConditionVariableCancelSleep();
goto retry;
}
@@ -1874,6 +1904,7 @@ MultiXactShmemInit(void)
/* Make sure we zero out the per-backend state */
MemSet(MultiXactState, 0, SHARED_MULTIXACT_STATE_SIZE);
+ ConditionVariableInit(&MultiXactState->nextoffCV);
}
else
Assert(found);
diff --git a/src/backend/postmaster/pgstat.c b/src/backend/postmaster/pgstat.c
index 7c75a25d21..20879fc90e 100644
--- a/src/backend/postmaster/pgstat.c
+++ b/src/backend/postmaster/pgstat.c
@@ -4041,6 +4041,8 @@ pgstat_get_wait_ipc(WaitEventIPC w)
break;
case WAIT_EVENT_XACT_GROUP_UPDATE:
event_name = "XactGroupUpdate";
+ case WAIT_EVENT_WAIT_NEXT_MXMEMBERS:
+ event_name = "MultiXactWaitNextMembers";
break;
/* no default case, so that compiler will warn */
}
diff --git a/src/include/pgstat.h b/src/include/pgstat.h
index 5954068dec..630ddcf1a5 100644
--- a/src/include/pgstat.h
+++ b/src/include/pgstat.h
@@ -964,6 +964,7 @@ typedef enum
WAIT_EVENT_REPLICATION_SLOT_DROP,
WAIT_EVENT_SAFE_SNAPSHOT,
WAIT_EVENT_SYNC_REP,
+ WAIT_EVENT_WAIT_NEXT_MXMEMBERS,
WAIT_EVENT_XACT_GROUP_UPDATE
} WaitEventIPC;
--
2.24.3 (Apple Git-128)
[application/octet-stream] v8-0002-Make-MultiXact-local-cache-size-configurable.patch (3.7K, ../../[email protected]/4-v8-0002-Make-MultiXact-local-cache-size-configurable.patch)
download | inline diff:
From 25f78196b00d5489924cb09e802ddda48c410d8d Mon Sep 17 00:00:00 2001
From: Alexander Korotkov <[email protected]>
Date: Mon, 26 Oct 2020 03:44:50 +0300
Subject: [PATCH v8 2/4] Make MultiXact local cache size configurable
---
doc/src/sgml/config.sgml | 16 ++++++++++++++++
src/backend/access/transam/multixact.c | 2 +-
src/backend/utils/init/globals.c | 2 ++
src/backend/utils/misc/guc.c | 10 ++++++++++
src/include/miscadmin.h | 2 ++
5 files changed, 31 insertions(+), 1 deletion(-)
diff --git a/doc/src/sgml/config.sgml b/doc/src/sgml/config.sgml
index 4b60382778..d9c27daa49 100644
--- a/doc/src/sgml/config.sgml
+++ b/doc/src/sgml/config.sgml
@@ -1845,6 +1845,22 @@ include_dir 'conf.d'
</listitem>
</varlistentry>
+ <varlistentry id="guc-multixact-local-cache-entries" xreflabel="multixact_local-cache-entries">
+ <term><varname>multixact_local_cache_entries</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>multixact_local_cache_entries</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the number cached MultiXact by backend. Any SLRU lookup is preceeded
+ by cache lookup. Higher numbers of cache size help to deduplicate lock sets, while
+ lower values make cache lookup faster.
+ It defaults to 256.
+ </para>
+ </listitem>
+ </varlistentry>
+
<varlistentry id="guc-max-stack-depth" xreflabel="max_stack_depth">
<term><varname>max_stack_depth</varname> (<type>integer</type>)
<indexterm>
diff --git a/src/backend/access/transam/multixact.c b/src/backend/access/transam/multixact.c
index 56bdd04364..171e2ab681 100644
--- a/src/backend/access/transam/multixact.c
+++ b/src/backend/access/transam/multixact.c
@@ -1614,7 +1614,7 @@ mXactCachePut(MultiXactId multi, int nmembers, MultiXactMember *members)
qsort(entry->members, nmembers, sizeof(MultiXactMember), mxactMemberComparator);
dlist_push_head(&MXactCache, &entry->node);
- if (MXactCacheMembers++ >= MAX_CACHE_ENTRIES)
+ if (MXactCacheMembers++ >= multixact_local_cache_entries)
{
dlist_node *node;
mXactCacheEnt *entry;
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
index 6ab8216839..9ca71933dc 100644
--- a/src/backend/utils/init/globals.c
+++ b/src/backend/utils/init/globals.c
@@ -149,3 +149,5 @@ int VacuumCostBalance = 0; /* working state for vacuum */
bool VacuumCostActive = false;
double vacuum_cleanup_index_scale_factor;
+
+int multixact_local_cache_entries = 256;
diff --git a/src/backend/utils/misc/guc.c b/src/backend/utils/misc/guc.c
index dabcbb0736..ae2a3e1304 100644
--- a/src/backend/utils/misc/guc.c
+++ b/src/backend/utils/misc/guc.c
@@ -2247,6 +2247,16 @@ static struct config_int ConfigureNamesInt[] =
NULL, NULL, NULL
},
+ {
+ {"multixact_local_cache_entries", PGC_SUSET, RESOURCES_MEM,
+ gettext_noop("Sets the number of cached MultiXact by backend."),
+ NULL
+ },
+ &multixact_local_cache_entries,
+ 256, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
{
{"temp_buffers", PGC_USERSET, RESOURCES_MEM,
gettext_noop("Sets the maximum number of temporary buffers used by each session."),
diff --git a/src/include/miscadmin.h b/src/include/miscadmin.h
index 72e3352398..01af61c963 100644
--- a/src/include/miscadmin.h
+++ b/src/include/miscadmin.h
@@ -162,6 +162,8 @@ extern PGDLLIMPORT int MaxConnections;
extern PGDLLIMPORT int max_worker_processes;
extern PGDLLIMPORT int max_parallel_workers;
+extern PGDLLIMPORT int multixact_local_cache_entries;
+
extern PGDLLIMPORT int MyProcPid;
extern PGDLLIMPORT pg_time_t MyStartTime;
extern PGDLLIMPORT TimestampTz MyStartTimestamp;
--
2.24.3 (Apple Git-128)
[application/octet-stream] v8-0001-Use-shared-lock-in-GetMultiXactIdMembers-for.patch (13.2K, ../../[email protected]/5-v8-0001-Use-shared-lock-in-GetMultiXactIdMembers-for.patch)
download | inline diff:
From 4c8c7d0f24fdce994412b6c104d72659cf6fc14f Mon Sep 17 00:00:00 2001
From: Andrey Borodin <[email protected]>
Date: Mon, 14 Dec 2020 11:24:46 +0500
Subject: [PATCH v8 1/4] Use shared lock in GetMultiXactIdMembers for offsets
and members
Previously the read of multixact required exclusive control locks for both
offsets and members SLRUs. This could lead to the excessive lock contention.
This commit we makes multixacts SLRU take advantage of SimpleLruReadPage_ReadOnly
similar to clog, commit_ts and subtrans.
In order to evade extra reacquiring of CLRU lock, we teach
SimpleLruReadPage_ReadOnly() to take into account the current lock mode and
report resulting lock mode back.
Discussion: https://postgr.es/m/a7f1c4e1-1015-92a4-2bd4-6736bd13d03e%40postgrespro.ru#c496c4e75fc0605094a0e1f763e6a6ec
Author: Andrey Borodin
Reviewed-by: Kyotaro Horiguchi, Daniel Gustafsson
Reviewed-by: Anastasia Lubennikova, Alexander Korotkov
---
src/backend/access/transam/clog.c | 4 +++-
src/backend/access/transam/commit_ts.c | 4 +++-
src/backend/access/transam/multixact.c | 23 ++++++++++++++++-------
src/backend/access/transam/slru.c | 23 +++++++++++++++++------
src/backend/access/transam/subtrans.c | 4 +++-
src/backend/commands/async.c | 4 +++-
src/backend/storage/lmgr/lwlock.c | 3 +++
src/backend/storage/lmgr/predicate.c | 4 +++-
src/include/access/slru.h | 2 +-
src/include/storage/lwlock.h | 2 ++
10 files changed, 54 insertions(+), 19 deletions(-)
diff --git a/src/backend/access/transam/clog.c b/src/backend/access/transam/clog.c
index 034349aa7b..4c372065de 100644
--- a/src/backend/access/transam/clog.c
+++ b/src/backend/access/transam/clog.c
@@ -640,10 +640,11 @@ TransactionIdGetStatus(TransactionId xid, XLogRecPtr *lsn)
int lsnindex;
char *byteptr;
XidStatus status;
+ LWLockMode lockmode = LW_NONE;
/* lock is acquired by SimpleLruReadPage_ReadOnly */
- slotno = SimpleLruReadPage_ReadOnly(XactCtl, pageno, xid);
+ slotno = SimpleLruReadPage_ReadOnly(XactCtl, pageno, xid, &lockmode);
byteptr = XactCtl->shared->page_buffer[slotno] + byteno;
status = (*byteptr >> bshift) & CLOG_XACT_BITMASK;
@@ -651,6 +652,7 @@ TransactionIdGetStatus(TransactionId xid, XLogRecPtr *lsn)
lsnindex = GetLSNIndex(slotno, xid);
*lsn = XactCtl->shared->group_lsn[lsnindex];
+ Assert(lockmode != LW_NONE);
LWLockRelease(XactSLRULock);
return status;
diff --git a/src/backend/access/transam/commit_ts.c b/src/backend/access/transam/commit_ts.c
index 2fe551f17e..2699de033d 100644
--- a/src/backend/access/transam/commit_ts.c
+++ b/src/backend/access/transam/commit_ts.c
@@ -288,6 +288,7 @@ TransactionIdGetCommitTsData(TransactionId xid, TimestampTz *ts,
CommitTimestampEntry entry;
TransactionId oldestCommitTsXid;
TransactionId newestCommitTsXid;
+ LWLockMode lockmode = LW_NONE;
if (!TransactionIdIsValid(xid))
ereport(ERROR,
@@ -342,7 +343,7 @@ TransactionIdGetCommitTsData(TransactionId xid, TimestampTz *ts,
}
/* lock is acquired by SimpleLruReadPage_ReadOnly */
- slotno = SimpleLruReadPage_ReadOnly(CommitTsCtl, pageno, xid);
+ slotno = SimpleLruReadPage_ReadOnly(CommitTsCtl, pageno, xid, &lockmode);
memcpy(&entry,
CommitTsCtl->shared->page_buffer[slotno] +
SizeOfCommitTimestampEntry * entryno,
@@ -352,6 +353,7 @@ TransactionIdGetCommitTsData(TransactionId xid, TimestampTz *ts,
if (nodeid)
*nodeid = entry.nodeid;
+ Assert(lockmode != LW_NONE);
LWLockRelease(CommitTsSLRULock);
return *ts != 0;
}
diff --git a/src/backend/access/transam/multixact.c b/src/backend/access/transam/multixact.c
index eb8de7cf32..56bdd04364 100644
--- a/src/backend/access/transam/multixact.c
+++ b/src/backend/access/transam/multixact.c
@@ -1237,6 +1237,7 @@ GetMultiXactIdMembers(MultiXactId multi, MultiXactMember **members,
MultiXactId tmpMXact;
MultiXactOffset nextOffset;
MultiXactMember *ptr;
+ LWLockMode lockmode = LW_NONE;
debug_elog3(DEBUG2, "GetMembers: asked for %u", multi);
@@ -1340,12 +1341,13 @@ GetMultiXactIdMembers(MultiXactId multi, MultiXactMember **members,
* time on every multixact creation.
*/
retry:
- LWLockAcquire(MultiXactOffsetSLRULock, LW_EXCLUSIVE);
pageno = MultiXactIdToOffsetPage(multi);
entryno = MultiXactIdToOffsetEntry(multi);
- slotno = SimpleLruReadPage(MultiXactOffsetCtl, pageno, true, multi);
+ /* lock is acquired by SimpleLruReadPage_ReadOnly */
+ slotno = SimpleLruReadPage_ReadOnly(MultiXactOffsetCtl, pageno,
+ multi, &lockmode);
offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
offptr += entryno;
offset = *offptr;
@@ -1377,7 +1379,8 @@ retry:
entryno = MultiXactIdToOffsetEntry(tmpMXact);
if (pageno != prev_pageno)
- slotno = SimpleLruReadPage(MultiXactOffsetCtl, pageno, true, tmpMXact);
+ slotno = SimpleLruReadPage_ReadOnly(MultiXactOffsetCtl, pageno,
+ tmpMXact, &lockmode);
offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
offptr += entryno;
@@ -1387,6 +1390,7 @@ retry:
{
/* Corner case 2: next multixact is still being filled in */
LWLockRelease(MultiXactOffsetSLRULock);
+ lockmode = LW_NONE;
CHECK_FOR_INTERRUPTS();
pg_usleep(1000L);
goto retry;
@@ -1395,14 +1399,14 @@ retry:
length = nextMXOffset - offset;
}
+ Assert(lockmode != LW_NONE);
LWLockRelease(MultiXactOffsetSLRULock);
ptr = (MultiXactMember *) palloc(length * sizeof(MultiXactMember));
*members = ptr;
/* Now get the members themselves. */
- LWLockAcquire(MultiXactMemberSLRULock, LW_EXCLUSIVE);
-
+ lockmode = LW_NONE;
truelength = 0;
prev_pageno = -1;
for (i = 0; i < length; i++, offset++)
@@ -1418,7 +1422,8 @@ retry:
if (pageno != prev_pageno)
{
- slotno = SimpleLruReadPage(MultiXactMemberCtl, pageno, true, multi);
+ slotno = SimpleLruReadPage_ReadOnly(MultiXactMemberCtl, pageno,
+ multi, &lockmode);
prev_pageno = pageno;
}
@@ -1441,6 +1446,7 @@ retry:
truelength++;
}
+ Assert(lockmode != LW_NONE);
LWLockRelease(MultiXactMemberSLRULock);
/*
@@ -2733,6 +2739,7 @@ find_multixact_start(MultiXactId multi, MultiXactOffset *result)
int entryno;
int slotno;
MultiXactOffset *offptr;
+ LWLockMode lockmode = LW_NONE;
Assert(MultiXactState->finishedStartup);
@@ -2749,10 +2756,12 @@ find_multixact_start(MultiXactId multi, MultiXactOffset *result)
return false;
/* lock is acquired by SimpleLruReadPage_ReadOnly */
- slotno = SimpleLruReadPage_ReadOnly(MultiXactOffsetCtl, pageno, multi);
+ slotno = SimpleLruReadPage_ReadOnly(MultiXactOffsetCtl, pageno,
+ multi, &lockmode);
offptr = (MultiXactOffset *) MultiXactOffsetCtl->shared->page_buffer[slotno];
offptr += entryno;
offset = *offptr;
+ Assert(lockmode != LW_NONE);
LWLockRelease(MultiXactOffsetSLRULock);
*result = offset;
diff --git a/src/backend/access/transam/slru.c b/src/backend/access/transam/slru.c
index cec17cb2ae..e271b37d7b 100644
--- a/src/backend/access/transam/slru.c
+++ b/src/backend/access/transam/slru.c
@@ -487,17 +487,23 @@ SimpleLruReadPage(SlruCtl ctl, int pageno, bool write_ok,
* Return value is the shared-buffer slot number now holding the page.
* The buffer's LRU access info is updated.
*
- * Control lock must NOT be held at entry, but will be held at exit.
- * It is unspecified whether the lock will be shared or exclusive.
+ * Control lock must be held in *lock_mode mode, which may be LW_NONE. Control
+ * lock will be held at exit in at least shared mode. Resulting control lock
+ * mode is set to *lock_mode.
*/
int
-SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno, TransactionId xid)
+SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno, TransactionId xid,
+ LWLockMode *lock_mode)
{
SlruShared shared = ctl->shared;
int slotno;
/* Try to find the page while holding only shared lock */
- LWLockAcquire(shared->ControlLock, LW_SHARED);
+ if (*lock_mode == LW_NONE)
+ {
+ LWLockAcquire(shared->ControlLock, LW_SHARED);
+ *lock_mode = LW_SHARED;
+ }
/* See if page is already in a buffer */
for (slotno = 0; slotno < shared->num_slots; slotno++)
@@ -517,8 +523,13 @@ SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno, TransactionId xid)
}
/* No luck, so switch to normal exclusive lock and do regular read */
- LWLockRelease(shared->ControlLock);
- LWLockAcquire(shared->ControlLock, LW_EXCLUSIVE);
+ if (*lock_mode != LW_EXCLUSIVE)
+ {
+ Assert(*lock_mode == LW_NONE);
+ LWLockRelease(shared->ControlLock);
+ LWLockAcquire(shared->ControlLock, LW_EXCLUSIVE);
+ *lock_mode = LW_EXCLUSIVE;
+ }
return SimpleLruReadPage(ctl, pageno, true, xid);
}
diff --git a/src/backend/access/transam/subtrans.c b/src/backend/access/transam/subtrans.c
index 0111e867c7..7bb1543189 100644
--- a/src/backend/access/transam/subtrans.c
+++ b/src/backend/access/transam/subtrans.c
@@ -113,6 +113,7 @@ SubTransGetParent(TransactionId xid)
int slotno;
TransactionId *ptr;
TransactionId parent;
+ LWLockMode lockmode = LW_NONE;
/* Can't ask about stuff that might not be around anymore */
Assert(TransactionIdFollowsOrEquals(xid, TransactionXmin));
@@ -123,12 +124,13 @@ SubTransGetParent(TransactionId xid)
/* lock is acquired by SimpleLruReadPage_ReadOnly */
- slotno = SimpleLruReadPage_ReadOnly(SubTransCtl, pageno, xid);
+ slotno = SimpleLruReadPage_ReadOnly(SubTransCtl, pageno, xid, &lockmode);
ptr = (TransactionId *) SubTransCtl->shared->page_buffer[slotno];
ptr += entryno;
parent = *ptr;
+ Assert(lockmode != LW_NONE);
LWLockRelease(SubtransSLRULock);
return parent;
diff --git a/src/backend/commands/async.c b/src/backend/commands/async.c
index c0763c63e2..ef1dbbe53d 100644
--- a/src/backend/commands/async.c
+++ b/src/backend/commands/async.c
@@ -2011,6 +2011,7 @@ asyncQueueReadAllNotifications(void)
int curoffset = QUEUE_POS_OFFSET(pos);
int slotno;
int copysize;
+ LWLockMode lockmode = LW_NONE;
/*
* We copy the data from SLRU into a local buffer, so as to avoid
@@ -2019,7 +2020,7 @@ asyncQueueReadAllNotifications(void)
* part of the page we will actually inspect.
*/
slotno = SimpleLruReadPage_ReadOnly(NotifyCtl, curpage,
- InvalidTransactionId);
+ InvalidTransactionId, &lockmode);
if (curpage == QUEUE_POS_PAGE(head))
{
/* we only want to read as far as head */
@@ -2036,6 +2037,7 @@ asyncQueueReadAllNotifications(void)
NotifyCtl->shared->page_buffer[slotno] + curoffset,
copysize);
/* Release lock that we got from SimpleLruReadPage_ReadOnly() */
+ Assert(lockmode != LW_NONE);
LWLockRelease(NotifySLRULock);
/*
diff --git a/src/backend/storage/lmgr/lwlock.c b/src/backend/storage/lmgr/lwlock.c
index 108e652179..ad1dcc7140 100644
--- a/src/backend/storage/lmgr/lwlock.c
+++ b/src/backend/storage/lmgr/lwlock.c
@@ -1067,6 +1067,8 @@ LWLockWakeup(LWLock *lock)
static void
LWLockQueueSelf(LWLock *lock, LWLockMode mode)
{
+ Assert(mode != LW_NONE);
+
/*
* If we don't have a PGPROC structure, there's no way to wait. This
* should never occur, since MyProc should only be null during shared
@@ -1827,6 +1829,7 @@ LWLockRelease(LWLock *lock)
elog(ERROR, "lock %s is not held", T_NAME(lock));
mode = held_lwlocks[i].mode;
+ Assert(mode == LW_EXCLUSIVE || mode == LW_SHARED);
num_held_lwlocks--;
for (; i < num_held_lwlocks; i++)
diff --git a/src/backend/storage/lmgr/predicate.c b/src/backend/storage/lmgr/predicate.c
index 8a365b400c..a4df90a8ae 100644
--- a/src/backend/storage/lmgr/predicate.c
+++ b/src/backend/storage/lmgr/predicate.c
@@ -924,6 +924,7 @@ SerialGetMinConflictCommitSeqNo(TransactionId xid)
TransactionId tailXid;
SerCommitSeqNo val;
int slotno;
+ LWLockMode lockmode = LW_NONE;
Assert(TransactionIdIsValid(xid));
@@ -946,8 +947,9 @@ SerialGetMinConflictCommitSeqNo(TransactionId xid)
* but will return with that lock held, which must then be released.
*/
slotno = SimpleLruReadPage_ReadOnly(SerialSlruCtl,
- SerialPage(xid), xid);
+ SerialPage(xid), xid, &lockmode);
val = SerialValue(slotno, xid);
+ Assert(lockmode != LW_NONE);
LWLockRelease(SerialSLRULock);
return val;
}
diff --git a/src/include/access/slru.h b/src/include/access/slru.h
index b39b43504d..4b66d3b592 100644
--- a/src/include/access/slru.h
+++ b/src/include/access/slru.h
@@ -142,7 +142,7 @@ extern int SimpleLruZeroPage(SlruCtl ctl, int pageno);
extern int SimpleLruReadPage(SlruCtl ctl, int pageno, bool write_ok,
TransactionId xid);
extern int SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno,
- TransactionId xid);
+ TransactionId xid, LWLockMode *lock_mode);
extern void SimpleLruWritePage(SlruCtl ctl, int slotno);
extern void SimpleLruWriteAll(SlruCtl ctl, bool allow_redirtied);
extern void SimpleLruTruncate(SlruCtl ctl, int cutoffPage);
diff --git a/src/include/storage/lwlock.h b/src/include/storage/lwlock.h
index af9b41795d..e680c6397a 100644
--- a/src/include/storage/lwlock.h
+++ b/src/include/storage/lwlock.h
@@ -129,6 +129,8 @@ extern PGDLLIMPORT int NamedLWLockTrancheRequests;
typedef enum LWLockMode
{
+ LW_NONE, /* Not a lock mode. Indicates that there is no
+ * lock. */
LW_EXCLUSIVE,
LW_SHARED,
LW_WAIT_UNTIL_FREE /* A special mode used in PGPROC->lwWaitMode,
--
2.24.3 (Apple Git-128)
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2020-12-23 16:31 Gilles Darold <[email protected]>
parent: Andrey Borodin <[email protected]>
1 sibling, 1 reply; 103+ messages in thread
From: Gilles Darold @ 2020-12-23 16:31 UTC (permalink / raw)
To: Andrey Borodin <[email protected]>; +Cc: Tomas Vondra <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
Le 13/12/2020 à 18:24, Andrey Borodin a écrit :
>
>> 13 дек. 2020 г., в 14:17, Gilles Darold <[email protected]> написал(а):
>>
>> I've done more review on these patches.
> Thanks, Gilles! I'll incorporate all your fixes to patchset.
> Can you also benchmark conditional variable sleep? The patch "Add conditional variable to wait for next MultXact offset in corner case"?
> The problem manifests on Standby when Primary is heavily loaded with MultiXactOffsetControlLock.
>
> Thanks!
>
> Best regards, Andrey Borodin.
Hi Andrey,
Sorry for the response delay, we have run several others tests trying to
figure out the performances gain per patch but unfortunately we have
very heratic results. With the same parameters and patches the test
doesn't returns the same results following the day or the hour of the
day. This is very frustrating and I suppose that this is related to the
Azure architecture. The only thing that I am sure is that we had the
best performances results with all patches and
multixact_offsets_slru_buffers = 256
multixact_members_slru_buffers = 512
multixact_local_cache_entries = 4096
but I can not say if all or part of the patches are improving the
performances. My feeling is that performances gain related to patches 1
(shared lock) and 3 (conditional variable) do not have much to do with
the performances gain compared to just tuning the multixact buffers.
This is when the multixact contention is observed but perhaps they are
delaying the contention. It's all the more frustrating that we had a
test case to reproduce the contention but not the architecture apparently.
Can't do much more at this point.
Best regards,
--
Gilles Darold
LzLabs GmbH
http://www.lzlabs.com/
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2021-02-15 17:17 Andrey Borodin <[email protected]>
parent: Gilles Darold <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Andrey Borodin @ 2021-02-15 17:17 UTC (permalink / raw)
To: Gilles Darold <[email protected]>; +Cc: Tomas Vondra <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
> 23 дек. 2020 г., в 21:31, Gilles Darold <[email protected]> написал(а):
>
> Sorry for the response delay, we have run several others tests trying to figure out the performances gain per patch but unfortunately we have very heratic results. With the same parameters and patches the test doesn't returns the same results following the day or the hour of the day. This is very frustrating and I suppose that this is related to the Azure architecture. The only thing that I am sure is that we had the best performances results with all patches and
>
> multixact_offsets_slru_buffers = 256
> multixact_members_slru_buffers = 512
> multixact_local_cache_entries = 4096
>
>
> but I can not say if all or part of the patches are improving the performances. My feeling is that performances gain related to patches 1 (shared lock) and 3 (conditional variable) do not have much to do with the performances gain compared to just tuning the multixact buffers. This is when the multixact contention is observed but perhaps they are delaying the contention. It's all the more frustrating that we had a test case to reproduce the contention but not the architecture apparently.
Hi! Thanks for the input.
I think we have a consensus here that configuring SLRU size is beneficial for MultiXacts.
There is proposal in nearby thread [0] on changing default value of commit_ts SLRU buffers.
In my experience from time to time there can be problems with subtransactions cured by extending subtrans SLRU.
Let's make all SLRUs configurable?
PFA patch with draft of these changes.
Best regards, Andrey Borodin.
[0] https://www.postgresql.org/message-id/flat/20210115220744.GA24457%40alvherre.pgsql
Attachments:
[application/octet-stream] v9-0001-Make-all-SLRU-buffer-sizes-configurable.patch (19.1K, ../../[email protected]/2-v9-0001-Make-all-SLRU-buffer-sizes-configurable.patch)
download | inline diff:
From 1b488106f7dfc142f219bb8ba0a0c5d2b6b94ba5 Mon Sep 17 00:00:00 2001
From: Andrey Borodin <[email protected]>
Date: Mon, 15 Feb 2021 21:51:56 +0500
Subject: [PATCH v9] Make all SLRU buffer sizes configurable
---
doc/src/sgml/config.sgml | 101 ++++++++++++++++++
src/backend/access/transam/clog.c | 6 ++
src/backend/access/transam/commit_ts.c | 5 +-
src/backend/access/transam/multixact.c | 8 +-
src/backend/access/transam/subtrans.c | 5 +-
src/backend/commands/async.c | 8 +-
src/backend/storage/lmgr/predicate.c | 4 +-
src/backend/utils/init/globals.c | 8 ++
src/backend/utils/misc/guc.c | 77 +++++++++++++
src/backend/utils/misc/postgresql.conf.sample | 16 +++
src/include/access/multixact.h | 4 -
src/include/access/subtrans.h | 3 -
src/include/commands/async.h | 5 -
src/include/miscadmin.h | 8 ++
src/include/storage/predicate.h | 4 -
15 files changed, 233 insertions(+), 29 deletions(-)
diff --git a/doc/src/sgml/config.sgml b/doc/src/sgml/config.sgml
index 4df1405d2e..5b24b435f3 100644
--- a/doc/src/sgml/config.sgml
+++ b/doc/src/sgml/config.sgml
@@ -1852,6 +1852,107 @@ include_dir 'conf.d'
</para>
</listitem>
</varlistentry>
+
+ <varlistentry id="guc-multixact-offsets-slru-buffers" xreflabel="multixact_offsets_slru_buffers">
+ <term><varname>multixact_offsets_slru_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>multixact_offsets_slru_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of memory to be used for MultiXact offsets. MultiXact offsets
+ are used to store information about offsets of multiple row lockers (caused by SELECT FOR UPDATE and others).
+ It defaults to 64 kilobytes (<literal>64KB</literal>).
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-multixact-members-slru-buffers" xreflabel="multixact_members_slru_buffers">
+ <term><varname>multixact_members_slru_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>multixact_members_slru_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of memory to be used for MultiXact members. MultiXact members
+ are used to store information about XIDs of multiple row lockers. Tipically <varname>multixact_members_slru_buffers</varname>
+ is twice more than <varname>multixact_offsets_slru_buffers</varname>.
+ It defaults to 128 kilobytes (<literal>128KB</literal>).
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-subtrans-buffers" xreflabel="subtrans_slru_buffers">
+ <term><varname>subtrans_slru_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>subtrans_slru_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of memory to be used for subtransactions.
+ It defaults to 256 kilobytes (<literal>256KB</literal>).
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-notify-buffers" xreflabel="notify_slru_buffers">
+ <term><varname>notify_slru_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>notify_slru_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of memory to be used for asincronous notifications (NOTIFY, LISTEN).
+ It defaults to 64 kilobytes (<literal>64KB</literal>).
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-serial-buffers" xreflabel="serial_slru_buffers">
+ <term><varname>serial_slru_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>serial_slru_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of memory to be used for predicate locks.
+ It defaults to 128 kilobytes (<literal>128KB</literal>).
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-clog-buffers" xreflabel="clog_slru_buffers">
+ <term><varname>clog_slru_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>clog_slru_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of memory to be used for CLOG.
+ It defaults to 0, in this case CLOG size is taken as <varname>shared_buffers</varname> / 512.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-commit-ts-buffers" xreflabel="commit_ts_slru_buffers">
+ <term><varname>commit_ts_slru_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>commit_ts_slru_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of memory to be used for commit timestamps.
+ It defaults to 0, in this case CLOG size is taken as <varname>shared_buffers</varname> / 512.
+ </para>
+ </listitem>
+ </varlistentry>
<varlistentry id="guc-max-stack-depth" xreflabel="max_stack_depth">
<term><varname>max_stack_depth</varname> (<type>integer</type>)
diff --git a/src/backend/access/transam/clog.c b/src/backend/access/transam/clog.c
index 6fa4713fb4..e57106b374 100644
--- a/src/backend/access/transam/clog.c
+++ b/src/backend/access/transam/clog.c
@@ -659,6 +659,9 @@ TransactionIdGetStatus(TransactionId xid, XLogRecPtr *lsn)
/*
* Number of shared CLOG buffers.
*
+ * If values is confugured via GUC - just use given value. Otherwise
+ * apply following euristics.
+ *
* On larger multi-processor systems, it is possible to have many CLOG page
* requests in flight at one time which could lead to disk access for CLOG
* page if the required page is not found in memory. Testing revealed that we
@@ -675,6 +678,9 @@ TransactionIdGetStatus(TransactionId xid, XLogRecPtr *lsn)
Size
CLOGShmemBuffers(void)
{
+ /* consider 0 and 1 as unset GUC */
+ if (clog_slru_buffers > 1)
+ return clog_slru_buffers;
return Min(128, Max(4, NBuffers / 512));
}
diff --git a/src/backend/access/transam/commit_ts.c b/src/backend/access/transam/commit_ts.c
index 48e8d66286..7de3bca63d 100644
--- a/src/backend/access/transam/commit_ts.c
+++ b/src/backend/access/transam/commit_ts.c
@@ -530,7 +530,10 @@ pg_xact_commit_timestamp_origin(PG_FUNCTION_ARGS)
Size
CommitTsShmemBuffers(void)
{
- return Min(16, Max(4, NBuffers / 1024));
+ /* consider 0 and 1 as unset GUC */
+ if (commit_ts_slru_buffers > 1)
+ return commit_ts_slru_buffers;
+ return Min(16, Max(4, NBuffers / 512));
}
/*
diff --git a/src/backend/access/transam/multixact.c b/src/backend/access/transam/multixact.c
index 7dcfa02323..dd7dd19ff4 100644
--- a/src/backend/access/transam/multixact.c
+++ b/src/backend/access/transam/multixact.c
@@ -1831,8 +1831,8 @@ MultiXactShmemSize(void)
mul_size(sizeof(MultiXactId) * 2, MaxOldestSlot))
size = SHARED_MULTIXACT_STATE_SIZE;
- size = add_size(size, SimpleLruShmemSize(NUM_MULTIXACTOFFSET_BUFFERS, 0));
- size = add_size(size, SimpleLruShmemSize(NUM_MULTIXACTMEMBER_BUFFERS, 0));
+ size = add_size(size, SimpleLruShmemSize(multixact_offsets_slru_buffers, 0));
+ size = add_size(size, SimpleLruShmemSize(multixact_members_slru_buffers, 0));
return size;
}
@@ -1848,13 +1848,13 @@ MultiXactShmemInit(void)
MultiXactMemberCtl->PagePrecedes = MultiXactMemberPagePrecedes;
SimpleLruInit(MultiXactOffsetCtl,
- "MultiXactOffset", NUM_MULTIXACTOFFSET_BUFFERS, 0,
+ "MultiXactOffset", multixact_offsets_slru_buffers, 0,
MultiXactOffsetSLRULock, "pg_multixact/offsets",
LWTRANCHE_MULTIXACTOFFSET_BUFFER,
SYNC_HANDLER_MULTIXACT_OFFSET);
SlruPagePrecedesUnitTests(MultiXactOffsetCtl, MULTIXACT_OFFSETS_PER_PAGE);
SimpleLruInit(MultiXactMemberCtl,
- "MultiXactMember", NUM_MULTIXACTMEMBER_BUFFERS, 0,
+ "MultiXactMember", multixact_offsets_slru_buffers, 0,
MultiXactMemberSLRULock, "pg_multixact/members",
LWTRANCHE_MULTIXACTMEMBER_BUFFER,
SYNC_HANDLER_MULTIXACT_MEMBER);
diff --git a/src/backend/access/transam/subtrans.c b/src/backend/access/transam/subtrans.c
index 6a8e521f89..0c24353d3a 100644
--- a/src/backend/access/transam/subtrans.c
+++ b/src/backend/access/transam/subtrans.c
@@ -31,6 +31,7 @@
#include "access/slru.h"
#include "access/subtrans.h"
#include "access/transam.h"
+#include "miscadmin.h"
#include "pg_trace.h"
#include "utils/snapmgr.h"
@@ -184,14 +185,14 @@ SubTransGetTopmostTransaction(TransactionId xid)
Size
SUBTRANSShmemSize(void)
{
- return SimpleLruShmemSize(NUM_SUBTRANS_BUFFERS, 0);
+ return SimpleLruShmemSize(subtrans_slru_buffers, 0);
}
void
SUBTRANSShmemInit(void)
{
SubTransCtl->PagePrecedes = SubTransPagePrecedes;
- SimpleLruInit(SubTransCtl, "Subtrans", NUM_SUBTRANS_BUFFERS, 0,
+ SimpleLruInit(SubTransCtl, "Subtrans", subtrans_slru_buffers, 0,
SubtransSLRULock, "pg_subtrans",
LWTRANCHE_SUBTRANS_BUFFER, SYNC_HANDLER_NONE);
SlruPagePrecedesUnitTests(SubTransCtl, SUBTRANS_XACTS_PER_PAGE);
diff --git a/src/backend/commands/async.c b/src/backend/commands/async.c
index 42b232d98b..a7dd92add0 100644
--- a/src/backend/commands/async.c
+++ b/src/backend/commands/async.c
@@ -107,7 +107,7 @@
* frontend during startup.) The above design guarantees that notifies from
* other backends will never be missed by ignoring self-notifies.
*
- * The amount of shared memory used for notify management (NUM_NOTIFY_BUFFERS)
+ * The amount of shared memory used for notify management (notify_slru_buffers)
* can be varied without affecting anything but performance. The maximum
* amount of notification data that can be queued at one time is determined
* by slru.c's wraparound limit; see QUEUE_MAX_PAGE below.
@@ -225,7 +225,7 @@ typedef struct QueuePosition
*
* Resist the temptation to make this really large. While that would save
* work in some places, it would add cost in others. In particular, this
- * should likely be less than NUM_NOTIFY_BUFFERS, to ensure that backends
+ * should likely be less than notify_slru_buffers, to ensure that backends
* catch up before the pages they'll need to read fall out of SLRU cache.
*/
#define QUEUE_CLEANUP_DELAY 4
@@ -514,7 +514,7 @@ AsyncShmemSize(void)
size = mul_size(MaxBackends + 1, sizeof(QueueBackendStatus));
size = add_size(size, offsetof(AsyncQueueControl, backend));
- size = add_size(size, SimpleLruShmemSize(NUM_NOTIFY_BUFFERS, 0));
+ size = add_size(size, SimpleLruShmemSize(notify_slru_buffers, 0));
return size;
}
@@ -562,7 +562,7 @@ AsyncShmemInit(void)
* Set up SLRU management of the pg_notify data.
*/
NotifyCtl->PagePrecedes = asyncQueuePagePrecedes;
- SimpleLruInit(NotifyCtl, "Notify", NUM_NOTIFY_BUFFERS, 0,
+ SimpleLruInit(NotifyCtl, "Notify", notify_slru_buffers, 0,
NotifySLRULock, "pg_notify", LWTRANCHE_NOTIFY_BUFFER,
SYNC_HANDLER_NONE);
diff --git a/src/backend/storage/lmgr/predicate.c b/src/backend/storage/lmgr/predicate.c
index 074df5b38c..ad6d800dbd 100644
--- a/src/backend/storage/lmgr/predicate.c
+++ b/src/backend/storage/lmgr/predicate.c
@@ -872,7 +872,7 @@ SerialInit(void)
*/
SerialSlruCtl->PagePrecedes = SerialPagePrecedesLogically;
SimpleLruInit(SerialSlruCtl, "Serial",
- NUM_SERIAL_BUFFERS, 0, SerialSLRULock, "pg_serial",
+ serial_slru_buffers, 0, SerialSLRULock, "pg_serial",
LWTRANCHE_SERIAL_BUFFER, SYNC_HANDLER_NONE);
#ifdef USE_ASSERT_CHECKING
SerialPagePrecedesLogicallyUnitTests();
@@ -1395,7 +1395,7 @@ PredicateLockShmemSize(void)
/* Shared memory structures for SLRU tracking of old committed xids. */
size = add_size(size, sizeof(SerialControlData));
- size = add_size(size, SimpleLruShmemSize(NUM_SERIAL_BUFFERS, 0));
+ size = add_size(size, SimpleLruShmemSize(serial_slru_buffers, 0));
return size;
}
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
index a5976ad5b1..88e785c246 100644
--- a/src/backend/utils/init/globals.c
+++ b/src/backend/utils/init/globals.c
@@ -150,3 +150,11 @@ int VacuumCostBalance = 0; /* working state for vacuum */
bool VacuumCostActive = false;
double vacuum_cleanup_index_scale_factor;
+
+int multixact_offsets_slru_buffers = 8;
+int multixact_members_slru_buffers = 16;
+int subtrans_slru_buffers = 32;
+int notify_slru_buffers = 8;
+int serial_slru_buffers = 16;
+int clog_slru_buffers = 0;
+int commit_ts_slru_buffers = 0;
diff --git a/src/backend/utils/misc/guc.c b/src/backend/utils/misc/guc.c
index eafdb1118e..d30f85bff8 100644
--- a/src/backend/utils/misc/guc.c
+++ b/src/backend/utils/misc/guc.c
@@ -2269,6 +2269,83 @@ static struct config_int ConfigureNamesInt[] =
NULL, NULL, NULL
},
+ {
+ {"multixact_offsets_slru_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for MultiXact offsets SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &multixact_offsets_slru_buffers,
+ 8, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"multixact_members_slru_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for MultiXact members SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &multixact_members_slru_buffers,
+ 16, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"subtrans_slru_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for substrnsactions SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &subtrans_slru_buffers,
+ 32, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"notify_slru_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for asyncronous notifications SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ ¬ify_slru_buffers,
+ 8, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"serial_slru_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for predicate locks SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &serial_slru_buffers,
+ 16, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"clog_slru_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for commit log SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &serial_slru_buffers,
+ 0, 0, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"commit_ts_slru_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for commit timestamps SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &commit_ts_slru_buffers,
+ 0, 0, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
{
{"temp_buffers", PGC_USERSET, RESOURCES_MEM,
gettext_noop("Sets the maximum number of temporary buffers used by each session."),
diff --git a/src/backend/utils/misc/postgresql.conf.sample b/src/backend/utils/misc/postgresql.conf.sample
index db6db376eb..e88271c048 100644
--- a/src/backend/utils/misc/postgresql.conf.sample
+++ b/src/backend/utils/misc/postgresql.conf.sample
@@ -189,6 +189,22 @@
# (change requires restart)
#backend_flush_after = 0 # measured in pages, 0 disables
+# - SLRU Buffers -
+
+#multixact_offsets_slru_buffers = 8 # memory used for MultiXact offsets
+ # (change requires restart)
+#multixact_members_slru_buffers = 16 # memory used for MultiXact members
+ # (change requires restart)
+#subtrans_slru_buffers = 32 # memory used for subtransactions
+ # (change requires restart)
+#notify_slru_buffers = 8 # memory used for asynchronous notifications
+ # (change requires restart)
+#serial_slru_buffers = 16 # memory used for predicate locks
+ # (change requires restart)
+#clog_slru_buffers = 0 # memory used for CLOG
+ # (change requires restart)
+#commit_ts_slru_buffers = 0 # memory used for commit timestamps
+ # (change requires restart)
#------------------------------------------------------------------------------
# WRITE-AHEAD LOG
diff --git a/src/include/access/multixact.h b/src/include/access/multixact.h
index 4bbb035eae..97c0a46376 100644
--- a/src/include/access/multixact.h
+++ b/src/include/access/multixact.h
@@ -29,10 +29,6 @@
#define MaxMultiXactOffset ((MultiXactOffset) 0xFFFFFFFF)
-/* Number of SLRU buffers to use for multixact */
-#define NUM_MULTIXACTOFFSET_BUFFERS 8
-#define NUM_MULTIXACTMEMBER_BUFFERS 16
-
/*
* Possible multixact lock modes ("status"). The first four modes are for
* tuple locks (FOR KEY SHARE, FOR SHARE, FOR NO KEY UPDATE, FOR UPDATE); the
diff --git a/src/include/access/subtrans.h b/src/include/access/subtrans.h
index d0ab44ae82..ca0999056e 100644
--- a/src/include/access/subtrans.h
+++ b/src/include/access/subtrans.h
@@ -11,9 +11,6 @@
#ifndef SUBTRANS_H
#define SUBTRANS_H
-/* Number of SLRU buffers to use for subtrans */
-#define NUM_SUBTRANS_BUFFERS 32
-
extern void SubTransSetParent(TransactionId xid, TransactionId parent);
extern TransactionId SubTransGetParent(TransactionId xid);
extern TransactionId SubTransGetTopmostTransaction(TransactionId xid);
diff --git a/src/include/commands/async.h b/src/include/commands/async.h
index 9217f66b91..fa831e3721 100644
--- a/src/include/commands/async.h
+++ b/src/include/commands/async.h
@@ -15,11 +15,6 @@
#include <signal.h>
-/*
- * The number of SLRU page buffers we use for the notification queue.
- */
-#define NUM_NOTIFY_BUFFERS 8
-
extern bool Trace_notify;
extern volatile sig_atomic_t notifyInterruptPending;
diff --git a/src/include/miscadmin.h b/src/include/miscadmin.h
index 1bdc97e308..51c7ec8d68 100644
--- a/src/include/miscadmin.h
+++ b/src/include/miscadmin.h
@@ -162,6 +162,14 @@ extern PGDLLIMPORT int MaxBackends;
extern PGDLLIMPORT int MaxConnections;
extern PGDLLIMPORT int max_worker_processes;
extern PGDLLIMPORT int max_parallel_workers;
+extern PGDLLIMPORT int multixact_offsets_slru_buffers;
+extern PGDLLIMPORT int multixact_members_slru_buffers;
+extern PGDLLIMPORT int multixact_members_slru_buffers;
+extern PGDLLIMPORT int subtrans_slru_buffers;
+extern PGDLLIMPORT int notify_slru_buffers;
+extern PGDLLIMPORT int serial_slru_buffers;
+extern PGDLLIMPORT int clog_slru_buffers;
+extern PGDLLIMPORT int commit_ts_slru_buffers;
extern PGDLLIMPORT int MyProcPid;
extern PGDLLIMPORT pg_time_t MyStartTime;
diff --git a/src/include/storage/predicate.h b/src/include/storage/predicate.h
index 152b698611..c72779bd88 100644
--- a/src/include/storage/predicate.h
+++ b/src/include/storage/predicate.h
@@ -26,10 +26,6 @@ extern int max_predicate_locks_per_xact;
extern int max_predicate_locks_per_relation;
extern int max_predicate_locks_per_page;
-
-/* Number of SLRU buffers to use for Serial SLRU */
-#define NUM_SERIAL_BUFFERS 16
-
/*
* A handle used for sharing SERIALIZABLEXACT objects between the participants
* in a parallel query.
--
2.24.3 (Apple Git-128)
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2021-03-11 15:50 Gilles Darold <[email protected]>
parent: Andrey Borodin <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Gilles Darold @ 2021-03-11 15:50 UTC (permalink / raw)
To: Andrey Borodin <[email protected]>; +Cc: Tomas Vondra <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
Le 15/02/2021 à 18:17, Andrey Borodin a écrit :
>
>> 23 дек. 2020 г., в 21:31, Gilles Darold <[email protected]> написал(а):
>>
>> Sorry for the response delay, we have run several others tests trying to figure out the performances gain per patch but unfortunately we have very heratic results. With the same parameters and patches the test doesn't returns the same results following the day or the hour of the day. This is very frustrating and I suppose that this is related to the Azure architecture. The only thing that I am sure is that we had the best performances results with all patches and
>>
>> multixact_offsets_slru_buffers = 256
>> multixact_members_slru_buffers = 512
>> multixact_local_cache_entries = 4096
>>
>>
>> but I can not say if all or part of the patches are improving the performances. My feeling is that performances gain related to patches 1 (shared lock) and 3 (conditional variable) do not have much to do with the performances gain compared to just tuning the multixact buffers. This is when the multixact contention is observed but perhaps they are delaying the contention. It's all the more frustrating that we had a test case to reproduce the contention but not the architecture apparently.
> Hi! Thanks for the input.
> I think we have a consensus here that configuring SLRU size is beneficial for MultiXacts.
> There is proposal in nearby thread [0] on changing default value of commit_ts SLRU buffers.
> In my experience from time to time there can be problems with subtransactions cured by extending subtrans SLRU.
>
> Let's make all SLRUs configurable?
> PFA patch with draft of these changes.
>
> Best regards, Andrey Borodin.
>
>
> [0] https://www.postgresql.org/message-id/flat/20210115220744.GA24457%40alvherre.pgsql
>
The patch doesn't apply anymore in master cause of error: patch failed:
src/backend/utils/init/globals.c:150
An other remark about this patch is that it should be mentionned in the
documentation (doc/src/sgml/config.sgml) that the new configuration
variables need a server restart, for example by adding "This parameter
can only be set at server start." like for shared_buffers. Patch on
postgresql.conf mention it.
And some typo to be fixed:
s/Tipically/Typically/
s/asincronous/asyncronous/
s/confugured/configured/
s/substrnsactions/substransactions/
--
Gilles Darold
LzLabs GmbH
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2021-03-12 12:44 Andrey Borodin <[email protected]>
parent: Gilles Darold <[email protected]>
0 siblings, 2 replies; 103+ messages in thread
From: Andrey Borodin @ 2021-03-12 12:44 UTC (permalink / raw)
To: Gilles Darold <[email protected]>; +Cc: Tomas Vondra <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
> 11 марта 2021 г., в 20:50, Gilles Darold <[email protected]> написал(а):
>
>
> The patch doesn't apply anymore in master cause of error: patch failed: src/backend/utils/init/globals.c:150
>
>
>
> An other remark about this patch is that it should be mentionned in the documentation (doc/src/sgml/config.sgml) that the new configuration variables need a server restart, for example by adding "This parameter can only be set at server start." like for shared_buffers. Patch on postgresql.conf mention it.
>
> And some typo to be fixed:
>
>
>
> s/Tipically/Typically/
>
> s/asincronous/asyncronous/
>
> s/confugured/configured/
>
> s/substrnsactions/substransactions/
>
>
Thanks, Gilles! Fixed.
Best regards, Andrey Borodin.
Attachments:
[application/octet-stream] v10-0001-Make-all-SLRU-buffer-sizes-configurable.patch (19.5K, ../../[email protected]/2-v10-0001-Make-all-SLRU-buffer-sizes-configurable.patch)
download | inline diff:
From 02c5d56a65a187b0795fac246a77f9cf3107ed3b Mon Sep 17 00:00:00 2001
From: Andrey Borodin <[email protected]>
Date: Mon, 15 Feb 2021 21:51:56 +0500
Subject: [PATCH v10] Make all SLRU buffer sizes configurable
---
doc/src/sgml/config.sgml | 108 ++++++++++++++++++
src/backend/access/transam/clog.c | 6 +
src/backend/access/transam/commit_ts.c | 5 +-
src/backend/access/transam/multixact.c | 8 +-
src/backend/access/transam/subtrans.c | 5 +-
src/backend/commands/async.c | 8 +-
src/backend/storage/lmgr/predicate.c | 4 +-
src/backend/utils/init/globals.c | 8 ++
src/backend/utils/misc/guc.c | 77 +++++++++++++
src/backend/utils/misc/postgresql.conf.sample | 16 +++
src/include/access/multixact.h | 4 -
src/include/access/subtrans.h | 3 -
src/include/commands/async.h | 5 -
src/include/miscadmin.h | 8 ++
src/include/storage/predicate.h | 4 -
15 files changed, 240 insertions(+), 29 deletions(-)
diff --git a/doc/src/sgml/config.sgml b/doc/src/sgml/config.sgml
index a218d78bef..eb7f8f3c49 100644
--- a/doc/src/sgml/config.sgml
+++ b/doc/src/sgml/config.sgml
@@ -1886,6 +1886,114 @@ include_dir 'conf.d'
</para>
</listitem>
</varlistentry>
+
+ <varlistentry id="guc-multixact-offsets-slru-buffers" xreflabel="multixact_offsets_slru_buffers">
+ <term><varname>multixact_offsets_slru_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>multixact_offsets_slru_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of memory to be used for MultiXact offsets. MultiXact offsets
+ are used to store information about offsets of multiple row lockers (caused by SELECT FOR UPDATE and others).
+ It defaults to 64 kilobytes (<literal>64KB</literal>).
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-multixact-members-slru-buffers" xreflabel="multixact_members_slru_buffers">
+ <term><varname>multixact_members_slru_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>multixact_members_slru_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of memory to be used for MultiXact members. MultiXact members
+ are used to store information about XIDs of multiple row lockers. Typically <varname>multixact_members_slru_buffers</varname>
+ is twice more than <varname>multixact_offsets_slru_buffers</varname>.
+ It defaults to 128 kilobytes (<literal>128KB</literal>).
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-subtrans-buffers" xreflabel="subtrans_slru_buffers">
+ <term><varname>subtrans_slru_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>subtrans_slru_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of memory to be used for subtransactions.
+ It defaults to 256 kilobytes (<literal>256KB</literal>).
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-notify-buffers" xreflabel="notify_slru_buffers">
+ <term><varname>notify_slru_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>notify_slru_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of memory to be used for asyncronous notifications (NOTIFY, LISTEN).
+ It defaults to 64 kilobytes (<literal>64KB</literal>).
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-serial-buffers" xreflabel="serial_slru_buffers">
+ <term><varname>serial_slru_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>serial_slru_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of memory to be used for predicate locks.
+ It defaults to 128 kilobytes (<literal>128KB</literal>).
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-clog-buffers" xreflabel="clog_slru_buffers">
+ <term><varname>clog_slru_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>clog_slru_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of memory to be used for CLOG.
+ It defaults to 0, in this case CLOG size is taken as <varname>shared_buffers</varname> / 512.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-commit-ts-buffers" xreflabel="commit_ts_slru_buffers">
+ <term><varname>commit_ts_slru_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>commit_ts_slru_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of memory to be used for commit timestamps.
+ It defaults to 0, in this case CLOG size is taken as <varname>shared_buffers</varname> / 512.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
<varlistentry id="guc-max-stack-depth" xreflabel="max_stack_depth">
<term><varname>max_stack_depth</varname> (<type>integer</type>)
diff --git a/src/backend/access/transam/clog.c b/src/backend/access/transam/clog.c
index 6fa4713fb4..e1d34aa361 100644
--- a/src/backend/access/transam/clog.c
+++ b/src/backend/access/transam/clog.c
@@ -659,6 +659,9 @@ TransactionIdGetStatus(TransactionId xid, XLogRecPtr *lsn)
/*
* Number of shared CLOG buffers.
*
+ * If values is configured via GUC - just use given value. Otherwise
+ * apply following euristics.
+ *
* On larger multi-processor systems, it is possible to have many CLOG page
* requests in flight at one time which could lead to disk access for CLOG
* page if the required page is not found in memory. Testing revealed that we
@@ -675,6 +678,9 @@ TransactionIdGetStatus(TransactionId xid, XLogRecPtr *lsn)
Size
CLOGShmemBuffers(void)
{
+ /* consider 0 and 1 as unset GUC */
+ if (clog_slru_buffers > 1)
+ return clog_slru_buffers;
return Min(128, Max(4, NBuffers / 512));
}
diff --git a/src/backend/access/transam/commit_ts.c b/src/backend/access/transam/commit_ts.c
index 48e8d66286..7de3bca63d 100644
--- a/src/backend/access/transam/commit_ts.c
+++ b/src/backend/access/transam/commit_ts.c
@@ -530,7 +530,10 @@ pg_xact_commit_timestamp_origin(PG_FUNCTION_ARGS)
Size
CommitTsShmemBuffers(void)
{
- return Min(16, Max(4, NBuffers / 1024));
+ /* consider 0 and 1 as unset GUC */
+ if (commit_ts_slru_buffers > 1)
+ return commit_ts_slru_buffers;
+ return Min(16, Max(4, NBuffers / 512));
}
/*
diff --git a/src/backend/access/transam/multixact.c b/src/backend/access/transam/multixact.c
index 1f9f1a1fa1..370c01e72b 100644
--- a/src/backend/access/transam/multixact.c
+++ b/src/backend/access/transam/multixact.c
@@ -1831,8 +1831,8 @@ MultiXactShmemSize(void)
mul_size(sizeof(MultiXactId) * 2, MaxOldestSlot))
size = SHARED_MULTIXACT_STATE_SIZE;
- size = add_size(size, SimpleLruShmemSize(NUM_MULTIXACTOFFSET_BUFFERS, 0));
- size = add_size(size, SimpleLruShmemSize(NUM_MULTIXACTMEMBER_BUFFERS, 0));
+ size = add_size(size, SimpleLruShmemSize(multixact_offsets_slru_buffers, 0));
+ size = add_size(size, SimpleLruShmemSize(multixact_members_slru_buffers, 0));
return size;
}
@@ -1848,13 +1848,13 @@ MultiXactShmemInit(void)
MultiXactMemberCtl->PagePrecedes = MultiXactMemberPagePrecedes;
SimpleLruInit(MultiXactOffsetCtl,
- "MultiXactOffset", NUM_MULTIXACTOFFSET_BUFFERS, 0,
+ "MultiXactOffset", multixact_offsets_slru_buffers, 0,
MultiXactOffsetSLRULock, "pg_multixact/offsets",
LWTRANCHE_MULTIXACTOFFSET_BUFFER,
SYNC_HANDLER_MULTIXACT_OFFSET);
SlruPagePrecedesUnitTests(MultiXactOffsetCtl, MULTIXACT_OFFSETS_PER_PAGE);
SimpleLruInit(MultiXactMemberCtl,
- "MultiXactMember", NUM_MULTIXACTMEMBER_BUFFERS, 0,
+ "MultiXactMember", multixact_offsets_slru_buffers, 0,
MultiXactMemberSLRULock, "pg_multixact/members",
LWTRANCHE_MULTIXACTMEMBER_BUFFER,
SYNC_HANDLER_MULTIXACT_MEMBER);
diff --git a/src/backend/access/transam/subtrans.c b/src/backend/access/transam/subtrans.c
index 6a8e521f89..0c24353d3a 100644
--- a/src/backend/access/transam/subtrans.c
+++ b/src/backend/access/transam/subtrans.c
@@ -31,6 +31,7 @@
#include "access/slru.h"
#include "access/subtrans.h"
#include "access/transam.h"
+#include "miscadmin.h"
#include "pg_trace.h"
#include "utils/snapmgr.h"
@@ -184,14 +185,14 @@ SubTransGetTopmostTransaction(TransactionId xid)
Size
SUBTRANSShmemSize(void)
{
- return SimpleLruShmemSize(NUM_SUBTRANS_BUFFERS, 0);
+ return SimpleLruShmemSize(subtrans_slru_buffers, 0);
}
void
SUBTRANSShmemInit(void)
{
SubTransCtl->PagePrecedes = SubTransPagePrecedes;
- SimpleLruInit(SubTransCtl, "Subtrans", NUM_SUBTRANS_BUFFERS, 0,
+ SimpleLruInit(SubTransCtl, "Subtrans", subtrans_slru_buffers, 0,
SubtransSLRULock, "pg_subtrans",
LWTRANCHE_SUBTRANS_BUFFER, SYNC_HANDLER_NONE);
SlruPagePrecedesUnitTests(SubTransCtl, SUBTRANS_XACTS_PER_PAGE);
diff --git a/src/backend/commands/async.c b/src/backend/commands/async.c
index 4b16fb5682..f5c5592057 100644
--- a/src/backend/commands/async.c
+++ b/src/backend/commands/async.c
@@ -107,7 +107,7 @@
* frontend during startup.) The above design guarantees that notifies from
* other backends will never be missed by ignoring self-notifies.
*
- * The amount of shared memory used for notify management (NUM_NOTIFY_BUFFERS)
+ * The amount of shared memory used for notify management (notify_slru_buffers)
* can be varied without affecting anything but performance. The maximum
* amount of notification data that can be queued at one time is determined
* by slru.c's wraparound limit; see QUEUE_MAX_PAGE below.
@@ -225,7 +225,7 @@ typedef struct QueuePosition
*
* Resist the temptation to make this really large. While that would save
* work in some places, it would add cost in others. In particular, this
- * should likely be less than NUM_NOTIFY_BUFFERS, to ensure that backends
+ * should likely be less than notify_slru_buffers, to ensure that backends
* catch up before the pages they'll need to read fall out of SLRU cache.
*/
#define QUEUE_CLEANUP_DELAY 4
@@ -514,7 +514,7 @@ AsyncShmemSize(void)
size = mul_size(MaxBackends + 1, sizeof(QueueBackendStatus));
size = add_size(size, offsetof(AsyncQueueControl, backend));
- size = add_size(size, SimpleLruShmemSize(NUM_NOTIFY_BUFFERS, 0));
+ size = add_size(size, SimpleLruShmemSize(notify_slru_buffers, 0));
return size;
}
@@ -562,7 +562,7 @@ AsyncShmemInit(void)
* Set up SLRU management of the pg_notify data.
*/
NotifyCtl->PagePrecedes = asyncQueuePagePrecedes;
- SimpleLruInit(NotifyCtl, "Notify", NUM_NOTIFY_BUFFERS, 0,
+ SimpleLruInit(NotifyCtl, "Notify", notify_slru_buffers, 0,
NotifySLRULock, "pg_notify", LWTRANCHE_NOTIFY_BUFFER,
SYNC_HANDLER_NONE);
diff --git a/src/backend/storage/lmgr/predicate.c b/src/backend/storage/lmgr/predicate.c
index d493aeef0f..fad8cc572e 100644
--- a/src/backend/storage/lmgr/predicate.c
+++ b/src/backend/storage/lmgr/predicate.c
@@ -872,7 +872,7 @@ SerialInit(void)
*/
SerialSlruCtl->PagePrecedes = SerialPagePrecedesLogically;
SimpleLruInit(SerialSlruCtl, "Serial",
- NUM_SERIAL_BUFFERS, 0, SerialSLRULock, "pg_serial",
+ serial_slru_buffers, 0, SerialSLRULock, "pg_serial",
LWTRANCHE_SERIAL_BUFFER, SYNC_HANDLER_NONE);
#ifdef USE_ASSERT_CHECKING
SerialPagePrecedesLogicallyUnitTests();
@@ -1395,7 +1395,7 @@ PredicateLockShmemSize(void)
/* Shared memory structures for SLRU tracking of old committed xids. */
size = add_size(size, sizeof(SerialControlData));
- size = add_size(size, SimpleLruShmemSize(NUM_SERIAL_BUFFERS, 0));
+ size = add_size(size, SimpleLruShmemSize(serial_slru_buffers, 0));
return size;
}
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
index 73e0a672ae..f163ca17e9 100644
--- a/src/backend/utils/init/globals.c
+++ b/src/backend/utils/init/globals.c
@@ -148,3 +148,11 @@ int64 VacuumPageDirty = 0;
int VacuumCostBalance = 0; /* working state for vacuum */
bool VacuumCostActive = false;
+
+int multixact_offsets_slru_buffers = 8;
+int multixact_members_slru_buffers = 16;
+int subtrans_slru_buffers = 32;
+int notify_slru_buffers = 8;
+int serial_slru_buffers = 16;
+int clog_slru_buffers = 0;
+int commit_ts_slru_buffers = 0;
diff --git a/src/backend/utils/misc/guc.c b/src/backend/utils/misc/guc.c
index 855076b1fd..080b3a9b35 100644
--- a/src/backend/utils/misc/guc.c
+++ b/src/backend/utils/misc/guc.c
@@ -2279,6 +2279,83 @@ static struct config_int ConfigureNamesInt[] =
NULL, NULL, NULL
},
+ {
+ {"multixact_offsets_slru_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for MultiXact offsets SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &multixact_offsets_slru_buffers,
+ 8, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"multixact_members_slru_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for MultiXact members SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &multixact_members_slru_buffers,
+ 16, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"subtrans_slru_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for substransactions SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &subtrans_slru_buffers,
+ 32, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"notify_slru_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for asyncronous notifications SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ ¬ify_slru_buffers,
+ 8, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"serial_slru_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for predicate locks SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &serial_slru_buffers,
+ 16, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"clog_slru_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for commit log SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &serial_slru_buffers,
+ 0, 0, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"commit_ts_slru_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for commit timestamps SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &commit_ts_slru_buffers,
+ 0, 0, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
{
{"temp_buffers", PGC_USERSET, RESOURCES_MEM,
gettext_noop("Sets the maximum number of temporary buffers used by each session."),
diff --git a/src/backend/utils/misc/postgresql.conf.sample b/src/backend/utils/misc/postgresql.conf.sample
index f46c2dd7a8..9ee7d17e8b 100644
--- a/src/backend/utils/misc/postgresql.conf.sample
+++ b/src/backend/utils/misc/postgresql.conf.sample
@@ -190,6 +190,22 @@
# (change requires restart)
#backend_flush_after = 0 # measured in pages, 0 disables
+# - SLRU Buffers -
+
+#multixact_offsets_slru_buffers = 8 # memory used for MultiXact offsets
+ # (change requires restart)
+#multixact_members_slru_buffers = 16 # memory used for MultiXact members
+ # (change requires restart)
+#subtrans_slru_buffers = 32 # memory used for subtransactions
+ # (change requires restart)
+#notify_slru_buffers = 8 # memory used for asynchronous notifications
+ # (change requires restart)
+#serial_slru_buffers = 16 # memory used for predicate locks
+ # (change requires restart)
+#clog_slru_buffers = 0 # memory used for CLOG
+ # (change requires restart)
+#commit_ts_slru_buffers = 0 # memory used for commit timestamps
+ # (change requires restart)
#------------------------------------------------------------------------------
# WRITE-AHEAD LOG
diff --git a/src/include/access/multixact.h b/src/include/access/multixact.h
index 4bbb035eae..97c0a46376 100644
--- a/src/include/access/multixact.h
+++ b/src/include/access/multixact.h
@@ -29,10 +29,6 @@
#define MaxMultiXactOffset ((MultiXactOffset) 0xFFFFFFFF)
-/* Number of SLRU buffers to use for multixact */
-#define NUM_MULTIXACTOFFSET_BUFFERS 8
-#define NUM_MULTIXACTMEMBER_BUFFERS 16
-
/*
* Possible multixact lock modes ("status"). The first four modes are for
* tuple locks (FOR KEY SHARE, FOR SHARE, FOR NO KEY UPDATE, FOR UPDATE); the
diff --git a/src/include/access/subtrans.h b/src/include/access/subtrans.h
index d0ab44ae82..ca0999056e 100644
--- a/src/include/access/subtrans.h
+++ b/src/include/access/subtrans.h
@@ -11,9 +11,6 @@
#ifndef SUBTRANS_H
#define SUBTRANS_H
-/* Number of SLRU buffers to use for subtrans */
-#define NUM_SUBTRANS_BUFFERS 32
-
extern void SubTransSetParent(TransactionId xid, TransactionId parent);
extern TransactionId SubTransGetParent(TransactionId xid);
extern TransactionId SubTransGetTopmostTransaction(TransactionId xid);
diff --git a/src/include/commands/async.h b/src/include/commands/async.h
index 9217f66b91..fa831e3721 100644
--- a/src/include/commands/async.h
+++ b/src/include/commands/async.h
@@ -15,11 +15,6 @@
#include <signal.h>
-/*
- * The number of SLRU page buffers we use for the notification queue.
- */
-#define NUM_NOTIFY_BUFFERS 8
-
extern bool Trace_notify;
extern volatile sig_atomic_t notifyInterruptPending;
diff --git a/src/include/miscadmin.h b/src/include/miscadmin.h
index 54693e047a..ca88b20ffa 100644
--- a/src/include/miscadmin.h
+++ b/src/include/miscadmin.h
@@ -162,6 +162,14 @@ extern PGDLLIMPORT int MaxBackends;
extern PGDLLIMPORT int MaxConnections;
extern PGDLLIMPORT int max_worker_processes;
extern PGDLLIMPORT int max_parallel_workers;
+extern PGDLLIMPORT int multixact_offsets_slru_buffers;
+extern PGDLLIMPORT int multixact_members_slru_buffers;
+extern PGDLLIMPORT int multixact_members_slru_buffers;
+extern PGDLLIMPORT int subtrans_slru_buffers;
+extern PGDLLIMPORT int notify_slru_buffers;
+extern PGDLLIMPORT int serial_slru_buffers;
+extern PGDLLIMPORT int clog_slru_buffers;
+extern PGDLLIMPORT int commit_ts_slru_buffers;
extern PGDLLIMPORT int MyProcPid;
extern PGDLLIMPORT pg_time_t MyStartTime;
diff --git a/src/include/storage/predicate.h b/src/include/storage/predicate.h
index 152b698611..c72779bd88 100644
--- a/src/include/storage/predicate.h
+++ b/src/include/storage/predicate.h
@@ -26,10 +26,6 @@ extern int max_predicate_locks_per_xact;
extern int max_predicate_locks_per_relation;
extern int max_predicate_locks_per_page;
-
-/* Number of SLRU buffers to use for Serial SLRU */
-#define NUM_SERIAL_BUFFERS 16
-
/*
* A handle used for sharing SERIALIZABLEXACT objects between the participants
* in a parallel query.
--
2.24.3 (Apple Git-128)
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2021-03-15 15:41 Gilles Darold <[email protected]>
parent: Andrey Borodin <[email protected]>
1 sibling, 0 replies; 103+ messages in thread
From: Gilles Darold @ 2021-03-15 15:41 UTC (permalink / raw)
To: Andrey Borodin <[email protected]>; Gilles Darold <[email protected]>; +Cc: Tomas Vondra <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
Le 12/03/2021 à 13:44, Andrey Borodin a écrit :
>
>> 11 марта 2021 г., в 20:50, Gilles Darold <[email protected]> написал(а):
>>
>>
>> The patch doesn't apply anymore in master cause of error: patch failed: src/backend/utils/init/globals.c:150
>>
>>
>>
>> An other remark about this patch is that it should be mentionned in the documentation (doc/src/sgml/config.sgml) that the new configuration variables need a server restart, for example by adding "This parameter can only be set at server start." like for shared_buffers. Patch on postgresql.conf mention it.
>>
>> And some typo to be fixed:
>>
>>
>>
>> s/Tipically/Typically/
>>
>> s/asincronous/asyncronous/
>>
>> s/confugured/configured/
>>
>> s/substrnsactions/substransactions/
>>
>>
> Thanks, Gilles! Fixed.
>
> Best regards, Andrey Borodin.
>
Hi Andrey,
I found two problems in this patch, first in src/include/miscadmin.h
multixact_members_slru_buffers is declared twice:
extern PGDLLIMPORT int max_parallel_workers;
+extern PGDLLIMPORT int multixact_offsets_slru_buffers;
+extern PGDLLIMPORT int multixact_members_slru_buffers;
+extern PGDLLIMPORT int multixact_members_slru_buffers; <---------
+extern PGDLLIMPORT int subtrans_slru_buffers;
In file src/backend/access/transam/multixact.c the second variable
should be multixact_buffers_slru_buffers and not
multixact_offsets_slru_buffers.
@@ -1848,13 +1848,13 @@ MultiXactShmemInit(void)
MultiXactMemberCtl->PagePrecedes = MultiXactMemberPagePrecedes;
SimpleLruInit(MultiXactOffsetCtl,
- "MultiXactOffset",
NUM_MULTIXACTOFFSET_BUFFERS, 0,
+ "MultiXactOffset",
multixact_offsets_slru_buffers, 0,
MultiXactOffsetSLRULock,
"pg_multixact/offsets",
LWTRANCHE_MULTIXACTOFFSET_BUFFER,
SYNC_HANDLER_MULTIXACT_OFFSET);
SlruPagePrecedesUnitTests(MultiXactOffsetCtl,
MULTIXACT_OFFSETS_PER_PAGE);
SimpleLruInit(MultiXactMemberCtl,
- "MultiXactMember",
NUM_MULTIXACTMEMBER_BUFFERS, 0,
+ "MultiXactMember",
multixact_offsets_slru_buffers, 0, <------------------
MultiXactMemberSLRULock,
"pg_multixact/members",
LWTRANCHE_MULTIXACTMEMBER_BUFFER,
SYNC_HANDLER_MULTIXACT_MEMBER);
Please fix them so that I can end the review.
--
Gilles Darold
LzLabs GmbH
http://www.lzlabs.com/
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2021-03-24 21:31 Thomas Munro <[email protected]>
parent: Andrey Borodin <[email protected]>
1 sibling, 1 reply; 103+ messages in thread
From: Thomas Munro @ 2021-03-24 21:31 UTC (permalink / raw)
To: Andrey Borodin <[email protected]>; +Cc: Gilles Darold <[email protected]>; Tomas Vondra <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
Hi Andrey,
On Sat, Mar 13, 2021 at 1:44 AM Andrey Borodin <[email protected]> wrote:
> [v10]
+int multixact_offsets_slru_buffers = 8;
+int multixact_members_slru_buffers = 16;
+int subtrans_slru_buffers = 32;
+int notify_slru_buffers = 8;
+int serial_slru_buffers = 16;
+int clog_slru_buffers = 0;
+int commit_ts_slru_buffers = 0;
I don't think we should put "slru" (the name of the buffer replacement
algorithm, implementation detail) in the GUC names.
+ It defaults to 0, in this case CLOG size is taken as
<varname>shared_buffers</varname> / 512.
We already know that increasing the number of CLOG buffers above the
current number hurts as the linear search begins to dominate
(according to the commit message for 5364b357), and it doesn't seem
great to ship a new feature that melts your CPU when you turn it up.
Perhaps, to ship this, we need to introduce a buffer mapping table? I
have attached a "one coffee" attempt at that, on top of your v10 patch
(unmodified), for discussion. It survives basic testing but I don't
know how it performs.
Attachments:
[text/x-patch] v10-0001-Make-all-SLRU-buffer-sizes-configurable.patch (19.5K, ../../CA+hUKGJ4TwipnMDmKdORu4zUo_rofsuu5FQsgV=UHJ3bY38ERw@mail.gmail.com/2-v10-0001-Make-all-SLRU-buffer-sizes-configurable.patch)
download | inline diff:
From 4817d16cfb6704d43a7bef12648e753d239c809c Mon Sep 17 00:00:00 2001
From: Andrey Borodin <[email protected]>
Date: Mon, 15 Feb 2021 21:51:56 +0500
Subject: [PATCH v10 1/2] Make all SLRU buffer sizes configurable
---
doc/src/sgml/config.sgml | 108 ++++++++++++++++++
src/backend/access/transam/clog.c | 6 +
src/backend/access/transam/commit_ts.c | 5 +-
src/backend/access/transam/multixact.c | 8 +-
src/backend/access/transam/subtrans.c | 5 +-
src/backend/commands/async.c | 8 +-
src/backend/storage/lmgr/predicate.c | 4 +-
src/backend/utils/init/globals.c | 8 ++
src/backend/utils/misc/guc.c | 77 +++++++++++++
src/backend/utils/misc/postgresql.conf.sample | 16 +++
src/include/access/multixact.h | 4 -
src/include/access/subtrans.h | 3 -
src/include/commands/async.h | 5 -
src/include/miscadmin.h | 8 ++
src/include/storage/predicate.h | 4 -
15 files changed, 240 insertions(+), 29 deletions(-)
diff --git a/doc/src/sgml/config.sgml b/doc/src/sgml/config.sgml
index ddc6d789d8..0adcf0efaf 100644
--- a/doc/src/sgml/config.sgml
+++ b/doc/src/sgml/config.sgml
@@ -1886,6 +1886,114 @@ include_dir 'conf.d'
</para>
</listitem>
</varlistentry>
+
+ <varlistentry id="guc-multixact-offsets-slru-buffers" xreflabel="multixact_offsets_slru_buffers">
+ <term><varname>multixact_offsets_slru_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>multixact_offsets_slru_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of memory to be used for MultiXact offsets. MultiXact offsets
+ are used to store information about offsets of multiple row lockers (caused by SELECT FOR UPDATE and others).
+ It defaults to 64 kilobytes (<literal>64KB</literal>).
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-multixact-members-slru-buffers" xreflabel="multixact_members_slru_buffers">
+ <term><varname>multixact_members_slru_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>multixact_members_slru_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of memory to be used for MultiXact members. MultiXact members
+ are used to store information about XIDs of multiple row lockers. Typically <varname>multixact_members_slru_buffers</varname>
+ is twice more than <varname>multixact_offsets_slru_buffers</varname>.
+ It defaults to 128 kilobytes (<literal>128KB</literal>).
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-subtrans-buffers" xreflabel="subtrans_slru_buffers">
+ <term><varname>subtrans_slru_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>subtrans_slru_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of memory to be used for subtransactions.
+ It defaults to 256 kilobytes (<literal>256KB</literal>).
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-notify-buffers" xreflabel="notify_slru_buffers">
+ <term><varname>notify_slru_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>notify_slru_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of memory to be used for asyncronous notifications (NOTIFY, LISTEN).
+ It defaults to 64 kilobytes (<literal>64KB</literal>).
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-serial-buffers" xreflabel="serial_slru_buffers">
+ <term><varname>serial_slru_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>serial_slru_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of memory to be used for predicate locks.
+ It defaults to 128 kilobytes (<literal>128KB</literal>).
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-clog-buffers" xreflabel="clog_slru_buffers">
+ <term><varname>clog_slru_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>clog_slru_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of memory to be used for CLOG.
+ It defaults to 0, in this case CLOG size is taken as <varname>shared_buffers</varname> / 512.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-commit-ts-buffers" xreflabel="commit_ts_slru_buffers">
+ <term><varname>commit_ts_slru_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>commit_ts_slru_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of memory to be used for commit timestamps.
+ It defaults to 0, in this case CLOG size is taken as <varname>shared_buffers</varname> / 512.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
<varlistentry id="guc-max-stack-depth" xreflabel="max_stack_depth">
<term><varname>max_stack_depth</varname> (<type>integer</type>)
diff --git a/src/backend/access/transam/clog.c b/src/backend/access/transam/clog.c
index 6fa4713fb4..e1d34aa361 100644
--- a/src/backend/access/transam/clog.c
+++ b/src/backend/access/transam/clog.c
@@ -659,6 +659,9 @@ TransactionIdGetStatus(TransactionId xid, XLogRecPtr *lsn)
/*
* Number of shared CLOG buffers.
*
+ * If values is configured via GUC - just use given value. Otherwise
+ * apply following euristics.
+ *
* On larger multi-processor systems, it is possible to have many CLOG page
* requests in flight at one time which could lead to disk access for CLOG
* page if the required page is not found in memory. Testing revealed that we
@@ -675,6 +678,9 @@ TransactionIdGetStatus(TransactionId xid, XLogRecPtr *lsn)
Size
CLOGShmemBuffers(void)
{
+ /* consider 0 and 1 as unset GUC */
+ if (clog_slru_buffers > 1)
+ return clog_slru_buffers;
return Min(128, Max(4, NBuffers / 512));
}
diff --git a/src/backend/access/transam/commit_ts.c b/src/backend/access/transam/commit_ts.c
index 48e8d66286..7de3bca63d 100644
--- a/src/backend/access/transam/commit_ts.c
+++ b/src/backend/access/transam/commit_ts.c
@@ -530,7 +530,10 @@ pg_xact_commit_timestamp_origin(PG_FUNCTION_ARGS)
Size
CommitTsShmemBuffers(void)
{
- return Min(16, Max(4, NBuffers / 1024));
+ /* consider 0 and 1 as unset GUC */
+ if (commit_ts_slru_buffers > 1)
+ return commit_ts_slru_buffers;
+ return Min(16, Max(4, NBuffers / 512));
}
/*
diff --git a/src/backend/access/transam/multixact.c b/src/backend/access/transam/multixact.c
index 1f9f1a1fa1..370c01e72b 100644
--- a/src/backend/access/transam/multixact.c
+++ b/src/backend/access/transam/multixact.c
@@ -1831,8 +1831,8 @@ MultiXactShmemSize(void)
mul_size(sizeof(MultiXactId) * 2, MaxOldestSlot))
size = SHARED_MULTIXACT_STATE_SIZE;
- size = add_size(size, SimpleLruShmemSize(NUM_MULTIXACTOFFSET_BUFFERS, 0));
- size = add_size(size, SimpleLruShmemSize(NUM_MULTIXACTMEMBER_BUFFERS, 0));
+ size = add_size(size, SimpleLruShmemSize(multixact_offsets_slru_buffers, 0));
+ size = add_size(size, SimpleLruShmemSize(multixact_members_slru_buffers, 0));
return size;
}
@@ -1848,13 +1848,13 @@ MultiXactShmemInit(void)
MultiXactMemberCtl->PagePrecedes = MultiXactMemberPagePrecedes;
SimpleLruInit(MultiXactOffsetCtl,
- "MultiXactOffset", NUM_MULTIXACTOFFSET_BUFFERS, 0,
+ "MultiXactOffset", multixact_offsets_slru_buffers, 0,
MultiXactOffsetSLRULock, "pg_multixact/offsets",
LWTRANCHE_MULTIXACTOFFSET_BUFFER,
SYNC_HANDLER_MULTIXACT_OFFSET);
SlruPagePrecedesUnitTests(MultiXactOffsetCtl, MULTIXACT_OFFSETS_PER_PAGE);
SimpleLruInit(MultiXactMemberCtl,
- "MultiXactMember", NUM_MULTIXACTMEMBER_BUFFERS, 0,
+ "MultiXactMember", multixact_offsets_slru_buffers, 0,
MultiXactMemberSLRULock, "pg_multixact/members",
LWTRANCHE_MULTIXACTMEMBER_BUFFER,
SYNC_HANDLER_MULTIXACT_MEMBER);
diff --git a/src/backend/access/transam/subtrans.c b/src/backend/access/transam/subtrans.c
index 6a8e521f89..0c24353d3a 100644
--- a/src/backend/access/transam/subtrans.c
+++ b/src/backend/access/transam/subtrans.c
@@ -31,6 +31,7 @@
#include "access/slru.h"
#include "access/subtrans.h"
#include "access/transam.h"
+#include "miscadmin.h"
#include "pg_trace.h"
#include "utils/snapmgr.h"
@@ -184,14 +185,14 @@ SubTransGetTopmostTransaction(TransactionId xid)
Size
SUBTRANSShmemSize(void)
{
- return SimpleLruShmemSize(NUM_SUBTRANS_BUFFERS, 0);
+ return SimpleLruShmemSize(subtrans_slru_buffers, 0);
}
void
SUBTRANSShmemInit(void)
{
SubTransCtl->PagePrecedes = SubTransPagePrecedes;
- SimpleLruInit(SubTransCtl, "Subtrans", NUM_SUBTRANS_BUFFERS, 0,
+ SimpleLruInit(SubTransCtl, "Subtrans", subtrans_slru_buffers, 0,
SubtransSLRULock, "pg_subtrans",
LWTRANCHE_SUBTRANS_BUFFER, SYNC_HANDLER_NONE);
SlruPagePrecedesUnitTests(SubTransCtl, SUBTRANS_XACTS_PER_PAGE);
diff --git a/src/backend/commands/async.c b/src/backend/commands/async.c
index 4b16fb5682..f5c5592057 100644
--- a/src/backend/commands/async.c
+++ b/src/backend/commands/async.c
@@ -107,7 +107,7 @@
* frontend during startup.) The above design guarantees that notifies from
* other backends will never be missed by ignoring self-notifies.
*
- * The amount of shared memory used for notify management (NUM_NOTIFY_BUFFERS)
+ * The amount of shared memory used for notify management (notify_slru_buffers)
* can be varied without affecting anything but performance. The maximum
* amount of notification data that can be queued at one time is determined
* by slru.c's wraparound limit; see QUEUE_MAX_PAGE below.
@@ -225,7 +225,7 @@ typedef struct QueuePosition
*
* Resist the temptation to make this really large. While that would save
* work in some places, it would add cost in others. In particular, this
- * should likely be less than NUM_NOTIFY_BUFFERS, to ensure that backends
+ * should likely be less than notify_slru_buffers, to ensure that backends
* catch up before the pages they'll need to read fall out of SLRU cache.
*/
#define QUEUE_CLEANUP_DELAY 4
@@ -514,7 +514,7 @@ AsyncShmemSize(void)
size = mul_size(MaxBackends + 1, sizeof(QueueBackendStatus));
size = add_size(size, offsetof(AsyncQueueControl, backend));
- size = add_size(size, SimpleLruShmemSize(NUM_NOTIFY_BUFFERS, 0));
+ size = add_size(size, SimpleLruShmemSize(notify_slru_buffers, 0));
return size;
}
@@ -562,7 +562,7 @@ AsyncShmemInit(void)
* Set up SLRU management of the pg_notify data.
*/
NotifyCtl->PagePrecedes = asyncQueuePagePrecedes;
- SimpleLruInit(NotifyCtl, "Notify", NUM_NOTIFY_BUFFERS, 0,
+ SimpleLruInit(NotifyCtl, "Notify", notify_slru_buffers, 0,
NotifySLRULock, "pg_notify", LWTRANCHE_NOTIFY_BUFFER,
SYNC_HANDLER_NONE);
diff --git a/src/backend/storage/lmgr/predicate.c b/src/backend/storage/lmgr/predicate.c
index d493aeef0f..fad8cc572e 100644
--- a/src/backend/storage/lmgr/predicate.c
+++ b/src/backend/storage/lmgr/predicate.c
@@ -872,7 +872,7 @@ SerialInit(void)
*/
SerialSlruCtl->PagePrecedes = SerialPagePrecedesLogically;
SimpleLruInit(SerialSlruCtl, "Serial",
- NUM_SERIAL_BUFFERS, 0, SerialSLRULock, "pg_serial",
+ serial_slru_buffers, 0, SerialSLRULock, "pg_serial",
LWTRANCHE_SERIAL_BUFFER, SYNC_HANDLER_NONE);
#ifdef USE_ASSERT_CHECKING
SerialPagePrecedesLogicallyUnitTests();
@@ -1395,7 +1395,7 @@ PredicateLockShmemSize(void)
/* Shared memory structures for SLRU tracking of old committed xids. */
size = add_size(size, sizeof(SerialControlData));
- size = add_size(size, SimpleLruShmemSize(NUM_SERIAL_BUFFERS, 0));
+ size = add_size(size, SimpleLruShmemSize(serial_slru_buffers, 0));
return size;
}
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
index 73e0a672ae..f163ca17e9 100644
--- a/src/backend/utils/init/globals.c
+++ b/src/backend/utils/init/globals.c
@@ -148,3 +148,11 @@ int64 VacuumPageDirty = 0;
int VacuumCostBalance = 0; /* working state for vacuum */
bool VacuumCostActive = false;
+
+int multixact_offsets_slru_buffers = 8;
+int multixact_members_slru_buffers = 16;
+int subtrans_slru_buffers = 32;
+int notify_slru_buffers = 8;
+int serial_slru_buffers = 16;
+int clog_slru_buffers = 0;
+int commit_ts_slru_buffers = 0;
diff --git a/src/backend/utils/misc/guc.c b/src/backend/utils/misc/guc.c
index 0c5dc4d3e8..b65a4ae9ce 100644
--- a/src/backend/utils/misc/guc.c
+++ b/src/backend/utils/misc/guc.c
@@ -2305,6 +2305,83 @@ static struct config_int ConfigureNamesInt[] =
NULL, NULL, NULL
},
+ {
+ {"multixact_offsets_slru_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for MultiXact offsets SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &multixact_offsets_slru_buffers,
+ 8, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"multixact_members_slru_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for MultiXact members SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &multixact_members_slru_buffers,
+ 16, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"subtrans_slru_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for substransactions SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &subtrans_slru_buffers,
+ 32, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"notify_slru_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for asyncronous notifications SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ ¬ify_slru_buffers,
+ 8, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"serial_slru_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for predicate locks SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &serial_slru_buffers,
+ 16, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"clog_slru_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for commit log SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &serial_slru_buffers,
+ 0, 0, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"commit_ts_slru_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for commit timestamps SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &commit_ts_slru_buffers,
+ 0, 0, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
{
{"temp_buffers", PGC_USERSET, RESOURCES_MEM,
gettext_noop("Sets the maximum number of temporary buffers used by each session."),
diff --git a/src/backend/utils/misc/postgresql.conf.sample b/src/backend/utils/misc/postgresql.conf.sample
index b234a6bfe6..308fd565d3 100644
--- a/src/backend/utils/misc/postgresql.conf.sample
+++ b/src/backend/utils/misc/postgresql.conf.sample
@@ -190,6 +190,22 @@
# (change requires restart)
#backend_flush_after = 0 # measured in pages, 0 disables
+# - SLRU Buffers -
+
+#multixact_offsets_slru_buffers = 8 # memory used for MultiXact offsets
+ # (change requires restart)
+#multixact_members_slru_buffers = 16 # memory used for MultiXact members
+ # (change requires restart)
+#subtrans_slru_buffers = 32 # memory used for subtransactions
+ # (change requires restart)
+#notify_slru_buffers = 8 # memory used for asynchronous notifications
+ # (change requires restart)
+#serial_slru_buffers = 16 # memory used for predicate locks
+ # (change requires restart)
+#clog_slru_buffers = 0 # memory used for CLOG
+ # (change requires restart)
+#commit_ts_slru_buffers = 0 # memory used for commit timestamps
+ # (change requires restart)
#------------------------------------------------------------------------------
# WRITE-AHEAD LOG
diff --git a/src/include/access/multixact.h b/src/include/access/multixact.h
index 4bbb035eae..97c0a46376 100644
--- a/src/include/access/multixact.h
+++ b/src/include/access/multixact.h
@@ -29,10 +29,6 @@
#define MaxMultiXactOffset ((MultiXactOffset) 0xFFFFFFFF)
-/* Number of SLRU buffers to use for multixact */
-#define NUM_MULTIXACTOFFSET_BUFFERS 8
-#define NUM_MULTIXACTMEMBER_BUFFERS 16
-
/*
* Possible multixact lock modes ("status"). The first four modes are for
* tuple locks (FOR KEY SHARE, FOR SHARE, FOR NO KEY UPDATE, FOR UPDATE); the
diff --git a/src/include/access/subtrans.h b/src/include/access/subtrans.h
index d0ab44ae82..ca0999056e 100644
--- a/src/include/access/subtrans.h
+++ b/src/include/access/subtrans.h
@@ -11,9 +11,6 @@
#ifndef SUBTRANS_H
#define SUBTRANS_H
-/* Number of SLRU buffers to use for subtrans */
-#define NUM_SUBTRANS_BUFFERS 32
-
extern void SubTransSetParent(TransactionId xid, TransactionId parent);
extern TransactionId SubTransGetParent(TransactionId xid);
extern TransactionId SubTransGetTopmostTransaction(TransactionId xid);
diff --git a/src/include/commands/async.h b/src/include/commands/async.h
index 9217f66b91..fa831e3721 100644
--- a/src/include/commands/async.h
+++ b/src/include/commands/async.h
@@ -15,11 +15,6 @@
#include <signal.h>
-/*
- * The number of SLRU page buffers we use for the notification queue.
- */
-#define NUM_NOTIFY_BUFFERS 8
-
extern bool Trace_notify;
extern volatile sig_atomic_t notifyInterruptPending;
diff --git a/src/include/miscadmin.h b/src/include/miscadmin.h
index 013850ac28..3d9f585fb9 100644
--- a/src/include/miscadmin.h
+++ b/src/include/miscadmin.h
@@ -162,6 +162,14 @@ extern PGDLLIMPORT int MaxBackends;
extern PGDLLIMPORT int MaxConnections;
extern PGDLLIMPORT int max_worker_processes;
extern PGDLLIMPORT int max_parallel_workers;
+extern PGDLLIMPORT int multixact_offsets_slru_buffers;
+extern PGDLLIMPORT int multixact_members_slru_buffers;
+extern PGDLLIMPORT int multixact_members_slru_buffers;
+extern PGDLLIMPORT int subtrans_slru_buffers;
+extern PGDLLIMPORT int notify_slru_buffers;
+extern PGDLLIMPORT int serial_slru_buffers;
+extern PGDLLIMPORT int clog_slru_buffers;
+extern PGDLLIMPORT int commit_ts_slru_buffers;
extern PGDLLIMPORT int MyProcPid;
extern PGDLLIMPORT pg_time_t MyStartTime;
diff --git a/src/include/storage/predicate.h b/src/include/storage/predicate.h
index 152b698611..c72779bd88 100644
--- a/src/include/storage/predicate.h
+++ b/src/include/storage/predicate.h
@@ -26,10 +26,6 @@ extern int max_predicate_locks_per_xact;
extern int max_predicate_locks_per_relation;
extern int max_predicate_locks_per_page;
-
-/* Number of SLRU buffers to use for Serial SLRU */
-#define NUM_SERIAL_BUFFERS 16
-
/*
* A handle used for sharing SERIALIZABLEXACT objects between the participants
* in a parallel query.
--
2.30.1
[text/x-patch] v10-0002-Add-buffer-mapping-table-for-SLRUs.patch (7.0K, ../../CA+hUKGJ4TwipnMDmKdORu4zUo_rofsuu5FQsgV=UHJ3bY38ERw@mail.gmail.com/3-v10-0002-Add-buffer-mapping-table-for-SLRUs.patch)
download | inline diff:
From 692d03167f4124e6daabdb8d4a72ff0b494efe96 Mon Sep 17 00:00:00 2001
From: Thomas Munro <[email protected]>
Date: Thu, 25 Mar 2021 10:11:31 +1300
Subject: [PATCH v10 2/2] Add buffer mapping table for SLRUs.
---
src/backend/access/transam/slru.c | 87 ++++++++++++++++++++++++++++---
src/include/access/slru.h | 2 +
2 files changed, 83 insertions(+), 6 deletions(-)
diff --git a/src/backend/access/transam/slru.c b/src/backend/access/transam/slru.c
index 82149ad782..2d6e84266e 100644
--- a/src/backend/access/transam/slru.c
+++ b/src/backend/access/transam/slru.c
@@ -58,6 +58,8 @@
#include "pgstat.h"
#include "storage/fd.h"
#include "storage/shmem.h"
+#include "utils/dynahash.h"
+#include "utils/hsearch.h"
#define SlruFileName(ctl, path, seg) \
snprintf(path, MAXPGPATH, "%s/%04X", (ctl)->Dir, seg)
@@ -79,6 +81,12 @@ typedef struct SlruWriteAllData
typedef struct SlruWriteAllData *SlruWriteAll;
+typedef struct SlruMappingTableEntry
+{
+ int pageno;
+ int slotno;
+} SlruMappingTableEntry;
+
/*
* Populate a file tag describing a segment file. We only use the segment
* number, since we can derive everything else we need by having separate
@@ -146,6 +154,9 @@ static int SlruSelectLRUPage(SlruCtl ctl, int pageno);
static bool SlruScanDirCbDeleteCutoff(SlruCtl ctl, char *filename,
int segpage, void *data);
static void SlruInternalDeleteSegment(SlruCtl ctl, int segno);
+static void SlruMappingAdd(SlruCtl ctl, int pageno, int slotno);
+static void SlruMappingRemove(SlruCtl ctl, int pageno);
+static int SlruMappingFind(SlruCtl ctl, int pageno);
/*
* Initialization of shared memory
@@ -168,7 +179,8 @@ SimpleLruShmemSize(int nslots, int nlsns)
if (nlsns > 0)
sz += MAXALIGN(nslots * nlsns * sizeof(XLogRecPtr)); /* group_lsn[] */
- return BUFFERALIGN(sz) + BLCKSZ * nslots;
+ return BUFFERALIGN(sz) + BLCKSZ * nslots +
+ hash_estimate_size(nslots, sizeof(SlruMappingTableEntry));
}
/*
@@ -187,6 +199,9 @@ SimpleLruInit(SlruCtl ctl, const char *name, int nslots, int nlsns,
LWLock *ctllock, const char *subdir, int tranche_id,
SyncRequestHandler sync_handler)
{
+ char mapping_table_name[SHMEM_INDEX_KEYSIZE];
+ HASHCTL mapping_table_info;
+ HTAB *mapping_table;
SlruShared shared;
bool found;
@@ -258,11 +273,21 @@ SimpleLruInit(SlruCtl ctl, const char *name, int nslots, int nlsns,
else
Assert(found);
+ /* Create or find the buffer mapping table. */
+ memset(&mapping_table_info, 0, sizeof(mapping_table_info));
+ mapping_table_info.keysize = sizeof(int);
+ mapping_table_info.entrysize = sizeof(SlruMappingTableEntry);
+ snprintf(mapping_table_name, sizeof(mapping_table_name),
+ "%s Mapping Table", name);
+ mapping_table = ShmemInitHash(mapping_table_name, nslots, nslots,
+ &mapping_table_info, HASH_ELEM | HASH_BLOBS);
+
/*
* Initialize the unshared control struct, including directory path. We
* assume caller set PagePrecedes.
*/
ctl->shared = shared;
+ ctl->mapping_table = mapping_table;
ctl->sync_handler = sync_handler;
strlcpy(ctl->Dir, subdir, sizeof(ctl->Dir));
}
@@ -289,6 +314,9 @@ SimpleLruZeroPage(SlruCtl ctl, int pageno)
shared->page_number[slotno] == pageno);
/* Mark the slot as containing this page */
+ if (shared->page_status[slotno] == SLRU_PAGE_VALID)
+ SlruMappingRemove(ctl, shared->page_number[slotno]);
+ SlruMappingAdd(ctl, pageno, slotno);
shared->page_number[slotno] = pageno;
shared->page_status[slotno] = SLRU_PAGE_VALID;
shared->page_dirty[slotno] = true;
@@ -362,7 +390,10 @@ SimpleLruWaitIO(SlruCtl ctl, int slotno)
{
/* indeed, the I/O must have failed */
if (shared->page_status[slotno] == SLRU_PAGE_READ_IN_PROGRESS)
+ {
+ SlruMappingRemove(ctl, shared->page_number[slotno]);
shared->page_status[slotno] = SLRU_PAGE_EMPTY;
+ }
else /* write_in_progress */
{
shared->page_status[slotno] = SLRU_PAGE_VALID;
@@ -436,6 +467,7 @@ SimpleLruReadPage(SlruCtl ctl, int pageno, bool write_ok,
!shared->page_dirty[slotno]));
/* Mark the slot read-busy */
+ SlruMappingAdd(ctl, pageno, slotno);
shared->page_number[slotno] = pageno;
shared->page_status[slotno] = SLRU_PAGE_READ_IN_PROGRESS;
shared->page_dirty[slotno] = false;
@@ -459,7 +491,13 @@ SimpleLruReadPage(SlruCtl ctl, int pageno, bool write_ok,
shared->page_status[slotno] == SLRU_PAGE_READ_IN_PROGRESS &&
!shared->page_dirty[slotno]);
- shared->page_status[slotno] = ok ? SLRU_PAGE_VALID : SLRU_PAGE_EMPTY;
+ if (ok)
+ shared->page_status[slotno] = SLRU_PAGE_VALID;
+ else
+ {
+ SlruMappingRemove(ctl, pageno);
+ shared->page_status[slotno] = SLRU_PAGE_EMPTY;
+ }
LWLockRelease(&shared->buffer_locks[slotno].lock);
@@ -1029,11 +1067,12 @@ SlruSelectLRUPage(SlruCtl ctl, int pageno)
int best_invalid_page_number = 0; /* keep compiler quiet */
/* See if page already has a buffer assigned */
- for (slotno = 0; slotno < shared->num_slots; slotno++)
+ slotno = SlruMappingFind(ctl, pageno);
+ if (slotno >= 0)
{
- if (shared->page_number[slotno] == pageno &&
- shared->page_status[slotno] != SLRU_PAGE_EMPTY)
- return slotno;
+ Assert(shared->page_number[slotno] == pageno);
+ Assert(shared->page_status[slotno] != SLRU_PAGE_EMPTY);
+ return slotno;
}
/*
@@ -1266,6 +1305,7 @@ restart:;
if (shared->page_status[slotno] == SLRU_PAGE_VALID &&
!shared->page_dirty[slotno])
{
+ SlruMappingRemove(ctl, shared->page_number[slotno]);
shared->page_status[slotno] = SLRU_PAGE_EMPTY;
continue;
}
@@ -1348,6 +1388,7 @@ restart:
if (shared->page_status[slotno] == SLRU_PAGE_VALID &&
!shared->page_dirty[slotno])
{
+ SlruMappingRemove(ctl, shared->page_number[slotno]);
shared->page_status[slotno] = SLRU_PAGE_EMPTY;
continue;
}
@@ -1609,3 +1650,37 @@ SlruSyncFileTag(SlruCtl ctl, const FileTag *ftag, char *path)
errno = save_errno;
return result;
}
+
+static int
+SlruMappingFind(SlruCtl ctl, int pageno)
+{
+ SlruMappingTableEntry *mapping;
+
+ mapping = hash_search(ctl->mapping_table, &pageno, HASH_FIND, NULL);
+ if (mapping)
+ return mapping->pageno;
+
+ return -1;
+}
+
+static void
+SlruMappingAdd(SlruCtl ctl, int pageno, int slotno)
+{
+ SlruMappingTableEntry *mapping;
+ bool found PG_USED_FOR_ASSERTS_ONLY;
+
+ mapping = hash_search(ctl->mapping_table, &pageno, HASH_ENTER, &found);
+ mapping->slotno = slotno;
+
+ Assert(!found);
+}
+
+static void
+SlruMappingRemove(SlruCtl ctl, int pageno)
+{
+ bool found PG_USED_FOR_ASSERTS_ONLY;
+
+ hash_search(ctl->mapping_table, &pageno, HASH_REMOVE, &found);
+
+ Assert(found);
+}
diff --git a/src/include/access/slru.h b/src/include/access/slru.h
index dd52e8cec7..8aa3efc0ee 100644
--- a/src/include/access/slru.h
+++ b/src/include/access/slru.h
@@ -16,6 +16,7 @@
#include "access/xlogdefs.h"
#include "storage/lwlock.h"
#include "storage/sync.h"
+#include "utils/hsearch.h"
/*
@@ -110,6 +111,7 @@ typedef SlruSharedData *SlruShared;
typedef struct SlruCtlData
{
SlruShared shared;
+ HTAB *mapping_table;
/*
* Which sync handler function to use when handing sync requests over to
--
2.30.1
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2021-03-25 01:03 Thomas Munro <[email protected]>
parent: Thomas Munro <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Thomas Munro @ 2021-03-25 01:03 UTC (permalink / raw)
To: Andrey Borodin <[email protected]>; +Cc: Gilles Darold <[email protected]>; Tomas Vondra <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
On Thu, Mar 25, 2021 at 10:31 AM Thomas Munro <[email protected]> wrote:
> We already know that increasing the number of CLOG buffers above the
> current number hurts as the linear search begins to dominate
> (according to the commit message for 5364b357), and it doesn't seem
> great to ship a new feature that melts your CPU when you turn it up.
> Perhaps, to ship this, we need to introduce a buffer mapping table? I
> have attached a "one coffee" attempt at that, on top of your v10 patch
> (unmodified), for discussion. It survives basic testing but I don't
> know how it performs.
Hrrr... Cfbot showed an assertion failure. Here's the two coffee
version with a couple of silly mistakes fixed.
Attachments:
[text/x-patch] v11-0001-Make-all-SLRU-buffer-sizes-configurable.patch (19.5K, ../../CA+hUKGK2bkx91=CWRuPqWdXAu2JSS0mDPBh8DNuSeKhy8eOLgg@mail.gmail.com/2-v11-0001-Make-all-SLRU-buffer-sizes-configurable.patch)
download | inline diff:
From 4817d16cfb6704d43a7bef12648e753d239c809c Mon Sep 17 00:00:00 2001
From: Andrey Borodin <[email protected]>
Date: Mon, 15 Feb 2021 21:51:56 +0500
Subject: [PATCH v11 1/2] Make all SLRU buffer sizes configurable
---
doc/src/sgml/config.sgml | 108 ++++++++++++++++++
src/backend/access/transam/clog.c | 6 +
src/backend/access/transam/commit_ts.c | 5 +-
src/backend/access/transam/multixact.c | 8 +-
src/backend/access/transam/subtrans.c | 5 +-
src/backend/commands/async.c | 8 +-
src/backend/storage/lmgr/predicate.c | 4 +-
src/backend/utils/init/globals.c | 8 ++
src/backend/utils/misc/guc.c | 77 +++++++++++++
src/backend/utils/misc/postgresql.conf.sample | 16 +++
src/include/access/multixact.h | 4 -
src/include/access/subtrans.h | 3 -
src/include/commands/async.h | 5 -
src/include/miscadmin.h | 8 ++
src/include/storage/predicate.h | 4 -
15 files changed, 240 insertions(+), 29 deletions(-)
diff --git a/doc/src/sgml/config.sgml b/doc/src/sgml/config.sgml
index ddc6d789d8..0adcf0efaf 100644
--- a/doc/src/sgml/config.sgml
+++ b/doc/src/sgml/config.sgml
@@ -1886,6 +1886,114 @@ include_dir 'conf.d'
</para>
</listitem>
</varlistentry>
+
+ <varlistentry id="guc-multixact-offsets-slru-buffers" xreflabel="multixact_offsets_slru_buffers">
+ <term><varname>multixact_offsets_slru_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>multixact_offsets_slru_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of memory to be used for MultiXact offsets. MultiXact offsets
+ are used to store information about offsets of multiple row lockers (caused by SELECT FOR UPDATE and others).
+ It defaults to 64 kilobytes (<literal>64KB</literal>).
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-multixact-members-slru-buffers" xreflabel="multixact_members_slru_buffers">
+ <term><varname>multixact_members_slru_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>multixact_members_slru_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of memory to be used for MultiXact members. MultiXact members
+ are used to store information about XIDs of multiple row lockers. Typically <varname>multixact_members_slru_buffers</varname>
+ is twice more than <varname>multixact_offsets_slru_buffers</varname>.
+ It defaults to 128 kilobytes (<literal>128KB</literal>).
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-subtrans-buffers" xreflabel="subtrans_slru_buffers">
+ <term><varname>subtrans_slru_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>subtrans_slru_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of memory to be used for subtransactions.
+ It defaults to 256 kilobytes (<literal>256KB</literal>).
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-notify-buffers" xreflabel="notify_slru_buffers">
+ <term><varname>notify_slru_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>notify_slru_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of memory to be used for asyncronous notifications (NOTIFY, LISTEN).
+ It defaults to 64 kilobytes (<literal>64KB</literal>).
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-serial-buffers" xreflabel="serial_slru_buffers">
+ <term><varname>serial_slru_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>serial_slru_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of memory to be used for predicate locks.
+ It defaults to 128 kilobytes (<literal>128KB</literal>).
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-clog-buffers" xreflabel="clog_slru_buffers">
+ <term><varname>clog_slru_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>clog_slru_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of memory to be used for CLOG.
+ It defaults to 0, in this case CLOG size is taken as <varname>shared_buffers</varname> / 512.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-commit-ts-buffers" xreflabel="commit_ts_slru_buffers">
+ <term><varname>commit_ts_slru_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>commit_ts_slru_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of memory to be used for commit timestamps.
+ It defaults to 0, in this case CLOG size is taken as <varname>shared_buffers</varname> / 512.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
<varlistentry id="guc-max-stack-depth" xreflabel="max_stack_depth">
<term><varname>max_stack_depth</varname> (<type>integer</type>)
diff --git a/src/backend/access/transam/clog.c b/src/backend/access/transam/clog.c
index 6fa4713fb4..e1d34aa361 100644
--- a/src/backend/access/transam/clog.c
+++ b/src/backend/access/transam/clog.c
@@ -659,6 +659,9 @@ TransactionIdGetStatus(TransactionId xid, XLogRecPtr *lsn)
/*
* Number of shared CLOG buffers.
*
+ * If values is configured via GUC - just use given value. Otherwise
+ * apply following euristics.
+ *
* On larger multi-processor systems, it is possible to have many CLOG page
* requests in flight at one time which could lead to disk access for CLOG
* page if the required page is not found in memory. Testing revealed that we
@@ -675,6 +678,9 @@ TransactionIdGetStatus(TransactionId xid, XLogRecPtr *lsn)
Size
CLOGShmemBuffers(void)
{
+ /* consider 0 and 1 as unset GUC */
+ if (clog_slru_buffers > 1)
+ return clog_slru_buffers;
return Min(128, Max(4, NBuffers / 512));
}
diff --git a/src/backend/access/transam/commit_ts.c b/src/backend/access/transam/commit_ts.c
index 48e8d66286..7de3bca63d 100644
--- a/src/backend/access/transam/commit_ts.c
+++ b/src/backend/access/transam/commit_ts.c
@@ -530,7 +530,10 @@ pg_xact_commit_timestamp_origin(PG_FUNCTION_ARGS)
Size
CommitTsShmemBuffers(void)
{
- return Min(16, Max(4, NBuffers / 1024));
+ /* consider 0 and 1 as unset GUC */
+ if (commit_ts_slru_buffers > 1)
+ return commit_ts_slru_buffers;
+ return Min(16, Max(4, NBuffers / 512));
}
/*
diff --git a/src/backend/access/transam/multixact.c b/src/backend/access/transam/multixact.c
index 1f9f1a1fa1..370c01e72b 100644
--- a/src/backend/access/transam/multixact.c
+++ b/src/backend/access/transam/multixact.c
@@ -1831,8 +1831,8 @@ MultiXactShmemSize(void)
mul_size(sizeof(MultiXactId) * 2, MaxOldestSlot))
size = SHARED_MULTIXACT_STATE_SIZE;
- size = add_size(size, SimpleLruShmemSize(NUM_MULTIXACTOFFSET_BUFFERS, 0));
- size = add_size(size, SimpleLruShmemSize(NUM_MULTIXACTMEMBER_BUFFERS, 0));
+ size = add_size(size, SimpleLruShmemSize(multixact_offsets_slru_buffers, 0));
+ size = add_size(size, SimpleLruShmemSize(multixact_members_slru_buffers, 0));
return size;
}
@@ -1848,13 +1848,13 @@ MultiXactShmemInit(void)
MultiXactMemberCtl->PagePrecedes = MultiXactMemberPagePrecedes;
SimpleLruInit(MultiXactOffsetCtl,
- "MultiXactOffset", NUM_MULTIXACTOFFSET_BUFFERS, 0,
+ "MultiXactOffset", multixact_offsets_slru_buffers, 0,
MultiXactOffsetSLRULock, "pg_multixact/offsets",
LWTRANCHE_MULTIXACTOFFSET_BUFFER,
SYNC_HANDLER_MULTIXACT_OFFSET);
SlruPagePrecedesUnitTests(MultiXactOffsetCtl, MULTIXACT_OFFSETS_PER_PAGE);
SimpleLruInit(MultiXactMemberCtl,
- "MultiXactMember", NUM_MULTIXACTMEMBER_BUFFERS, 0,
+ "MultiXactMember", multixact_offsets_slru_buffers, 0,
MultiXactMemberSLRULock, "pg_multixact/members",
LWTRANCHE_MULTIXACTMEMBER_BUFFER,
SYNC_HANDLER_MULTIXACT_MEMBER);
diff --git a/src/backend/access/transam/subtrans.c b/src/backend/access/transam/subtrans.c
index 6a8e521f89..0c24353d3a 100644
--- a/src/backend/access/transam/subtrans.c
+++ b/src/backend/access/transam/subtrans.c
@@ -31,6 +31,7 @@
#include "access/slru.h"
#include "access/subtrans.h"
#include "access/transam.h"
+#include "miscadmin.h"
#include "pg_trace.h"
#include "utils/snapmgr.h"
@@ -184,14 +185,14 @@ SubTransGetTopmostTransaction(TransactionId xid)
Size
SUBTRANSShmemSize(void)
{
- return SimpleLruShmemSize(NUM_SUBTRANS_BUFFERS, 0);
+ return SimpleLruShmemSize(subtrans_slru_buffers, 0);
}
void
SUBTRANSShmemInit(void)
{
SubTransCtl->PagePrecedes = SubTransPagePrecedes;
- SimpleLruInit(SubTransCtl, "Subtrans", NUM_SUBTRANS_BUFFERS, 0,
+ SimpleLruInit(SubTransCtl, "Subtrans", subtrans_slru_buffers, 0,
SubtransSLRULock, "pg_subtrans",
LWTRANCHE_SUBTRANS_BUFFER, SYNC_HANDLER_NONE);
SlruPagePrecedesUnitTests(SubTransCtl, SUBTRANS_XACTS_PER_PAGE);
diff --git a/src/backend/commands/async.c b/src/backend/commands/async.c
index 4b16fb5682..f5c5592057 100644
--- a/src/backend/commands/async.c
+++ b/src/backend/commands/async.c
@@ -107,7 +107,7 @@
* frontend during startup.) The above design guarantees that notifies from
* other backends will never be missed by ignoring self-notifies.
*
- * The amount of shared memory used for notify management (NUM_NOTIFY_BUFFERS)
+ * The amount of shared memory used for notify management (notify_slru_buffers)
* can be varied without affecting anything but performance. The maximum
* amount of notification data that can be queued at one time is determined
* by slru.c's wraparound limit; see QUEUE_MAX_PAGE below.
@@ -225,7 +225,7 @@ typedef struct QueuePosition
*
* Resist the temptation to make this really large. While that would save
* work in some places, it would add cost in others. In particular, this
- * should likely be less than NUM_NOTIFY_BUFFERS, to ensure that backends
+ * should likely be less than notify_slru_buffers, to ensure that backends
* catch up before the pages they'll need to read fall out of SLRU cache.
*/
#define QUEUE_CLEANUP_DELAY 4
@@ -514,7 +514,7 @@ AsyncShmemSize(void)
size = mul_size(MaxBackends + 1, sizeof(QueueBackendStatus));
size = add_size(size, offsetof(AsyncQueueControl, backend));
- size = add_size(size, SimpleLruShmemSize(NUM_NOTIFY_BUFFERS, 0));
+ size = add_size(size, SimpleLruShmemSize(notify_slru_buffers, 0));
return size;
}
@@ -562,7 +562,7 @@ AsyncShmemInit(void)
* Set up SLRU management of the pg_notify data.
*/
NotifyCtl->PagePrecedes = asyncQueuePagePrecedes;
- SimpleLruInit(NotifyCtl, "Notify", NUM_NOTIFY_BUFFERS, 0,
+ SimpleLruInit(NotifyCtl, "Notify", notify_slru_buffers, 0,
NotifySLRULock, "pg_notify", LWTRANCHE_NOTIFY_BUFFER,
SYNC_HANDLER_NONE);
diff --git a/src/backend/storage/lmgr/predicate.c b/src/backend/storage/lmgr/predicate.c
index d493aeef0f..fad8cc572e 100644
--- a/src/backend/storage/lmgr/predicate.c
+++ b/src/backend/storage/lmgr/predicate.c
@@ -872,7 +872,7 @@ SerialInit(void)
*/
SerialSlruCtl->PagePrecedes = SerialPagePrecedesLogically;
SimpleLruInit(SerialSlruCtl, "Serial",
- NUM_SERIAL_BUFFERS, 0, SerialSLRULock, "pg_serial",
+ serial_slru_buffers, 0, SerialSLRULock, "pg_serial",
LWTRANCHE_SERIAL_BUFFER, SYNC_HANDLER_NONE);
#ifdef USE_ASSERT_CHECKING
SerialPagePrecedesLogicallyUnitTests();
@@ -1395,7 +1395,7 @@ PredicateLockShmemSize(void)
/* Shared memory structures for SLRU tracking of old committed xids. */
size = add_size(size, sizeof(SerialControlData));
- size = add_size(size, SimpleLruShmemSize(NUM_SERIAL_BUFFERS, 0));
+ size = add_size(size, SimpleLruShmemSize(serial_slru_buffers, 0));
return size;
}
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
index 73e0a672ae..f163ca17e9 100644
--- a/src/backend/utils/init/globals.c
+++ b/src/backend/utils/init/globals.c
@@ -148,3 +148,11 @@ int64 VacuumPageDirty = 0;
int VacuumCostBalance = 0; /* working state for vacuum */
bool VacuumCostActive = false;
+
+int multixact_offsets_slru_buffers = 8;
+int multixact_members_slru_buffers = 16;
+int subtrans_slru_buffers = 32;
+int notify_slru_buffers = 8;
+int serial_slru_buffers = 16;
+int clog_slru_buffers = 0;
+int commit_ts_slru_buffers = 0;
diff --git a/src/backend/utils/misc/guc.c b/src/backend/utils/misc/guc.c
index 0c5dc4d3e8..b65a4ae9ce 100644
--- a/src/backend/utils/misc/guc.c
+++ b/src/backend/utils/misc/guc.c
@@ -2305,6 +2305,83 @@ static struct config_int ConfigureNamesInt[] =
NULL, NULL, NULL
},
+ {
+ {"multixact_offsets_slru_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for MultiXact offsets SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &multixact_offsets_slru_buffers,
+ 8, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"multixact_members_slru_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for MultiXact members SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &multixact_members_slru_buffers,
+ 16, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"subtrans_slru_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for substransactions SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &subtrans_slru_buffers,
+ 32, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"notify_slru_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for asyncronous notifications SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ ¬ify_slru_buffers,
+ 8, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"serial_slru_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for predicate locks SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &serial_slru_buffers,
+ 16, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"clog_slru_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for commit log SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &serial_slru_buffers,
+ 0, 0, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"commit_ts_slru_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for commit timestamps SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &commit_ts_slru_buffers,
+ 0, 0, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
{
{"temp_buffers", PGC_USERSET, RESOURCES_MEM,
gettext_noop("Sets the maximum number of temporary buffers used by each session."),
diff --git a/src/backend/utils/misc/postgresql.conf.sample b/src/backend/utils/misc/postgresql.conf.sample
index b234a6bfe6..308fd565d3 100644
--- a/src/backend/utils/misc/postgresql.conf.sample
+++ b/src/backend/utils/misc/postgresql.conf.sample
@@ -190,6 +190,22 @@
# (change requires restart)
#backend_flush_after = 0 # measured in pages, 0 disables
+# - SLRU Buffers -
+
+#multixact_offsets_slru_buffers = 8 # memory used for MultiXact offsets
+ # (change requires restart)
+#multixact_members_slru_buffers = 16 # memory used for MultiXact members
+ # (change requires restart)
+#subtrans_slru_buffers = 32 # memory used for subtransactions
+ # (change requires restart)
+#notify_slru_buffers = 8 # memory used for asynchronous notifications
+ # (change requires restart)
+#serial_slru_buffers = 16 # memory used for predicate locks
+ # (change requires restart)
+#clog_slru_buffers = 0 # memory used for CLOG
+ # (change requires restart)
+#commit_ts_slru_buffers = 0 # memory used for commit timestamps
+ # (change requires restart)
#------------------------------------------------------------------------------
# WRITE-AHEAD LOG
diff --git a/src/include/access/multixact.h b/src/include/access/multixact.h
index 4bbb035eae..97c0a46376 100644
--- a/src/include/access/multixact.h
+++ b/src/include/access/multixact.h
@@ -29,10 +29,6 @@
#define MaxMultiXactOffset ((MultiXactOffset) 0xFFFFFFFF)
-/* Number of SLRU buffers to use for multixact */
-#define NUM_MULTIXACTOFFSET_BUFFERS 8
-#define NUM_MULTIXACTMEMBER_BUFFERS 16
-
/*
* Possible multixact lock modes ("status"). The first four modes are for
* tuple locks (FOR KEY SHARE, FOR SHARE, FOR NO KEY UPDATE, FOR UPDATE); the
diff --git a/src/include/access/subtrans.h b/src/include/access/subtrans.h
index d0ab44ae82..ca0999056e 100644
--- a/src/include/access/subtrans.h
+++ b/src/include/access/subtrans.h
@@ -11,9 +11,6 @@
#ifndef SUBTRANS_H
#define SUBTRANS_H
-/* Number of SLRU buffers to use for subtrans */
-#define NUM_SUBTRANS_BUFFERS 32
-
extern void SubTransSetParent(TransactionId xid, TransactionId parent);
extern TransactionId SubTransGetParent(TransactionId xid);
extern TransactionId SubTransGetTopmostTransaction(TransactionId xid);
diff --git a/src/include/commands/async.h b/src/include/commands/async.h
index 9217f66b91..fa831e3721 100644
--- a/src/include/commands/async.h
+++ b/src/include/commands/async.h
@@ -15,11 +15,6 @@
#include <signal.h>
-/*
- * The number of SLRU page buffers we use for the notification queue.
- */
-#define NUM_NOTIFY_BUFFERS 8
-
extern bool Trace_notify;
extern volatile sig_atomic_t notifyInterruptPending;
diff --git a/src/include/miscadmin.h b/src/include/miscadmin.h
index 013850ac28..3d9f585fb9 100644
--- a/src/include/miscadmin.h
+++ b/src/include/miscadmin.h
@@ -162,6 +162,14 @@ extern PGDLLIMPORT int MaxBackends;
extern PGDLLIMPORT int MaxConnections;
extern PGDLLIMPORT int max_worker_processes;
extern PGDLLIMPORT int max_parallel_workers;
+extern PGDLLIMPORT int multixact_offsets_slru_buffers;
+extern PGDLLIMPORT int multixact_members_slru_buffers;
+extern PGDLLIMPORT int multixact_members_slru_buffers;
+extern PGDLLIMPORT int subtrans_slru_buffers;
+extern PGDLLIMPORT int notify_slru_buffers;
+extern PGDLLIMPORT int serial_slru_buffers;
+extern PGDLLIMPORT int clog_slru_buffers;
+extern PGDLLIMPORT int commit_ts_slru_buffers;
extern PGDLLIMPORT int MyProcPid;
extern PGDLLIMPORT pg_time_t MyStartTime;
diff --git a/src/include/storage/predicate.h b/src/include/storage/predicate.h
index 152b698611..c72779bd88 100644
--- a/src/include/storage/predicate.h
+++ b/src/include/storage/predicate.h
@@ -26,10 +26,6 @@ extern int max_predicate_locks_per_xact;
extern int max_predicate_locks_per_relation;
extern int max_predicate_locks_per_page;
-
-/* Number of SLRU buffers to use for Serial SLRU */
-#define NUM_SERIAL_BUFFERS 16
-
/*
* A handle used for sharing SERIALIZABLEXACT objects between the participants
* in a parallel query.
--
2.30.1
[text/x-patch] v11-0002-Add-buffer-mapping-table-for-SLRUs.patch (7.0K, ../../CA+hUKGK2bkx91=CWRuPqWdXAu2JSS0mDPBh8DNuSeKhy8eOLgg@mail.gmail.com/3-v11-0002-Add-buffer-mapping-table-for-SLRUs.patch)
download | inline diff:
From 65600b53939c34abf43e62f3f59be5671c43d301 Mon Sep 17 00:00:00 2001
From: Thomas Munro <[email protected]>
Date: Thu, 25 Mar 2021 10:11:31 +1300
Subject: [PATCH v11 2/2] Add buffer mapping table for SLRUs.
---
src/backend/access/transam/slru.c | 87 ++++++++++++++++++++++++++++---
src/include/access/slru.h | 2 +
2 files changed, 83 insertions(+), 6 deletions(-)
diff --git a/src/backend/access/transam/slru.c b/src/backend/access/transam/slru.c
index 82149ad782..487585bb60 100644
--- a/src/backend/access/transam/slru.c
+++ b/src/backend/access/transam/slru.c
@@ -58,6 +58,8 @@
#include "pgstat.h"
#include "storage/fd.h"
#include "storage/shmem.h"
+#include "utils/dynahash.h"
+#include "utils/hsearch.h"
#define SlruFileName(ctl, path, seg) \
snprintf(path, MAXPGPATH, "%s/%04X", (ctl)->Dir, seg)
@@ -79,6 +81,12 @@ typedef struct SlruWriteAllData
typedef struct SlruWriteAllData *SlruWriteAll;
+typedef struct SlruMappingTableEntry
+{
+ int pageno;
+ int slotno;
+} SlruMappingTableEntry;
+
/*
* Populate a file tag describing a segment file. We only use the segment
* number, since we can derive everything else we need by having separate
@@ -146,6 +154,9 @@ static int SlruSelectLRUPage(SlruCtl ctl, int pageno);
static bool SlruScanDirCbDeleteCutoff(SlruCtl ctl, char *filename,
int segpage, void *data);
static void SlruInternalDeleteSegment(SlruCtl ctl, int segno);
+static void SlruMappingAdd(SlruCtl ctl, int pageno, int slotno);
+static void SlruMappingRemove(SlruCtl ctl, int pageno);
+static int SlruMappingFind(SlruCtl ctl, int pageno);
/*
* Initialization of shared memory
@@ -168,7 +179,8 @@ SimpleLruShmemSize(int nslots, int nlsns)
if (nlsns > 0)
sz += MAXALIGN(nslots * nlsns * sizeof(XLogRecPtr)); /* group_lsn[] */
- return BUFFERALIGN(sz) + BLCKSZ * nslots;
+ return BUFFERALIGN(sz) + BLCKSZ * nslots +
+ hash_estimate_size(nslots, sizeof(SlruMappingTableEntry));
}
/*
@@ -187,6 +199,9 @@ SimpleLruInit(SlruCtl ctl, const char *name, int nslots, int nlsns,
LWLock *ctllock, const char *subdir, int tranche_id,
SyncRequestHandler sync_handler)
{
+ char mapping_table_name[SHMEM_INDEX_KEYSIZE];
+ HASHCTL mapping_table_info;
+ HTAB *mapping_table;
SlruShared shared;
bool found;
@@ -258,11 +273,21 @@ SimpleLruInit(SlruCtl ctl, const char *name, int nslots, int nlsns,
else
Assert(found);
+ /* Create or find the buffer mapping table. */
+ memset(&mapping_table_info, 0, sizeof(mapping_table_info));
+ mapping_table_info.keysize = sizeof(int);
+ mapping_table_info.entrysize = sizeof(SlruMappingTableEntry);
+ snprintf(mapping_table_name, sizeof(mapping_table_name),
+ "%s Mapping Table", name);
+ mapping_table = ShmemInitHash(mapping_table_name, nslots, nslots,
+ &mapping_table_info, HASH_ELEM | HASH_BLOBS);
+
/*
* Initialize the unshared control struct, including directory path. We
* assume caller set PagePrecedes.
*/
ctl->shared = shared;
+ ctl->mapping_table = mapping_table;
ctl->sync_handler = sync_handler;
strlcpy(ctl->Dir, subdir, sizeof(ctl->Dir));
}
@@ -289,6 +314,9 @@ SimpleLruZeroPage(SlruCtl ctl, int pageno)
shared->page_number[slotno] == pageno);
/* Mark the slot as containing this page */
+ if (shared->page_status[slotno] != SLRU_PAGE_EMPTY)
+ SlruMappingRemove(ctl, shared->page_number[slotno]);
+ SlruMappingAdd(ctl, pageno, slotno);
shared->page_number[slotno] = pageno;
shared->page_status[slotno] = SLRU_PAGE_VALID;
shared->page_dirty[slotno] = true;
@@ -362,7 +390,10 @@ SimpleLruWaitIO(SlruCtl ctl, int slotno)
{
/* indeed, the I/O must have failed */
if (shared->page_status[slotno] == SLRU_PAGE_READ_IN_PROGRESS)
+ {
+ SlruMappingRemove(ctl, shared->page_number[slotno]);
shared->page_status[slotno] = SLRU_PAGE_EMPTY;
+ }
else /* write_in_progress */
{
shared->page_status[slotno] = SLRU_PAGE_VALID;
@@ -436,6 +467,7 @@ SimpleLruReadPage(SlruCtl ctl, int pageno, bool write_ok,
!shared->page_dirty[slotno]));
/* Mark the slot read-busy */
+ SlruMappingAdd(ctl, pageno, slotno);
shared->page_number[slotno] = pageno;
shared->page_status[slotno] = SLRU_PAGE_READ_IN_PROGRESS;
shared->page_dirty[slotno] = false;
@@ -459,7 +491,13 @@ SimpleLruReadPage(SlruCtl ctl, int pageno, bool write_ok,
shared->page_status[slotno] == SLRU_PAGE_READ_IN_PROGRESS &&
!shared->page_dirty[slotno]);
- shared->page_status[slotno] = ok ? SLRU_PAGE_VALID : SLRU_PAGE_EMPTY;
+ if (ok)
+ shared->page_status[slotno] = SLRU_PAGE_VALID;
+ else
+ {
+ SlruMappingRemove(ctl, pageno);
+ shared->page_status[slotno] = SLRU_PAGE_EMPTY;
+ }
LWLockRelease(&shared->buffer_locks[slotno].lock);
@@ -1029,11 +1067,12 @@ SlruSelectLRUPage(SlruCtl ctl, int pageno)
int best_invalid_page_number = 0; /* keep compiler quiet */
/* See if page already has a buffer assigned */
- for (slotno = 0; slotno < shared->num_slots; slotno++)
+ slotno = SlruMappingFind(ctl, pageno);
+ if (slotno >= 0)
{
- if (shared->page_number[slotno] == pageno &&
- shared->page_status[slotno] != SLRU_PAGE_EMPTY)
- return slotno;
+ Assert(shared->page_number[slotno] == pageno);
+ Assert(shared->page_status[slotno] != SLRU_PAGE_EMPTY);
+ return slotno;
}
/*
@@ -1266,6 +1305,7 @@ restart:;
if (shared->page_status[slotno] == SLRU_PAGE_VALID &&
!shared->page_dirty[slotno])
{
+ SlruMappingRemove(ctl, shared->page_number[slotno]);
shared->page_status[slotno] = SLRU_PAGE_EMPTY;
continue;
}
@@ -1348,6 +1388,7 @@ restart:
if (shared->page_status[slotno] == SLRU_PAGE_VALID &&
!shared->page_dirty[slotno])
{
+ SlruMappingRemove(ctl, shared->page_number[slotno]);
shared->page_status[slotno] = SLRU_PAGE_EMPTY;
continue;
}
@@ -1609,3 +1650,37 @@ SlruSyncFileTag(SlruCtl ctl, const FileTag *ftag, char *path)
errno = save_errno;
return result;
}
+
+static int
+SlruMappingFind(SlruCtl ctl, int pageno)
+{
+ SlruMappingTableEntry *mapping;
+
+ mapping = hash_search(ctl->mapping_table, &pageno, HASH_FIND, NULL);
+ if (mapping)
+ return mapping->slotno;
+
+ return -1;
+}
+
+static void
+SlruMappingAdd(SlruCtl ctl, int pageno, int slotno)
+{
+ SlruMappingTableEntry *mapping;
+ bool found PG_USED_FOR_ASSERTS_ONLY;
+
+ mapping = hash_search(ctl->mapping_table, &pageno, HASH_ENTER, &found);
+ mapping->slotno = slotno;
+
+ Assert(!found);
+}
+
+static void
+SlruMappingRemove(SlruCtl ctl, int pageno)
+{
+ bool found PG_USED_FOR_ASSERTS_ONLY;
+
+ hash_search(ctl->mapping_table, &pageno, HASH_REMOVE, &found);
+
+ Assert(found);
+}
diff --git a/src/include/access/slru.h b/src/include/access/slru.h
index dd52e8cec7..8aa3efc0ee 100644
--- a/src/include/access/slru.h
+++ b/src/include/access/slru.h
@@ -16,6 +16,7 @@
#include "access/xlogdefs.h"
#include "storage/lwlock.h"
#include "storage/sync.h"
+#include "utils/hsearch.h"
/*
@@ -110,6 +111,7 @@ typedef SlruSharedData *SlruShared;
typedef struct SlruCtlData
{
SlruShared shared;
+ HTAB *mapping_table;
/*
* Which sync handler function to use when handing sync requests over to
--
2.30.1
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2021-03-26 03:46 Thomas Munro <[email protected]>
parent: Thomas Munro <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Thomas Munro @ 2021-03-26 03:46 UTC (permalink / raw)
To: Andrey Borodin <[email protected]>; +Cc: Gilles Darold <[email protected]>; Tomas Vondra <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
Hi Andrey, all,
I propose some changes, and I'm attaching a new version:
I renamed the GUCs as clog_buffers etc (no "_slru_"). I fixed some
copy/paste mistakes where the different GUCs were mixed up. I made
some changes to the .conf.sample. I rewrote the documentation so that
it states the correct unit and defaults, and refers to the
subdirectories that are cached by these buffers instead of trying to
give a new definition of each of the SLRUs.
Do you like those changes?
Some things I thought about but didn't change:
I'm not entirely sure if we should use the internal and historical
names well known to hackers (CLOG), or the visible directory names (I
mean, we could use pg_xact_buffers instead of clog_buffers). I am not
sure why these GUCs need to be PGDLLIMPORT, but I see that NBuffers is
like that.
I wanted to do some very simple smoke testing of CLOG sizes on my
local development machine:
pgbench -i -s1000 postgres
pgbench -t4000000 -c8 -j8 -Mprepared postgres
I disabled autovacuum after running that just to be sure it wouldn't
interfere with my experiment:
alter table pgbench_accounts set (autovacuum_enabled = off);
Then I shut the cluster down and made a copy, so I could do some
repeated experiments from the same initial conditions each time. At
this point I had 30 files 0000-001E under pg_xact, holding 256kB = ~1
million transactions each. It'd take ~960 buffers to cache it all.
So how long does VACUUM FREEZE pgbench_accounts take?
I tested with just the 0001 patch, and also with the 0002 patch
(improved version, attached):
clog_buffers=128: 0001=2:28.499, 0002=2:17:891
clog_buffers=1024: 0001=1:38.485, 0002=1:29.701
I'm sure the speedup of the 0002 patch can be amplified by increasing
the number of transactions referenced in the table OR number of
clog_buffers, considering that the linear search produces
O(transactions * clog_buffers) work. That was 32M transactions and
8MB of CLOG, but I bet if you double both of those numbers once or
twice things start to get hot. I don't see why you shouldn't be able
to opt to cache literally all of CLOG if you want (something like 50MB
assuming default autovacuum_freeze_max_age, scale to taste, up to
512MB for the theoretical maximum useful value).
I'm not saying the 0002 patch is bug-free yet though, it's a bit finickity.
Attachments:
[text/x-patch] v12-0001-Make-all-SLRU-buffer-sizes-configurable.patch (20.6K, ../../CA+hUKGL0g+Nmf=2XePaXtw82BkW9CRzoxxXB4gqZOk7tHxNJXg@mail.gmail.com/2-v12-0001-Make-all-SLRU-buffer-sizes-configurable.patch)
download | inline diff:
From e7a8b9a0e27de80bab13f004b4d7e3fa1da677c4 Mon Sep 17 00:00:00 2001
From: Andrey Borodin <[email protected]>
Date: Mon, 15 Feb 2021 21:51:56 +0500
Subject: [PATCH v12 1/2] Make all SLRU buffer sizes configurable.
Provide new GUCs to set the number of buffers, instead of using hard
coded defaults.
Author: Andrey M. Borodin <[email protected]>
Reviewed-by: Anastasia Lubennikova <[email protected]>
Reviewed-by: Tomas Vondra <[email protected]>
Reviewed-by: Alexander Korotkov <[email protected]>
Reviewed-by: Gilles Darold <[email protected]>
Reviewed-by: Thomas Munro <[email protected]>
Discussion: https://postgr.es/m/2BEC2B3F-9B61-4C1D-9FB5-5FAB0F05EF86%40yandex-team.ru
---
doc/src/sgml/config.sgml | 137 ++++++++++++++++++
src/backend/access/transam/clog.c | 6 +
src/backend/access/transam/commit_ts.c | 5 +-
src/backend/access/transam/multixact.c | 8 +-
src/backend/access/transam/subtrans.c | 5 +-
src/backend/commands/async.c | 8 +-
src/backend/storage/lmgr/predicate.c | 4 +-
src/backend/utils/init/globals.c | 8 +
src/backend/utils/misc/guc.c | 77 ++++++++++
src/backend/utils/misc/postgresql.conf.sample | 9 ++
src/include/access/multixact.h | 4 -
src/include/access/subtrans.h | 3 -
src/include/commands/async.h | 5 -
src/include/miscadmin.h | 7 +
src/include/storage/predicate.h | 4 -
15 files changed, 261 insertions(+), 29 deletions(-)
diff --git a/doc/src/sgml/config.sgml b/doc/src/sgml/config.sgml
index ddc6d789d8..f1112bfa9c 100644
--- a/doc/src/sgml/config.sgml
+++ b/doc/src/sgml/config.sgml
@@ -1886,6 +1886,143 @@ include_dir 'conf.d'
</para>
</listitem>
</varlistentry>
+
+ <varlistentry id="guc-multixact-offsets-buffers" xreflabel="multixact_offsets_buffers">
+ <term><varname>multixact_offsets_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>multixact_offsets_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to used to cache the contents
+ of <literal>pg_multixact/offsets</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>8</literal>.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-multixact-members-buffers" xreflabel="multixact_members_buffers">
+ <term><varname>multixact_members_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>multixact_members_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to used to cache the contents
+ of <literal>pg_multixact/members</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>16</literal>.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-subtrans-buffers" xreflabel="subtrans_buffers">
+ <term><varname>subtrans_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>subtrans_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to used to cache the contents
+ of <literal>pg_subtrans</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>8</literal>.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-notify-buffers" xreflabel="notify_buffers">
+ <term><varname>notify_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>notify_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to used to cache the contents
+ of <literal>pg_notify</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>8</literal>.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-serial-buffers" xreflabel="serial_buffers">
+ <term><varname>serial_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>serial_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to used to cache the contents
+ of <literal>pg_serial</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>16</literal>.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-clog-buffers" xreflabel="clog_buffers">
+ <term><varname>clog_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>clog_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to used to cache the contents
+ of <literal>pg_xact</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>0</literal>, which requests
+ <varname>shared_buffers</varname> / 512, but not more than 128 or
+ fewer than 4 blocks.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-commit-ts-buffers" xreflabel="commit_ts_buffers">
+ <term><varname>commit_ts_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>commit_ts_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of memory to be used to cache the cotents of
+ <literal>pg_commit_ts</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>0</literal>, which requests
+ <varname>shared_buffers</varname> / 512, but not more than 128 or
+ fewer than 16 blocks.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
<varlistentry id="guc-max-stack-depth" xreflabel="max_stack_depth">
<term><varname>max_stack_depth</varname> (<type>integer</type>)
diff --git a/src/backend/access/transam/clog.c b/src/backend/access/transam/clog.c
index 6fa4713fb4..0318e8ff59 100644
--- a/src/backend/access/transam/clog.c
+++ b/src/backend/access/transam/clog.c
@@ -659,6 +659,9 @@ TransactionIdGetStatus(TransactionId xid, XLogRecPtr *lsn)
/*
* Number of shared CLOG buffers.
*
+ * If values is configured via GUC - just use given value. Otherwise
+ * apply following euristics.
+ *
* On larger multi-processor systems, it is possible to have many CLOG page
* requests in flight at one time which could lead to disk access for CLOG
* page if the required page is not found in memory. Testing revealed that we
@@ -675,6 +678,9 @@ TransactionIdGetStatus(TransactionId xid, XLogRecPtr *lsn)
Size
CLOGShmemBuffers(void)
{
+ /* consider 0 and 1 as unset GUC */
+ if (clog_buffers > 1)
+ return clog_buffers;
return Min(128, Max(4, NBuffers / 512));
}
diff --git a/src/backend/access/transam/commit_ts.c b/src/backend/access/transam/commit_ts.c
index 268bdba339..0d2632b90e 100644
--- a/src/backend/access/transam/commit_ts.c
+++ b/src/backend/access/transam/commit_ts.c
@@ -530,7 +530,10 @@ pg_xact_commit_timestamp_origin(PG_FUNCTION_ARGS)
Size
CommitTsShmemBuffers(void)
{
- return Min(16, Max(4, NBuffers / 1024));
+ /* consider 0 and 1 as unset GUC */
+ if (commit_ts_buffers > 1)
+ return commit_ts_buffers;
+ return Min(16, Max(4, NBuffers / 512));
}
/*
diff --git a/src/backend/access/transam/multixact.c b/src/backend/access/transam/multixact.c
index 1f9f1a1fa1..21787765e2 100644
--- a/src/backend/access/transam/multixact.c
+++ b/src/backend/access/transam/multixact.c
@@ -1831,8 +1831,8 @@ MultiXactShmemSize(void)
mul_size(sizeof(MultiXactId) * 2, MaxOldestSlot))
size = SHARED_MULTIXACT_STATE_SIZE;
- size = add_size(size, SimpleLruShmemSize(NUM_MULTIXACTOFFSET_BUFFERS, 0));
- size = add_size(size, SimpleLruShmemSize(NUM_MULTIXACTMEMBER_BUFFERS, 0));
+ size = add_size(size, SimpleLruShmemSize(multixact_offsets_buffers, 0));
+ size = add_size(size, SimpleLruShmemSize(multixact_members_buffers, 0));
return size;
}
@@ -1848,13 +1848,13 @@ MultiXactShmemInit(void)
MultiXactMemberCtl->PagePrecedes = MultiXactMemberPagePrecedes;
SimpleLruInit(MultiXactOffsetCtl,
- "MultiXactOffset", NUM_MULTIXACTOFFSET_BUFFERS, 0,
+ "MultiXactOffset", multixact_offsets_buffers, 0,
MultiXactOffsetSLRULock, "pg_multixact/offsets",
LWTRANCHE_MULTIXACTOFFSET_BUFFER,
SYNC_HANDLER_MULTIXACT_OFFSET);
SlruPagePrecedesUnitTests(MultiXactOffsetCtl, MULTIXACT_OFFSETS_PER_PAGE);
SimpleLruInit(MultiXactMemberCtl,
- "MultiXactMember", NUM_MULTIXACTMEMBER_BUFFERS, 0,
+ "MultiXactMember", multixact_offsets_buffers, 0,
MultiXactMemberSLRULock, "pg_multixact/members",
LWTRANCHE_MULTIXACTMEMBER_BUFFER,
SYNC_HANDLER_MULTIXACT_MEMBER);
diff --git a/src/backend/access/transam/subtrans.c b/src/backend/access/transam/subtrans.c
index 6a8e521f89..785f2520fd 100644
--- a/src/backend/access/transam/subtrans.c
+++ b/src/backend/access/transam/subtrans.c
@@ -31,6 +31,7 @@
#include "access/slru.h"
#include "access/subtrans.h"
#include "access/transam.h"
+#include "miscadmin.h"
#include "pg_trace.h"
#include "utils/snapmgr.h"
@@ -184,14 +185,14 @@ SubTransGetTopmostTransaction(TransactionId xid)
Size
SUBTRANSShmemSize(void)
{
- return SimpleLruShmemSize(NUM_SUBTRANS_BUFFERS, 0);
+ return SimpleLruShmemSize(subtrans_buffers, 0);
}
void
SUBTRANSShmemInit(void)
{
SubTransCtl->PagePrecedes = SubTransPagePrecedes;
- SimpleLruInit(SubTransCtl, "Subtrans", NUM_SUBTRANS_BUFFERS, 0,
+ SimpleLruInit(SubTransCtl, "Subtrans", subtrans_buffers, 0,
SubtransSLRULock, "pg_subtrans",
LWTRANCHE_SUBTRANS_BUFFER, SYNC_HANDLER_NONE);
SlruPagePrecedesUnitTests(SubTransCtl, SUBTRANS_XACTS_PER_PAGE);
diff --git a/src/backend/commands/async.c b/src/backend/commands/async.c
index 4b16fb5682..de17f52cd7 100644
--- a/src/backend/commands/async.c
+++ b/src/backend/commands/async.c
@@ -107,7 +107,7 @@
* frontend during startup.) The above design guarantees that notifies from
* other backends will never be missed by ignoring self-notifies.
*
- * The amount of shared memory used for notify management (NUM_NOTIFY_BUFFERS)
+ * The amount of shared memory used for notify management (notify_buffers)
* can be varied without affecting anything but performance. The maximum
* amount of notification data that can be queued at one time is determined
* by slru.c's wraparound limit; see QUEUE_MAX_PAGE below.
@@ -225,7 +225,7 @@ typedef struct QueuePosition
*
* Resist the temptation to make this really large. While that would save
* work in some places, it would add cost in others. In particular, this
- * should likely be less than NUM_NOTIFY_BUFFERS, to ensure that backends
+ * should likely be less than notify_buffers, to ensure that backends
* catch up before the pages they'll need to read fall out of SLRU cache.
*/
#define QUEUE_CLEANUP_DELAY 4
@@ -514,7 +514,7 @@ AsyncShmemSize(void)
size = mul_size(MaxBackends + 1, sizeof(QueueBackendStatus));
size = add_size(size, offsetof(AsyncQueueControl, backend));
- size = add_size(size, SimpleLruShmemSize(NUM_NOTIFY_BUFFERS, 0));
+ size = add_size(size, SimpleLruShmemSize(notify_buffers, 0));
return size;
}
@@ -562,7 +562,7 @@ AsyncShmemInit(void)
* Set up SLRU management of the pg_notify data.
*/
NotifyCtl->PagePrecedes = asyncQueuePagePrecedes;
- SimpleLruInit(NotifyCtl, "Notify", NUM_NOTIFY_BUFFERS, 0,
+ SimpleLruInit(NotifyCtl, "Notify", notify_buffers, 0,
NotifySLRULock, "pg_notify", LWTRANCHE_NOTIFY_BUFFER,
SYNC_HANDLER_NONE);
diff --git a/src/backend/storage/lmgr/predicate.c b/src/backend/storage/lmgr/predicate.c
index d493aeef0f..b1f4f1651d 100644
--- a/src/backend/storage/lmgr/predicate.c
+++ b/src/backend/storage/lmgr/predicate.c
@@ -872,7 +872,7 @@ SerialInit(void)
*/
SerialSlruCtl->PagePrecedes = SerialPagePrecedesLogically;
SimpleLruInit(SerialSlruCtl, "Serial",
- NUM_SERIAL_BUFFERS, 0, SerialSLRULock, "pg_serial",
+ serial_buffers, 0, SerialSLRULock, "pg_serial",
LWTRANCHE_SERIAL_BUFFER, SYNC_HANDLER_NONE);
#ifdef USE_ASSERT_CHECKING
SerialPagePrecedesLogicallyUnitTests();
@@ -1395,7 +1395,7 @@ PredicateLockShmemSize(void)
/* Shared memory structures for SLRU tracking of old committed xids. */
size = add_size(size, sizeof(SerialControlData));
- size = add_size(size, SimpleLruShmemSize(NUM_SERIAL_BUFFERS, 0));
+ size = add_size(size, SimpleLruShmemSize(serial_buffers, 0));
return size;
}
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
index 73e0a672ae..e90275be6c 100644
--- a/src/backend/utils/init/globals.c
+++ b/src/backend/utils/init/globals.c
@@ -148,3 +148,11 @@ int64 VacuumPageDirty = 0;
int VacuumCostBalance = 0; /* working state for vacuum */
bool VacuumCostActive = false;
+
+int multixact_offsets_buffers = 8;
+int multixact_members_buffers = 16;
+int subtrans_buffers = 32;
+int notify_buffers = 8;
+int serial_buffers = 16;
+int clog_buffers = 0;
+int commit_ts_buffers = 0;
diff --git a/src/backend/utils/misc/guc.c b/src/backend/utils/misc/guc.c
index 0c5dc4d3e8..003bc820d2 100644
--- a/src/backend/utils/misc/guc.c
+++ b/src/backend/utils/misc/guc.c
@@ -2305,6 +2305,83 @@ static struct config_int ConfigureNamesInt[] =
NULL, NULL, NULL
},
+ {
+ {"multixact_offsets_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for MultiXact offsets SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &multixact_offsets_buffers,
+ 8, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"multixact_members_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for MultiXact members SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &multixact_members_buffers,
+ 16, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"subtrans_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for substransactions SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &subtrans_buffers,
+ 32, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"notify_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for asyncronous notifications SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ ¬ify_buffers,
+ 8, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"serial_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for predicate locks SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &serial_buffers,
+ 16, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"clog_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for commit log SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &clog_buffers,
+ 0, 0, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"commit_ts_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for commit timestamps SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &commit_ts_buffers,
+ 0, 0, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
{
{"temp_buffers", PGC_USERSET, RESOURCES_MEM,
gettext_noop("Sets the maximum number of temporary buffers used by each session."),
diff --git a/src/backend/utils/misc/postgresql.conf.sample b/src/backend/utils/misc/postgresql.conf.sample
index b234a6bfe6..1b8515989b 100644
--- a/src/backend/utils/misc/postgresql.conf.sample
+++ b/src/backend/utils/misc/postgresql.conf.sample
@@ -190,6 +190,15 @@
# (change requires restart)
#backend_flush_after = 0 # measured in pages, 0 disables
+# - SLRU Buffers (change requires restart) -
+
+#clog_buffers = 0 # memory for pg_xact (0 = auto)
+#subtrans_buffers = 32 # memory for pg_subtrans
+#multixact_offsets_buffers = 8 # memory for pg_multixact/offsets
+#multixact_members_buffers = 16 # memory for pg_multixact/members
+#notify_buffers = 8 # memory for pg_nofity
+#serial_buffers = 16 # memory for pg_serial
+#commit_ts_buffers = 0 # memory for pg_commit_ts (0 = auto)
#------------------------------------------------------------------------------
# WRITE-AHEAD LOG
diff --git a/src/include/access/multixact.h b/src/include/access/multixact.h
index 4bbb035eae..97c0a46376 100644
--- a/src/include/access/multixact.h
+++ b/src/include/access/multixact.h
@@ -29,10 +29,6 @@
#define MaxMultiXactOffset ((MultiXactOffset) 0xFFFFFFFF)
-/* Number of SLRU buffers to use for multixact */
-#define NUM_MULTIXACTOFFSET_BUFFERS 8
-#define NUM_MULTIXACTMEMBER_BUFFERS 16
-
/*
* Possible multixact lock modes ("status"). The first four modes are for
* tuple locks (FOR KEY SHARE, FOR SHARE, FOR NO KEY UPDATE, FOR UPDATE); the
diff --git a/src/include/access/subtrans.h b/src/include/access/subtrans.h
index d0ab44ae82..ca0999056e 100644
--- a/src/include/access/subtrans.h
+++ b/src/include/access/subtrans.h
@@ -11,9 +11,6 @@
#ifndef SUBTRANS_H
#define SUBTRANS_H
-/* Number of SLRU buffers to use for subtrans */
-#define NUM_SUBTRANS_BUFFERS 32
-
extern void SubTransSetParent(TransactionId xid, TransactionId parent);
extern TransactionId SubTransGetParent(TransactionId xid);
extern TransactionId SubTransGetTopmostTransaction(TransactionId xid);
diff --git a/src/include/commands/async.h b/src/include/commands/async.h
index 9217f66b91..fa831e3721 100644
--- a/src/include/commands/async.h
+++ b/src/include/commands/async.h
@@ -15,11 +15,6 @@
#include <signal.h>
-/*
- * The number of SLRU page buffers we use for the notification queue.
- */
-#define NUM_NOTIFY_BUFFERS 8
-
extern bool Trace_notify;
extern volatile sig_atomic_t notifyInterruptPending;
diff --git a/src/include/miscadmin.h b/src/include/miscadmin.h
index 013850ac28..9c325b4312 100644
--- a/src/include/miscadmin.h
+++ b/src/include/miscadmin.h
@@ -162,6 +162,13 @@ extern PGDLLIMPORT int MaxBackends;
extern PGDLLIMPORT int MaxConnections;
extern PGDLLIMPORT int max_worker_processes;
extern PGDLLIMPORT int max_parallel_workers;
+extern PGDLLIMPORT int multixact_offsets_buffers;
+extern PGDLLIMPORT int multixact_members_buffers;
+extern PGDLLIMPORT int subtrans_buffers;
+extern PGDLLIMPORT int notify_buffers;
+extern PGDLLIMPORT int serial_buffers;
+extern PGDLLIMPORT int clog_buffers;
+extern PGDLLIMPORT int commit_ts_buffers;
extern PGDLLIMPORT int MyProcPid;
extern PGDLLIMPORT pg_time_t MyStartTime;
diff --git a/src/include/storage/predicate.h b/src/include/storage/predicate.h
index 152b698611..c72779bd88 100644
--- a/src/include/storage/predicate.h
+++ b/src/include/storage/predicate.h
@@ -26,10 +26,6 @@ extern int max_predicate_locks_per_xact;
extern int max_predicate_locks_per_relation;
extern int max_predicate_locks_per_page;
-
-/* Number of SLRU buffers to use for Serial SLRU */
-#define NUM_SERIAL_BUFFERS 16
-
/*
* A handle used for sharing SERIALIZABLEXACT objects between the participants
* in a parallel query.
--
2.30.1
[text/x-patch] v12-0002-Add-buffer-mapping-table-for-SLRUs.patch (8.5K, ../../CA+hUKGL0g+Nmf=2XePaXtw82BkW9CRzoxxXB4gqZOk7tHxNJXg@mail.gmail.com/3-v12-0002-Add-buffer-mapping-table-for-SLRUs.patch)
download | inline diff:
From 98092b5b4e7a9a11a5bbd50c36ad8f635fdfab6f Mon Sep 17 00:00:00 2001
From: Thomas Munro <[email protected]>
Date: Thu, 25 Mar 2021 10:11:31 +1300
Subject: [PATCH v12 2/2] Add buffer mapping table for SLRUs.
Instead of doing a linear search for the buffer holding a given page
number, use a hash table.
Discussion: https://postgr.es/m/2BEC2B3F-9B61-4C1D-9FB5-5FAB0F05EF86%40yandex-team.ru
---
src/backend/access/transam/slru.c | 111 +++++++++++++++++++++++++-----
src/include/access/slru.h | 2 +
2 files changed, 96 insertions(+), 17 deletions(-)
diff --git a/src/backend/access/transam/slru.c b/src/backend/access/transam/slru.c
index 82149ad782..f823c2a0c8 100644
--- a/src/backend/access/transam/slru.c
+++ b/src/backend/access/transam/slru.c
@@ -58,6 +58,8 @@
#include "pgstat.h"
#include "storage/fd.h"
#include "storage/shmem.h"
+#include "utils/dynahash.h"
+#include "utils/hsearch.h"
#define SlruFileName(ctl, path, seg) \
snprintf(path, MAXPGPATH, "%s/%04X", (ctl)->Dir, seg)
@@ -79,6 +81,12 @@ typedef struct SlruWriteAllData
typedef struct SlruWriteAllData *SlruWriteAll;
+typedef struct SlruMappingTableEntry
+{
+ int pageno;
+ int slotno;
+} SlruMappingTableEntry;
+
/*
* Populate a file tag describing a segment file. We only use the segment
* number, since we can derive everything else we need by having separate
@@ -146,6 +154,9 @@ static int SlruSelectLRUPage(SlruCtl ctl, int pageno);
static bool SlruScanDirCbDeleteCutoff(SlruCtl ctl, char *filename,
int segpage, void *data);
static void SlruInternalDeleteSegment(SlruCtl ctl, int segno);
+static void SlruMappingAdd(SlruCtl ctl, int pageno, int slotno);
+static void SlruMappingRemove(SlruCtl ctl, int pageno);
+static int SlruMappingFind(SlruCtl ctl, int pageno);
/*
* Initialization of shared memory
@@ -168,7 +179,8 @@ SimpleLruShmemSize(int nslots, int nlsns)
if (nlsns > 0)
sz += MAXALIGN(nslots * nlsns * sizeof(XLogRecPtr)); /* group_lsn[] */
- return BUFFERALIGN(sz) + BLCKSZ * nslots;
+ return BUFFERALIGN(sz) + BLCKSZ * nslots +
+ hash_estimate_size(nslots, sizeof(SlruMappingTableEntry));
}
/*
@@ -187,6 +199,9 @@ SimpleLruInit(SlruCtl ctl, const char *name, int nslots, int nlsns,
LWLock *ctllock, const char *subdir, int tranche_id,
SyncRequestHandler sync_handler)
{
+ char mapping_table_name[SHMEM_INDEX_KEYSIZE];
+ HASHCTL mapping_table_info;
+ HTAB *mapping_table;
SlruShared shared;
bool found;
@@ -258,11 +273,21 @@ SimpleLruInit(SlruCtl ctl, const char *name, int nslots, int nlsns,
else
Assert(found);
+ /* Create or find the buffer mapping table. */
+ memset(&mapping_table_info, 0, sizeof(mapping_table_info));
+ mapping_table_info.keysize = sizeof(int);
+ mapping_table_info.entrysize = sizeof(SlruMappingTableEntry);
+ snprintf(mapping_table_name, sizeof(mapping_table_name),
+ "%s Mapping Table", name);
+ mapping_table = ShmemInitHash(mapping_table_name, nslots, nslots,
+ &mapping_table_info, HASH_ELEM | HASH_BLOBS);
+
/*
* Initialize the unshared control struct, including directory path. We
* assume caller set PagePrecedes.
*/
ctl->shared = shared;
+ ctl->mapping_table = mapping_table;
ctl->sync_handler = sync_handler;
strlcpy(ctl->Dir, subdir, sizeof(ctl->Dir));
}
@@ -289,6 +314,9 @@ SimpleLruZeroPage(SlruCtl ctl, int pageno)
shared->page_number[slotno] == pageno);
/* Mark the slot as containing this page */
+ if (shared->page_status[slotno] != SLRU_PAGE_EMPTY)
+ SlruMappingRemove(ctl, shared->page_number[slotno]);
+ SlruMappingAdd(ctl, pageno, slotno);
shared->page_number[slotno] = pageno;
shared->page_status[slotno] = SLRU_PAGE_VALID;
shared->page_dirty[slotno] = true;
@@ -362,7 +390,10 @@ SimpleLruWaitIO(SlruCtl ctl, int slotno)
{
/* indeed, the I/O must have failed */
if (shared->page_status[slotno] == SLRU_PAGE_READ_IN_PROGRESS)
+ {
+ SlruMappingRemove(ctl, shared->page_number[slotno]);
shared->page_status[slotno] = SLRU_PAGE_EMPTY;
+ }
else /* write_in_progress */
{
shared->page_status[slotno] = SLRU_PAGE_VALID;
@@ -436,6 +467,9 @@ SimpleLruReadPage(SlruCtl ctl, int pageno, bool write_ok,
!shared->page_dirty[slotno]));
/* Mark the slot read-busy */
+ if (shared->page_status[slotno] != SLRU_PAGE_EMPTY)
+ SlruMappingRemove(ctl, shared->page_number[slotno]);
+ SlruMappingAdd(ctl, pageno, slotno);
shared->page_number[slotno] = pageno;
shared->page_status[slotno] = SLRU_PAGE_READ_IN_PROGRESS;
shared->page_dirty[slotno] = false;
@@ -459,7 +493,13 @@ SimpleLruReadPage(SlruCtl ctl, int pageno, bool write_ok,
shared->page_status[slotno] == SLRU_PAGE_READ_IN_PROGRESS &&
!shared->page_dirty[slotno]);
- shared->page_status[slotno] = ok ? SLRU_PAGE_VALID : SLRU_PAGE_EMPTY;
+ if (ok)
+ shared->page_status[slotno] = SLRU_PAGE_VALID;
+ else
+ {
+ SlruMappingRemove(ctl, pageno);
+ shared->page_status[slotno] = SLRU_PAGE_EMPTY;
+ }
LWLockRelease(&shared->buffer_locks[slotno].lock);
@@ -500,20 +540,20 @@ SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno, TransactionId xid)
LWLockAcquire(shared->ControlLock, LW_SHARED);
/* See if page is already in a buffer */
- for (slotno = 0; slotno < shared->num_slots; slotno++)
+ slotno = SlruMappingFind(ctl, pageno);
+ if (slotno >= 0 &&
+ shared->page_status[slotno] != SLRU_PAGE_READ_IN_PROGRESS)
{
- if (shared->page_number[slotno] == pageno &&
- shared->page_status[slotno] != SLRU_PAGE_EMPTY &&
- shared->page_status[slotno] != SLRU_PAGE_READ_IN_PROGRESS)
- {
- /* See comments for SlruRecentlyUsed macro */
- SlruRecentlyUsed(shared, slotno);
+ Assert(shared->page_status[slotno] != SLRU_PAGE_EMPTY);
+ Assert(shared->page_number[slotno] == pageno);
- /* update the stats counter of pages found in the SLRU */
- pgstat_count_slru_page_hit(shared->slru_stats_idx);
+ /* See comments for SlruRecentlyUsed macro */
+ SlruRecentlyUsed(shared, slotno);
- return slotno;
- }
+ /* update the stats counter of pages found in the SLRU */
+ pgstat_count_slru_page_hit(shared->slru_stats_idx);
+
+ return slotno;
}
/* No luck, so switch to normal exclusive lock and do regular read */
@@ -1029,11 +1069,12 @@ SlruSelectLRUPage(SlruCtl ctl, int pageno)
int best_invalid_page_number = 0; /* keep compiler quiet */
/* See if page already has a buffer assigned */
- for (slotno = 0; slotno < shared->num_slots; slotno++)
+ slotno = SlruMappingFind(ctl, pageno);
+ if (slotno >= 0)
{
- if (shared->page_number[slotno] == pageno &&
- shared->page_status[slotno] != SLRU_PAGE_EMPTY)
- return slotno;
+ Assert(shared->page_number[slotno] == pageno);
+ Assert(shared->page_status[slotno] != SLRU_PAGE_EMPTY);
+ return slotno;
}
/*
@@ -1266,6 +1307,7 @@ restart:;
if (shared->page_status[slotno] == SLRU_PAGE_VALID &&
!shared->page_dirty[slotno])
{
+ SlruMappingRemove(ctl, shared->page_number[slotno]);
shared->page_status[slotno] = SLRU_PAGE_EMPTY;
continue;
}
@@ -1348,6 +1390,7 @@ restart:
if (shared->page_status[slotno] == SLRU_PAGE_VALID &&
!shared->page_dirty[slotno])
{
+ SlruMappingRemove(ctl, shared->page_number[slotno]);
shared->page_status[slotno] = SLRU_PAGE_EMPTY;
continue;
}
@@ -1609,3 +1652,37 @@ SlruSyncFileTag(SlruCtl ctl, const FileTag *ftag, char *path)
errno = save_errno;
return result;
}
+
+static int
+SlruMappingFind(SlruCtl ctl, int pageno)
+{
+ SlruMappingTableEntry *mapping;
+
+ mapping = hash_search(ctl->mapping_table, &pageno, HASH_FIND, NULL);
+ if (mapping)
+ return mapping->slotno;
+
+ return -1;
+}
+
+static void
+SlruMappingAdd(SlruCtl ctl, int pageno, int slotno)
+{
+ SlruMappingTableEntry *mapping;
+ bool found PG_USED_FOR_ASSERTS_ONLY;
+
+ mapping = hash_search(ctl->mapping_table, &pageno, HASH_ENTER, &found);
+ mapping->slotno = slotno;
+
+ Assert(!found);
+}
+
+static void
+SlruMappingRemove(SlruCtl ctl, int pageno)
+{
+ bool found PG_USED_FOR_ASSERTS_ONLY;
+
+ hash_search(ctl->mapping_table, &pageno, HASH_REMOVE, &found);
+
+ Assert(found);
+}
diff --git a/src/include/access/slru.h b/src/include/access/slru.h
index dd52e8cec7..8aa3efc0ee 100644
--- a/src/include/access/slru.h
+++ b/src/include/access/slru.h
@@ -16,6 +16,7 @@
#include "access/xlogdefs.h"
#include "storage/lwlock.h"
#include "storage/sync.h"
+#include "utils/hsearch.h"
/*
@@ -110,6 +111,7 @@ typedef SlruSharedData *SlruShared;
typedef struct SlruCtlData
{
SlruShared shared;
+ HTAB *mapping_table;
/*
* Which sync handler function to use when handing sync requests over to
--
2.30.1
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2021-03-26 06:00 Andrey Borodin <[email protected]>
parent: Thomas Munro <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Andrey Borodin @ 2021-03-26 06:00 UTC (permalink / raw)
To: Thomas Munro <[email protected]>; +Cc: Gilles Darold <[email protected]>; Tomas Vondra <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
Hi Thomas, Gilles, all!
Thanks for reviewing this!
> 25 марта 2021 г., в 02:31, Thomas Munro <[email protected]> написал(а):
>
> I don't think we should put "slru" (the name of the buffer replacement
> algorithm, implementation detail) in the GUC names.
+1
> + It defaults to 0, in this case CLOG size is taken as
> <varname>shared_buffers</varname> / 512.
>
> We already know that increasing the number of CLOG buffers above the
> current number hurts as the linear search begins to dominate
Uh, my intent was to copy original approach of CLOG SLRU size, I just missed that Min(,) thing in shared_buffers logic.
> 26 марта 2021 г., в 08:46, Thomas Munro <[email protected]> написал(а):
>
> Hi Andrey, all,
>
> I propose some changes, and I'm attaching a new version:
>
> I renamed the GUCs as clog_buffers etc (no "_slru_"). I fixed some
> copy/paste mistakes where the different GUCs were mixed up. I made
> some changes to the .conf.sample. I rewrote the documentation so that
> it states the correct unit and defaults, and refers to the
> subdirectories that are cached by these buffers instead of trying to
> give a new definition of each of the SLRUs.
>
> Do you like those changes?
Yes!
> Some things I thought about but didn't change:
>
> I'm not entirely sure if we should use the internal and historical
> names well known to hackers (CLOG), or the visible directory names (I
> mean, we could use pg_xact_buffers instead of clog_buffers).
While it is good idea to make notes about directory name, I think the real naming criteria is to help find this GUC when user encounters wait events in pg_stat_activity. I think there is no CLOG mentions in docs [0], only XactBuffer, XactSLRU et c.
> I'm not saying the 0002 patch is bug-free yet though, it's a bit finickity.
I think the idea of speeding up linear search is really really good for scaling SLRUs. It's not even about improving normal performance of the cluster, but it's important from preventing pathological degradation under certain circumstances. Bigger cache really saves SLAs :) I'll look into the patch more closely this weekend. Thank you!
Best regards, Andrey Borodin.
[0] https://www.postgresql.org/docs/current/monitoring-stats.html#MONITORING-PG-STAT-ACTIVITY-VIEW
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2021-03-26 15:52 Andrey Borodin <[email protected]>
parent: Andrey Borodin <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Andrey Borodin @ 2021-03-26 15:52 UTC (permalink / raw)
To: Thomas Munro <[email protected]>; +Cc: Gilles Darold <[email protected]>; Tomas Vondra <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
> 26 марта 2021 г., в 11:00, Andrey Borodin <[email protected]> написал(а):
>
>> I'm not saying the 0002 patch is bug-free yet though, it's a bit finickity.
> I think the idea of speeding up linear search is really really good for scaling SLRUs. It's not even about improving normal performance of the cluster, but it's important from preventing pathological degradation under certain circumstances. Bigger cache really saves SLAs :) I'll look into the patch more closely this weekend. Thank you!
Some thoughts on HashTable patch:
1. Can we allocate bigger hashtable to reduce probability of collisions?
2. Can we use specialised hashtable for this case? I'm afraid hash_search() does comparable number of CPU cycles as simple cycle from 0 to 128. We could inline everything and avoid hashp->hash(keyPtr, hashp->keysize) call. I'm not insisting on special hash though, just an idea.
3. pageno in SlruMappingTableEntry seems to be unused.
Thanks!
Best regards, Andrey Borodin.
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2021-03-26 20:26 Thomas Munro <[email protected]>
parent: Andrey Borodin <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Thomas Munro @ 2021-03-26 20:26 UTC (permalink / raw)
To: Andrey Borodin <[email protected]>; +Cc: Gilles Darold <[email protected]>; Tomas Vondra <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
On Sat, Mar 27, 2021 at 4:52 AM Andrey Borodin <[email protected]> wrote:
> Some thoughts on HashTable patch:
> 1. Can we allocate bigger hashtable to reduce probability of collisions?
Yeah, good idea, might require some study.
> 2. Can we use specialised hashtable for this case? I'm afraid hash_search() does comparable number of CPU cycles as simple cycle from 0 to 128. We could inline everything and avoid hashp->hash(keyPtr, hashp->keysize) call. I'm not insisting on special hash though, just an idea.
I tried really hard to not fall into this rabbit h.... [hack hack
hack], OK, here's a first attempt to use simplehash, Andres's
steampunk macro-based robinhood template that we're already using for
several other things, and murmurhash which is inlineable and
branch-free. I had to tweak it to support "in-place" creation and
fixed size (in other words, no allocators, for use in shared memory).
Then I was annoyed that I had to add a "status" member to our struct,
so I tried to fix that. Definitely needs more work to think about
failure modes when running out of memory, how much spare space you
need, etc.
I have not experimented with this much beyond hacking until the tests
pass, but it *should* be more efficient...
> 3. pageno in SlruMappingTableEntry seems to be unused.
It's the key (dynahash uses the first N bytes of your struct as the
key, but in this new simplehash version it's more explicit).
Attachments:
[application/x-patch] v13-0001-Make-all-SLRU-buffer-sizes-configurable.patch (20.6K, ../../CA+hUKGKVqrxOp82zER1=XN=yPwV_-OCGAg=ez=1iz9rG+A7Smw@mail.gmail.com/2-v13-0001-Make-all-SLRU-buffer-sizes-configurable.patch)
download | inline diff:
From 5f5d4ed8ae2808766ac1fd48f68602ef530e3833 Mon Sep 17 00:00:00 2001
From: Andrey Borodin <[email protected]>
Date: Mon, 15 Feb 2021 21:51:56 +0500
Subject: [PATCH v13 1/5] Make all SLRU buffer sizes configurable.
Provide new GUCs to set the number of buffers, instead of using hard
coded defaults.
Author: Andrey M. Borodin <[email protected]>
Reviewed-by: Anastasia Lubennikova <[email protected]>
Reviewed-by: Tomas Vondra <[email protected]>
Reviewed-by: Alexander Korotkov <[email protected]>
Reviewed-by: Gilles Darold <[email protected]>
Reviewed-by: Thomas Munro <[email protected]>
Discussion: https://postgr.es/m/2BEC2B3F-9B61-4C1D-9FB5-5FAB0F05EF86%40yandex-team.ru
---
doc/src/sgml/config.sgml | 137 ++++++++++++++++++
src/backend/access/transam/clog.c | 6 +
src/backend/access/transam/commit_ts.c | 5 +-
src/backend/access/transam/multixact.c | 8 +-
src/backend/access/transam/subtrans.c | 5 +-
src/backend/commands/async.c | 8 +-
src/backend/storage/lmgr/predicate.c | 4 +-
src/backend/utils/init/globals.c | 8 +
src/backend/utils/misc/guc.c | 77 ++++++++++
src/backend/utils/misc/postgresql.conf.sample | 9 ++
src/include/access/multixact.h | 4 -
src/include/access/subtrans.h | 3 -
src/include/commands/async.h | 5 -
src/include/miscadmin.h | 7 +
src/include/storage/predicate.h | 4 -
15 files changed, 261 insertions(+), 29 deletions(-)
diff --git a/doc/src/sgml/config.sgml b/doc/src/sgml/config.sgml
index ddc6d789d8..f1112bfa9c 100644
--- a/doc/src/sgml/config.sgml
+++ b/doc/src/sgml/config.sgml
@@ -1886,6 +1886,143 @@ include_dir 'conf.d'
</para>
</listitem>
</varlistentry>
+
+ <varlistentry id="guc-multixact-offsets-buffers" xreflabel="multixact_offsets_buffers">
+ <term><varname>multixact_offsets_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>multixact_offsets_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to used to cache the contents
+ of <literal>pg_multixact/offsets</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>8</literal>.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-multixact-members-buffers" xreflabel="multixact_members_buffers">
+ <term><varname>multixact_members_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>multixact_members_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to used to cache the contents
+ of <literal>pg_multixact/members</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>16</literal>.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-subtrans-buffers" xreflabel="subtrans_buffers">
+ <term><varname>subtrans_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>subtrans_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to used to cache the contents
+ of <literal>pg_subtrans</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>8</literal>.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-notify-buffers" xreflabel="notify_buffers">
+ <term><varname>notify_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>notify_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to used to cache the contents
+ of <literal>pg_notify</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>8</literal>.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-serial-buffers" xreflabel="serial_buffers">
+ <term><varname>serial_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>serial_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to used to cache the contents
+ of <literal>pg_serial</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>16</literal>.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-clog-buffers" xreflabel="clog_buffers">
+ <term><varname>clog_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>clog_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to used to cache the contents
+ of <literal>pg_xact</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>0</literal>, which requests
+ <varname>shared_buffers</varname> / 512, but not more than 128 or
+ fewer than 4 blocks.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-commit-ts-buffers" xreflabel="commit_ts_buffers">
+ <term><varname>commit_ts_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>commit_ts_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of memory to be used to cache the cotents of
+ <literal>pg_commit_ts</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>0</literal>, which requests
+ <varname>shared_buffers</varname> / 512, but not more than 128 or
+ fewer than 16 blocks.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
<varlistentry id="guc-max-stack-depth" xreflabel="max_stack_depth">
<term><varname>max_stack_depth</varname> (<type>integer</type>)
diff --git a/src/backend/access/transam/clog.c b/src/backend/access/transam/clog.c
index 6fa4713fb4..0318e8ff59 100644
--- a/src/backend/access/transam/clog.c
+++ b/src/backend/access/transam/clog.c
@@ -659,6 +659,9 @@ TransactionIdGetStatus(TransactionId xid, XLogRecPtr *lsn)
/*
* Number of shared CLOG buffers.
*
+ * If values is configured via GUC - just use given value. Otherwise
+ * apply following euristics.
+ *
* On larger multi-processor systems, it is possible to have many CLOG page
* requests in flight at one time which could lead to disk access for CLOG
* page if the required page is not found in memory. Testing revealed that we
@@ -675,6 +678,9 @@ TransactionIdGetStatus(TransactionId xid, XLogRecPtr *lsn)
Size
CLOGShmemBuffers(void)
{
+ /* consider 0 and 1 as unset GUC */
+ if (clog_buffers > 1)
+ return clog_buffers;
return Min(128, Max(4, NBuffers / 512));
}
diff --git a/src/backend/access/transam/commit_ts.c b/src/backend/access/transam/commit_ts.c
index 268bdba339..0d2632b90e 100644
--- a/src/backend/access/transam/commit_ts.c
+++ b/src/backend/access/transam/commit_ts.c
@@ -530,7 +530,10 @@ pg_xact_commit_timestamp_origin(PG_FUNCTION_ARGS)
Size
CommitTsShmemBuffers(void)
{
- return Min(16, Max(4, NBuffers / 1024));
+ /* consider 0 and 1 as unset GUC */
+ if (commit_ts_buffers > 1)
+ return commit_ts_buffers;
+ return Min(16, Max(4, NBuffers / 512));
}
/*
diff --git a/src/backend/access/transam/multixact.c b/src/backend/access/transam/multixact.c
index 1f9f1a1fa1..21787765e2 100644
--- a/src/backend/access/transam/multixact.c
+++ b/src/backend/access/transam/multixact.c
@@ -1831,8 +1831,8 @@ MultiXactShmemSize(void)
mul_size(sizeof(MultiXactId) * 2, MaxOldestSlot))
size = SHARED_MULTIXACT_STATE_SIZE;
- size = add_size(size, SimpleLruShmemSize(NUM_MULTIXACTOFFSET_BUFFERS, 0));
- size = add_size(size, SimpleLruShmemSize(NUM_MULTIXACTMEMBER_BUFFERS, 0));
+ size = add_size(size, SimpleLruShmemSize(multixact_offsets_buffers, 0));
+ size = add_size(size, SimpleLruShmemSize(multixact_members_buffers, 0));
return size;
}
@@ -1848,13 +1848,13 @@ MultiXactShmemInit(void)
MultiXactMemberCtl->PagePrecedes = MultiXactMemberPagePrecedes;
SimpleLruInit(MultiXactOffsetCtl,
- "MultiXactOffset", NUM_MULTIXACTOFFSET_BUFFERS, 0,
+ "MultiXactOffset", multixact_offsets_buffers, 0,
MultiXactOffsetSLRULock, "pg_multixact/offsets",
LWTRANCHE_MULTIXACTOFFSET_BUFFER,
SYNC_HANDLER_MULTIXACT_OFFSET);
SlruPagePrecedesUnitTests(MultiXactOffsetCtl, MULTIXACT_OFFSETS_PER_PAGE);
SimpleLruInit(MultiXactMemberCtl,
- "MultiXactMember", NUM_MULTIXACTMEMBER_BUFFERS, 0,
+ "MultiXactMember", multixact_offsets_buffers, 0,
MultiXactMemberSLRULock, "pg_multixact/members",
LWTRANCHE_MULTIXACTMEMBER_BUFFER,
SYNC_HANDLER_MULTIXACT_MEMBER);
diff --git a/src/backend/access/transam/subtrans.c b/src/backend/access/transam/subtrans.c
index 6a8e521f89..785f2520fd 100644
--- a/src/backend/access/transam/subtrans.c
+++ b/src/backend/access/transam/subtrans.c
@@ -31,6 +31,7 @@
#include "access/slru.h"
#include "access/subtrans.h"
#include "access/transam.h"
+#include "miscadmin.h"
#include "pg_trace.h"
#include "utils/snapmgr.h"
@@ -184,14 +185,14 @@ SubTransGetTopmostTransaction(TransactionId xid)
Size
SUBTRANSShmemSize(void)
{
- return SimpleLruShmemSize(NUM_SUBTRANS_BUFFERS, 0);
+ return SimpleLruShmemSize(subtrans_buffers, 0);
}
void
SUBTRANSShmemInit(void)
{
SubTransCtl->PagePrecedes = SubTransPagePrecedes;
- SimpleLruInit(SubTransCtl, "Subtrans", NUM_SUBTRANS_BUFFERS, 0,
+ SimpleLruInit(SubTransCtl, "Subtrans", subtrans_buffers, 0,
SubtransSLRULock, "pg_subtrans",
LWTRANCHE_SUBTRANS_BUFFER, SYNC_HANDLER_NONE);
SlruPagePrecedesUnitTests(SubTransCtl, SUBTRANS_XACTS_PER_PAGE);
diff --git a/src/backend/commands/async.c b/src/backend/commands/async.c
index 4b16fb5682..de17f52cd7 100644
--- a/src/backend/commands/async.c
+++ b/src/backend/commands/async.c
@@ -107,7 +107,7 @@
* frontend during startup.) The above design guarantees that notifies from
* other backends will never be missed by ignoring self-notifies.
*
- * The amount of shared memory used for notify management (NUM_NOTIFY_BUFFERS)
+ * The amount of shared memory used for notify management (notify_buffers)
* can be varied without affecting anything but performance. The maximum
* amount of notification data that can be queued at one time is determined
* by slru.c's wraparound limit; see QUEUE_MAX_PAGE below.
@@ -225,7 +225,7 @@ typedef struct QueuePosition
*
* Resist the temptation to make this really large. While that would save
* work in some places, it would add cost in others. In particular, this
- * should likely be less than NUM_NOTIFY_BUFFERS, to ensure that backends
+ * should likely be less than notify_buffers, to ensure that backends
* catch up before the pages they'll need to read fall out of SLRU cache.
*/
#define QUEUE_CLEANUP_DELAY 4
@@ -514,7 +514,7 @@ AsyncShmemSize(void)
size = mul_size(MaxBackends + 1, sizeof(QueueBackendStatus));
size = add_size(size, offsetof(AsyncQueueControl, backend));
- size = add_size(size, SimpleLruShmemSize(NUM_NOTIFY_BUFFERS, 0));
+ size = add_size(size, SimpleLruShmemSize(notify_buffers, 0));
return size;
}
@@ -562,7 +562,7 @@ AsyncShmemInit(void)
* Set up SLRU management of the pg_notify data.
*/
NotifyCtl->PagePrecedes = asyncQueuePagePrecedes;
- SimpleLruInit(NotifyCtl, "Notify", NUM_NOTIFY_BUFFERS, 0,
+ SimpleLruInit(NotifyCtl, "Notify", notify_buffers, 0,
NotifySLRULock, "pg_notify", LWTRANCHE_NOTIFY_BUFFER,
SYNC_HANDLER_NONE);
diff --git a/src/backend/storage/lmgr/predicate.c b/src/backend/storage/lmgr/predicate.c
index d493aeef0f..b1f4f1651d 100644
--- a/src/backend/storage/lmgr/predicate.c
+++ b/src/backend/storage/lmgr/predicate.c
@@ -872,7 +872,7 @@ SerialInit(void)
*/
SerialSlruCtl->PagePrecedes = SerialPagePrecedesLogically;
SimpleLruInit(SerialSlruCtl, "Serial",
- NUM_SERIAL_BUFFERS, 0, SerialSLRULock, "pg_serial",
+ serial_buffers, 0, SerialSLRULock, "pg_serial",
LWTRANCHE_SERIAL_BUFFER, SYNC_HANDLER_NONE);
#ifdef USE_ASSERT_CHECKING
SerialPagePrecedesLogicallyUnitTests();
@@ -1395,7 +1395,7 @@ PredicateLockShmemSize(void)
/* Shared memory structures for SLRU tracking of old committed xids. */
size = add_size(size, sizeof(SerialControlData));
- size = add_size(size, SimpleLruShmemSize(NUM_SERIAL_BUFFERS, 0));
+ size = add_size(size, SimpleLruShmemSize(serial_buffers, 0));
return size;
}
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
index 73e0a672ae..e90275be6c 100644
--- a/src/backend/utils/init/globals.c
+++ b/src/backend/utils/init/globals.c
@@ -148,3 +148,11 @@ int64 VacuumPageDirty = 0;
int VacuumCostBalance = 0; /* working state for vacuum */
bool VacuumCostActive = false;
+
+int multixact_offsets_buffers = 8;
+int multixact_members_buffers = 16;
+int subtrans_buffers = 32;
+int notify_buffers = 8;
+int serial_buffers = 16;
+int clog_buffers = 0;
+int commit_ts_buffers = 0;
diff --git a/src/backend/utils/misc/guc.c b/src/backend/utils/misc/guc.c
index 0c5dc4d3e8..003bc820d2 100644
--- a/src/backend/utils/misc/guc.c
+++ b/src/backend/utils/misc/guc.c
@@ -2305,6 +2305,83 @@ static struct config_int ConfigureNamesInt[] =
NULL, NULL, NULL
},
+ {
+ {"multixact_offsets_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for MultiXact offsets SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &multixact_offsets_buffers,
+ 8, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"multixact_members_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for MultiXact members SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &multixact_members_buffers,
+ 16, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"subtrans_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for substransactions SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &subtrans_buffers,
+ 32, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"notify_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for asyncronous notifications SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ ¬ify_buffers,
+ 8, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"serial_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for predicate locks SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &serial_buffers,
+ 16, 2, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"clog_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for commit log SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &clog_buffers,
+ 0, 0, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"commit_ts_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for commit timestamps SLRU."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &commit_ts_buffers,
+ 0, 0, INT_MAX / 2,
+ NULL, NULL, NULL
+ },
+
{
{"temp_buffers", PGC_USERSET, RESOURCES_MEM,
gettext_noop("Sets the maximum number of temporary buffers used by each session."),
diff --git a/src/backend/utils/misc/postgresql.conf.sample b/src/backend/utils/misc/postgresql.conf.sample
index b234a6bfe6..1b8515989b 100644
--- a/src/backend/utils/misc/postgresql.conf.sample
+++ b/src/backend/utils/misc/postgresql.conf.sample
@@ -190,6 +190,15 @@
# (change requires restart)
#backend_flush_after = 0 # measured in pages, 0 disables
+# - SLRU Buffers (change requires restart) -
+
+#clog_buffers = 0 # memory for pg_xact (0 = auto)
+#subtrans_buffers = 32 # memory for pg_subtrans
+#multixact_offsets_buffers = 8 # memory for pg_multixact/offsets
+#multixact_members_buffers = 16 # memory for pg_multixact/members
+#notify_buffers = 8 # memory for pg_nofity
+#serial_buffers = 16 # memory for pg_serial
+#commit_ts_buffers = 0 # memory for pg_commit_ts (0 = auto)
#------------------------------------------------------------------------------
# WRITE-AHEAD LOG
diff --git a/src/include/access/multixact.h b/src/include/access/multixact.h
index 4bbb035eae..97c0a46376 100644
--- a/src/include/access/multixact.h
+++ b/src/include/access/multixact.h
@@ -29,10 +29,6 @@
#define MaxMultiXactOffset ((MultiXactOffset) 0xFFFFFFFF)
-/* Number of SLRU buffers to use for multixact */
-#define NUM_MULTIXACTOFFSET_BUFFERS 8
-#define NUM_MULTIXACTMEMBER_BUFFERS 16
-
/*
* Possible multixact lock modes ("status"). The first four modes are for
* tuple locks (FOR KEY SHARE, FOR SHARE, FOR NO KEY UPDATE, FOR UPDATE); the
diff --git a/src/include/access/subtrans.h b/src/include/access/subtrans.h
index d0ab44ae82..ca0999056e 100644
--- a/src/include/access/subtrans.h
+++ b/src/include/access/subtrans.h
@@ -11,9 +11,6 @@
#ifndef SUBTRANS_H
#define SUBTRANS_H
-/* Number of SLRU buffers to use for subtrans */
-#define NUM_SUBTRANS_BUFFERS 32
-
extern void SubTransSetParent(TransactionId xid, TransactionId parent);
extern TransactionId SubTransGetParent(TransactionId xid);
extern TransactionId SubTransGetTopmostTransaction(TransactionId xid);
diff --git a/src/include/commands/async.h b/src/include/commands/async.h
index 9217f66b91..fa831e3721 100644
--- a/src/include/commands/async.h
+++ b/src/include/commands/async.h
@@ -15,11 +15,6 @@
#include <signal.h>
-/*
- * The number of SLRU page buffers we use for the notification queue.
- */
-#define NUM_NOTIFY_BUFFERS 8
-
extern bool Trace_notify;
extern volatile sig_atomic_t notifyInterruptPending;
diff --git a/src/include/miscadmin.h b/src/include/miscadmin.h
index 013850ac28..9c325b4312 100644
--- a/src/include/miscadmin.h
+++ b/src/include/miscadmin.h
@@ -162,6 +162,13 @@ extern PGDLLIMPORT int MaxBackends;
extern PGDLLIMPORT int MaxConnections;
extern PGDLLIMPORT int max_worker_processes;
extern PGDLLIMPORT int max_parallel_workers;
+extern PGDLLIMPORT int multixact_offsets_buffers;
+extern PGDLLIMPORT int multixact_members_buffers;
+extern PGDLLIMPORT int subtrans_buffers;
+extern PGDLLIMPORT int notify_buffers;
+extern PGDLLIMPORT int serial_buffers;
+extern PGDLLIMPORT int clog_buffers;
+extern PGDLLIMPORT int commit_ts_buffers;
extern PGDLLIMPORT int MyProcPid;
extern PGDLLIMPORT pg_time_t MyStartTime;
diff --git a/src/include/storage/predicate.h b/src/include/storage/predicate.h
index 152b698611..c72779bd88 100644
--- a/src/include/storage/predicate.h
+++ b/src/include/storage/predicate.h
@@ -26,10 +26,6 @@ extern int max_predicate_locks_per_xact;
extern int max_predicate_locks_per_relation;
extern int max_predicate_locks_per_page;
-
-/* Number of SLRU buffers to use for Serial SLRU */
-#define NUM_SERIAL_BUFFERS 16
-
/*
* A handle used for sharing SERIALIZABLEXACT objects between the participants
* in a parallel query.
--
2.30.1
[application/x-patch] v13-0002-Make-simplehash-easy-to-use-in-shmem.patch (5.9K, ../../CA+hUKGKVqrxOp82zER1=XN=yPwV_-OCGAg=ez=1iz9rG+A7Smw@mail.gmail.com/3-v13-0002-Make-simplehash-easy-to-use-in-shmem.patch)
download | inline diff:
From 3441027b9ac2f135dea7ec155503be9d6331dfc1 Mon Sep 17 00:00:00 2001
From: Thomas Munro <[email protected]>
Date: Sat, 27 Mar 2021 08:34:58 +1300
Subject: [PATCH v13 2/5] Make simplehash easy to use in shmem.
Allow "in-place" creation of a simplehash hash table of fixed size,
suitable for use in shared memory. No calling out to allocators, and no
ability to grow.
---
src/include/lib/simplehash.h | 56 +++++++++++++++++++++++++++++++++---
1 file changed, 52 insertions(+), 4 deletions(-)
diff --git a/src/include/lib/simplehash.h b/src/include/lib/simplehash.h
index 395be1ca9a..32d3fa58fe 100644
--- a/src/include/lib/simplehash.h
+++ b/src/include/lib/simplehash.h
@@ -120,6 +120,7 @@
#define SH_ALLOCATE SH_MAKE_NAME(allocate)
#define SH_FREE SH_MAKE_NAME(free)
#define SH_STAT SH_MAKE_NAME(stat)
+#define SH_ESTIMATE_SIZE SH_MAKE_NAME(estimate_size)
/* internal helper functions (no externally visible prototypes) */
#define SH_COMPUTE_PARAMETERS SH_MAKE_NAME(compute_parameters)
@@ -153,16 +154,22 @@ typedef struct SH_TYPE
/* boundary after which to grow hashtable */
uint32 grow_threshold;
+#ifndef SH_IN_PLACE
/* hash buckets */
SH_ELEMENT_TYPE *data;
+#endif
-#ifndef SH_RAW_ALLOCATOR
+#if !defined(SH_RAW_ALLOCATOR) && !defined(SH_IN_PLACE)
/* memory context to use for allocations */
MemoryContext ctx;
#endif
/* user defined data, useful for callbacks */
void *private_data;
+
+#ifdef SH_IN_PLACE
+ SH_ELEMENT_TYPE data[FLEXIBLE_ARRAY_MEMBER];
+#endif
} SH_TYPE;
typedef enum SH_STATUS
@@ -182,6 +189,11 @@ typedef struct SH_ITERATOR
#ifdef SH_RAW_ALLOCATOR
/* <prefix>_hash <prefix>_create(uint32 nelements, void *private_data) */
SH_SCOPE SH_TYPE *SH_CREATE(uint32 nelements, void *private_data);
+#elif defined(SH_IN_PLACE)
+/* size_t <prefix>_estimate_size(uint32 nelements) */
+SH_SCOPE size_t SH_ESTIMATE_SIZE(uint32 nelements);
+/* void <prefix>_create(<prefix>_hash *place, uint32 nelements, void *private_data) */
+SH_SCOPE void SH_CREATE(SH_TYPE *place, uint32 nelements, void *private_data);
#else
/*
* <prefix>_hash <prefix>_create(MemoryContext ctx, uint32 nelements,
@@ -191,14 +203,18 @@ SH_SCOPE SH_TYPE *SH_CREATE(MemoryContext ctx, uint32 nelements,
void *private_data);
#endif
+#ifndef SH_IN_PLACE
/* void <prefix>_destroy(<prefix>_hash *tb) */
SH_SCOPE void SH_DESTROY(SH_TYPE * tb);
+#endif
/* void <prefix>_reset(<prefix>_hash *tb) */
SH_SCOPE void SH_RESET(SH_TYPE * tb);
+#ifndef SH_IN_PLACE
/* void <prefix>_grow(<prefix>_hash *tb) */
SH_SCOPE void SH_GROW(SH_TYPE * tb, uint32 newsize);
+#endif
/* <element> *<prefix>_insert(<prefix>_hash *tb, <key> key, bool *found) */
SH_SCOPE SH_ELEMENT_TYPE *SH_INSERT(SH_TYPE * tb, SH_KEY_TYPE key, bool *found);
@@ -241,7 +257,7 @@ SH_SCOPE void SH_STAT(SH_TYPE * tb);
/* generate implementation of the hash table */
#ifdef SH_DEFINE
-#ifndef SH_RAW_ALLOCATOR
+#if !defined(SH_RAW_ALLOCATOR) && !defined(SH_IN_PLACE)
#include "utils/memutils.h"
#endif
@@ -383,11 +399,13 @@ SH_ENTRY_HASH(SH_TYPE * tb, SH_ELEMENT_TYPE * entry)
#endif
}
+#ifndef SH_IN_PLACE
/* default memory allocator function */
static inline void *SH_ALLOCATE(SH_TYPE * type, Size size);
static inline void SH_FREE(SH_TYPE * type, void *pointer);
+#endif
-#ifndef SH_USE_NONDEFAULT_ALLOCATOR
+#if !defined(SH_USE_NONDEFAULT_ALLOCATOR) && !defined(SH_IN_PLACE)
/* default memory allocator function */
static inline void *
@@ -410,6 +428,22 @@ SH_FREE(SH_TYPE * type, void *pointer)
#endif
+#ifdef SH_IN_PLACE
+/*
+ * Compute the amount of memory required for a fixed sized in-place hash table.
+ */
+SH_SCOPE size_t
+SH_ESTIMATE_SIZE(uint32 nelements)
+{
+ size_t size;
+
+ size = Max(nelements, 2);
+ size = pg_nextpower2_64(size);
+
+ return offsetof(SH_TYPE, data) + sizeof(SH_ELEMENT_TYPE) * size;
+}
+#endif
+
/*
* Create a hash table with enough space for `nelements` distinct members.
* Memory for the hash table is allocated from the passed-in context. If
@@ -422,6 +456,9 @@ SH_FREE(SH_TYPE * type, void *pointer)
#ifdef SH_RAW_ALLOCATOR
SH_SCOPE SH_TYPE *
SH_CREATE(uint32 nelements, void *private_data)
+#elif defined(SH_IN_PLACE)
+SH_SCOPE void
+SH_CREATE(SH_TYPE *place, uint32 nelements, void *private_data)
#else
SH_SCOPE SH_TYPE *
SH_CREATE(MemoryContext ctx, uint32 nelements, void *private_data)
@@ -432,6 +469,8 @@ SH_CREATE(MemoryContext ctx, uint32 nelements, void *private_data)
#ifdef SH_RAW_ALLOCATOR
tb = SH_RAW_ALLOCATOR(sizeof(SH_TYPE));
+#elif defined(SH_IN_PLACE)
+ tb = place;
#else
tb = MemoryContextAllocZero(ctx, sizeof(SH_TYPE));
tb->ctx = ctx;
@@ -443,11 +482,15 @@ SH_CREATE(MemoryContext ctx, uint32 nelements, void *private_data)
SH_COMPUTE_PARAMETERS(tb, size);
+#if defined(SH_IN_PLACE)
+ memset(&tb->data, 0, sizeof(SH_ELEMENT_TYPE) * tb->size);
+#else
tb->data = SH_ALLOCATE(tb, sizeof(SH_ELEMENT_TYPE) * tb->size);
-
return tb;
+#endif
}
+#ifndef SH_IN_PLACE
/* destroy a previously created hash table */
SH_SCOPE void
SH_DESTROY(SH_TYPE * tb)
@@ -455,6 +498,7 @@ SH_DESTROY(SH_TYPE * tb)
SH_FREE(tb, tb->data);
pfree(tb);
}
+#endif
/* reset the contents of a previously created hash table */
SH_SCOPE void
@@ -464,6 +508,7 @@ SH_RESET(SH_TYPE * tb)
tb->members = 0;
}
+#ifndef SH_IN_PLACE
/*
* Grow a hash table to at least `newsize` buckets.
*
@@ -576,6 +621,7 @@ SH_GROW(SH_TYPE * tb, uint32 newsize)
SH_FREE(tb, olddata);
}
+#endif
/*
* This is a separate static inline function, so it can be reliably be inlined
@@ -592,6 +638,7 @@ SH_INSERT_HASH_INTERNAL(SH_TYPE * tb, SH_KEY_TYPE key, uint32 hash, bool *found)
restart:
insertdist = 0;
+#ifndef SH_IN_PLACE
/*
* We do the grow check even if the key is actually present, to avoid
* doing the check inside the loop. This also lets us avoid having to
@@ -614,6 +661,7 @@ restart:
SH_GROW(tb, tb->size * 2);
/* SH_STAT(tb); */
}
+#endif
/* perform insert, start bucket search at optimal location */
data = tb->data;
--
2.30.1
[application/x-patch] v13-0003-Support-intrusive-status-flag-in-simplehash.patch (5.7K, ../../CA+hUKGKVqrxOp82zER1=XN=yPwV_-OCGAg=ez=1iz9rG+A7Smw@mail.gmail.com/4-v13-0003-Support-intrusive-status-flag-in-simplehash.patch)
download | inline diff:
From 3bbaf8d558e1e405217e17d1a502ff68556b21e3 Mon Sep 17 00:00:00 2001
From: Thomas Munro <[email protected]>
Date: Sat, 27 Mar 2021 09:04:56 +1300
Subject: [PATCH v13 3/5] Support intrusive status flag in simplehash.
Before, you had to include a "status" member in the element type, which
simplehash.h could use to detect free space. Allow the user to specify
a special key value to use instead, for more compact representation.
---
src/include/lib/simplehash.h | 66 ++++++++++++++++++++++++++----------
1 file changed, 48 insertions(+), 18 deletions(-)
diff --git a/src/include/lib/simplehash.h b/src/include/lib/simplehash.h
index 32d3fa58fe..05c7ca8a47 100644
--- a/src/include/lib/simplehash.h
+++ b/src/include/lib/simplehash.h
@@ -131,6 +131,8 @@
#define SH_ENTRY_HASH SH_MAKE_NAME(entry_hash)
#define SH_INSERT_HASH_INTERNAL SH_MAKE_NAME(insert_hash_internal)
#define SH_LOOKUP_HASH_INTERNAL SH_MAKE_NAME(lookup_hash_internal)
+#define SH_ENTRY_IS_EMPTY SH_MAKE_NAME(entry_is_empty)
+#define SH_SET_ENTRY_EMPTY SH_MAKE_NAME(set_entry_empty)
/* generate forward declarations necessary to use the hash table */
#ifdef SH_DECLARE
@@ -172,11 +174,13 @@ typedef struct SH_TYPE
#endif
} SH_TYPE;
+#ifndef SH_IS_EMPTY_KEY
typedef enum SH_STATUS
{
SH_STATUS_EMPTY = 0x00,
SH_STATUS_IN_USE = 0x01
} SH_STATUS;
+#endif
typedef struct SH_ITERATOR
{
@@ -309,6 +313,26 @@ SH_SCOPE void SH_STAT(SH_TYPE * tb);
#endif
+static inline bool
+SH_ENTRY_IS_EMPTY(SH_TYPE * tb, SH_ELEMENT_TYPE * entry)
+{
+#ifdef SH_IS_EMPTY_KEY
+ return SH_IS_EMPTY_KEY(tb, entry->SH_KEY);
+#else
+ return entry->status == SH_STATUS_EMPTY;
+#endif
+}
+
+static inline void
+SH_SET_ENTRY_EMPTY(SH_TYPE * tb, SH_ELEMENT_TYPE *entry)
+{
+#ifdef SH_EMPTY_KEY
+ entry->SH_KEY = SH_EMPTY_KEY(tb);
+#else
+ entry->status = SH_STATUS_EMPTY;
+#endif
+}
+
/*
* Compute sizing parameters for hashtable. Called when creating and growing
* the hashtable.
@@ -483,7 +507,7 @@ SH_CREATE(MemoryContext ctx, uint32 nelements, void *private_data)
SH_COMPUTE_PARAMETERS(tb, size);
#if defined(SH_IN_PLACE)
- memset(&tb->data, 0, sizeof(SH_ELEMENT_TYPE) * tb->size);
+ SH_RESET(tb);
#else
tb->data = SH_ALLOCATE(tb, sizeof(SH_ELEMENT_TYPE) * tb->size);
return tb;
@@ -504,7 +528,12 @@ SH_DESTROY(SH_TYPE * tb)
SH_SCOPE void
SH_RESET(SH_TYPE * tb)
{
+#ifdef SH_EMPTY_KEY
+ for (size_t i = 0; i < tb->size; ++i)
+ tb->data[i].SH_KEY = SH_EMPTY_KEY(tb);
+#else
memset(tb->data, 0, sizeof(SH_ELEMENT_TYPE) * tb->size);
+#endif
tb->members = 0;
}
@@ -675,14 +704,17 @@ restart:
SH_ELEMENT_TYPE *entry = &data[curelem];
/* any empty bucket can directly be used */
- if (entry->status == SH_STATUS_EMPTY)
+ if (SH_ENTRY_IS_EMPTY(tb, entry))
{
tb->members++;
entry->SH_KEY = key;
#ifdef SH_STORE_HASH
SH_GET_HASH(tb, entry) = hash;
#endif
+#ifndef SH_IS_EMPTY_KEY
entry->status = SH_STATUS_IN_USE;
+#endif
+ Assert(!SH_ENTRY_IS_EMPTY(tb, entry));
*found = false;
return entry;
}
@@ -697,7 +729,7 @@ restart:
if (SH_COMPARE_KEYS(tb, hash, key, entry))
{
- Assert(entry->status == SH_STATUS_IN_USE);
+ Assert(!SH_ENTRY_IS_EMPTY(tb, entry));
*found = true;
return entry;
}
@@ -721,7 +753,7 @@ restart:
emptyelem = SH_NEXT(tb, emptyelem, startelem);
emptyentry = &data[emptyelem];
- if (emptyentry->status == SH_STATUS_EMPTY)
+ if (SH_ENTRY_IS_EMPTY(tb, emptyentry))
{
lastentry = emptyentry;
break;
@@ -770,7 +802,10 @@ restart:
#ifdef SH_STORE_HASH
SH_GET_HASH(tb, entry) = hash;
#endif
+#ifndef SH_IS_EMPTY_KEY
entry->status = SH_STATUS_IN_USE;
+#endif
+ Assert(!SH_ENTRY_IS_EMPTY(tb, entry));
*found = false;
return entry;
}
@@ -833,12 +868,8 @@ SH_LOOKUP_HASH_INTERNAL(SH_TYPE * tb, SH_KEY_TYPE key, uint32 hash)
{
SH_ELEMENT_TYPE *entry = &tb->data[curelem];
- if (entry->status == SH_STATUS_EMPTY)
- {
+ if (SH_ENTRY_IS_EMPTY(tb, entry))
return NULL;
- }
-
- Assert(entry->status == SH_STATUS_IN_USE);
if (SH_COMPARE_KEYS(tb, hash, key, entry))
return entry;
@@ -891,11 +922,10 @@ SH_DELETE(SH_TYPE * tb, SH_KEY_TYPE key)
{
SH_ELEMENT_TYPE *entry = &tb->data[curelem];
- if (entry->status == SH_STATUS_EMPTY)
+ if (SH_ENTRY_IS_EMPTY(tb, entry))
return false;
- if (entry->status == SH_STATUS_IN_USE &&
- SH_COMPARE_KEYS(tb, hash, key, entry))
+ if (SH_COMPARE_KEYS(tb, hash, key, entry))
{
SH_ELEMENT_TYPE *lastentry = entry;
@@ -917,9 +947,9 @@ SH_DELETE(SH_TYPE * tb, SH_KEY_TYPE key)
curelem = SH_NEXT(tb, curelem, startelem);
curentry = &tb->data[curelem];
- if (curentry->status != SH_STATUS_IN_USE)
+ if (SH_ENTRY_IS_EMPTY(tb, curentry))
{
- lastentry->status = SH_STATUS_EMPTY;
+ SH_SET_ENTRY_EMPTY(tb, lastentry);
break;
}
@@ -929,7 +959,7 @@ SH_DELETE(SH_TYPE * tb, SH_KEY_TYPE key)
/* current is at optimal position, done */
if (curoptimal == curelem)
{
- lastentry->status = SH_STATUS_EMPTY;
+ SH_SET_ENTRY_EMPTY(tb, lastentry);
break;
}
@@ -966,7 +996,7 @@ SH_START_ITERATE(SH_TYPE * tb, SH_ITERATOR * iter)
{
SH_ELEMENT_TYPE *entry = &tb->data[i];
- if (entry->status != SH_STATUS_IN_USE)
+ if (SH_ENTRY_IS_EMPTY(tb, entry))
{
startelem = i;
break;
@@ -1027,7 +1057,7 @@ SH_ITERATE(SH_TYPE * tb, SH_ITERATOR * iter)
if ((iter->cur & tb->sizemask) == (iter->end & tb->sizemask))
iter->done = true;
- if (elem->status == SH_STATUS_IN_USE)
+ if (!SH_ENTRY_IS_EMPTY(tb, elem))
{
return elem;
}
@@ -1063,7 +1093,7 @@ SH_STAT(SH_TYPE * tb)
elem = &tb->data[i];
- if (elem->status != SH_STATUS_IN_USE)
+ if (SH_ENTRY_IS_EMPTY(tb, elem))
continue;
hash = SH_ENTRY_HASH(tb, elem);
--
2.30.1
[application/x-patch] v13-0004-Add-buffer-mapping-table-for-SLRUs.patch (8.5K, ../../CA+hUKGKVqrxOp82zER1=XN=yPwV_-OCGAg=ez=1iz9rG+A7Smw@mail.gmail.com/5-v13-0004-Add-buffer-mapping-table-for-SLRUs.patch)
download | inline diff:
From 970504d304af469e494de6ca83123caf1a683ecf Mon Sep 17 00:00:00 2001
From: Thomas Munro <[email protected]>
Date: Thu, 25 Mar 2021 10:11:31 +1300
Subject: [PATCH v13 4/5] Add buffer mapping table for SLRUs.
Instead of doing a linear search for the buffer holding a given page
number, use a hash table.
Discussion: https://postgr.es/m/2BEC2B3F-9B61-4C1D-9FB5-5FAB0F05EF86%40yandex-team.ru
---
src/backend/access/transam/slru.c | 111 +++++++++++++++++++++++++-----
src/include/access/slru.h | 2 +
2 files changed, 96 insertions(+), 17 deletions(-)
diff --git a/src/backend/access/transam/slru.c b/src/backend/access/transam/slru.c
index 82149ad782..f823c2a0c8 100644
--- a/src/backend/access/transam/slru.c
+++ b/src/backend/access/transam/slru.c
@@ -58,6 +58,8 @@
#include "pgstat.h"
#include "storage/fd.h"
#include "storage/shmem.h"
+#include "utils/dynahash.h"
+#include "utils/hsearch.h"
#define SlruFileName(ctl, path, seg) \
snprintf(path, MAXPGPATH, "%s/%04X", (ctl)->Dir, seg)
@@ -79,6 +81,12 @@ typedef struct SlruWriteAllData
typedef struct SlruWriteAllData *SlruWriteAll;
+typedef struct SlruMappingTableEntry
+{
+ int pageno;
+ int slotno;
+} SlruMappingTableEntry;
+
/*
* Populate a file tag describing a segment file. We only use the segment
* number, since we can derive everything else we need by having separate
@@ -146,6 +154,9 @@ static int SlruSelectLRUPage(SlruCtl ctl, int pageno);
static bool SlruScanDirCbDeleteCutoff(SlruCtl ctl, char *filename,
int segpage, void *data);
static void SlruInternalDeleteSegment(SlruCtl ctl, int segno);
+static void SlruMappingAdd(SlruCtl ctl, int pageno, int slotno);
+static void SlruMappingRemove(SlruCtl ctl, int pageno);
+static int SlruMappingFind(SlruCtl ctl, int pageno);
/*
* Initialization of shared memory
@@ -168,7 +179,8 @@ SimpleLruShmemSize(int nslots, int nlsns)
if (nlsns > 0)
sz += MAXALIGN(nslots * nlsns * sizeof(XLogRecPtr)); /* group_lsn[] */
- return BUFFERALIGN(sz) + BLCKSZ * nslots;
+ return BUFFERALIGN(sz) + BLCKSZ * nslots +
+ hash_estimate_size(nslots, sizeof(SlruMappingTableEntry));
}
/*
@@ -187,6 +199,9 @@ SimpleLruInit(SlruCtl ctl, const char *name, int nslots, int nlsns,
LWLock *ctllock, const char *subdir, int tranche_id,
SyncRequestHandler sync_handler)
{
+ char mapping_table_name[SHMEM_INDEX_KEYSIZE];
+ HASHCTL mapping_table_info;
+ HTAB *mapping_table;
SlruShared shared;
bool found;
@@ -258,11 +273,21 @@ SimpleLruInit(SlruCtl ctl, const char *name, int nslots, int nlsns,
else
Assert(found);
+ /* Create or find the buffer mapping table. */
+ memset(&mapping_table_info, 0, sizeof(mapping_table_info));
+ mapping_table_info.keysize = sizeof(int);
+ mapping_table_info.entrysize = sizeof(SlruMappingTableEntry);
+ snprintf(mapping_table_name, sizeof(mapping_table_name),
+ "%s Mapping Table", name);
+ mapping_table = ShmemInitHash(mapping_table_name, nslots, nslots,
+ &mapping_table_info, HASH_ELEM | HASH_BLOBS);
+
/*
* Initialize the unshared control struct, including directory path. We
* assume caller set PagePrecedes.
*/
ctl->shared = shared;
+ ctl->mapping_table = mapping_table;
ctl->sync_handler = sync_handler;
strlcpy(ctl->Dir, subdir, sizeof(ctl->Dir));
}
@@ -289,6 +314,9 @@ SimpleLruZeroPage(SlruCtl ctl, int pageno)
shared->page_number[slotno] == pageno);
/* Mark the slot as containing this page */
+ if (shared->page_status[slotno] != SLRU_PAGE_EMPTY)
+ SlruMappingRemove(ctl, shared->page_number[slotno]);
+ SlruMappingAdd(ctl, pageno, slotno);
shared->page_number[slotno] = pageno;
shared->page_status[slotno] = SLRU_PAGE_VALID;
shared->page_dirty[slotno] = true;
@@ -362,7 +390,10 @@ SimpleLruWaitIO(SlruCtl ctl, int slotno)
{
/* indeed, the I/O must have failed */
if (shared->page_status[slotno] == SLRU_PAGE_READ_IN_PROGRESS)
+ {
+ SlruMappingRemove(ctl, shared->page_number[slotno]);
shared->page_status[slotno] = SLRU_PAGE_EMPTY;
+ }
else /* write_in_progress */
{
shared->page_status[slotno] = SLRU_PAGE_VALID;
@@ -436,6 +467,9 @@ SimpleLruReadPage(SlruCtl ctl, int pageno, bool write_ok,
!shared->page_dirty[slotno]));
/* Mark the slot read-busy */
+ if (shared->page_status[slotno] != SLRU_PAGE_EMPTY)
+ SlruMappingRemove(ctl, shared->page_number[slotno]);
+ SlruMappingAdd(ctl, pageno, slotno);
shared->page_number[slotno] = pageno;
shared->page_status[slotno] = SLRU_PAGE_READ_IN_PROGRESS;
shared->page_dirty[slotno] = false;
@@ -459,7 +493,13 @@ SimpleLruReadPage(SlruCtl ctl, int pageno, bool write_ok,
shared->page_status[slotno] == SLRU_PAGE_READ_IN_PROGRESS &&
!shared->page_dirty[slotno]);
- shared->page_status[slotno] = ok ? SLRU_PAGE_VALID : SLRU_PAGE_EMPTY;
+ if (ok)
+ shared->page_status[slotno] = SLRU_PAGE_VALID;
+ else
+ {
+ SlruMappingRemove(ctl, pageno);
+ shared->page_status[slotno] = SLRU_PAGE_EMPTY;
+ }
LWLockRelease(&shared->buffer_locks[slotno].lock);
@@ -500,20 +540,20 @@ SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno, TransactionId xid)
LWLockAcquire(shared->ControlLock, LW_SHARED);
/* See if page is already in a buffer */
- for (slotno = 0; slotno < shared->num_slots; slotno++)
+ slotno = SlruMappingFind(ctl, pageno);
+ if (slotno >= 0 &&
+ shared->page_status[slotno] != SLRU_PAGE_READ_IN_PROGRESS)
{
- if (shared->page_number[slotno] == pageno &&
- shared->page_status[slotno] != SLRU_PAGE_EMPTY &&
- shared->page_status[slotno] != SLRU_PAGE_READ_IN_PROGRESS)
- {
- /* See comments for SlruRecentlyUsed macro */
- SlruRecentlyUsed(shared, slotno);
+ Assert(shared->page_status[slotno] != SLRU_PAGE_EMPTY);
+ Assert(shared->page_number[slotno] == pageno);
- /* update the stats counter of pages found in the SLRU */
- pgstat_count_slru_page_hit(shared->slru_stats_idx);
+ /* See comments for SlruRecentlyUsed macro */
+ SlruRecentlyUsed(shared, slotno);
- return slotno;
- }
+ /* update the stats counter of pages found in the SLRU */
+ pgstat_count_slru_page_hit(shared->slru_stats_idx);
+
+ return slotno;
}
/* No luck, so switch to normal exclusive lock and do regular read */
@@ -1029,11 +1069,12 @@ SlruSelectLRUPage(SlruCtl ctl, int pageno)
int best_invalid_page_number = 0; /* keep compiler quiet */
/* See if page already has a buffer assigned */
- for (slotno = 0; slotno < shared->num_slots; slotno++)
+ slotno = SlruMappingFind(ctl, pageno);
+ if (slotno >= 0)
{
- if (shared->page_number[slotno] == pageno &&
- shared->page_status[slotno] != SLRU_PAGE_EMPTY)
- return slotno;
+ Assert(shared->page_number[slotno] == pageno);
+ Assert(shared->page_status[slotno] != SLRU_PAGE_EMPTY);
+ return slotno;
}
/*
@@ -1266,6 +1307,7 @@ restart:;
if (shared->page_status[slotno] == SLRU_PAGE_VALID &&
!shared->page_dirty[slotno])
{
+ SlruMappingRemove(ctl, shared->page_number[slotno]);
shared->page_status[slotno] = SLRU_PAGE_EMPTY;
continue;
}
@@ -1348,6 +1390,7 @@ restart:
if (shared->page_status[slotno] == SLRU_PAGE_VALID &&
!shared->page_dirty[slotno])
{
+ SlruMappingRemove(ctl, shared->page_number[slotno]);
shared->page_status[slotno] = SLRU_PAGE_EMPTY;
continue;
}
@@ -1609,3 +1652,37 @@ SlruSyncFileTag(SlruCtl ctl, const FileTag *ftag, char *path)
errno = save_errno;
return result;
}
+
+static int
+SlruMappingFind(SlruCtl ctl, int pageno)
+{
+ SlruMappingTableEntry *mapping;
+
+ mapping = hash_search(ctl->mapping_table, &pageno, HASH_FIND, NULL);
+ if (mapping)
+ return mapping->slotno;
+
+ return -1;
+}
+
+static void
+SlruMappingAdd(SlruCtl ctl, int pageno, int slotno)
+{
+ SlruMappingTableEntry *mapping;
+ bool found PG_USED_FOR_ASSERTS_ONLY;
+
+ mapping = hash_search(ctl->mapping_table, &pageno, HASH_ENTER, &found);
+ mapping->slotno = slotno;
+
+ Assert(!found);
+}
+
+static void
+SlruMappingRemove(SlruCtl ctl, int pageno)
+{
+ bool found PG_USED_FOR_ASSERTS_ONLY;
+
+ hash_search(ctl->mapping_table, &pageno, HASH_REMOVE, &found);
+
+ Assert(found);
+}
diff --git a/src/include/access/slru.h b/src/include/access/slru.h
index dd52e8cec7..8aa3efc0ee 100644
--- a/src/include/access/slru.h
+++ b/src/include/access/slru.h
@@ -16,6 +16,7 @@
#include "access/xlogdefs.h"
#include "storage/lwlock.h"
#include "storage/sync.h"
+#include "utils/hsearch.h"
/*
@@ -110,6 +111,7 @@ typedef SlruSharedData *SlruShared;
typedef struct SlruCtlData
{
SlruShared shared;
+ HTAB *mapping_table;
/*
* Which sync handler function to use when handing sync requests over to
--
2.30.1
[application/x-patch] v13-0005-fixup-use-simplehash-instead-of-dynahash.patch (4.6K, ../../CA+hUKGKVqrxOp82zER1=XN=yPwV_-OCGAg=ez=1iz9rG+A7Smw@mail.gmail.com/6-v13-0005-fixup-use-simplehash-instead-of-dynahash.patch)
download | inline diff:
From ef3c1e45a1f869544280f65d1eeba8544b028735 Mon Sep 17 00:00:00 2001
From: Thomas Munro <[email protected]>
Date: Sat, 27 Mar 2021 09:07:22 +1300
Subject: [PATCH v13 5/5] fixup: use simplehash instead of dynahash
---
src/backend/access/transam/slru.c | 41 ++++++++++++++++++++-----------
src/include/access/slru.h | 4 ++-
2 files changed, 30 insertions(+), 15 deletions(-)
diff --git a/src/backend/access/transam/slru.c b/src/backend/access/transam/slru.c
index f823c2a0c8..ba9c0efc81 100644
--- a/src/backend/access/transam/slru.c
+++ b/src/backend/access/transam/slru.c
@@ -54,12 +54,11 @@
#include "access/slru.h"
#include "access/transam.h"
#include "access/xlog.h"
+#include "common/hashfn.h"
#include "miscadmin.h"
#include "pgstat.h"
#include "storage/fd.h"
#include "storage/shmem.h"
-#include "utils/dynahash.h"
-#include "utils/hsearch.h"
#define SlruFileName(ctl, path, seg) \
snprintf(path, MAXPGPATH, "%s/%04X", (ctl)->Dir, seg)
@@ -85,8 +84,24 @@ typedef struct SlruMappingTableEntry
{
int pageno;
int slotno;
+ char status;
} SlruMappingTableEntry;
+/* Instantiate specialized hash table routines. */
+#define SH_PREFIX smte
+#define SH_ELEMENT_TYPE SlruMappingTableEntry
+#define SH_KEY_TYPE int
+#define SH_KEY pageno
+#define SH_HASH_KEY(table, key) murmurhash32(key)
+#define SH_EQUAL(table, a, b) a == b
+#define SH_IS_EMPTY_KEY(table, pageno) pageno == -1
+#define SH_EMPTY_KEY(table) -1
+#define SH_DECLARE
+#define SH_DEFINE
+#define SH_SCOPE static inline
+#define SH_IN_PLACE
+#include "lib/simplehash.h"
+
/*
* Populate a file tag describing a segment file. We only use the segment
* number, since we can derive everything else we need by having separate
@@ -179,8 +194,7 @@ SimpleLruShmemSize(int nslots, int nlsns)
if (nlsns > 0)
sz += MAXALIGN(nslots * nlsns * sizeof(XLogRecPtr)); /* group_lsn[] */
- return BUFFERALIGN(sz) + BLCKSZ * nslots +
- hash_estimate_size(nslots, sizeof(SlruMappingTableEntry));
+ return BUFFERALIGN(sz) + BLCKSZ * nslots + smte_estimate_size(nslots);
}
/*
@@ -200,8 +214,7 @@ SimpleLruInit(SlruCtl ctl, const char *name, int nslots, int nlsns,
SyncRequestHandler sync_handler)
{
char mapping_table_name[SHMEM_INDEX_KEYSIZE];
- HASHCTL mapping_table_info;
- HTAB *mapping_table;
+ smte_hash *mapping_table;
SlruShared shared;
bool found;
@@ -274,13 +287,13 @@ SimpleLruInit(SlruCtl ctl, const char *name, int nslots, int nlsns,
Assert(found);
/* Create or find the buffer mapping table. */
- memset(&mapping_table_info, 0, sizeof(mapping_table_info));
- mapping_table_info.keysize = sizeof(int);
- mapping_table_info.entrysize = sizeof(SlruMappingTableEntry);
snprintf(mapping_table_name, sizeof(mapping_table_name),
"%s Mapping Table", name);
- mapping_table = ShmemInitHash(mapping_table_name, nslots, nslots,
- &mapping_table_info, HASH_ELEM | HASH_BLOBS);
+ mapping_table = ShmemInitStruct(mapping_table_name,
+ smte_estimate_size(nslots),
+ &found);
+ if (!found)
+ smte_create(mapping_table, nslots, NULL);
/*
* Initialize the unshared control struct, including directory path. We
@@ -1658,7 +1671,7 @@ SlruMappingFind(SlruCtl ctl, int pageno)
{
SlruMappingTableEntry *mapping;
- mapping = hash_search(ctl->mapping_table, &pageno, HASH_FIND, NULL);
+ mapping = smte_lookup(ctl->mapping_table, pageno);
if (mapping)
return mapping->slotno;
@@ -1671,7 +1684,7 @@ SlruMappingAdd(SlruCtl ctl, int pageno, int slotno)
SlruMappingTableEntry *mapping;
bool found PG_USED_FOR_ASSERTS_ONLY;
- mapping = hash_search(ctl->mapping_table, &pageno, HASH_ENTER, &found);
+ mapping = smte_insert(ctl->mapping_table, pageno, &found);
mapping->slotno = slotno;
Assert(!found);
@@ -1682,7 +1695,7 @@ SlruMappingRemove(SlruCtl ctl, int pageno)
{
bool found PG_USED_FOR_ASSERTS_ONLY;
- hash_search(ctl->mapping_table, &pageno, HASH_REMOVE, &found);
+ found = smte_delete(ctl->mapping_table, pageno);
Assert(found);
}
diff --git a/src/include/access/slru.h b/src/include/access/slru.h
index 8aa3efc0ee..b3d28ff135 100644
--- a/src/include/access/slru.h
+++ b/src/include/access/slru.h
@@ -104,6 +104,8 @@ typedef struct SlruSharedData
typedef SlruSharedData *SlruShared;
+struct smte_hash;
+
/*
* SlruCtlData is an unshared structure that points to the active information
* in shared memory.
@@ -111,7 +113,7 @@ typedef SlruSharedData *SlruShared;
typedef struct SlruCtlData
{
SlruShared shared;
- HTAB *mapping_table;
+ struct smte_hash *mapping_table;
/*
* Which sync handler function to use when handing sync requests over to
--
2.30.1
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2021-03-27 05:31 Andrey Borodin <[email protected]>
parent: Thomas Munro <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Andrey Borodin @ 2021-03-27 05:31 UTC (permalink / raw)
To: Thomas Munro <[email protected]>; +Cc: Gilles Darold <[email protected]>; Tomas Vondra <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
> 27 марта 2021 г., в 01:26, Thomas Munro <[email protected]> написал(а):
>
> On Sat, Mar 27, 2021 at 4:52 AM Andrey Borodin <[email protected]> wrote:
>> Some thoughts on HashTable patch:
>> 1. Can we allocate bigger hashtable to reduce probability of collisions?
>
> Yeah, good idea, might require some study.
In a long run we always have this table filled with nslots. But the keys will be usually consecutive numbers (current working set of CLOG\Multis\etc). So in a happy hashing scenario collisions will only appear for some random backward jumps. I think just size = nslots * 2 will produce results which cannot be improved significantly.
And this reflects original growth strategy SH_GROW(tb, tb->size * 2).
>> 2. Can we use specialised hashtable for this case? I'm afraid hash_search() does comparable number of CPU cycles as simple cycle from 0 to 128. We could inline everything and avoid hashp->hash(keyPtr, hashp->keysize) call. I'm not insisting on special hash though, just an idea.
>
> I tried really hard to not fall into this rabbit h.... [hack hack
> hack], OK, here's a first attempt to use simplehash,
> Andres's
> steampunk macro-based robinhood template
Sounds magnificent.
> that we're already using for
> several other things
I could not find much tests to be sure that we do not break something...
> , and murmurhash which is inlineable and
> branch-free.
I think pageno is a hash already. Why hash any further? And pages accessed together will have smaller access time due to colocation.
> I had to tweak it to support "in-place" creation and
> fixed size (in other words, no allocators, for use in shared memory).
We really need to have a test to know what happens when this structure goes out of memory, as you mentioned below. What would be apropriate place for simplehash tests?
> Then I was annoyed that I had to add a "status" member to our struct,
> so I tried to fix that.
Indeed, sizeof(SlruMappingTableEntry) == 9 seems strange. Will simplehash align it well?
Thanks!
Best regards, Andrey Borodin.
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2021-03-28 21:15 Thomas Munro <[email protected]>
parent: Andrey Borodin <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Thomas Munro @ 2021-03-28 21:15 UTC (permalink / raw)
To: Andrey Borodin <[email protected]>; +Cc: Gilles Darold <[email protected]>; Tomas Vondra <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
On Sat, Mar 27, 2021 at 6:31 PM Andrey Borodin <[email protected]> wrote:
> > 27 марта 2021 г., в 01:26, Thomas Munro <[email protected]> написал(а):
> > , and murmurhash which is inlineable and
> > branch-free.
> I think pageno is a hash already. Why hash any further? And pages accessed together will have smaller access time due to colocation.
Yeah, if clog_buffers is large enough then it's already a "perfect
hash", but if it's not then you might get some weird "harmonic"
effects (not sure if that's the right word), basically higher or lower
collision rate depending on coincidences in the data. If you apply a
hash, the collisions should be evenly spread out so at least it'll be
somewhat consistent. Does that make sense?
(At some point I figured out that the syscaches have lower collision
rates and perform better if you use oids directly instead of hashing
them... but then it's easy to create a pathological pattern of DDL
that turns your hash table into a linked list. Not sure what to think
about that.)
> > I had to tweak it to support "in-place" creation and
> > fixed size (in other words, no allocators, for use in shared memory).
> We really need to have a test to know what happens when this structure goes out of memory, as you mentioned below. What would be apropriate place for simplehash tests?
Good questions. This has to be based on being guaranteed to have
enough space for all of the entries, so the question is really just
"how bad can performance get with different load factors". FWIW there
were some interesting cases with clustering when simplehash was first
used in the executor (see commits ab9f2c42 and parent) which required
some work on hashing quality to fix.
> > Then I was annoyed that I had to add a "status" member to our struct,
> > so I tried to fix that.
> Indeed, sizeof(SlruMappingTableEntry) == 9 seems strange. Will simplehash align it well?
With that "intrusive status" patch, the size is back to 8. But I
think I made a mistake: I made it steal some key space to indicate
presence, but I think the presence test should really get access to
the whole entry so that you can encode it in more ways. For example,
with slotno == -1.
Alright, considering the date, if we want to get this into PostgreSQL
14 it's time to make some decisions.
1. Do we want customisable SLRU sizes in PG14?
+1 from me, we have multiple reports of performance gains from
increasing various different SLRUs, and it's easy to find workloads
that go faster.
One thought: it'd be nice if the user could *see* the current size,
when using the default. SHOW clog_buffers -> 0 isn't very helpful if
you want to increase it, but don't know what it's currently set to.
Not sure off the top of my head how best to do that.
2. What names do we want the GUCs to have? Here's what we have:
Proposed GUC Directory System views
clog_buffers pg_xact Xact
multixact_offsets_buffers pg_multixact/offsets MultiXactOffset
multixact_members_buffers pg_multixact/members MultiXactMember
notify_buffers pg_notify Notify
serial_buffers pg_serial Serial
subtrans_buffers pg_subtrans Subtrans
commit_ts_buffers pg_commit_ts CommitTs
By system views, I mean pg_stat_slru, pg_shmem_allocations and
pg_stat_activity (lock names add "SLRU" on the end).
Observations:
It seems obvious that "clog_buffers" should be renamed to "xact_buffers".
It's not clear whether the multixact GUCs should have the extra "s"
like the directories, or not, like the system views.
It see that we have "Shared Buffer Lookup Table" in
pg_shmem_allocations, so where I generated names like "Subtrans
Mapping Table" I should change that to "Lookup" to match.
3. What recommendations should we make about how to set it?
I think the answer depends partially on the next questions! I think
we should probably at least say something short about the pg_stat_slru
view (cache miss rate) and pg_stat_actitity view (waits on locks), and
how to tell if you might need to increase it. I think this probably
needs a new paragraph, separate from the docs for the individual GUC.
4. Do we want to ship the dynahash patch?
+0.9. The slight hesitation is that it's new code written very late
in the cycle, so it may still have bugs or unintended consequences,
and as you said, at small sizes the linear search must be faster than
the hash computation. Could you help test it, and try to break it?
Can we quantify the scaling effect for some interesting workloads, to
see at what size the dynahash beats the linear search, so that we can
make an informed decision? Of course, without a hash table, large
sizes will surely work badly, so it'd be tempting to restrict the size
you can set the GUC to.
If we do include the dynahash patch, then I think it would also be
reasonable to change the formula for the default, to make it higher on
large systems. The restriction to 128 buffers (= 1MB) doesn't make
much sense on a high frequency OLTP system with 128GB of shared
buffers or even 4GB. I think "unleashing better defaults" would
actually be bigger news than the GUC for typical users, because
they'll just see PG14 use a few extra MB and go faster without having
to learn about these obscure new settings.
5. Do we want to ship the simplehash patch?
-0.5. It's a bit too exciting for the last minute, so I'd be inclined
to wait until the next cycle to do some more research and testing. I
know it's a better idea in the long run.
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2021-03-29 08:26 Andrey Borodin <[email protected]>
parent: Thomas Munro <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Andrey Borodin @ 2021-03-29 08:26 UTC (permalink / raw)
To: Thomas Munro <[email protected]>; +Cc: Gilles Darold <[email protected]>; Tomas Vondra <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
> 29 марта 2021 г., в 02:15, Thomas Munro <[email protected]> написал(а):
>
> On Sat, Mar 27, 2021 at 6:31 PM Andrey Borodin <[email protected]> wrote:
>>> 27 марта 2021 г., в 01:26, Thomas Munro <[email protected]> написал(а):
>>> , and murmurhash which is inlineable and
>>> branch-free.
>
>> I think pageno is a hash already. Why hash any further? And pages accessed together will have smaller access time due to colocation.
>
> Yeah, if clog_buffers is large enough then it's already a "perfect
> hash", but if it's not then you might get some weird "harmonic"
> effects (not sure if that's the right word), basically higher or lower
> collision rate depending on coincidences in the data. If you apply a
> hash, the collisions should be evenly spread out so at least it'll be
> somewhat consistent. Does that make sense?
As far as I understand "Harmonic" effects only make sense if the distribution is unknown. Hash protects from "periodic" data when periods are equal to hash table size. I don't think we need to protect from this case, SLRU data is expected to be localised...
Cost of this protection is necessity to calculate murmur hash on each SLRU lookup. Probably, 10-100ns. Seems like not a big deal.
> (At some point I figured out that the syscaches have lower collision
> rates and perform better if you use oids directly instead of hashing
> them... but then it's easy to create a pathological pattern of DDL
> that turns your hash table into a linked list. Not sure what to think
> about that.)
>
>>> I had to tweak it to support "in-place" creation and
>>> fixed size (in other words, no allocators, for use in shared memory).
>
>> We really need to have a test to know what happens when this structure goes out of memory, as you mentioned below. What would be apropriate place for simplehash tests?
>
> Good questions. This has to be based on being guaranteed to have
> enough space for all of the entries, so the question is really just
> "how bad can performance get with different load factors". FWIW there
> were some interesting cases with clustering when simplehash was first
> used in the executor (see commits ab9f2c42 and parent) which required
> some work on hashing quality to fix.
Interesting read, I didn't know much about simple hash, but seems like there is still many cases where it can be used for good. I always wondered why Postgres uses only Larson's linear hash.
>
>>> Then I was annoyed that I had to add a "status" member to our struct,
>>> so I tried to fix that.
>
>> Indeed, sizeof(SlruMappingTableEntry) == 9 seems strange. Will simplehash align it well?
>
> With that "intrusive status" patch, the size is back to 8. But I
> think I made a mistake: I made it steal some key space to indicate
> presence, but I think the presence test should really get access to
> the whole entry so that you can encode it in more ways. For example,
> with slotno == -1.
>
> Alright, considering the date, if we want to get this into PostgreSQL
> 14 it's time to make some decisions.
>
> 1. Do we want customisable SLRU sizes in PG14?
>
> +1 from me, we have multiple reports of performance gains from
> increasing various different SLRUs, and it's easy to find workloads
> that go faster.
Yes, this is main point of this discussion. So +1 from me too.
>
> One thought: it'd be nice if the user could *see* the current size,
> when using the default. SHOW clog_buffers -> 0 isn't very helpful if
> you want to increase it, but don't know what it's currently set to.
> Not sure off the top of my head how best to do that.
Don't we expect that SHOW command indicate exactly same value as in config or SET command? If this convention does not exist - probably showing effective value is a good idea.
> 2. What names do we want the GUCs to have? Here's what we have:
>
> Proposed GUC Directory System views
> clog_buffers pg_xact Xact
> multixact_offsets_buffers pg_multixact/offsets MultiXactOffset
> multixact_members_buffers pg_multixact/members MultiXactMember
> notify_buffers pg_notify Notify
> serial_buffers pg_serial Serial
> subtrans_buffers pg_subtrans Subtrans
> commit_ts_buffers pg_commit_ts CommitTs
>
> By system views, I mean pg_stat_slru, pg_shmem_allocations and
> pg_stat_activity (lock names add "SLRU" on the end).
>
> Observations:
>
> It seems obvious that "clog_buffers" should be renamed to "xact_buffers".
+1
> It's not clear whether the multixact GUCs should have the extra "s"
> like the directories, or not, like the system views.
I think we show break the ties by native English speaker's ears or typing habits. I'm not a native speaker.
> It see that we have "Shared Buffer Lookup Table" in
> pg_shmem_allocations, so where I generated names like "Subtrans
> Mapping Table" I should change that to "Lookup" to match.
>
> 3. What recommendations should we make about how to set it?
>
> I think the answer depends partially on the next questions! I think
> we should probably at least say something short about the pg_stat_slru
> view (cache miss rate) and pg_stat_actitity view (waits on locks), and
> how to tell if you might need to increase it. I think this probably
> needs a new paragraph, separate from the docs for the individual GUC.
I can only suggest incident-driven approach.
1. Observe ridiculous amount of backends waiting on particular SLRU.
2. Double SLRU buffers for that SLRU.
3. Goto 1.
I don't think we should mention this approach in docs.
> 4. Do we want to ship the dynahash patch?
This patch allows to throw infinite amount of memory on a problem of SLRU waiting for IO. So the scale of improvement is much higher. Do I want that we ship this patch? Definitely. Does this change much? I don't know.
>
> +0.9. The slight hesitation is that it's new code written very late
> in the cycle, so it may still have bugs or unintended consequences,
> and as you said, at small sizes the linear search must be faster than
> the hash computation. Could you help test it, and try to break it?
I'll test it and try to break.
> Can we quantify the scaling effect for some interesting workloads, to
> see at what size the dynahash beats the linear search, so that we can
> make an informed decision?
I think we cannot statistically distinguish linear search from hash search by means of SLRU. But we can create some synthetic benchmarks.
> Of course, without a hash table, large
> sizes will surely work badly, so it'd be tempting to restrict the size
> you can set the GUC to.
>
> If we do include the dynahash patch, then I think it would also be
> reasonable to change the formula for the default, to make it higher on
> large systems. The restriction to 128 buffers (= 1MB) doesn't make
> much sense on a high frequency OLTP system with 128GB of shared
> buffers or even 4GB. I think "unleashing better defaults" would
> actually be bigger news than the GUC for typical users, because
> they'll just see PG14 use a few extra MB and go faster without having
> to learn about these obscure new settings.
I agree. I don't see why we would need to limit buffers to 128 in presence of hash search.
> 5. Do we want to ship the simplehash patch?
>
> -0.5. It's a bit too exciting for the last minute, so I'd be inclined
> to wait until the next cycle to do some more research and testing. I
> know it's a better idea in the long run.
OK, obviously, it's safer decision.
My TODO list:
1. Try to break patch set v13-[0001-0004]
2. Think how to measure performance of linear search versus hash search in SLRU buffer mapping.
Thanks!
Best regards, Andrey Borodin.
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2021-03-31 21:09 Andrey Borodin <[email protected]>
parent: Andrey Borodin <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Andrey Borodin @ 2021-03-31 21:09 UTC (permalink / raw)
To: Thomas Munro <[email protected]>; +Cc: Gilles Darold <[email protected]>; Tomas Vondra <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
> 29 марта 2021 г., в 11:26, Andrey Borodin <[email protected]> написал(а):
>
> My TODO list:
> 1. Try to break patch set v13-[0001-0004]
> 2. Think how to measure performance of linear search versus hash search in SLRU buffer mapping.
Hi Thomas!
I'm still doing my homework. And to this moment all my catch is that "utils/dynahash.h" is not necessary.
I'm thinking about hashtables and measuring performance near optimum of linear search does not seem a good idea now.
It's impossible to prove that difference is statistically significant on all platforms. But even on one platform measurements are just too noisy.
Shared buffers lookup table is indeed very similar to this SLRU lookup table. And it does not try to use more memory than needed. I could not find pgbench-visible impact of growing shared buffer lookup table. Obviously, because it's not a bottleneck on regular workload. And it's hard to guess representative pathological workload.
In fact, this thread started with proposal to use reader-writer lock for multis (instead of exclusive lock), and this proposal encountered same problem. It's very hard to create stable reproduction of pathological workload when this lock is heavily contented. Many people observed the problem, but still there is no open repro.
I bet it will be hard to prove that simplehash is any better then HTAB. But if it is really better, shared buffers could benefit from the same technique.
I think its just fine to use HTAB with normal size, as long as shared buffers do so. But there we allocate slightly more space InitBufTable(NBuffers + NUM_BUFFER_PARTITIONS). Don't we need to allocate nslots + 1 ? It seems that we always do SlruMappingRemove() before SlruMappingAdd() and it is not necessary.
Thanks!
Best regards, Andrey Borodin.
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2021-04-01 03:40 Thomas Munro <[email protected]>
parent: Andrey Borodin <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Thomas Munro @ 2021-04-01 03:40 UTC (permalink / raw)
To: Andrey Borodin <[email protected]>; +Cc: Gilles Darold <[email protected]>; Tomas Vondra <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
On Thu, Apr 1, 2021 at 10:09 AM Andrey Borodin <[email protected]> wrote:
> > 29 марта 2021 г., в 11:26, Andrey Borodin <[email protected]> написал(а):
> > My TODO list:
> > 1. Try to break patch set v13-[0001-0004]
> > 2. Think how to measure performance of linear search versus hash search in SLRU buffer mapping.
>
> Hi Thomas!
> I'm still doing my homework. And to this moment all my catch is that "utils/dynahash.h" is not necessary.
Thanks. Here's a new patch with that fixed, and also:
1. New names ("... Mapping Table" -> "... Lookup Table" in
pg_shmem_allocations view, "clog_buffers" -> "xact_buffers") and a
couple of typos fixed here and there.
2. Remove the cap of 128 buffers for xact_buffers as agreed. We
still need a cap though, to avoid a couple of kinds of overflow inside
slru.c, both when computing the default value and accepting a
user-provided number. I introduced SLRU_MAX_ALLOWED_BUFFERS to keep
it <= 1GB, and tested this on a 32 bit build with extreme block sizes.
Likewise, I removed the cap of 16 buffers for commit_ts_buffers, but
only if you have track_commit_timestamp enabled. It seems silly to
waste 1MB per 1GB of shared_buffers on a feature you're not using. So
the default is capped at 16 in that case to preserve existing
behaviour, but otherwise can be set very large if you want.
I think it's plausible that we'll want to make the multixact sizes
adaptive too, but I that might have to be a job for later. Likewise,
I am sure that substransaction-heavy workloads might be slower than
they need to be due to the current default, but I have not done the
research, With these new GUCs, people are free to experiment and
develop theories about what the defaults should be in v15.
3. In the special case of xact_buffers, there is a lower cap of
512MB, because there's no point trying to cache more xids than there
can be in existence, and that is computed by working backwards from
CLOG_XACTS_PER_PAGE etc. It's not possible to do the same sort of
thing for the other SLRUs without overflow problems, and it doesn't
seem worth trying to fix that right now (1GB of cached commit
timestamps ought to be enough for anyone™, while the theoretical
maximum is 10 bytes for 2b xids = 20GB).
To make this more explicit for people not following our discussion in
detail: with shared_buffers=0...512MB, the behaviour doesn't change.
But for shared_buffers=1GB you'll get twice as many xact_buffers as
today (2MB instead of being capped at 1MB), and it keeps scaling
linearly from there at 0.2%. In other words, all real world databases
will get a boost in this department.
4. Change the default for commit_ts_buffers back to shared_buffers /
1024 (with a minimum of 4), because I think you might have changed it
by a copy and paste error -- or did you intend to make the default
higher?
5. Improve descriptions for the GUCs, visible in pg_settings view, to
match the documentation for related wait events. So "for commit log
SLRU" -> "for the transaction status SLRU cache", and similar
corrections elsewhere. (I am tempted to try to find a better word
than "SLRU", which doesn't seem relevant to users, but for now
consistency is good.)
6. Added a callback so that SHOW xact_buffers and SHOW
commit_ts_buffers display the real size in effect (instead of "0" for
default).
I tried running with xact_buffers=1 and soon saw why you change it to
interpret 1 the same as 0; with 1 you hit buffer starvation and get
stuck. I wish there were some way to say "the default for this GUC is
0, but if it's not zero then it's got to be at least 4". I didn't
study the theoretical basis for the previous minimum value of 4, so I
think we should keep it that way, so that if you say 3 you get 4. I
thought it was better to express that like so:
/* Use configured value if provided. */
if (xact_buffers > 0)
return Max(4, xact_buffers);
return Min(CLOG_MAX_ALLOWED_BUFFERS, Max(4, NBuffers / 512));
> I'm thinking about hashtables and measuring performance near optimum of linear search does not seem a good idea now.
> It's impossible to prove that difference is statistically significant on all platforms. But even on one platform measurements are just too noisy.
>
> Shared buffers lookup table is indeed very similar to this SLRU lookup table. And it does not try to use more memory than needed. I could not find pgbench-visible impact of growing shared buffer lookup table. Obviously, because it's not a bottleneck on regular workload. And it's hard to guess representative pathological workload.
Thanks for testing. I agree that it's a good idea to follow the main
buffer pool's approach for our first version of this. One small
difference is that the SLRU patch performs the hash computation while
it holds the lock. If we computed the hash first and used
hash_search_with_hash_value(), we could compute it before we obtain
the lock, like the main buffer pool.
If we computed the hash value first, we could also ignore the rule in
the documentation for hash_search_with_hash_value() that says that you
must calculate it with get_hash_value(), and just call the hash
function ourselves, so that it's fully inlinable. The same
opportunity exists for the main buffer pool. That'd get you one of
the micro-optimisations that simplehash.h offers. Whether it's worth
bothering with, I don't know.
> In fact, this thread started with proposal to use reader-writer lock for multis (instead of exclusive lock), and this proposal encountered same problem. It's very hard to create stable reproduction of pathological workload when this lock is heavily contented. Many people observed the problem, but still there is no open repro.
>
> I bet it will be hard to prove that simplehash is any better then HTAB. But if it is really better, shared buffers could benefit from the same technique.
Agreed, there are a lot of interesting future projects in this area,
when you compare the main buffer pool, these special buffer pools, and
maybe also the "shared relfilenode" pool patch I have proposed for v15
(CF entry 2933). All have mapping tables and buffer replacement
algorithms (why should they be different?), one has partitions, some
have atomic-based header locks, they interlock with WAL differently
(on page LSN, FPIs and checksum support), ... etc etc.
> I think its just fine to use HTAB with normal size, as long as shared buffers do so. But there we allocate slightly more space InitBufTable(NBuffers + NUM_BUFFER_PARTITIONS). Don't we need to allocate nslots + 1 ? It seems that we always do SlruMappingRemove() before SlruMappingAdd() and it is not necessary.
Yeah, we never try to add more elements than allowed, because we have
a big lock around the mapping. The main buffer mapping table has a
more concurrent design and might temporarily have one extra entry per
partition.
Attachments:
[text/x-patch] v14-0001-Add-a-buffer-mapping-table-for-SLRUs.patch (8.6K, ../../CA+hUKGKEdeVKpOqBxzRF+Z4=j0T+CA7ERrXni5De71Mm6-dBWA@mail.gmail.com/2-v14-0001-Add-a-buffer-mapping-table-for-SLRUs.patch)
download | inline diff:
From 05ce938fef85a5f23407d6977165319650b05d63 Mon Sep 17 00:00:00 2001
From: Thomas Munro <[email protected]>
Date: Thu, 25 Mar 2021 10:11:31 +1300
Subject: [PATCH v14 1/2] Add a buffer mapping table for SLRUs.
Instead of doing a linear search for the buffer holding a given page
number, use a hash table. This will allow us to increase the size of
these caches.
Reviewed-by: Andrey M. Borodin <[email protected]>
Discussion: https://postgr.es/m/2BEC2B3F-9B61-4C1D-9FB5-5FAB0F05EF86%40yandex-team.ru
---
src/backend/access/transam/slru.c | 110 +++++++++++++++++++++++++-----
src/include/access/slru.h | 2 +
2 files changed, 95 insertions(+), 17 deletions(-)
diff --git a/src/backend/access/transam/slru.c b/src/backend/access/transam/slru.c
index 82149ad782..a88e7a38a3 100644
--- a/src/backend/access/transam/slru.c
+++ b/src/backend/access/transam/slru.c
@@ -58,6 +58,7 @@
#include "pgstat.h"
#include "storage/fd.h"
#include "storage/shmem.h"
+#include "utils/hsearch.h"
#define SlruFileName(ctl, path, seg) \
snprintf(path, MAXPGPATH, "%s/%04X", (ctl)->Dir, seg)
@@ -79,6 +80,12 @@ typedef struct SlruWriteAllData
typedef struct SlruWriteAllData *SlruWriteAll;
+typedef struct SlruMappingTableEntry
+{
+ int pageno;
+ int slotno;
+} SlruMappingTableEntry;
+
/*
* Populate a file tag describing a segment file. We only use the segment
* number, since we can derive everything else we need by having separate
@@ -146,6 +153,9 @@ static int SlruSelectLRUPage(SlruCtl ctl, int pageno);
static bool SlruScanDirCbDeleteCutoff(SlruCtl ctl, char *filename,
int segpage, void *data);
static void SlruInternalDeleteSegment(SlruCtl ctl, int segno);
+static void SlruMappingAdd(SlruCtl ctl, int pageno, int slotno);
+static void SlruMappingRemove(SlruCtl ctl, int pageno);
+static int SlruMappingFind(SlruCtl ctl, int pageno);
/*
* Initialization of shared memory
@@ -168,7 +178,8 @@ SimpleLruShmemSize(int nslots, int nlsns)
if (nlsns > 0)
sz += MAXALIGN(nslots * nlsns * sizeof(XLogRecPtr)); /* group_lsn[] */
- return BUFFERALIGN(sz) + BLCKSZ * nslots;
+ return BUFFERALIGN(sz) + BLCKSZ * nslots +
+ hash_estimate_size(nslots, sizeof(SlruMappingTableEntry));
}
/*
@@ -187,6 +198,9 @@ SimpleLruInit(SlruCtl ctl, const char *name, int nslots, int nlsns,
LWLock *ctllock, const char *subdir, int tranche_id,
SyncRequestHandler sync_handler)
{
+ char mapping_table_name[SHMEM_INDEX_KEYSIZE];
+ HASHCTL mapping_table_info;
+ HTAB *mapping_table;
SlruShared shared;
bool found;
@@ -258,11 +272,21 @@ SimpleLruInit(SlruCtl ctl, const char *name, int nslots, int nlsns,
else
Assert(found);
+ /* Create or find the buffer mapping table. */
+ memset(&mapping_table_info, 0, sizeof(mapping_table_info));
+ mapping_table_info.keysize = sizeof(int);
+ mapping_table_info.entrysize = sizeof(SlruMappingTableEntry);
+ snprintf(mapping_table_name, sizeof(mapping_table_name),
+ "%s Lookup Table", name);
+ mapping_table = ShmemInitHash(mapping_table_name, nslots, nslots,
+ &mapping_table_info, HASH_ELEM | HASH_BLOBS);
+
/*
* Initialize the unshared control struct, including directory path. We
* assume caller set PagePrecedes.
*/
ctl->shared = shared;
+ ctl->mapping_table = mapping_table;
ctl->sync_handler = sync_handler;
strlcpy(ctl->Dir, subdir, sizeof(ctl->Dir));
}
@@ -289,6 +313,9 @@ SimpleLruZeroPage(SlruCtl ctl, int pageno)
shared->page_number[slotno] == pageno);
/* Mark the slot as containing this page */
+ if (shared->page_status[slotno] != SLRU_PAGE_EMPTY)
+ SlruMappingRemove(ctl, shared->page_number[slotno]);
+ SlruMappingAdd(ctl, pageno, slotno);
shared->page_number[slotno] = pageno;
shared->page_status[slotno] = SLRU_PAGE_VALID;
shared->page_dirty[slotno] = true;
@@ -362,7 +389,10 @@ SimpleLruWaitIO(SlruCtl ctl, int slotno)
{
/* indeed, the I/O must have failed */
if (shared->page_status[slotno] == SLRU_PAGE_READ_IN_PROGRESS)
+ {
+ SlruMappingRemove(ctl, shared->page_number[slotno]);
shared->page_status[slotno] = SLRU_PAGE_EMPTY;
+ }
else /* write_in_progress */
{
shared->page_status[slotno] = SLRU_PAGE_VALID;
@@ -436,6 +466,9 @@ SimpleLruReadPage(SlruCtl ctl, int pageno, bool write_ok,
!shared->page_dirty[slotno]));
/* Mark the slot read-busy */
+ if (shared->page_status[slotno] != SLRU_PAGE_EMPTY)
+ SlruMappingRemove(ctl, shared->page_number[slotno]);
+ SlruMappingAdd(ctl, pageno, slotno);
shared->page_number[slotno] = pageno;
shared->page_status[slotno] = SLRU_PAGE_READ_IN_PROGRESS;
shared->page_dirty[slotno] = false;
@@ -459,7 +492,13 @@ SimpleLruReadPage(SlruCtl ctl, int pageno, bool write_ok,
shared->page_status[slotno] == SLRU_PAGE_READ_IN_PROGRESS &&
!shared->page_dirty[slotno]);
- shared->page_status[slotno] = ok ? SLRU_PAGE_VALID : SLRU_PAGE_EMPTY;
+ if (ok)
+ shared->page_status[slotno] = SLRU_PAGE_VALID;
+ else
+ {
+ SlruMappingRemove(ctl, pageno);
+ shared->page_status[slotno] = SLRU_PAGE_EMPTY;
+ }
LWLockRelease(&shared->buffer_locks[slotno].lock);
@@ -500,20 +539,20 @@ SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno, TransactionId xid)
LWLockAcquire(shared->ControlLock, LW_SHARED);
/* See if page is already in a buffer */
- for (slotno = 0; slotno < shared->num_slots; slotno++)
+ slotno = SlruMappingFind(ctl, pageno);
+ if (slotno >= 0 &&
+ shared->page_status[slotno] != SLRU_PAGE_READ_IN_PROGRESS)
{
- if (shared->page_number[slotno] == pageno &&
- shared->page_status[slotno] != SLRU_PAGE_EMPTY &&
- shared->page_status[slotno] != SLRU_PAGE_READ_IN_PROGRESS)
- {
- /* See comments for SlruRecentlyUsed macro */
- SlruRecentlyUsed(shared, slotno);
+ Assert(shared->page_status[slotno] != SLRU_PAGE_EMPTY);
+ Assert(shared->page_number[slotno] == pageno);
- /* update the stats counter of pages found in the SLRU */
- pgstat_count_slru_page_hit(shared->slru_stats_idx);
+ /* See comments for SlruRecentlyUsed macro */
+ SlruRecentlyUsed(shared, slotno);
- return slotno;
- }
+ /* update the stats counter of pages found in the SLRU */
+ pgstat_count_slru_page_hit(shared->slru_stats_idx);
+
+ return slotno;
}
/* No luck, so switch to normal exclusive lock and do regular read */
@@ -1029,11 +1068,12 @@ SlruSelectLRUPage(SlruCtl ctl, int pageno)
int best_invalid_page_number = 0; /* keep compiler quiet */
/* See if page already has a buffer assigned */
- for (slotno = 0; slotno < shared->num_slots; slotno++)
+ slotno = SlruMappingFind(ctl, pageno);
+ if (slotno >= 0)
{
- if (shared->page_number[slotno] == pageno &&
- shared->page_status[slotno] != SLRU_PAGE_EMPTY)
- return slotno;
+ Assert(shared->page_number[slotno] == pageno);
+ Assert(shared->page_status[slotno] != SLRU_PAGE_EMPTY);
+ return slotno;
}
/*
@@ -1266,6 +1306,7 @@ restart:;
if (shared->page_status[slotno] == SLRU_PAGE_VALID &&
!shared->page_dirty[slotno])
{
+ SlruMappingRemove(ctl, shared->page_number[slotno]);
shared->page_status[slotno] = SLRU_PAGE_EMPTY;
continue;
}
@@ -1348,6 +1389,7 @@ restart:
if (shared->page_status[slotno] == SLRU_PAGE_VALID &&
!shared->page_dirty[slotno])
{
+ SlruMappingRemove(ctl, shared->page_number[slotno]);
shared->page_status[slotno] = SLRU_PAGE_EMPTY;
continue;
}
@@ -1609,3 +1651,37 @@ SlruSyncFileTag(SlruCtl ctl, const FileTag *ftag, char *path)
errno = save_errno;
return result;
}
+
+static int
+SlruMappingFind(SlruCtl ctl, int pageno)
+{
+ SlruMappingTableEntry *mapping;
+
+ mapping = hash_search(ctl->mapping_table, &pageno, HASH_FIND, NULL);
+ if (mapping)
+ return mapping->slotno;
+
+ return -1;
+}
+
+static void
+SlruMappingAdd(SlruCtl ctl, int pageno, int slotno)
+{
+ SlruMappingTableEntry *mapping;
+ bool found PG_USED_FOR_ASSERTS_ONLY;
+
+ mapping = hash_search(ctl->mapping_table, &pageno, HASH_ENTER, &found);
+ mapping->slotno = slotno;
+
+ Assert(!found);
+}
+
+static void
+SlruMappingRemove(SlruCtl ctl, int pageno)
+{
+ bool found PG_USED_FOR_ASSERTS_ONLY;
+
+ hash_search(ctl->mapping_table, &pageno, HASH_REMOVE, &found);
+
+ Assert(found);
+}
diff --git a/src/include/access/slru.h b/src/include/access/slru.h
index dd52e8cec7..8aa3efc0ee 100644
--- a/src/include/access/slru.h
+++ b/src/include/access/slru.h
@@ -16,6 +16,7 @@
#include "access/xlogdefs.h"
#include "storage/lwlock.h"
#include "storage/sync.h"
+#include "utils/hsearch.h"
/*
@@ -110,6 +111,7 @@ typedef SlruSharedData *SlruShared;
typedef struct SlruCtlData
{
SlruShared shared;
+ HTAB *mapping_table;
/*
* Which sync handler function to use when handing sync requests over to
--
2.30.2
[text/x-patch] v14-0002-Make-all-SLRU-buffer-sizes-configurable.patch (25.7K, ../../CA+hUKGKEdeVKpOqBxzRF+Z4=j0T+CA7ERrXni5De71Mm6-dBWA@mail.gmail.com/3-v14-0002-Make-all-SLRU-buffer-sizes-configurable.patch)
download | inline diff:
From edd07254e3f018ee3afe3f4251ab2cc0406d8b7a Mon Sep 17 00:00:00 2001
From: Andrey Borodin <[email protected]>
Date: Mon, 15 Feb 2021 21:51:56 +0500
Subject: [PATCH v14 2/2] Make all SLRU buffer sizes configurable.
Provide new GUCs to set the number of buffers, instead of using hard
coded defaults.
Remove the limits on xact_buffers and commit_ts_buffers. The default
sizes for those caches are ~0.2% and ~0.1% of shared_buffers, as before,
but now there is no cap at 128 and 16 buffers respectively (unless
track_commit_timestamp is disabled, in the latter case, then we might as
well keep it tiny). Sizes much larger than the old limits have been
shown to be useful on modern systems, and an earlier commit replaced a
linear search with a hash table to avoid problems with extreme cases.
Author: Andrey M. Borodin <[email protected]>
Reviewed-by: Anastasia Lubennikova <[email protected]>
Reviewed-by: Tomas Vondra <[email protected]>
Reviewed-by: Alexander Korotkov <[email protected]>
Reviewed-by: Gilles Darold <[email protected]>
Reviewed-by: Thomas Munro <[email protected]>
Discussion: https://postgr.es/m/2BEC2B3F-9B61-4C1D-9FB5-5FAB0F05EF86%40yandex-team.ru
---
doc/src/sgml/config.sgml | 135 ++++++++++++++++++
src/backend/access/transam/clog.c | 23 ++-
src/backend/access/transam/commit_ts.c | 10 +-
src/backend/access/transam/multixact.c | 8 +-
src/backend/access/transam/subtrans.c | 5 +-
src/backend/commands/async.c | 8 +-
src/backend/storage/lmgr/predicate.c | 4 +-
src/backend/utils/init/globals.c | 8 ++
src/backend/utils/misc/guc.c | 99 +++++++++++++
src/backend/utils/misc/postgresql.conf.sample | 9 ++
src/include/access/clog.h | 10 ++
src/include/access/commit_ts.h | 1 -
src/include/access/multixact.h | 4 -
src/include/access/slru.h | 5 +
src/include/access/subtrans.h | 2 -
src/include/commands/async.h | 5 -
src/include/miscadmin.h | 7 +
src/include/storage/predicate.h | 4 -
18 files changed, 301 insertions(+), 46 deletions(-)
diff --git a/doc/src/sgml/config.sgml b/doc/src/sgml/config.sgml
index 701cb65cc7..58388f144f 100644
--- a/doc/src/sgml/config.sgml
+++ b/doc/src/sgml/config.sgml
@@ -1886,6 +1886,141 @@ include_dir 'conf.d'
</para>
</listitem>
</varlistentry>
+
+ <varlistentry id="guc-multixact-offsets-buffers" xreflabel="multixact_offsets_buffers">
+ <term><varname>multixact_offsets_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>multixact_offsets_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to use to cache the contents
+ of <literal>pg_multixact/offsets</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>8</literal>.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-multixact-members-buffers" xreflabel="multixact_members_buffers">
+ <term><varname>multixact_members_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>multixact_members_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to use to cache the contents
+ of <literal>pg_multixact/members</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>16</literal>.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-subtrans-buffers" xreflabel="subtrans_buffers">
+ <term><varname>subtrans_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>subtrans_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to use to cache the contents
+ of <literal>pg_subtrans</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>8</literal>.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-notify-buffers" xreflabel="notify_buffers">
+ <term><varname>notify_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>notify_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to use to cache the contents
+ of <literal>pg_notify</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>8</literal>.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-serial-buffers" xreflabel="serial_buffers">
+ <term><varname>serial_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>serial_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to use to cache the contents
+ of <literal>pg_serial</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>16</literal>.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-xact-buffers" xreflabel="xact_buffers">
+ <term><varname>xact_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>xact_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to use to cache the contents
+ of <literal>pg_xact</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>0</literal>, which requests
+ <varname>shared_buffers</varname> / 512, but not fewer than 4 blocks.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-commit-ts-buffers" xreflabel="commit_ts_buffers">
+ <term><varname>commit_ts_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>commit_ts_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of memory to use to cache the cotents of
+ <literal>pg_commit_ts</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>0</literal>, which requests
+ <varname>shared_buffers</varname> / 1024, but not fewer than 4 blocks.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
<varlistentry id="guc-max-stack-depth" xreflabel="max_stack_depth">
<term><varname>max_stack_depth</varname> (<type>integer</type>)
diff --git a/src/backend/access/transam/clog.c b/src/backend/access/transam/clog.c
index 6fa4713fb4..dd2d7a5184 100644
--- a/src/backend/access/transam/clog.c
+++ b/src/backend/access/transam/clog.c
@@ -58,8 +58,8 @@
/* We need two bits per xact, so four xacts fit in a byte */
#define CLOG_BITS_PER_XACT 2
-#define CLOG_XACTS_PER_BYTE 4
-#define CLOG_XACTS_PER_PAGE (BLCKSZ * CLOG_XACTS_PER_BYTE)
+StaticAssertDecl((CLOG_BITS_PER_XACT * CLOG_XACTS_PER_BYTE) == BITS_PER_BYTE,
+ "CLOG_BITS_PER_XACT and CLOG_XACTS_PER_BYTE are inconsistent");
#define CLOG_XACT_BITMASK ((1 << CLOG_BITS_PER_XACT) - 1)
#define TransactionIdToPage(xid) ((xid) / (TransactionId) CLOG_XACTS_PER_PAGE)
@@ -659,23 +659,16 @@ TransactionIdGetStatus(TransactionId xid, XLogRecPtr *lsn)
/*
* Number of shared CLOG buffers.
*
- * On larger multi-processor systems, it is possible to have many CLOG page
- * requests in flight at one time which could lead to disk access for CLOG
- * page if the required page is not found in memory. Testing revealed that we
- * can get the best performance by having 128 CLOG buffers, more than that it
- * doesn't improve performance.
- *
- * Unconditionally keeping the number of CLOG buffers to 128 did not seem like
- * a good idea, because it would increase the minimum amount of shared memory
- * required to start, which could be a problem for people running very small
- * configurations. The following formula seems to represent a reasonable
- * compromise: people with very low values for shared_buffers will get fewer
- * CLOG buffers as well, and everyone else will get 128.
+ * By default, we'll use 2MB of for every 1GB of shared buffers, up to the
+ * theoretical maximum useful value, but always at least 4 buffers.
*/
Size
CLOGShmemBuffers(void)
{
- return Min(128, Max(4, NBuffers / 512));
+ /* Use configured value if provided. */
+ if (xact_buffers > 0)
+ return Max(4, xact_buffers);
+ return Min(CLOG_MAX_ALLOWED_BUFFERS, Max(4, NBuffers / 512));
}
/*
diff --git a/src/backend/access/transam/commit_ts.c b/src/backend/access/transam/commit_ts.c
index 268bdba339..729a4b9212 100644
--- a/src/backend/access/transam/commit_ts.c
+++ b/src/backend/access/transam/commit_ts.c
@@ -524,13 +524,17 @@ pg_xact_commit_timestamp_origin(PG_FUNCTION_ARGS)
/*
* Number of shared CommitTS buffers.
*
- * We use a very similar logic as for the number of CLOG buffers; see comments
- * in CLOGShmemBuffers.
+ * By default, we'll use 1MB of for every 1GB of shared buffers, up to the
+ * maximum value that slru.c will allow, but always at least 4 buffers.
*/
Size
CommitTsShmemBuffers(void)
{
- return Min(16, Max(4, NBuffers / 1024));
+ /* Use configured value if provided. */
+ if (commit_ts_buffers > 0)
+ return Max(4, commit_ts_buffers);
+ return Min(track_commit_timestamp ? SLRU_MAX_ALLOWED_BUFFERS : 16,
+ Max(4, NBuffers / 1024));
}
/*
diff --git a/src/backend/access/transam/multixact.c b/src/backend/access/transam/multixact.c
index 1f9f1a1fa1..21787765e2 100644
--- a/src/backend/access/transam/multixact.c
+++ b/src/backend/access/transam/multixact.c
@@ -1831,8 +1831,8 @@ MultiXactShmemSize(void)
mul_size(sizeof(MultiXactId) * 2, MaxOldestSlot))
size = SHARED_MULTIXACT_STATE_SIZE;
- size = add_size(size, SimpleLruShmemSize(NUM_MULTIXACTOFFSET_BUFFERS, 0));
- size = add_size(size, SimpleLruShmemSize(NUM_MULTIXACTMEMBER_BUFFERS, 0));
+ size = add_size(size, SimpleLruShmemSize(multixact_offsets_buffers, 0));
+ size = add_size(size, SimpleLruShmemSize(multixact_members_buffers, 0));
return size;
}
@@ -1848,13 +1848,13 @@ MultiXactShmemInit(void)
MultiXactMemberCtl->PagePrecedes = MultiXactMemberPagePrecedes;
SimpleLruInit(MultiXactOffsetCtl,
- "MultiXactOffset", NUM_MULTIXACTOFFSET_BUFFERS, 0,
+ "MultiXactOffset", multixact_offsets_buffers, 0,
MultiXactOffsetSLRULock, "pg_multixact/offsets",
LWTRANCHE_MULTIXACTOFFSET_BUFFER,
SYNC_HANDLER_MULTIXACT_OFFSET);
SlruPagePrecedesUnitTests(MultiXactOffsetCtl, MULTIXACT_OFFSETS_PER_PAGE);
SimpleLruInit(MultiXactMemberCtl,
- "MultiXactMember", NUM_MULTIXACTMEMBER_BUFFERS, 0,
+ "MultiXactMember", multixact_offsets_buffers, 0,
MultiXactMemberSLRULock, "pg_multixact/members",
LWTRANCHE_MULTIXACTMEMBER_BUFFER,
SYNC_HANDLER_MULTIXACT_MEMBER);
diff --git a/src/backend/access/transam/subtrans.c b/src/backend/access/transam/subtrans.c
index 6a8e521f89..785f2520fd 100644
--- a/src/backend/access/transam/subtrans.c
+++ b/src/backend/access/transam/subtrans.c
@@ -31,6 +31,7 @@
#include "access/slru.h"
#include "access/subtrans.h"
#include "access/transam.h"
+#include "miscadmin.h"
#include "pg_trace.h"
#include "utils/snapmgr.h"
@@ -184,14 +185,14 @@ SubTransGetTopmostTransaction(TransactionId xid)
Size
SUBTRANSShmemSize(void)
{
- return SimpleLruShmemSize(NUM_SUBTRANS_BUFFERS, 0);
+ return SimpleLruShmemSize(subtrans_buffers, 0);
}
void
SUBTRANSShmemInit(void)
{
SubTransCtl->PagePrecedes = SubTransPagePrecedes;
- SimpleLruInit(SubTransCtl, "Subtrans", NUM_SUBTRANS_BUFFERS, 0,
+ SimpleLruInit(SubTransCtl, "Subtrans", subtrans_buffers, 0,
SubtransSLRULock, "pg_subtrans",
LWTRANCHE_SUBTRANS_BUFFER, SYNC_HANDLER_NONE);
SlruPagePrecedesUnitTests(SubTransCtl, SUBTRANS_XACTS_PER_PAGE);
diff --git a/src/backend/commands/async.c b/src/backend/commands/async.c
index 4b16fb5682..de17f52cd7 100644
--- a/src/backend/commands/async.c
+++ b/src/backend/commands/async.c
@@ -107,7 +107,7 @@
* frontend during startup.) The above design guarantees that notifies from
* other backends will never be missed by ignoring self-notifies.
*
- * The amount of shared memory used for notify management (NUM_NOTIFY_BUFFERS)
+ * The amount of shared memory used for notify management (notify_buffers)
* can be varied without affecting anything but performance. The maximum
* amount of notification data that can be queued at one time is determined
* by slru.c's wraparound limit; see QUEUE_MAX_PAGE below.
@@ -225,7 +225,7 @@ typedef struct QueuePosition
*
* Resist the temptation to make this really large. While that would save
* work in some places, it would add cost in others. In particular, this
- * should likely be less than NUM_NOTIFY_BUFFERS, to ensure that backends
+ * should likely be less than notify_buffers, to ensure that backends
* catch up before the pages they'll need to read fall out of SLRU cache.
*/
#define QUEUE_CLEANUP_DELAY 4
@@ -514,7 +514,7 @@ AsyncShmemSize(void)
size = mul_size(MaxBackends + 1, sizeof(QueueBackendStatus));
size = add_size(size, offsetof(AsyncQueueControl, backend));
- size = add_size(size, SimpleLruShmemSize(NUM_NOTIFY_BUFFERS, 0));
+ size = add_size(size, SimpleLruShmemSize(notify_buffers, 0));
return size;
}
@@ -562,7 +562,7 @@ AsyncShmemInit(void)
* Set up SLRU management of the pg_notify data.
*/
NotifyCtl->PagePrecedes = asyncQueuePagePrecedes;
- SimpleLruInit(NotifyCtl, "Notify", NUM_NOTIFY_BUFFERS, 0,
+ SimpleLruInit(NotifyCtl, "Notify", notify_buffers, 0,
NotifySLRULock, "pg_notify", LWTRANCHE_NOTIFY_BUFFER,
SYNC_HANDLER_NONE);
diff --git a/src/backend/storage/lmgr/predicate.c b/src/backend/storage/lmgr/predicate.c
index d493aeef0f..b1f4f1651d 100644
--- a/src/backend/storage/lmgr/predicate.c
+++ b/src/backend/storage/lmgr/predicate.c
@@ -872,7 +872,7 @@ SerialInit(void)
*/
SerialSlruCtl->PagePrecedes = SerialPagePrecedesLogically;
SimpleLruInit(SerialSlruCtl, "Serial",
- NUM_SERIAL_BUFFERS, 0, SerialSLRULock, "pg_serial",
+ serial_buffers, 0, SerialSLRULock, "pg_serial",
LWTRANCHE_SERIAL_BUFFER, SYNC_HANDLER_NONE);
#ifdef USE_ASSERT_CHECKING
SerialPagePrecedesLogicallyUnitTests();
@@ -1395,7 +1395,7 @@ PredicateLockShmemSize(void)
/* Shared memory structures for SLRU tracking of old committed xids. */
size = add_size(size, sizeof(SerialControlData));
- size = add_size(size, SimpleLruShmemSize(NUM_SERIAL_BUFFERS, 0));
+ size = add_size(size, SimpleLruShmemSize(serial_buffers, 0));
return size;
}
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
index 73e0a672ae..3de58042d3 100644
--- a/src/backend/utils/init/globals.c
+++ b/src/backend/utils/init/globals.c
@@ -148,3 +148,11 @@ int64 VacuumPageDirty = 0;
int VacuumCostBalance = 0; /* working state for vacuum */
bool VacuumCostActive = false;
+
+int multixact_offsets_buffers = 8;
+int multixact_members_buffers = 16;
+int subtrans_buffers = 32;
+int notify_buffers = 8;
+int serial_buffers = 16;
+int xact_buffers = 0;
+int commit_ts_buffers = 0;
diff --git a/src/backend/utils/misc/guc.c b/src/backend/utils/misc/guc.c
index 03daec9a08..e05f05ffea 100644
--- a/src/backend/utils/misc/guc.c
+++ b/src/backend/utils/misc/guc.c
@@ -29,9 +29,11 @@
#endif
#include <unistd.h>
+#include "access/clog.h"
#include "access/commit_ts.h"
#include "access/gin.h"
#include "access/rmgr.h"
+#include "access/slru.h"
#include "access/tableam.h"
#include "access/toast_compression.h"
#include "access/transam.h"
@@ -196,6 +198,8 @@ static const char *show_tcp_keepalives_idle(void);
static const char *show_tcp_keepalives_interval(void);
static const char *show_tcp_keepalives_count(void);
static const char *show_tcp_user_timeout(void);
+static const char *show_xact_buffers(void);
+static const char *show_commit_ts_buffers(void);
static bool check_maxconnections(int *newval, void **extra, GucSource source);
static bool check_max_worker_processes(int *newval, void **extra, GucSource source);
static bool check_autovacuum_max_workers(int *newval, void **extra, GucSource source);
@@ -2315,6 +2319,83 @@ static struct config_int ConfigureNamesInt[] =
NULL, NULL, NULL
},
+ {
+ {"multixact_offsets_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for the MultiXact offset SLRU cache."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &multixact_offsets_buffers,
+ 8, 2, SLRU_MAX_ALLOWED_BUFFERS,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"multixact_members_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for the MultiXact member SLRU cache."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &multixact_members_buffers,
+ 16, 2, SLRU_MAX_ALLOWED_BUFFERS,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"subtrans_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for the sub-transaction SLRU cache."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &subtrans_buffers,
+ 32, 2, SLRU_MAX_ALLOWED_BUFFERS,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"notify_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for the NOTIFY message SLRU cache."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ ¬ify_buffers,
+ 8, 2, SLRU_MAX_ALLOWED_BUFFERS,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"serial_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for the serializable transaction SLRU cache."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &serial_buffers,
+ 16, 2, SLRU_MAX_ALLOWED_BUFFERS,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"xact_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for the transaction status SLRU cache."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &xact_buffers,
+ 0, 0, CLOG_MAX_ALLOWED_BUFFERS,
+ NULL, NULL, show_xact_buffers
+ },
+
+ {
+ {"commit_ts_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the size of the dedicated buffer pool used for the commit timestamp SLRU cache."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &commit_ts_buffers,
+ 0, 0, SLRU_MAX_ALLOWED_BUFFERS,
+ NULL, NULL, show_commit_ts_buffers
+ },
+
{
{"temp_buffers", PGC_USERSET, RESOURCES_MEM,
gettext_noop("Sets the maximum number of temporary buffers used by each session."),
@@ -11871,6 +11952,24 @@ show_tcp_user_timeout(void)
return nbuf;
}
+static const char *
+show_xact_buffers(void)
+{
+ static char nbuf[16];
+
+ snprintf(nbuf, sizeof(nbuf), "%zu", CLOGShmemBuffers());
+ return nbuf;
+}
+
+static const char *
+show_commit_ts_buffers(void)
+{
+ static char nbuf[16];
+
+ snprintf(nbuf, sizeof(nbuf), "%zu", CommitTsShmemBuffers());
+ return nbuf;
+}
+
static bool
check_maxconnections(int *newval, void **extra, GucSource source)
{
diff --git a/src/backend/utils/misc/postgresql.conf.sample b/src/backend/utils/misc/postgresql.conf.sample
index 791d39cf07..a380d063da 100644
--- a/src/backend/utils/misc/postgresql.conf.sample
+++ b/src/backend/utils/misc/postgresql.conf.sample
@@ -190,6 +190,15 @@
# (change requires restart)
#backend_flush_after = 0 # measured in pages, 0 disables
+# - SLRU Buffers (change requires restart) -
+
+#xact_buffers = 0 # memory for pg_xact (0 = auto)
+#subtrans_buffers = 32 # memory for pg_subtrans
+#multixact_offsets_buffers = 8 # memory for pg_multixact/offsets
+#multixact_members_buffers = 16 # memory for pg_multixact/members
+#notify_buffers = 8 # memory for pg_notify
+#serial_buffers = 16 # memory for pg_serial
+#commit_ts_buffers = 0 # memory for pg_commit_ts (0 = auto)
#------------------------------------------------------------------------------
# WRITE-AHEAD LOG
diff --git a/src/include/access/clog.h b/src/include/access/clog.h
index 39b8e4afa8..739a292f7f 100644
--- a/src/include/access/clog.h
+++ b/src/include/access/clog.h
@@ -15,6 +15,16 @@
#include "storage/sync.h"
#include "lib/stringinfo.h"
+/*
+ * Don't allow xact_buffers to be set higher than could possibly be useful or
+ * SLRU would allow.
+ */
+#define CLOG_XACTS_PER_BYTE 4
+#define CLOG_XACTS_PER_PAGE (BLCKSZ * CLOG_XACTS_PER_BYTE)
+#define CLOG_MAX_ALLOWED_BUFFERS \
+ Min(SLRU_MAX_ALLOWED_BUFFERS, \
+ (((MaxTransactionId / 2) + (CLOG_XACTS_PER_PAGE - 1)) / CLOG_XACTS_PER_PAGE))
+
/*
* Possible transaction statuses --- note that all-zeroes is the initial
* state.
diff --git a/src/include/access/commit_ts.h b/src/include/access/commit_ts.h
index 750369104a..e4cf988609 100644
--- a/src/include/access/commit_ts.h
+++ b/src/include/access/commit_ts.h
@@ -17,7 +17,6 @@
#include "storage/sync.h"
#include "utils/guc.h"
-
extern PGDLLIMPORT bool track_commit_timestamp;
extern bool check_track_commit_timestamp(bool *newval, void **extra,
diff --git a/src/include/access/multixact.h b/src/include/access/multixact.h
index 4bbb035eae..97c0a46376 100644
--- a/src/include/access/multixact.h
+++ b/src/include/access/multixact.h
@@ -29,10 +29,6 @@
#define MaxMultiXactOffset ((MultiXactOffset) 0xFFFFFFFF)
-/* Number of SLRU buffers to use for multixact */
-#define NUM_MULTIXACTOFFSET_BUFFERS 8
-#define NUM_MULTIXACTMEMBER_BUFFERS 16
-
/*
* Possible multixact lock modes ("status"). The first four modes are for
* tuple locks (FOR KEY SHARE, FOR SHARE, FOR NO KEY UPDATE, FOR UPDATE); the
diff --git a/src/include/access/slru.h b/src/include/access/slru.h
index 8aa3efc0ee..97ea837646 100644
--- a/src/include/access/slru.h
+++ b/src/include/access/slru.h
@@ -18,6 +18,11 @@
#include "storage/sync.h"
#include "utils/hsearch.h"
+/*
+ * To avoid overflowing internal arithmetic and the size_t data type, the
+ * number of buffers should not exceed this number.
+ */
+#define SLRU_MAX_ALLOWED_BUFFERS ((1024 * 1024 * 1024) / BLCKSZ)
/*
* Define SLRU segment size. A page is the same BLCKSZ as is used everywhere
diff --git a/src/include/access/subtrans.h b/src/include/access/subtrans.h
index d0ab44ae82..64fa86938e 100644
--- a/src/include/access/subtrans.h
+++ b/src/include/access/subtrans.h
@@ -11,8 +11,6 @@
#ifndef SUBTRANS_H
#define SUBTRANS_H
-/* Number of SLRU buffers to use for subtrans */
-#define NUM_SUBTRANS_BUFFERS 32
extern void SubTransSetParent(TransactionId xid, TransactionId parent);
extern TransactionId SubTransGetParent(TransactionId xid);
diff --git a/src/include/commands/async.h b/src/include/commands/async.h
index 9217f66b91..fa831e3721 100644
--- a/src/include/commands/async.h
+++ b/src/include/commands/async.h
@@ -15,11 +15,6 @@
#include <signal.h>
-/*
- * The number of SLRU page buffers we use for the notification queue.
- */
-#define NUM_NOTIFY_BUFFERS 8
-
extern bool Trace_notify;
extern volatile sig_atomic_t notifyInterruptPending;
diff --git a/src/include/miscadmin.h b/src/include/miscadmin.h
index 013850ac28..eb2c1f18f3 100644
--- a/src/include/miscadmin.h
+++ b/src/include/miscadmin.h
@@ -162,6 +162,13 @@ extern PGDLLIMPORT int MaxBackends;
extern PGDLLIMPORT int MaxConnections;
extern PGDLLIMPORT int max_worker_processes;
extern PGDLLIMPORT int max_parallel_workers;
+extern PGDLLIMPORT int multixact_offsets_buffers;
+extern PGDLLIMPORT int multixact_members_buffers;
+extern PGDLLIMPORT int subtrans_buffers;
+extern PGDLLIMPORT int notify_buffers;
+extern PGDLLIMPORT int serial_buffers;
+extern PGDLLIMPORT int xact_buffers;
+extern PGDLLIMPORT int commit_ts_buffers;
extern PGDLLIMPORT int MyProcPid;
extern PGDLLIMPORT pg_time_t MyStartTime;
diff --git a/src/include/storage/predicate.h b/src/include/storage/predicate.h
index 152b698611..c72779bd88 100644
--- a/src/include/storage/predicate.h
+++ b/src/include/storage/predicate.h
@@ -26,10 +26,6 @@ extern int max_predicate_locks_per_xact;
extern int max_predicate_locks_per_relation;
extern int max_predicate_locks_per_page;
-
-/* Number of SLRU buffers to use for Serial SLRU */
-#define NUM_SERIAL_BUFFERS 16
-
/*
* A handle used for sharing SERIALIZABLEXACT objects between the participants
* in a parallel query.
--
2.30.2
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2021-04-03 19:57 Andrey Borodin <[email protected]>
parent: Thomas Munro <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Andrey Borodin @ 2021-04-03 19:57 UTC (permalink / raw)
To: Thomas Munro <[email protected]>; +Cc: Gilles Darold <[email protected]>; Tomas Vondra <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
> 1 апр. 2021 г., в 06:40, Thomas Munro <[email protected]> написал(а):
>
> 2. Remove the cap of 128 buffers for xact_buffers as agreed. We
> still need a cap though, to avoid a couple of kinds of overflow inside
> slru.c, both when computing the default value and accepting a
> user-provided number. I introduced SLRU_MAX_ALLOWED_BUFFERS to keep
> it <= 1GB, and tested this on a 32 bit build with extreme block sizes.
BTW we do not document maximum values right now.
I was toying around with big values. For example if we set different big xact_buffers we can get something like
FATAL: not enough shared memory for data structure "Notify" (72768 bytes requested)
FATAL: not enough shared memory for data structure "Async Queue Control" (2492 bytes requested)
FATAL: not enough shared memory for data structure "Checkpointer Data" (393280 bytes requested)
But never anything about xact_buffers. I don't think it's important, though.
>
> Likewise, I removed the cap of 16 buffers for commit_ts_buffers, but
> only if you have track_commit_timestamp enabled.
Is there a reason to leave 16 pages if commit_ts is disabled? They might be useful for some artefacts of previously enabled commit_ts?
> 4. Change the default for commit_ts_buffers back to shared_buffers /
> 1024 (with a minimum of 4), because I think you might have changed it
> by a copy and paste error -- or did you intend to make the default
> higher?
I changed default due to some experiments with https://www.postgresql.org/message-id/flat/20210115220744.GA24457%40alvherre.pgsql
In fact most important part of that thread was removing the cap, which is done by the patchset now.
Thanks!
Best regards, Andrey Borodin.
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2021-04-07 05:59 Thomas Munro <[email protected]>
parent: Andrey Borodin <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Thomas Munro @ 2021-04-07 05:59 UTC (permalink / raw)
To: Andrey Borodin <[email protected]>; +Cc: Gilles Darold <[email protected]>; Tomas Vondra <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
On Sun, Apr 4, 2021 at 7:57 AM Andrey Borodin <[email protected]> wrote:
> I was toying around with big values. For example if we set different big
xact_buffers we can get something like
> FATAL: not enough shared memory for data structure "Notify" (72768 bytes
requested)
> FATAL: not enough shared memory for data structure "Async Queue Control"
(2492 bytes requested)
> FATAL: not enough shared memory for data structure "Checkpointer Data"
(393280 bytes requested)
>
> But never anything about xact_buffers. I don't think it's important,
though.
I had added the hash table size in SimpleLruShmemSize(), but then
SimpleLruInit() passed that same value in when allocating the struct, so
the struct was oversized. Oops. Fixed.
> > Likewise, I removed the cap of 16 buffers for commit_ts_buffers, but
> > only if you have track_commit_timestamp enabled.
> Is there a reason to leave 16 pages if commit_ts is disabled? They might
be useful for some artefacts of previously enabled commit_ts?
Alvaro, do you have an opinion on that?
The remaining thing that bothers me about this patch set is that there is
still a linear search in the replacement algorithm, and it runs with an
exclusive lock. That creates a serious problem for large caches that still
aren't large enough. I wonder if we can do something to improve that
situation in the time we have. I considered a bunch of ideas but could
only find one that fits with slru.c's simplistic locking while tracking
recency. What do you think about a hybrid of SLRU and random replacement,
that retains some characteristics of both? You could think of it as being
a bit like the tournament selection of the genetic algorithm, with a
tournament size of (say) 8 or 16. Any ideas on how to evaluate this and
choose the number? See attached.
Attachments:
[text/x-patch] v15-0001-Add-a-buffer-mapping-table-for-SLRUs.patch (9.0K, ../../CA+hUKGKTESKn9-BkUT8T3vwcgZQJqsV4PSQKpE+M1qUgfSTLMg@mail.gmail.com/3-v15-0001-Add-a-buffer-mapping-table-for-SLRUs.patch)
download | inline diff:
From 5f61bd7d227077f86649a890be603eae01c828f8 Mon Sep 17 00:00:00 2001
From: Thomas Munro <[email protected]>
Date: Thu, 25 Mar 2021 10:11:31 +1300
Subject: [PATCH v15 1/3] Add a buffer mapping table for SLRUs.
Instead of doing a linear search for the buffer holding a given page
number, use a hash table. This will allow us to increase the size of
these caches.
Reviewed-by: Andrey M. Borodin <[email protected]>
Discussion: https://postgr.es/m/2BEC2B3F-9B61-4C1D-9FB5-5FAB0F05EF86%40yandex-team.ru
---
src/backend/access/transam/slru.c | 121 +++++++++++++++++++++++++-----
src/include/access/slru.h | 2 +
2 files changed, 103 insertions(+), 20 deletions(-)
diff --git a/src/backend/access/transam/slru.c b/src/backend/access/transam/slru.c
index 82149ad782..82c61c475b 100644
--- a/src/backend/access/transam/slru.c
+++ b/src/backend/access/transam/slru.c
@@ -58,6 +58,7 @@
#include "pgstat.h"
#include "storage/fd.h"
#include "storage/shmem.h"
+#include "utils/hsearch.h"
#define SlruFileName(ctl, path, seg) \
snprintf(path, MAXPGPATH, "%s/%04X", (ctl)->Dir, seg)
@@ -79,6 +80,12 @@ typedef struct SlruWriteAllData
typedef struct SlruWriteAllData *SlruWriteAll;
+typedef struct SlruMappingTableEntry
+{
+ int pageno;
+ int slotno;
+} SlruMappingTableEntry;
+
/*
* Populate a file tag describing a segment file. We only use the segment
* number, since we can derive everything else we need by having separate
@@ -146,13 +153,16 @@ static int SlruSelectLRUPage(SlruCtl ctl, int pageno);
static bool SlruScanDirCbDeleteCutoff(SlruCtl ctl, char *filename,
int segpage, void *data);
static void SlruInternalDeleteSegment(SlruCtl ctl, int segno);
+static void SlruMappingAdd(SlruCtl ctl, int pageno, int slotno);
+static void SlruMappingRemove(SlruCtl ctl, int pageno);
+static int SlruMappingFind(SlruCtl ctl, int pageno);
/*
* Initialization of shared memory
*/
-Size
-SimpleLruShmemSize(int nslots, int nlsns)
+static Size
+SimpleLruStructSize(int nslots, int nlsns)
{
Size sz;
@@ -167,10 +177,16 @@ SimpleLruShmemSize(int nslots, int nlsns)
if (nlsns > 0)
sz += MAXALIGN(nslots * nlsns * sizeof(XLogRecPtr)); /* group_lsn[] */
-
return BUFFERALIGN(sz) + BLCKSZ * nslots;
}
+Size
+SimpleLruShmemSize(int nslots, int nlsns)
+{
+ return SimpleLruStructSize(nslots, nlsns) +
+ hash_estimate_size(nslots, sizeof(SlruMappingTableEntry));
+}
+
/*
* Initialize, or attach to, a simple LRU cache in shared memory.
*
@@ -187,11 +203,14 @@ SimpleLruInit(SlruCtl ctl, const char *name, int nslots, int nlsns,
LWLock *ctllock, const char *subdir, int tranche_id,
SyncRequestHandler sync_handler)
{
+ char mapping_table_name[SHMEM_INDEX_KEYSIZE];
+ HASHCTL mapping_table_info;
+ HTAB *mapping_table;
SlruShared shared;
bool found;
shared = (SlruShared) ShmemInitStruct(name,
- SimpleLruShmemSize(nslots, nlsns),
+ SimpleLruStructSize(nslots, nlsns),
&found);
if (!IsUnderPostmaster)
@@ -258,11 +277,21 @@ SimpleLruInit(SlruCtl ctl, const char *name, int nslots, int nlsns,
else
Assert(found);
+ /* Create or find the buffer mapping table. */
+ memset(&mapping_table_info, 0, sizeof(mapping_table_info));
+ mapping_table_info.keysize = sizeof(int);
+ mapping_table_info.entrysize = sizeof(SlruMappingTableEntry);
+ snprintf(mapping_table_name, sizeof(mapping_table_name),
+ "%s Lookup Table", name);
+ mapping_table = ShmemInitHash(mapping_table_name, nslots, nslots,
+ &mapping_table_info, HASH_ELEM | HASH_BLOBS);
+
/*
* Initialize the unshared control struct, including directory path. We
* assume caller set PagePrecedes.
*/
ctl->shared = shared;
+ ctl->mapping_table = mapping_table;
ctl->sync_handler = sync_handler;
strlcpy(ctl->Dir, subdir, sizeof(ctl->Dir));
}
@@ -289,6 +318,9 @@ SimpleLruZeroPage(SlruCtl ctl, int pageno)
shared->page_number[slotno] == pageno);
/* Mark the slot as containing this page */
+ if (shared->page_status[slotno] != SLRU_PAGE_EMPTY)
+ SlruMappingRemove(ctl, shared->page_number[slotno]);
+ SlruMappingAdd(ctl, pageno, slotno);
shared->page_number[slotno] = pageno;
shared->page_status[slotno] = SLRU_PAGE_VALID;
shared->page_dirty[slotno] = true;
@@ -362,7 +394,10 @@ SimpleLruWaitIO(SlruCtl ctl, int slotno)
{
/* indeed, the I/O must have failed */
if (shared->page_status[slotno] == SLRU_PAGE_READ_IN_PROGRESS)
+ {
+ SlruMappingRemove(ctl, shared->page_number[slotno]);
shared->page_status[slotno] = SLRU_PAGE_EMPTY;
+ }
else /* write_in_progress */
{
shared->page_status[slotno] = SLRU_PAGE_VALID;
@@ -436,6 +471,9 @@ SimpleLruReadPage(SlruCtl ctl, int pageno, bool write_ok,
!shared->page_dirty[slotno]));
/* Mark the slot read-busy */
+ if (shared->page_status[slotno] != SLRU_PAGE_EMPTY)
+ SlruMappingRemove(ctl, shared->page_number[slotno]);
+ SlruMappingAdd(ctl, pageno, slotno);
shared->page_number[slotno] = pageno;
shared->page_status[slotno] = SLRU_PAGE_READ_IN_PROGRESS;
shared->page_dirty[slotno] = false;
@@ -459,7 +497,13 @@ SimpleLruReadPage(SlruCtl ctl, int pageno, bool write_ok,
shared->page_status[slotno] == SLRU_PAGE_READ_IN_PROGRESS &&
!shared->page_dirty[slotno]);
- shared->page_status[slotno] = ok ? SLRU_PAGE_VALID : SLRU_PAGE_EMPTY;
+ if (ok)
+ shared->page_status[slotno] = SLRU_PAGE_VALID;
+ else
+ {
+ SlruMappingRemove(ctl, pageno);
+ shared->page_status[slotno] = SLRU_PAGE_EMPTY;
+ }
LWLockRelease(&shared->buffer_locks[slotno].lock);
@@ -500,20 +544,20 @@ SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno, TransactionId xid)
LWLockAcquire(shared->ControlLock, LW_SHARED);
/* See if page is already in a buffer */
- for (slotno = 0; slotno < shared->num_slots; slotno++)
+ slotno = SlruMappingFind(ctl, pageno);
+ if (slotno >= 0 &&
+ shared->page_status[slotno] != SLRU_PAGE_READ_IN_PROGRESS)
{
- if (shared->page_number[slotno] == pageno &&
- shared->page_status[slotno] != SLRU_PAGE_EMPTY &&
- shared->page_status[slotno] != SLRU_PAGE_READ_IN_PROGRESS)
- {
- /* See comments for SlruRecentlyUsed macro */
- SlruRecentlyUsed(shared, slotno);
+ Assert(shared->page_status[slotno] != SLRU_PAGE_EMPTY);
+ Assert(shared->page_number[slotno] == pageno);
- /* update the stats counter of pages found in the SLRU */
- pgstat_count_slru_page_hit(shared->slru_stats_idx);
+ /* See comments for SlruRecentlyUsed macro */
+ SlruRecentlyUsed(shared, slotno);
- return slotno;
- }
+ /* update the stats counter of pages found in the SLRU */
+ pgstat_count_slru_page_hit(shared->slru_stats_idx);
+
+ return slotno;
}
/* No luck, so switch to normal exclusive lock and do regular read */
@@ -1029,11 +1073,12 @@ SlruSelectLRUPage(SlruCtl ctl, int pageno)
int best_invalid_page_number = 0; /* keep compiler quiet */
/* See if page already has a buffer assigned */
- for (slotno = 0; slotno < shared->num_slots; slotno++)
+ slotno = SlruMappingFind(ctl, pageno);
+ if (slotno >= 0)
{
- if (shared->page_number[slotno] == pageno &&
- shared->page_status[slotno] != SLRU_PAGE_EMPTY)
- return slotno;
+ Assert(shared->page_number[slotno] == pageno);
+ Assert(shared->page_status[slotno] != SLRU_PAGE_EMPTY);
+ return slotno;
}
/*
@@ -1266,6 +1311,7 @@ restart:;
if (shared->page_status[slotno] == SLRU_PAGE_VALID &&
!shared->page_dirty[slotno])
{
+ SlruMappingRemove(ctl, shared->page_number[slotno]);
shared->page_status[slotno] = SLRU_PAGE_EMPTY;
continue;
}
@@ -1348,6 +1394,7 @@ restart:
if (shared->page_status[slotno] == SLRU_PAGE_VALID &&
!shared->page_dirty[slotno])
{
+ SlruMappingRemove(ctl, shared->page_number[slotno]);
shared->page_status[slotno] = SLRU_PAGE_EMPTY;
continue;
}
@@ -1609,3 +1656,37 @@ SlruSyncFileTag(SlruCtl ctl, const FileTag *ftag, char *path)
errno = save_errno;
return result;
}
+
+static int
+SlruMappingFind(SlruCtl ctl, int pageno)
+{
+ SlruMappingTableEntry *mapping;
+
+ mapping = hash_search(ctl->mapping_table, &pageno, HASH_FIND, NULL);
+ if (mapping)
+ return mapping->slotno;
+
+ return -1;
+}
+
+static void
+SlruMappingAdd(SlruCtl ctl, int pageno, int slotno)
+{
+ SlruMappingTableEntry *mapping;
+ bool found PG_USED_FOR_ASSERTS_ONLY;
+
+ mapping = hash_search(ctl->mapping_table, &pageno, HASH_ENTER, &found);
+ mapping->slotno = slotno;
+
+ Assert(!found);
+}
+
+static void
+SlruMappingRemove(SlruCtl ctl, int pageno)
+{
+ bool found PG_USED_FOR_ASSERTS_ONLY;
+
+ hash_search(ctl->mapping_table, &pageno, HASH_REMOVE, &found);
+
+ Assert(found);
+}
diff --git a/src/include/access/slru.h b/src/include/access/slru.h
index dd52e8cec7..8aa3efc0ee 100644
--- a/src/include/access/slru.h
+++ b/src/include/access/slru.h
@@ -16,6 +16,7 @@
#include "access/xlogdefs.h"
#include "storage/lwlock.h"
#include "storage/sync.h"
+#include "utils/hsearch.h"
/*
@@ -110,6 +111,7 @@ typedef SlruSharedData *SlruShared;
typedef struct SlruCtlData
{
SlruShared shared;
+ HTAB *mapping_table;
/*
* Which sync handler function to use when handing sync requests over to
--
2.27.0
[text/x-patch] v15-0002-Make-all-SLRU-buffer-sizes-configurable.patch (25.7K, ../../CA+hUKGKTESKn9-BkUT8T3vwcgZQJqsV4PSQKpE+M1qUgfSTLMg@mail.gmail.com/4-v15-0002-Make-all-SLRU-buffer-sizes-configurable.patch)
download | inline diff:
From 3462ffe7d740ef0a8d6b217a338fa5cb55a1fbcc Mon Sep 17 00:00:00 2001
From: Andrey Borodin <[email protected]>
Date: Mon, 15 Feb 2021 21:51:56 +0500
Subject: [PATCH v15 2/3] Make all SLRU buffer sizes configurable.
Provide new GUCs to set the number of buffers, instead of using hard
coded defaults.
Remove the limits on xact_buffers and commit_ts_buffers. The default
sizes for those caches are ~0.2% and ~0.1% of shared_buffers, as before,
but now there is no cap at 128 and 16 buffers respectively (unless
track_commit_timestamp is disabled, in the latter case, then we might as
well keep it tiny). Sizes much larger than the old limits have been
shown to be useful on modern systems, and an earlier commit replaced a
linear search with a hash table to avoid problems with extreme cases.
Author: Andrey M. Borodin <[email protected]>
Reviewed-by: Anastasia Lubennikova <[email protected]>
Reviewed-by: Tomas Vondra <[email protected]>
Reviewed-by: Alexander Korotkov <[email protected]>
Reviewed-by: Gilles Darold <[email protected]>
Reviewed-by: Thomas Munro <[email protected]>
Discussion: https://postgr.es/m/2BEC2B3F-9B61-4C1D-9FB5-5FAB0F05EF86%40yandex-team.ru
---
doc/src/sgml/config.sgml | 135 ++++++++++++++++++
src/backend/access/transam/clog.c | 23 ++-
src/backend/access/transam/commit_ts.c | 10 +-
src/backend/access/transam/multixact.c | 8 +-
src/backend/access/transam/subtrans.c | 5 +-
src/backend/commands/async.c | 8 +-
src/backend/storage/lmgr/predicate.c | 4 +-
src/backend/utils/init/globals.c | 8 ++
src/backend/utils/misc/guc.c | 99 +++++++++++++
src/backend/utils/misc/postgresql.conf.sample | 9 ++
src/include/access/clog.h | 10 ++
src/include/access/commit_ts.h | 1 -
src/include/access/multixact.h | 4 -
src/include/access/slru.h | 5 +
src/include/access/subtrans.h | 2 -
src/include/commands/async.h | 5 -
src/include/miscadmin.h | 7 +
src/include/storage/predicate.h | 4 -
18 files changed, 301 insertions(+), 46 deletions(-)
diff --git a/doc/src/sgml/config.sgml b/doc/src/sgml/config.sgml
index e51639d56c..79f46a3df8 100644
--- a/doc/src/sgml/config.sgml
+++ b/doc/src/sgml/config.sgml
@@ -1924,6 +1924,141 @@ include_dir 'conf.d'
</para>
</listitem>
</varlistentry>
+
+ <varlistentry id="guc-multixact-offsets-buffers" xreflabel="multixact_offsets_buffers">
+ <term><varname>multixact_offsets_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>multixact_offsets_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to use to cache the contents
+ of <literal>pg_multixact/offsets</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>8</literal>.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-multixact-members-buffers" xreflabel="multixact_members_buffers">
+ <term><varname>multixact_members_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>multixact_members_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to use to cache the contents
+ of <literal>pg_multixact/members</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>16</literal>.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-subtrans-buffers" xreflabel="subtrans_buffers">
+ <term><varname>subtrans_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>subtrans_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to use to cache the contents
+ of <literal>pg_subtrans</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>8</literal>.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-notify-buffers" xreflabel="notify_buffers">
+ <term><varname>notify_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>notify_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to use to cache the contents
+ of <literal>pg_notify</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>8</literal>.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-serial-buffers" xreflabel="serial_buffers">
+ <term><varname>serial_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>serial_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to use to cache the contents
+ of <literal>pg_serial</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>16</literal>.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-xact-buffers" xreflabel="xact_buffers">
+ <term><varname>xact_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>xact_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to use to cache the contents
+ of <literal>pg_xact</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>0</literal>, which requests
+ <varname>shared_buffers</varname> / 512, but not fewer than 4 blocks.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-commit-ts-buffers" xreflabel="commit_ts_buffers">
+ <term><varname>commit_ts_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>commit_ts_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of memory to use to cache the cotents of
+ <literal>pg_commit_ts</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>0</literal>, which requests
+ <varname>shared_buffers</varname> / 1024, but not fewer than 4 blocks.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
<varlistentry id="guc-max-stack-depth" xreflabel="max_stack_depth">
<term><varname>max_stack_depth</varname> (<type>integer</type>)
diff --git a/src/backend/access/transam/clog.c b/src/backend/access/transam/clog.c
index 6fa4713fb4..dd2d7a5184 100644
--- a/src/backend/access/transam/clog.c
+++ b/src/backend/access/transam/clog.c
@@ -58,8 +58,8 @@
/* We need two bits per xact, so four xacts fit in a byte */
#define CLOG_BITS_PER_XACT 2
-#define CLOG_XACTS_PER_BYTE 4
-#define CLOG_XACTS_PER_PAGE (BLCKSZ * CLOG_XACTS_PER_BYTE)
+StaticAssertDecl((CLOG_BITS_PER_XACT * CLOG_XACTS_PER_BYTE) == BITS_PER_BYTE,
+ "CLOG_BITS_PER_XACT and CLOG_XACTS_PER_BYTE are inconsistent");
#define CLOG_XACT_BITMASK ((1 << CLOG_BITS_PER_XACT) - 1)
#define TransactionIdToPage(xid) ((xid) / (TransactionId) CLOG_XACTS_PER_PAGE)
@@ -659,23 +659,16 @@ TransactionIdGetStatus(TransactionId xid, XLogRecPtr *lsn)
/*
* Number of shared CLOG buffers.
*
- * On larger multi-processor systems, it is possible to have many CLOG page
- * requests in flight at one time which could lead to disk access for CLOG
- * page if the required page is not found in memory. Testing revealed that we
- * can get the best performance by having 128 CLOG buffers, more than that it
- * doesn't improve performance.
- *
- * Unconditionally keeping the number of CLOG buffers to 128 did not seem like
- * a good idea, because it would increase the minimum amount of shared memory
- * required to start, which could be a problem for people running very small
- * configurations. The following formula seems to represent a reasonable
- * compromise: people with very low values for shared_buffers will get fewer
- * CLOG buffers as well, and everyone else will get 128.
+ * By default, we'll use 2MB of for every 1GB of shared buffers, up to the
+ * theoretical maximum useful value, but always at least 4 buffers.
*/
Size
CLOGShmemBuffers(void)
{
- return Min(128, Max(4, NBuffers / 512));
+ /* Use configured value if provided. */
+ if (xact_buffers > 0)
+ return Max(4, xact_buffers);
+ return Min(CLOG_MAX_ALLOWED_BUFFERS, Max(4, NBuffers / 512));
}
/*
diff --git a/src/backend/access/transam/commit_ts.c b/src/backend/access/transam/commit_ts.c
index 268bdba339..729a4b9212 100644
--- a/src/backend/access/transam/commit_ts.c
+++ b/src/backend/access/transam/commit_ts.c
@@ -524,13 +524,17 @@ pg_xact_commit_timestamp_origin(PG_FUNCTION_ARGS)
/*
* Number of shared CommitTS buffers.
*
- * We use a very similar logic as for the number of CLOG buffers; see comments
- * in CLOGShmemBuffers.
+ * By default, we'll use 1MB of for every 1GB of shared buffers, up to the
+ * maximum value that slru.c will allow, but always at least 4 buffers.
*/
Size
CommitTsShmemBuffers(void)
{
- return Min(16, Max(4, NBuffers / 1024));
+ /* Use configured value if provided. */
+ if (commit_ts_buffers > 0)
+ return Max(4, commit_ts_buffers);
+ return Min(track_commit_timestamp ? SLRU_MAX_ALLOWED_BUFFERS : 16,
+ Max(4, NBuffers / 1024));
}
/*
diff --git a/src/backend/access/transam/multixact.c b/src/backend/access/transam/multixact.c
index 1f9f1a1fa1..21787765e2 100644
--- a/src/backend/access/transam/multixact.c
+++ b/src/backend/access/transam/multixact.c
@@ -1831,8 +1831,8 @@ MultiXactShmemSize(void)
mul_size(sizeof(MultiXactId) * 2, MaxOldestSlot))
size = SHARED_MULTIXACT_STATE_SIZE;
- size = add_size(size, SimpleLruShmemSize(NUM_MULTIXACTOFFSET_BUFFERS, 0));
- size = add_size(size, SimpleLruShmemSize(NUM_MULTIXACTMEMBER_BUFFERS, 0));
+ size = add_size(size, SimpleLruShmemSize(multixact_offsets_buffers, 0));
+ size = add_size(size, SimpleLruShmemSize(multixact_members_buffers, 0));
return size;
}
@@ -1848,13 +1848,13 @@ MultiXactShmemInit(void)
MultiXactMemberCtl->PagePrecedes = MultiXactMemberPagePrecedes;
SimpleLruInit(MultiXactOffsetCtl,
- "MultiXactOffset", NUM_MULTIXACTOFFSET_BUFFERS, 0,
+ "MultiXactOffset", multixact_offsets_buffers, 0,
MultiXactOffsetSLRULock, "pg_multixact/offsets",
LWTRANCHE_MULTIXACTOFFSET_BUFFER,
SYNC_HANDLER_MULTIXACT_OFFSET);
SlruPagePrecedesUnitTests(MultiXactOffsetCtl, MULTIXACT_OFFSETS_PER_PAGE);
SimpleLruInit(MultiXactMemberCtl,
- "MultiXactMember", NUM_MULTIXACTMEMBER_BUFFERS, 0,
+ "MultiXactMember", multixact_offsets_buffers, 0,
MultiXactMemberSLRULock, "pg_multixact/members",
LWTRANCHE_MULTIXACTMEMBER_BUFFER,
SYNC_HANDLER_MULTIXACT_MEMBER);
diff --git a/src/backend/access/transam/subtrans.c b/src/backend/access/transam/subtrans.c
index 6a8e521f89..785f2520fd 100644
--- a/src/backend/access/transam/subtrans.c
+++ b/src/backend/access/transam/subtrans.c
@@ -31,6 +31,7 @@
#include "access/slru.h"
#include "access/subtrans.h"
#include "access/transam.h"
+#include "miscadmin.h"
#include "pg_trace.h"
#include "utils/snapmgr.h"
@@ -184,14 +185,14 @@ SubTransGetTopmostTransaction(TransactionId xid)
Size
SUBTRANSShmemSize(void)
{
- return SimpleLruShmemSize(NUM_SUBTRANS_BUFFERS, 0);
+ return SimpleLruShmemSize(subtrans_buffers, 0);
}
void
SUBTRANSShmemInit(void)
{
SubTransCtl->PagePrecedes = SubTransPagePrecedes;
- SimpleLruInit(SubTransCtl, "Subtrans", NUM_SUBTRANS_BUFFERS, 0,
+ SimpleLruInit(SubTransCtl, "Subtrans", subtrans_buffers, 0,
SubtransSLRULock, "pg_subtrans",
LWTRANCHE_SUBTRANS_BUFFER, SYNC_HANDLER_NONE);
SlruPagePrecedesUnitTests(SubTransCtl, SUBTRANS_XACTS_PER_PAGE);
diff --git a/src/backend/commands/async.c b/src/backend/commands/async.c
index 4b16fb5682..de17f52cd7 100644
--- a/src/backend/commands/async.c
+++ b/src/backend/commands/async.c
@@ -107,7 +107,7 @@
* frontend during startup.) The above design guarantees that notifies from
* other backends will never be missed by ignoring self-notifies.
*
- * The amount of shared memory used for notify management (NUM_NOTIFY_BUFFERS)
+ * The amount of shared memory used for notify management (notify_buffers)
* can be varied without affecting anything but performance. The maximum
* amount of notification data that can be queued at one time is determined
* by slru.c's wraparound limit; see QUEUE_MAX_PAGE below.
@@ -225,7 +225,7 @@ typedef struct QueuePosition
*
* Resist the temptation to make this really large. While that would save
* work in some places, it would add cost in others. In particular, this
- * should likely be less than NUM_NOTIFY_BUFFERS, to ensure that backends
+ * should likely be less than notify_buffers, to ensure that backends
* catch up before the pages they'll need to read fall out of SLRU cache.
*/
#define QUEUE_CLEANUP_DELAY 4
@@ -514,7 +514,7 @@ AsyncShmemSize(void)
size = mul_size(MaxBackends + 1, sizeof(QueueBackendStatus));
size = add_size(size, offsetof(AsyncQueueControl, backend));
- size = add_size(size, SimpleLruShmemSize(NUM_NOTIFY_BUFFERS, 0));
+ size = add_size(size, SimpleLruShmemSize(notify_buffers, 0));
return size;
}
@@ -562,7 +562,7 @@ AsyncShmemInit(void)
* Set up SLRU management of the pg_notify data.
*/
NotifyCtl->PagePrecedes = asyncQueuePagePrecedes;
- SimpleLruInit(NotifyCtl, "Notify", NUM_NOTIFY_BUFFERS, 0,
+ SimpleLruInit(NotifyCtl, "Notify", notify_buffers, 0,
NotifySLRULock, "pg_notify", LWTRANCHE_NOTIFY_BUFFER,
SYNC_HANDLER_NONE);
diff --git a/src/backend/storage/lmgr/predicate.c b/src/backend/storage/lmgr/predicate.c
index d493aeef0f..b1f4f1651d 100644
--- a/src/backend/storage/lmgr/predicate.c
+++ b/src/backend/storage/lmgr/predicate.c
@@ -872,7 +872,7 @@ SerialInit(void)
*/
SerialSlruCtl->PagePrecedes = SerialPagePrecedesLogically;
SimpleLruInit(SerialSlruCtl, "Serial",
- NUM_SERIAL_BUFFERS, 0, SerialSLRULock, "pg_serial",
+ serial_buffers, 0, SerialSLRULock, "pg_serial",
LWTRANCHE_SERIAL_BUFFER, SYNC_HANDLER_NONE);
#ifdef USE_ASSERT_CHECKING
SerialPagePrecedesLogicallyUnitTests();
@@ -1395,7 +1395,7 @@ PredicateLockShmemSize(void)
/* Shared memory structures for SLRU tracking of old committed xids. */
size = add_size(size, sizeof(SerialControlData));
- size = add_size(size, SimpleLruShmemSize(NUM_SERIAL_BUFFERS, 0));
+ size = add_size(size, SimpleLruShmemSize(serial_buffers, 0));
return size;
}
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
index 381d9e548d..c83151d5ab 100644
--- a/src/backend/utils/init/globals.c
+++ b/src/backend/utils/init/globals.c
@@ -150,3 +150,11 @@ int64 VacuumPageDirty = 0;
int VacuumCostBalance = 0; /* working state for vacuum */
bool VacuumCostActive = false;
+
+int multixact_offsets_buffers = 8;
+int multixact_members_buffers = 16;
+int subtrans_buffers = 32;
+int notify_buffers = 8;
+int serial_buffers = 16;
+int xact_buffers = 0;
+int commit_ts_buffers = 0;
diff --git a/src/backend/utils/misc/guc.c b/src/backend/utils/misc/guc.c
index c9c9da85f3..d0f0532daa 100644
--- a/src/backend/utils/misc/guc.c
+++ b/src/backend/utils/misc/guc.c
@@ -32,9 +32,11 @@
#endif
#include <unistd.h>
+#include "access/clog.h"
#include "access/commit_ts.h"
#include "access/gin.h"
#include "access/rmgr.h"
+#include "access/slru.h"
#include "access/tableam.h"
#include "access/toast_compression.h"
#include "access/transam.h"
@@ -200,6 +202,8 @@ static const char *show_tcp_keepalives_idle(void);
static const char *show_tcp_keepalives_interval(void);
static const char *show_tcp_keepalives_count(void);
static const char *show_tcp_user_timeout(void);
+static const char *show_xact_buffers(void);
+static const char *show_commit_ts_buffers(void);
static bool check_maxconnections(int *newval, void **extra, GucSource source);
static bool check_max_worker_processes(int *newval, void **extra, GucSource source);
static bool check_autovacuum_max_workers(int *newval, void **extra, GucSource source);
@@ -2330,6 +2334,83 @@ static struct config_int ConfigureNamesInt[] =
NULL, NULL, NULL
},
+ {
+ {"multixact_offsets_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for the MultiXact offset SLRU cache."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &multixact_offsets_buffers,
+ 8, 2, SLRU_MAX_ALLOWED_BUFFERS,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"multixact_members_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for the MultiXact member SLRU cache."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &multixact_members_buffers,
+ 16, 2, SLRU_MAX_ALLOWED_BUFFERS,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"subtrans_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for the sub-transaction SLRU cache."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &subtrans_buffers,
+ 32, 2, SLRU_MAX_ALLOWED_BUFFERS,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"notify_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for the NOTIFY message SLRU cache."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ ¬ify_buffers,
+ 8, 2, SLRU_MAX_ALLOWED_BUFFERS,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"serial_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for the serializable transaction SLRU cache."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &serial_buffers,
+ 16, 2, SLRU_MAX_ALLOWED_BUFFERS,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"xact_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for the transaction status SLRU cache."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &xact_buffers,
+ 0, 0, CLOG_MAX_ALLOWED_BUFFERS,
+ NULL, NULL, show_xact_buffers
+ },
+
+ {
+ {"commit_ts_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the size of the dedicated buffer pool used for the commit timestamp SLRU cache."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &commit_ts_buffers,
+ 0, 0, SLRU_MAX_ALLOWED_BUFFERS,
+ NULL, NULL, show_commit_ts_buffers
+ },
+
{
{"temp_buffers", PGC_USERSET, RESOURCES_MEM,
gettext_noop("Sets the maximum number of temporary buffers used by each session."),
@@ -11897,6 +11978,24 @@ show_tcp_user_timeout(void)
return nbuf;
}
+static const char *
+show_xact_buffers(void)
+{
+ static char nbuf[16];
+
+ snprintf(nbuf, sizeof(nbuf), "%zu", CLOGShmemBuffers());
+ return nbuf;
+}
+
+static const char *
+show_commit_ts_buffers(void)
+{
+ static char nbuf[16];
+
+ snprintf(nbuf, sizeof(nbuf), "%zu", CommitTsShmemBuffers());
+ return nbuf;
+}
+
static bool
check_maxconnections(int *newval, void **extra, GucSource source)
{
diff --git a/src/backend/utils/misc/postgresql.conf.sample b/src/backend/utils/misc/postgresql.conf.sample
index 39da7cc942..5bbbf85923 100644
--- a/src/backend/utils/misc/postgresql.conf.sample
+++ b/src/backend/utils/misc/postgresql.conf.sample
@@ -190,6 +190,15 @@
# (change requires restart)
#backend_flush_after = 0 # measured in pages, 0 disables
+# - SLRU Buffers (change requires restart) -
+
+#xact_buffers = 0 # memory for pg_xact (0 = auto)
+#subtrans_buffers = 32 # memory for pg_subtrans
+#multixact_offsets_buffers = 8 # memory for pg_multixact/offsets
+#multixact_members_buffers = 16 # memory for pg_multixact/members
+#notify_buffers = 8 # memory for pg_notify
+#serial_buffers = 16 # memory for pg_serial
+#commit_ts_buffers = 0 # memory for pg_commit_ts (0 = auto)
#------------------------------------------------------------------------------
# WRITE-AHEAD LOG
diff --git a/src/include/access/clog.h b/src/include/access/clog.h
index 39b8e4afa8..739a292f7f 100644
--- a/src/include/access/clog.h
+++ b/src/include/access/clog.h
@@ -15,6 +15,16 @@
#include "storage/sync.h"
#include "lib/stringinfo.h"
+/*
+ * Don't allow xact_buffers to be set higher than could possibly be useful or
+ * SLRU would allow.
+ */
+#define CLOG_XACTS_PER_BYTE 4
+#define CLOG_XACTS_PER_PAGE (BLCKSZ * CLOG_XACTS_PER_BYTE)
+#define CLOG_MAX_ALLOWED_BUFFERS \
+ Min(SLRU_MAX_ALLOWED_BUFFERS, \
+ (((MaxTransactionId / 2) + (CLOG_XACTS_PER_PAGE - 1)) / CLOG_XACTS_PER_PAGE))
+
/*
* Possible transaction statuses --- note that all-zeroes is the initial
* state.
diff --git a/src/include/access/commit_ts.h b/src/include/access/commit_ts.h
index 750369104a..e4cf988609 100644
--- a/src/include/access/commit_ts.h
+++ b/src/include/access/commit_ts.h
@@ -17,7 +17,6 @@
#include "storage/sync.h"
#include "utils/guc.h"
-
extern PGDLLIMPORT bool track_commit_timestamp;
extern bool check_track_commit_timestamp(bool *newval, void **extra,
diff --git a/src/include/access/multixact.h b/src/include/access/multixact.h
index 4bbb035eae..97c0a46376 100644
--- a/src/include/access/multixact.h
+++ b/src/include/access/multixact.h
@@ -29,10 +29,6 @@
#define MaxMultiXactOffset ((MultiXactOffset) 0xFFFFFFFF)
-/* Number of SLRU buffers to use for multixact */
-#define NUM_MULTIXACTOFFSET_BUFFERS 8
-#define NUM_MULTIXACTMEMBER_BUFFERS 16
-
/*
* Possible multixact lock modes ("status"). The first four modes are for
* tuple locks (FOR KEY SHARE, FOR SHARE, FOR NO KEY UPDATE, FOR UPDATE); the
diff --git a/src/include/access/slru.h b/src/include/access/slru.h
index 8aa3efc0ee..97ea837646 100644
--- a/src/include/access/slru.h
+++ b/src/include/access/slru.h
@@ -18,6 +18,11 @@
#include "storage/sync.h"
#include "utils/hsearch.h"
+/*
+ * To avoid overflowing internal arithmetic and the size_t data type, the
+ * number of buffers should not exceed this number.
+ */
+#define SLRU_MAX_ALLOWED_BUFFERS ((1024 * 1024 * 1024) / BLCKSZ)
/*
* Define SLRU segment size. A page is the same BLCKSZ as is used everywhere
diff --git a/src/include/access/subtrans.h b/src/include/access/subtrans.h
index d0ab44ae82..64fa86938e 100644
--- a/src/include/access/subtrans.h
+++ b/src/include/access/subtrans.h
@@ -11,8 +11,6 @@
#ifndef SUBTRANS_H
#define SUBTRANS_H
-/* Number of SLRU buffers to use for subtrans */
-#define NUM_SUBTRANS_BUFFERS 32
extern void SubTransSetParent(TransactionId xid, TransactionId parent);
extern TransactionId SubTransGetParent(TransactionId xid);
diff --git a/src/include/commands/async.h b/src/include/commands/async.h
index 9217f66b91..fa831e3721 100644
--- a/src/include/commands/async.h
+++ b/src/include/commands/async.h
@@ -15,11 +15,6 @@
#include <signal.h>
-/*
- * The number of SLRU page buffers we use for the notification queue.
- */
-#define NUM_NOTIFY_BUFFERS 8
-
extern bool Trace_notify;
extern volatile sig_atomic_t notifyInterruptPending;
diff --git a/src/include/miscadmin.h b/src/include/miscadmin.h
index 95202d37af..495c1bf901 100644
--- a/src/include/miscadmin.h
+++ b/src/include/miscadmin.h
@@ -164,6 +164,13 @@ extern PGDLLIMPORT int MaxBackends;
extern PGDLLIMPORT int MaxConnections;
extern PGDLLIMPORT int max_worker_processes;
extern PGDLLIMPORT int max_parallel_workers;
+extern PGDLLIMPORT int multixact_offsets_buffers;
+extern PGDLLIMPORT int multixact_members_buffers;
+extern PGDLLIMPORT int subtrans_buffers;
+extern PGDLLIMPORT int notify_buffers;
+extern PGDLLIMPORT int serial_buffers;
+extern PGDLLIMPORT int xact_buffers;
+extern PGDLLIMPORT int commit_ts_buffers;
extern PGDLLIMPORT int MyProcPid;
extern PGDLLIMPORT pg_time_t MyStartTime;
diff --git a/src/include/storage/predicate.h b/src/include/storage/predicate.h
index 152b698611..c72779bd88 100644
--- a/src/include/storage/predicate.h
+++ b/src/include/storage/predicate.h
@@ -26,10 +26,6 @@ extern int max_predicate_locks_per_xact;
extern int max_predicate_locks_per_relation;
extern int max_predicate_locks_per_page;
-
-/* Number of SLRU buffers to use for Serial SLRU */
-#define NUM_SERIAL_BUFFERS 16
-
/*
* A handle used for sharing SERIALIZABLEXACT objects between the participants
* in a parallel query.
--
2.27.0
[text/x-patch] v15-0003-Use-hybrid-random-SLRU-replacement-for-SLRUs.patch (3.3K, ../../CA+hUKGKTESKn9-BkUT8T3vwcgZQJqsV4PSQKpE+M1qUgfSTLMg@mail.gmail.com/5-v15-0003-Use-hybrid-random-SLRU-replacement-for-SLRUs.patch)
download | inline diff:
From cea3142af73c434d11206dd39a7edaa8718269b4 Mon Sep 17 00:00:00 2001
From: Thomas Munro <[email protected]>
Date: Wed, 7 Apr 2021 16:39:09 +1200
Subject: [PATCH v15 3/3] Use hybrid random/SLRU replacement for SLRUs.
The previous algorithm would search the entire SLRU to find the least
recently used page. That doesn't seem like a good idea now that the
user can set SLRUs to large sizes. Instead, apply the existing algorithm
to a small number of randomly chosen buffers, for a cheap hybrid of SLRU
and random replacement that doesn't require a complete rewrite of
slru.c's locking.
XXX experiment
---
src/backend/access/transam/slru.c | 22 +++++++++++++++++-----
1 file changed, 17 insertions(+), 5 deletions(-)
diff --git a/src/backend/access/transam/slru.c b/src/backend/access/transam/slru.c
index 82c61c475b..95e404e856 100644
--- a/src/backend/access/transam/slru.c
+++ b/src/backend/access/transam/slru.c
@@ -71,6 +71,14 @@
*/
#define MAX_WRITEALL_BUFFERS 16
+/*
+ * We need a fast way to choose a buffer to replace, because we hold a coarse
+ * grained exclusive lock while searching. As an improvement over simple
+ * random replacement that still has some LRU tendancies, we'll compare the
+ * recency of a smallish number of randomly selected buffers.
+ */
+#define MAX_RANDOM_SEARCH 8
+
typedef struct SlruWriteAllData
{
int num_files; /* # files actually open */
@@ -1071,6 +1079,7 @@ SlruSelectLRUPage(SlruCtl ctl, int pageno)
int bestinvalidslot = 0; /* keep compiler quiet */
int best_invalid_delta = -1;
int best_invalid_page_number = 0; /* keep compiler quiet */
+ int candidates = Min(shared->num_slots, MAX_RANDOM_SEARCH);
/* See if page already has a buffer assigned */
slotno = SlruMappingFind(ctl, pageno);
@@ -1082,9 +1091,10 @@ SlruSelectLRUPage(SlruCtl ctl, int pageno)
}
/*
- * If we find any EMPTY slot, just select that one. Else choose a
- * victim page to replace. We normally take the least recently used
- * valid page, but we will never take the slot containing
+ * If we find any EMPTY slot, just select that one. Else choose a
+ * victim page to replace. We take the least recently used of a small
+ * number of randomly chosen pages, to avoid having to scan a
+ * potentially large number of pages. We skip the slot containing
* latest_page_number, even if it appears least recently used. We
* will select a slot that is already I/O busy only if there is no
* other choice: a read-busy slot will not be least recently used once
@@ -1109,11 +1119,13 @@ SlruSelectLRUPage(SlruCtl ctl, int pageno)
* multiple pages with the same lru_count.
*/
cur_count = (shared->cur_lru_count)++;
- for (slotno = 0; slotno < shared->num_slots; slotno++)
+ for (int i = 0; i < candidates; ++i)
{
int this_delta;
int this_page_number;
+retry:
+ slotno = (random() & LONG_MAX) % shared->num_slots;
if (shared->page_status[slotno] == SLRU_PAGE_EMPTY)
return slotno;
this_delta = cur_count - shared->page_lru_count[slotno];
@@ -1131,7 +1143,7 @@ SlruSelectLRUPage(SlruCtl ctl, int pageno)
}
this_page_number = shared->page_number[slotno];
if (this_page_number == shared->latest_page_number)
- continue;
+ goto retry;
if (shared->page_status[slotno] == SLRU_PAGE_VALID)
{
if (this_delta > best_valid_delta ||
--
2.27.0
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2021-04-07 11:44 Andrey Borodin <[email protected]>
parent: Thomas Munro <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Andrey Borodin @ 2021-04-07 11:44 UTC (permalink / raw)
To: Thomas Munro <[email protected]>; +Cc: Gilles Darold <[email protected]>; Tomas Vondra <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
> 7 апр. 2021 г., в 08:59, Thomas Munro <[email protected]> написал(а):
>
> The remaining thing that bothers me about this patch set is that there is still a linear search in the replacement algorithm, and it runs with an exclusive lock. That creates a serious problem for large caches that still aren't large enough. I wonder if we can do something to improve that situation in the time we have. I considered a bunch of ideas but could only find one that fits with slru.c's simplistic locking while tracking recency. What do you think about a hybrid of SLRU and random replacement, that retains some characteristics of both? You could think of it as being a bit like the tournament selection of the genetic algorithm, with a tournament size of (say) 8 or 16. Any ideas on how to evaluate this and choose the number? See attached.
> <v15-0001-Add-a-buffer-mapping-table-for-SLRUs.patch><v15-0002-Make-all-SLRU-buffer-sizes-configurable.patch><v15-0003-Use-hybrid-random-SLRU-replacement-for-SLRUs.patch>
Maybe instead of fully associative cache with random replacement we could use 1-associative cache?
i.e. each page can reside only in one spcific buffer slot. If there's something else - evict it.
I think this would be as efficient as RR cache. And it's soooo fast.
Best regards, Andrey Borodin.
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2021-04-07 12:13 Andrey Borodin <[email protected]>
parent: Andrey Borodin <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Andrey Borodin @ 2021-04-07 12:13 UTC (permalink / raw)
To: Thomas Munro <[email protected]>; +Cc: Gilles Darold <[email protected]>; Tomas Vondra <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
> 7 апр. 2021 г., в 14:44, Andrey Borodin <[email protected]> написал(а):
>
> Maybe instead of fully associative cache with random replacement we could use 1-associative cache?
> i.e. each page can reside only in one spcific buffer slot. If there's something else - evict it.
> I think this would be as efficient as RR cache. And it's soooo fast.
I thought a bit more and understood that RR is protected from two competing pages in working set, while 1-associative cache is not. So, discard that idea.
Best regards, Andrey Borodin.
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2021-04-08 00:30 Thomas Munro <[email protected]>
parent: Andrey Borodin <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Thomas Munro @ 2021-04-08 00:30 UTC (permalink / raw)
To: Andrey Borodin <[email protected]>; +Cc: Gilles Darold <[email protected]>; Tomas Vondra <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
On Thu, Apr 8, 2021 at 12:13 AM Andrey Borodin <[email protected]> wrote:
> > 7 апр. 2021 г., в 14:44, Andrey Borodin <[email protected]> написал(а):
> > Maybe instead of fully associative cache with random replacement we could use 1-associative cache?
> > i.e. each page can reside only in one spcific buffer slot. If there's something else - evict it.
> > I think this would be as efficient as RR cache. And it's soooo fast.
>
> I thought a bit more and understood that RR is protected from two competing pages in working set, while 1-associative cache is not. So, discard that idea.
It's an interesting idea. I know that at least one proprietary fork
just puts the whole CLOG in memory for direct indexing, which is what
we'd have here if we said "oh, your xact_buffers setting is so large
I'm just going to use slotno = pageno & mask".
Here's another approach that is a little less exciting than
"tournament RR" (or whatever that should be called; I couldn't find an
established name for it). This version is just our traditional linear
search, except that it stops at 128, and remembers where to start from
next time (like a sort of Fisher-Price GCLOCK hand). This feels more
committable to me. You can argue that all buffers above 128 are bonus
buffers that PostgreSQL 13 didn't have, so the fact that we can no
longer find the globally least recently used page when you set
xact_buffers > 128 doesn't seem too bad to me, as an incremental step
(but to be clear, of course we can do better than this with more work
in later releases).
Attachments:
[text/x-patch] v16-0001-Add-a-buffer-mapping-table-for-SLRUs.patch (9.0K, ../../CA+hUKGLCLDtgDj2Xsf0uBk5WXDCeHxBDDJPsyY7m65Fde-=pyg@mail.gmail.com/2-v16-0001-Add-a-buffer-mapping-table-for-SLRUs.patch)
download | inline diff:
From 72ebd5052851aa4b9aa281df30b9bf42c0ad5de4 Mon Sep 17 00:00:00 2001
From: Thomas Munro <[email protected]>
Date: Thu, 25 Mar 2021 10:11:31 +1300
Subject: [PATCH v16 1/3] Add a buffer mapping table for SLRUs.
Instead of doing a linear search for the buffer holding a given page
number, use a hash table. This will allow us to increase the size of
these caches.
Reviewed-by: Andrey M. Borodin <[email protected]>
Discussion: https://postgr.es/m/2BEC2B3F-9B61-4C1D-9FB5-5FAB0F05EF86%40yandex-team.ru
---
src/backend/access/transam/slru.c | 121 +++++++++++++++++++++++++-----
src/include/access/slru.h | 2 +
2 files changed, 103 insertions(+), 20 deletions(-)
diff --git a/src/backend/access/transam/slru.c b/src/backend/access/transam/slru.c
index 82149ad782..82c61c475b 100644
--- a/src/backend/access/transam/slru.c
+++ b/src/backend/access/transam/slru.c
@@ -58,6 +58,7 @@
#include "pgstat.h"
#include "storage/fd.h"
#include "storage/shmem.h"
+#include "utils/hsearch.h"
#define SlruFileName(ctl, path, seg) \
snprintf(path, MAXPGPATH, "%s/%04X", (ctl)->Dir, seg)
@@ -79,6 +80,12 @@ typedef struct SlruWriteAllData
typedef struct SlruWriteAllData *SlruWriteAll;
+typedef struct SlruMappingTableEntry
+{
+ int pageno;
+ int slotno;
+} SlruMappingTableEntry;
+
/*
* Populate a file tag describing a segment file. We only use the segment
* number, since we can derive everything else we need by having separate
@@ -146,13 +153,16 @@ static int SlruSelectLRUPage(SlruCtl ctl, int pageno);
static bool SlruScanDirCbDeleteCutoff(SlruCtl ctl, char *filename,
int segpage, void *data);
static void SlruInternalDeleteSegment(SlruCtl ctl, int segno);
+static void SlruMappingAdd(SlruCtl ctl, int pageno, int slotno);
+static void SlruMappingRemove(SlruCtl ctl, int pageno);
+static int SlruMappingFind(SlruCtl ctl, int pageno);
/*
* Initialization of shared memory
*/
-Size
-SimpleLruShmemSize(int nslots, int nlsns)
+static Size
+SimpleLruStructSize(int nslots, int nlsns)
{
Size sz;
@@ -167,10 +177,16 @@ SimpleLruShmemSize(int nslots, int nlsns)
if (nlsns > 0)
sz += MAXALIGN(nslots * nlsns * sizeof(XLogRecPtr)); /* group_lsn[] */
-
return BUFFERALIGN(sz) + BLCKSZ * nslots;
}
+Size
+SimpleLruShmemSize(int nslots, int nlsns)
+{
+ return SimpleLruStructSize(nslots, nlsns) +
+ hash_estimate_size(nslots, sizeof(SlruMappingTableEntry));
+}
+
/*
* Initialize, or attach to, a simple LRU cache in shared memory.
*
@@ -187,11 +203,14 @@ SimpleLruInit(SlruCtl ctl, const char *name, int nslots, int nlsns,
LWLock *ctllock, const char *subdir, int tranche_id,
SyncRequestHandler sync_handler)
{
+ char mapping_table_name[SHMEM_INDEX_KEYSIZE];
+ HASHCTL mapping_table_info;
+ HTAB *mapping_table;
SlruShared shared;
bool found;
shared = (SlruShared) ShmemInitStruct(name,
- SimpleLruShmemSize(nslots, nlsns),
+ SimpleLruStructSize(nslots, nlsns),
&found);
if (!IsUnderPostmaster)
@@ -258,11 +277,21 @@ SimpleLruInit(SlruCtl ctl, const char *name, int nslots, int nlsns,
else
Assert(found);
+ /* Create or find the buffer mapping table. */
+ memset(&mapping_table_info, 0, sizeof(mapping_table_info));
+ mapping_table_info.keysize = sizeof(int);
+ mapping_table_info.entrysize = sizeof(SlruMappingTableEntry);
+ snprintf(mapping_table_name, sizeof(mapping_table_name),
+ "%s Lookup Table", name);
+ mapping_table = ShmemInitHash(mapping_table_name, nslots, nslots,
+ &mapping_table_info, HASH_ELEM | HASH_BLOBS);
+
/*
* Initialize the unshared control struct, including directory path. We
* assume caller set PagePrecedes.
*/
ctl->shared = shared;
+ ctl->mapping_table = mapping_table;
ctl->sync_handler = sync_handler;
strlcpy(ctl->Dir, subdir, sizeof(ctl->Dir));
}
@@ -289,6 +318,9 @@ SimpleLruZeroPage(SlruCtl ctl, int pageno)
shared->page_number[slotno] == pageno);
/* Mark the slot as containing this page */
+ if (shared->page_status[slotno] != SLRU_PAGE_EMPTY)
+ SlruMappingRemove(ctl, shared->page_number[slotno]);
+ SlruMappingAdd(ctl, pageno, slotno);
shared->page_number[slotno] = pageno;
shared->page_status[slotno] = SLRU_PAGE_VALID;
shared->page_dirty[slotno] = true;
@@ -362,7 +394,10 @@ SimpleLruWaitIO(SlruCtl ctl, int slotno)
{
/* indeed, the I/O must have failed */
if (shared->page_status[slotno] == SLRU_PAGE_READ_IN_PROGRESS)
+ {
+ SlruMappingRemove(ctl, shared->page_number[slotno]);
shared->page_status[slotno] = SLRU_PAGE_EMPTY;
+ }
else /* write_in_progress */
{
shared->page_status[slotno] = SLRU_PAGE_VALID;
@@ -436,6 +471,9 @@ SimpleLruReadPage(SlruCtl ctl, int pageno, bool write_ok,
!shared->page_dirty[slotno]));
/* Mark the slot read-busy */
+ if (shared->page_status[slotno] != SLRU_PAGE_EMPTY)
+ SlruMappingRemove(ctl, shared->page_number[slotno]);
+ SlruMappingAdd(ctl, pageno, slotno);
shared->page_number[slotno] = pageno;
shared->page_status[slotno] = SLRU_PAGE_READ_IN_PROGRESS;
shared->page_dirty[slotno] = false;
@@ -459,7 +497,13 @@ SimpleLruReadPage(SlruCtl ctl, int pageno, bool write_ok,
shared->page_status[slotno] == SLRU_PAGE_READ_IN_PROGRESS &&
!shared->page_dirty[slotno]);
- shared->page_status[slotno] = ok ? SLRU_PAGE_VALID : SLRU_PAGE_EMPTY;
+ if (ok)
+ shared->page_status[slotno] = SLRU_PAGE_VALID;
+ else
+ {
+ SlruMappingRemove(ctl, pageno);
+ shared->page_status[slotno] = SLRU_PAGE_EMPTY;
+ }
LWLockRelease(&shared->buffer_locks[slotno].lock);
@@ -500,20 +544,20 @@ SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno, TransactionId xid)
LWLockAcquire(shared->ControlLock, LW_SHARED);
/* See if page is already in a buffer */
- for (slotno = 0; slotno < shared->num_slots; slotno++)
+ slotno = SlruMappingFind(ctl, pageno);
+ if (slotno >= 0 &&
+ shared->page_status[slotno] != SLRU_PAGE_READ_IN_PROGRESS)
{
- if (shared->page_number[slotno] == pageno &&
- shared->page_status[slotno] != SLRU_PAGE_EMPTY &&
- shared->page_status[slotno] != SLRU_PAGE_READ_IN_PROGRESS)
- {
- /* See comments for SlruRecentlyUsed macro */
- SlruRecentlyUsed(shared, slotno);
+ Assert(shared->page_status[slotno] != SLRU_PAGE_EMPTY);
+ Assert(shared->page_number[slotno] == pageno);
- /* update the stats counter of pages found in the SLRU */
- pgstat_count_slru_page_hit(shared->slru_stats_idx);
+ /* See comments for SlruRecentlyUsed macro */
+ SlruRecentlyUsed(shared, slotno);
- return slotno;
- }
+ /* update the stats counter of pages found in the SLRU */
+ pgstat_count_slru_page_hit(shared->slru_stats_idx);
+
+ return slotno;
}
/* No luck, so switch to normal exclusive lock and do regular read */
@@ -1029,11 +1073,12 @@ SlruSelectLRUPage(SlruCtl ctl, int pageno)
int best_invalid_page_number = 0; /* keep compiler quiet */
/* See if page already has a buffer assigned */
- for (slotno = 0; slotno < shared->num_slots; slotno++)
+ slotno = SlruMappingFind(ctl, pageno);
+ if (slotno >= 0)
{
- if (shared->page_number[slotno] == pageno &&
- shared->page_status[slotno] != SLRU_PAGE_EMPTY)
- return slotno;
+ Assert(shared->page_number[slotno] == pageno);
+ Assert(shared->page_status[slotno] != SLRU_PAGE_EMPTY);
+ return slotno;
}
/*
@@ -1266,6 +1311,7 @@ restart:;
if (shared->page_status[slotno] == SLRU_PAGE_VALID &&
!shared->page_dirty[slotno])
{
+ SlruMappingRemove(ctl, shared->page_number[slotno]);
shared->page_status[slotno] = SLRU_PAGE_EMPTY;
continue;
}
@@ -1348,6 +1394,7 @@ restart:
if (shared->page_status[slotno] == SLRU_PAGE_VALID &&
!shared->page_dirty[slotno])
{
+ SlruMappingRemove(ctl, shared->page_number[slotno]);
shared->page_status[slotno] = SLRU_PAGE_EMPTY;
continue;
}
@@ -1609,3 +1656,37 @@ SlruSyncFileTag(SlruCtl ctl, const FileTag *ftag, char *path)
errno = save_errno;
return result;
}
+
+static int
+SlruMappingFind(SlruCtl ctl, int pageno)
+{
+ SlruMappingTableEntry *mapping;
+
+ mapping = hash_search(ctl->mapping_table, &pageno, HASH_FIND, NULL);
+ if (mapping)
+ return mapping->slotno;
+
+ return -1;
+}
+
+static void
+SlruMappingAdd(SlruCtl ctl, int pageno, int slotno)
+{
+ SlruMappingTableEntry *mapping;
+ bool found PG_USED_FOR_ASSERTS_ONLY;
+
+ mapping = hash_search(ctl->mapping_table, &pageno, HASH_ENTER, &found);
+ mapping->slotno = slotno;
+
+ Assert(!found);
+}
+
+static void
+SlruMappingRemove(SlruCtl ctl, int pageno)
+{
+ bool found PG_USED_FOR_ASSERTS_ONLY;
+
+ hash_search(ctl->mapping_table, &pageno, HASH_REMOVE, &found);
+
+ Assert(found);
+}
diff --git a/src/include/access/slru.h b/src/include/access/slru.h
index dd52e8cec7..8aa3efc0ee 100644
--- a/src/include/access/slru.h
+++ b/src/include/access/slru.h
@@ -16,6 +16,7 @@
#include "access/xlogdefs.h"
#include "storage/lwlock.h"
#include "storage/sync.h"
+#include "utils/hsearch.h"
/*
@@ -110,6 +111,7 @@ typedef SlruSharedData *SlruShared;
typedef struct SlruCtlData
{
SlruShared shared;
+ HTAB *mapping_table;
/*
* Which sync handler function to use when handing sync requests over to
--
2.30.1
[text/x-patch] v16-0002-Make-all-SLRU-buffer-sizes-configurable.patch (25.7K, ../../CA+hUKGLCLDtgDj2Xsf0uBk5WXDCeHxBDDJPsyY7m65Fde-=pyg@mail.gmail.com/3-v16-0002-Make-all-SLRU-buffer-sizes-configurable.patch)
download | inline diff:
From 97c5078f661a41617450826173dd252fc4d0c856 Mon Sep 17 00:00:00 2001
From: Andrey Borodin <[email protected]>
Date: Mon, 15 Feb 2021 21:51:56 +0500
Subject: [PATCH v16 2/3] Make all SLRU buffer sizes configurable.
Provide new GUCs to set the number of buffers, instead of using hard
coded defaults.
Remove the limits on xact_buffers and commit_ts_buffers. The default
sizes for those caches are ~0.2% and ~0.1% of shared_buffers, as before,
but now there is no cap at 128 and 16 buffers respectively (unless
track_commit_timestamp is disabled, in the latter case, then we might as
well keep it tiny). Sizes much larger than the old limits have been
shown to be useful on modern systems, and an earlier commit replaced a
linear search with a hash table to avoid problems with extreme cases.
Author: Andrey M. Borodin <[email protected]>
Reviewed-by: Anastasia Lubennikova <[email protected]>
Reviewed-by: Tomas Vondra <[email protected]>
Reviewed-by: Alexander Korotkov <[email protected]>
Reviewed-by: Gilles Darold <[email protected]>
Reviewed-by: Thomas Munro <[email protected]>
Discussion: https://postgr.es/m/2BEC2B3F-9B61-4C1D-9FB5-5FAB0F05EF86%40yandex-team.ru
---
doc/src/sgml/config.sgml | 135 ++++++++++++++++++
src/backend/access/transam/clog.c | 23 ++-
src/backend/access/transam/commit_ts.c | 10 +-
src/backend/access/transam/multixact.c | 8 +-
src/backend/access/transam/subtrans.c | 5 +-
src/backend/commands/async.c | 8 +-
src/backend/storage/lmgr/predicate.c | 4 +-
src/backend/utils/init/globals.c | 8 ++
src/backend/utils/misc/guc.c | 99 +++++++++++++
src/backend/utils/misc/postgresql.conf.sample | 9 ++
src/include/access/clog.h | 10 ++
src/include/access/commit_ts.h | 1 -
src/include/access/multixact.h | 4 -
src/include/access/slru.h | 5 +
src/include/access/subtrans.h | 2 -
src/include/commands/async.h | 5 -
src/include/miscadmin.h | 7 +
src/include/storage/predicate.h | 4 -
18 files changed, 301 insertions(+), 46 deletions(-)
diff --git a/doc/src/sgml/config.sgml b/doc/src/sgml/config.sgml
index 963824d050..58c46edf62 100644
--- a/doc/src/sgml/config.sgml
+++ b/doc/src/sgml/config.sgml
@@ -1924,6 +1924,141 @@ include_dir 'conf.d'
</para>
</listitem>
</varlistentry>
+
+ <varlistentry id="guc-multixact-offsets-buffers" xreflabel="multixact_offsets_buffers">
+ <term><varname>multixact_offsets_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>multixact_offsets_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to use to cache the contents
+ of <literal>pg_multixact/offsets</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>8</literal>.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-multixact-members-buffers" xreflabel="multixact_members_buffers">
+ <term><varname>multixact_members_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>multixact_members_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to use to cache the contents
+ of <literal>pg_multixact/members</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>16</literal>.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-subtrans-buffers" xreflabel="subtrans_buffers">
+ <term><varname>subtrans_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>subtrans_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to use to cache the contents
+ of <literal>pg_subtrans</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>8</literal>.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-notify-buffers" xreflabel="notify_buffers">
+ <term><varname>notify_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>notify_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to use to cache the contents
+ of <literal>pg_notify</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>8</literal>.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-serial-buffers" xreflabel="serial_buffers">
+ <term><varname>serial_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>serial_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to use to cache the contents
+ of <literal>pg_serial</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>16</literal>.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-xact-buffers" xreflabel="xact_buffers">
+ <term><varname>xact_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>xact_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to use to cache the contents
+ of <literal>pg_xact</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>0</literal>, which requests
+ <varname>shared_buffers</varname> / 512, but not fewer than 4 blocks.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-commit-ts-buffers" xreflabel="commit_ts_buffers">
+ <term><varname>commit_ts_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>commit_ts_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of memory to use to cache the cotents of
+ <literal>pg_commit_ts</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>0</literal>, which requests
+ <varname>shared_buffers</varname> / 1024, but not fewer than 4 blocks.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
<varlistentry id="guc-max-stack-depth" xreflabel="max_stack_depth">
<term><varname>max_stack_depth</varname> (<type>integer</type>)
diff --git a/src/backend/access/transam/clog.c b/src/backend/access/transam/clog.c
index 6fa4713fb4..dd2d7a5184 100644
--- a/src/backend/access/transam/clog.c
+++ b/src/backend/access/transam/clog.c
@@ -58,8 +58,8 @@
/* We need two bits per xact, so four xacts fit in a byte */
#define CLOG_BITS_PER_XACT 2
-#define CLOG_XACTS_PER_BYTE 4
-#define CLOG_XACTS_PER_PAGE (BLCKSZ * CLOG_XACTS_PER_BYTE)
+StaticAssertDecl((CLOG_BITS_PER_XACT * CLOG_XACTS_PER_BYTE) == BITS_PER_BYTE,
+ "CLOG_BITS_PER_XACT and CLOG_XACTS_PER_BYTE are inconsistent");
#define CLOG_XACT_BITMASK ((1 << CLOG_BITS_PER_XACT) - 1)
#define TransactionIdToPage(xid) ((xid) / (TransactionId) CLOG_XACTS_PER_PAGE)
@@ -659,23 +659,16 @@ TransactionIdGetStatus(TransactionId xid, XLogRecPtr *lsn)
/*
* Number of shared CLOG buffers.
*
- * On larger multi-processor systems, it is possible to have many CLOG page
- * requests in flight at one time which could lead to disk access for CLOG
- * page if the required page is not found in memory. Testing revealed that we
- * can get the best performance by having 128 CLOG buffers, more than that it
- * doesn't improve performance.
- *
- * Unconditionally keeping the number of CLOG buffers to 128 did not seem like
- * a good idea, because it would increase the minimum amount of shared memory
- * required to start, which could be a problem for people running very small
- * configurations. The following formula seems to represent a reasonable
- * compromise: people with very low values for shared_buffers will get fewer
- * CLOG buffers as well, and everyone else will get 128.
+ * By default, we'll use 2MB of for every 1GB of shared buffers, up to the
+ * theoretical maximum useful value, but always at least 4 buffers.
*/
Size
CLOGShmemBuffers(void)
{
- return Min(128, Max(4, NBuffers / 512));
+ /* Use configured value if provided. */
+ if (xact_buffers > 0)
+ return Max(4, xact_buffers);
+ return Min(CLOG_MAX_ALLOWED_BUFFERS, Max(4, NBuffers / 512));
}
/*
diff --git a/src/backend/access/transam/commit_ts.c b/src/backend/access/transam/commit_ts.c
index 268bdba339..729a4b9212 100644
--- a/src/backend/access/transam/commit_ts.c
+++ b/src/backend/access/transam/commit_ts.c
@@ -524,13 +524,17 @@ pg_xact_commit_timestamp_origin(PG_FUNCTION_ARGS)
/*
* Number of shared CommitTS buffers.
*
- * We use a very similar logic as for the number of CLOG buffers; see comments
- * in CLOGShmemBuffers.
+ * By default, we'll use 1MB of for every 1GB of shared buffers, up to the
+ * maximum value that slru.c will allow, but always at least 4 buffers.
*/
Size
CommitTsShmemBuffers(void)
{
- return Min(16, Max(4, NBuffers / 1024));
+ /* Use configured value if provided. */
+ if (commit_ts_buffers > 0)
+ return Max(4, commit_ts_buffers);
+ return Min(track_commit_timestamp ? SLRU_MAX_ALLOWED_BUFFERS : 16,
+ Max(4, NBuffers / 1024));
}
/*
diff --git a/src/backend/access/transam/multixact.c b/src/backend/access/transam/multixact.c
index 1f9f1a1fa1..21787765e2 100644
--- a/src/backend/access/transam/multixact.c
+++ b/src/backend/access/transam/multixact.c
@@ -1831,8 +1831,8 @@ MultiXactShmemSize(void)
mul_size(sizeof(MultiXactId) * 2, MaxOldestSlot))
size = SHARED_MULTIXACT_STATE_SIZE;
- size = add_size(size, SimpleLruShmemSize(NUM_MULTIXACTOFFSET_BUFFERS, 0));
- size = add_size(size, SimpleLruShmemSize(NUM_MULTIXACTMEMBER_BUFFERS, 0));
+ size = add_size(size, SimpleLruShmemSize(multixact_offsets_buffers, 0));
+ size = add_size(size, SimpleLruShmemSize(multixact_members_buffers, 0));
return size;
}
@@ -1848,13 +1848,13 @@ MultiXactShmemInit(void)
MultiXactMemberCtl->PagePrecedes = MultiXactMemberPagePrecedes;
SimpleLruInit(MultiXactOffsetCtl,
- "MultiXactOffset", NUM_MULTIXACTOFFSET_BUFFERS, 0,
+ "MultiXactOffset", multixact_offsets_buffers, 0,
MultiXactOffsetSLRULock, "pg_multixact/offsets",
LWTRANCHE_MULTIXACTOFFSET_BUFFER,
SYNC_HANDLER_MULTIXACT_OFFSET);
SlruPagePrecedesUnitTests(MultiXactOffsetCtl, MULTIXACT_OFFSETS_PER_PAGE);
SimpleLruInit(MultiXactMemberCtl,
- "MultiXactMember", NUM_MULTIXACTMEMBER_BUFFERS, 0,
+ "MultiXactMember", multixact_offsets_buffers, 0,
MultiXactMemberSLRULock, "pg_multixact/members",
LWTRANCHE_MULTIXACTMEMBER_BUFFER,
SYNC_HANDLER_MULTIXACT_MEMBER);
diff --git a/src/backend/access/transam/subtrans.c b/src/backend/access/transam/subtrans.c
index 6a8e521f89..785f2520fd 100644
--- a/src/backend/access/transam/subtrans.c
+++ b/src/backend/access/transam/subtrans.c
@@ -31,6 +31,7 @@
#include "access/slru.h"
#include "access/subtrans.h"
#include "access/transam.h"
+#include "miscadmin.h"
#include "pg_trace.h"
#include "utils/snapmgr.h"
@@ -184,14 +185,14 @@ SubTransGetTopmostTransaction(TransactionId xid)
Size
SUBTRANSShmemSize(void)
{
- return SimpleLruShmemSize(NUM_SUBTRANS_BUFFERS, 0);
+ return SimpleLruShmemSize(subtrans_buffers, 0);
}
void
SUBTRANSShmemInit(void)
{
SubTransCtl->PagePrecedes = SubTransPagePrecedes;
- SimpleLruInit(SubTransCtl, "Subtrans", NUM_SUBTRANS_BUFFERS, 0,
+ SimpleLruInit(SubTransCtl, "Subtrans", subtrans_buffers, 0,
SubtransSLRULock, "pg_subtrans",
LWTRANCHE_SUBTRANS_BUFFER, SYNC_HANDLER_NONE);
SlruPagePrecedesUnitTests(SubTransCtl, SUBTRANS_XACTS_PER_PAGE);
diff --git a/src/backend/commands/async.c b/src/backend/commands/async.c
index 4b16fb5682..de17f52cd7 100644
--- a/src/backend/commands/async.c
+++ b/src/backend/commands/async.c
@@ -107,7 +107,7 @@
* frontend during startup.) The above design guarantees that notifies from
* other backends will never be missed by ignoring self-notifies.
*
- * The amount of shared memory used for notify management (NUM_NOTIFY_BUFFERS)
+ * The amount of shared memory used for notify management (notify_buffers)
* can be varied without affecting anything but performance. The maximum
* amount of notification data that can be queued at one time is determined
* by slru.c's wraparound limit; see QUEUE_MAX_PAGE below.
@@ -225,7 +225,7 @@ typedef struct QueuePosition
*
* Resist the temptation to make this really large. While that would save
* work in some places, it would add cost in others. In particular, this
- * should likely be less than NUM_NOTIFY_BUFFERS, to ensure that backends
+ * should likely be less than notify_buffers, to ensure that backends
* catch up before the pages they'll need to read fall out of SLRU cache.
*/
#define QUEUE_CLEANUP_DELAY 4
@@ -514,7 +514,7 @@ AsyncShmemSize(void)
size = mul_size(MaxBackends + 1, sizeof(QueueBackendStatus));
size = add_size(size, offsetof(AsyncQueueControl, backend));
- size = add_size(size, SimpleLruShmemSize(NUM_NOTIFY_BUFFERS, 0));
+ size = add_size(size, SimpleLruShmemSize(notify_buffers, 0));
return size;
}
@@ -562,7 +562,7 @@ AsyncShmemInit(void)
* Set up SLRU management of the pg_notify data.
*/
NotifyCtl->PagePrecedes = asyncQueuePagePrecedes;
- SimpleLruInit(NotifyCtl, "Notify", NUM_NOTIFY_BUFFERS, 0,
+ SimpleLruInit(NotifyCtl, "Notify", notify_buffers, 0,
NotifySLRULock, "pg_notify", LWTRANCHE_NOTIFY_BUFFER,
SYNC_HANDLER_NONE);
diff --git a/src/backend/storage/lmgr/predicate.c b/src/backend/storage/lmgr/predicate.c
index d493aeef0f..b1f4f1651d 100644
--- a/src/backend/storage/lmgr/predicate.c
+++ b/src/backend/storage/lmgr/predicate.c
@@ -872,7 +872,7 @@ SerialInit(void)
*/
SerialSlruCtl->PagePrecedes = SerialPagePrecedesLogically;
SimpleLruInit(SerialSlruCtl, "Serial",
- NUM_SERIAL_BUFFERS, 0, SerialSLRULock, "pg_serial",
+ serial_buffers, 0, SerialSLRULock, "pg_serial",
LWTRANCHE_SERIAL_BUFFER, SYNC_HANDLER_NONE);
#ifdef USE_ASSERT_CHECKING
SerialPagePrecedesLogicallyUnitTests();
@@ -1395,7 +1395,7 @@ PredicateLockShmemSize(void)
/* Shared memory structures for SLRU tracking of old committed xids. */
size = add_size(size, sizeof(SerialControlData));
- size = add_size(size, SimpleLruShmemSize(NUM_SERIAL_BUFFERS, 0));
+ size = add_size(size, SimpleLruShmemSize(serial_buffers, 0));
return size;
}
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
index 381d9e548d..c83151d5ab 100644
--- a/src/backend/utils/init/globals.c
+++ b/src/backend/utils/init/globals.c
@@ -150,3 +150,11 @@ int64 VacuumPageDirty = 0;
int VacuumCostBalance = 0; /* working state for vacuum */
bool VacuumCostActive = false;
+
+int multixact_offsets_buffers = 8;
+int multixact_members_buffers = 16;
+int subtrans_buffers = 32;
+int notify_buffers = 8;
+int serial_buffers = 16;
+int xact_buffers = 0;
+int commit_ts_buffers = 0;
diff --git a/src/backend/utils/misc/guc.c b/src/backend/utils/misc/guc.c
index bee976bae8..250cef80bd 100644
--- a/src/backend/utils/misc/guc.c
+++ b/src/backend/utils/misc/guc.c
@@ -32,9 +32,11 @@
#endif
#include <unistd.h>
+#include "access/clog.h"
#include "access/commit_ts.h"
#include "access/gin.h"
#include "access/rmgr.h"
+#include "access/slru.h"
#include "access/tableam.h"
#include "access/toast_compression.h"
#include "access/transam.h"
@@ -200,6 +202,8 @@ static const char *show_tcp_keepalives_idle(void);
static const char *show_tcp_keepalives_interval(void);
static const char *show_tcp_keepalives_count(void);
static const char *show_tcp_user_timeout(void);
+static const char *show_xact_buffers(void);
+static const char *show_commit_ts_buffers(void);
static bool check_maxconnections(int *newval, void **extra, GucSource source);
static bool check_max_worker_processes(int *newval, void **extra, GucSource source);
static bool check_autovacuum_max_workers(int *newval, void **extra, GucSource source);
@@ -2340,6 +2344,83 @@ static struct config_int ConfigureNamesInt[] =
NULL, NULL, NULL
},
+ {
+ {"multixact_offsets_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for the MultiXact offset SLRU cache."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &multixact_offsets_buffers,
+ 8, 2, SLRU_MAX_ALLOWED_BUFFERS,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"multixact_members_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for the MultiXact member SLRU cache."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &multixact_members_buffers,
+ 16, 2, SLRU_MAX_ALLOWED_BUFFERS,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"subtrans_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for the sub-transaction SLRU cache."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &subtrans_buffers,
+ 32, 2, SLRU_MAX_ALLOWED_BUFFERS,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"notify_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for the NOTIFY message SLRU cache."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ ¬ify_buffers,
+ 8, 2, SLRU_MAX_ALLOWED_BUFFERS,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"serial_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for the serializable transaction SLRU cache."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &serial_buffers,
+ 16, 2, SLRU_MAX_ALLOWED_BUFFERS,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"xact_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for the transaction status SLRU cache."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &xact_buffers,
+ 0, 0, CLOG_MAX_ALLOWED_BUFFERS,
+ NULL, NULL, show_xact_buffers
+ },
+
+ {
+ {"commit_ts_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the size of the dedicated buffer pool used for the commit timestamp SLRU cache."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &commit_ts_buffers,
+ 0, 0, SLRU_MAX_ALLOWED_BUFFERS,
+ NULL, NULL, show_commit_ts_buffers
+ },
+
{
{"temp_buffers", PGC_USERSET, RESOURCES_MEM,
gettext_noop("Sets the maximum number of temporary buffers used by each session."),
@@ -11959,6 +12040,24 @@ show_tcp_user_timeout(void)
return nbuf;
}
+static const char *
+show_xact_buffers(void)
+{
+ static char nbuf[16];
+
+ snprintf(nbuf, sizeof(nbuf), "%zu", CLOGShmemBuffers());
+ return nbuf;
+}
+
+static const char *
+show_commit_ts_buffers(void)
+{
+ static char nbuf[16];
+
+ snprintf(nbuf, sizeof(nbuf), "%zu", CommitTsShmemBuffers());
+ return nbuf;
+}
+
static bool
check_maxconnections(int *newval, void **extra, GucSource source)
{
diff --git a/src/backend/utils/misc/postgresql.conf.sample b/src/backend/utils/misc/postgresql.conf.sample
index ff9fa006fe..7e14df3b51 100644
--- a/src/backend/utils/misc/postgresql.conf.sample
+++ b/src/backend/utils/misc/postgresql.conf.sample
@@ -190,6 +190,15 @@
# (change requires restart)
#backend_flush_after = 0 # measured in pages, 0 disables
+# - SLRU Buffers (change requires restart) -
+
+#xact_buffers = 0 # memory for pg_xact (0 = auto)
+#subtrans_buffers = 32 # memory for pg_subtrans
+#multixact_offsets_buffers = 8 # memory for pg_multixact/offsets
+#multixact_members_buffers = 16 # memory for pg_multixact/members
+#notify_buffers = 8 # memory for pg_notify
+#serial_buffers = 16 # memory for pg_serial
+#commit_ts_buffers = 0 # memory for pg_commit_ts (0 = auto)
#------------------------------------------------------------------------------
# WRITE-AHEAD LOG
diff --git a/src/include/access/clog.h b/src/include/access/clog.h
index 39b8e4afa8..739a292f7f 100644
--- a/src/include/access/clog.h
+++ b/src/include/access/clog.h
@@ -15,6 +15,16 @@
#include "storage/sync.h"
#include "lib/stringinfo.h"
+/*
+ * Don't allow xact_buffers to be set higher than could possibly be useful or
+ * SLRU would allow.
+ */
+#define CLOG_XACTS_PER_BYTE 4
+#define CLOG_XACTS_PER_PAGE (BLCKSZ * CLOG_XACTS_PER_BYTE)
+#define CLOG_MAX_ALLOWED_BUFFERS \
+ Min(SLRU_MAX_ALLOWED_BUFFERS, \
+ (((MaxTransactionId / 2) + (CLOG_XACTS_PER_PAGE - 1)) / CLOG_XACTS_PER_PAGE))
+
/*
* Possible transaction statuses --- note that all-zeroes is the initial
* state.
diff --git a/src/include/access/commit_ts.h b/src/include/access/commit_ts.h
index 750369104a..e4cf988609 100644
--- a/src/include/access/commit_ts.h
+++ b/src/include/access/commit_ts.h
@@ -17,7 +17,6 @@
#include "storage/sync.h"
#include "utils/guc.h"
-
extern PGDLLIMPORT bool track_commit_timestamp;
extern bool check_track_commit_timestamp(bool *newval, void **extra,
diff --git a/src/include/access/multixact.h b/src/include/access/multixact.h
index 4bbb035eae..97c0a46376 100644
--- a/src/include/access/multixact.h
+++ b/src/include/access/multixact.h
@@ -29,10 +29,6 @@
#define MaxMultiXactOffset ((MultiXactOffset) 0xFFFFFFFF)
-/* Number of SLRU buffers to use for multixact */
-#define NUM_MULTIXACTOFFSET_BUFFERS 8
-#define NUM_MULTIXACTMEMBER_BUFFERS 16
-
/*
* Possible multixact lock modes ("status"). The first four modes are for
* tuple locks (FOR KEY SHARE, FOR SHARE, FOR NO KEY UPDATE, FOR UPDATE); the
diff --git a/src/include/access/slru.h b/src/include/access/slru.h
index 8aa3efc0ee..97ea837646 100644
--- a/src/include/access/slru.h
+++ b/src/include/access/slru.h
@@ -18,6 +18,11 @@
#include "storage/sync.h"
#include "utils/hsearch.h"
+/*
+ * To avoid overflowing internal arithmetic and the size_t data type, the
+ * number of buffers should not exceed this number.
+ */
+#define SLRU_MAX_ALLOWED_BUFFERS ((1024 * 1024 * 1024) / BLCKSZ)
/*
* Define SLRU segment size. A page is the same BLCKSZ as is used everywhere
diff --git a/src/include/access/subtrans.h b/src/include/access/subtrans.h
index d0ab44ae82..64fa86938e 100644
--- a/src/include/access/subtrans.h
+++ b/src/include/access/subtrans.h
@@ -11,8 +11,6 @@
#ifndef SUBTRANS_H
#define SUBTRANS_H
-/* Number of SLRU buffers to use for subtrans */
-#define NUM_SUBTRANS_BUFFERS 32
extern void SubTransSetParent(TransactionId xid, TransactionId parent);
extern TransactionId SubTransGetParent(TransactionId xid);
diff --git a/src/include/commands/async.h b/src/include/commands/async.h
index 9217f66b91..fa831e3721 100644
--- a/src/include/commands/async.h
+++ b/src/include/commands/async.h
@@ -15,11 +15,6 @@
#include <signal.h>
-/*
- * The number of SLRU page buffers we use for the notification queue.
- */
-#define NUM_NOTIFY_BUFFERS 8
-
extern bool Trace_notify;
extern volatile sig_atomic_t notifyInterruptPending;
diff --git a/src/include/miscadmin.h b/src/include/miscadmin.h
index 95202d37af..495c1bf901 100644
--- a/src/include/miscadmin.h
+++ b/src/include/miscadmin.h
@@ -164,6 +164,13 @@ extern PGDLLIMPORT int MaxBackends;
extern PGDLLIMPORT int MaxConnections;
extern PGDLLIMPORT int max_worker_processes;
extern PGDLLIMPORT int max_parallel_workers;
+extern PGDLLIMPORT int multixact_offsets_buffers;
+extern PGDLLIMPORT int multixact_members_buffers;
+extern PGDLLIMPORT int subtrans_buffers;
+extern PGDLLIMPORT int notify_buffers;
+extern PGDLLIMPORT int serial_buffers;
+extern PGDLLIMPORT int xact_buffers;
+extern PGDLLIMPORT int commit_ts_buffers;
extern PGDLLIMPORT int MyProcPid;
extern PGDLLIMPORT pg_time_t MyStartTime;
diff --git a/src/include/storage/predicate.h b/src/include/storage/predicate.h
index 152b698611..c72779bd88 100644
--- a/src/include/storage/predicate.h
+++ b/src/include/storage/predicate.h
@@ -26,10 +26,6 @@ extern int max_predicate_locks_per_xact;
extern int max_predicate_locks_per_relation;
extern int max_predicate_locks_per_page;
-
-/* Number of SLRU buffers to use for Serial SLRU */
-#define NUM_SERIAL_BUFFERS 16
-
/*
* A handle used for sharing SERIALIZABLEXACT objects between the participants
* in a parallel query.
--
2.30.1
[text/x-patch] v16-0003-Limit-SLRU-buffer-replacement-search.patch (3.0K, ../../CA+hUKGLCLDtgDj2Xsf0uBk5WXDCeHxBDDJPsyY7m65Fde-=pyg@mail.gmail.com/4-v16-0003-Limit-SLRU-buffer-replacement-search.patch)
download | inline diff:
From 79c15f73a21cfb67cb6157fc5b6822bcdc696383 Mon Sep 17 00:00:00 2001
From: Thomas Munro <[email protected]>
Date: Thu, 8 Apr 2021 12:18:25 +1200
Subject: [PATCH v16 3/3] Limit SLRU buffer replacement search.
Now that users can configure large SLRU caches, slru.c's simple buffer
replacement algorithm needs some adjustment. For now, limit its linear
search for the least recently accessed buffer to an arbitrary cap. This
means it won't find the globally least recently used buffer, just the
least recently used in a given range of pages. The cap is initially set
as large as the previous hard-coded search size.
Discussion: https://postgr.es/m/2BEC2B3F-9B61-4C1D-9FB5-5FAB0F05EF86%40yandex-team.ru
---
src/backend/access/transam/slru.c | 15 ++++++++++++++-
src/include/access/slru.h | 3 +++
2 files changed, 17 insertions(+), 1 deletion(-)
diff --git a/src/backend/access/transam/slru.c b/src/backend/access/transam/slru.c
index 82c61c475b..131f7a48d7 100644
--- a/src/backend/access/transam/slru.c
+++ b/src/backend/access/transam/slru.c
@@ -71,6 +71,12 @@
*/
#define MAX_WRITEALL_BUFFERS 16
+/*
+ * When searching for buffers to replace, we will limit the scope of our search
+ * for now, to avoid holding an exclusive lock for too long.
+ */
+#define MAX_REPLACEMENT_SEARCH 128
+
typedef struct SlruWriteAllData
{
int num_files; /* # files actually open */
@@ -1060,6 +1066,7 @@ SlruSelectLRUPage(SlruCtl ctl, int pageno)
{
SlruShared shared = ctl->shared;
+
/* Outer loop handles restart after I/O */
for (;;)
{
@@ -1071,6 +1078,7 @@ SlruSelectLRUPage(SlruCtl ctl, int pageno)
int bestinvalidslot = 0; /* keep compiler quiet */
int best_invalid_delta = -1;
int best_invalid_page_number = 0; /* keep compiler quiet */
+ int max_search;
/* See if page already has a buffer assigned */
slotno = SlruMappingFind(ctl, pageno);
@@ -1108,12 +1116,17 @@ SlruSelectLRUPage(SlruCtl ctl, int pageno)
* That gets us back on the path to having good data when there are
* multiple pages with the same lru_count.
*/
+ max_search = Min(shared->num_slots, MAX_REPLACEMENT_SEARCH);
cur_count = (shared->cur_lru_count)++;
- for (slotno = 0; slotno < shared->num_slots; slotno++)
+ for (int i = 0; i < max_search; ++i)
{
int this_delta;
int this_page_number;
+ slotno = shared->search_slotno++;
+ if (shared->search_slotno == shared->num_slots)
+ shared->search_slotno = 0;
+
if (shared->page_status[slotno] == SLRU_PAGE_EMPTY)
return slotno;
this_delta = cur_count - shared->page_lru_count[slotno];
diff --git a/src/include/access/slru.h b/src/include/access/slru.h
index 97ea837646..fb8a03972d 100644
--- a/src/include/access/slru.h
+++ b/src/include/access/slru.h
@@ -63,6 +63,9 @@ typedef struct SlruSharedData
/* Number of buffers managed by this SLRU structure */
int num_slots;
+ /* Where to start buffer replacement search. */
+ int search_slotno;
+
/*
* Arrays holding info for each buffer slot. Page number is undefined
* when status is EMPTY, as is page_lru_count.
--
2.30.1
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2021-04-08 07:24 Andrey Borodin <[email protected]>
parent: Thomas Munro <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Andrey Borodin @ 2021-04-08 07:24 UTC (permalink / raw)
To: Thomas Munro <[email protected]>; +Cc: Gilles Darold <[email protected]>; Tomas Vondra <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
> 8 апр. 2021 г., в 03:30, Thomas Munro <[email protected]> написал(а):
>
> Here's another approach that is a little less exciting than
> "tournament RR" (or whatever that should be called; I couldn't find an
> established name for it). This version is just our traditional linear
> search, except that it stops at 128, and remembers where to start from
> next time (like a sort of Fisher-Price GCLOCK hand). This feels more
> committable to me. You can argue that all buffers above 128 are bonus
> buffers that PostgreSQL 13 didn't have, so the fact that we can no
> longer find the globally least recently used page when you set
> xact_buffers > 128 doesn't seem too bad to me, as an incremental step
> (but to be clear, of course we can do better than this with more work
> in later releases).
I agree that this version of eviction seems much more effective and less intrusive than RR. And it's still LRU, which is important for subsystem that is called SLRU.
shared->search_slotno is initialized implicitly with memset(). But this seems like a common practice.
Also comment above "max_search = Min(shared->num_slots, MAX_REPLACEMENT_SEARCH);" does not reflect changes.
Besides this patch looks good to me.
Thanks!
Best regards, Andrey Borodin.
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2021-04-08 12:22 Thomas Munro <[email protected]>
parent: Andrey Borodin <[email protected]>
0 siblings, 2 replies; 103+ messages in thread
From: Thomas Munro @ 2021-04-08 12:22 UTC (permalink / raw)
To: Andrey Borodin <[email protected]>; +Cc: Gilles Darold <[email protected]>; Tomas Vondra <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
On Thu, Apr 8, 2021 at 7:24 PM Andrey Borodin <[email protected]> wrote:
> I agree that this version of eviction seems much more effective and less intrusive than RR. And it's still LRU, which is important for subsystem that is called SLRU.
> shared->search_slotno is initialized implicitly with memset(). But this seems like a common practice.
> Also comment above "max_search = Min(shared->num_slots, MAX_REPLACEMENT_SEARCH);" does not reflect changes.
>
> Besides this patch looks good to me.
Thanks! I chickened out of committing a buffer replacement algorithm
patch written 11 hours before the feature freeze, but I also didn't
really want to commit the GUC patch without that. Ahh, if only we'd
latched onto the real problems here just a little sooner, but there is
always PostgreSQL 15, I heard it's going to be amazing. Moved to next
CF.
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2021-04-11 18:37 Andrey Borodin <[email protected]>
parent: Thomas Munro <[email protected]>
1 sibling, 2 replies; 103+ messages in thread
From: Andrey Borodin @ 2021-04-11 18:37 UTC (permalink / raw)
To: Thomas Munro <[email protected]>; +Cc: Gilles Darold <[email protected]>; Tomas Vondra <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
> 8 апр. 2021 г., в 15:22, Thomas Munro <[email protected]> написал(а):
>
> On Thu, Apr 8, 2021 at 7:24 PM Andrey Borodin <[email protected]> wrote:
>> I agree that this version of eviction seems much more effective and less intrusive than RR. And it's still LRU, which is important for subsystem that is called SLRU.
>> shared->search_slotno is initialized implicitly with memset(). But this seems like a common practice.
>> Also comment above "max_search = Min(shared->num_slots, MAX_REPLACEMENT_SEARCH);" does not reflect changes.
>>
>> Besides this patch looks good to me.
>
> Thanks! I chickened out of committing a buffer replacement algorithm
> patch written 11 hours before the feature freeze, but I also didn't
> really want to commit the GUC patch without that. Ahh, if only we'd
> latched onto the real problems here just a little sooner, but there is
> always PostgreSQL 15, I heard it's going to be amazing. Moved to next
> CF.
I have one more idea inspired by CPU caches.
Let's make SLRU n-associative, where n ~ 8.
We can divide buffers into "banks", number of banks must be power of 2.
All banks are of equal size. We choose bank size to approximately satisfy user's configured buffer size.
Each page can live only within one bank. We use same search and eviction algorithms as we used in SLRU, but we only need to search\evict over 8 elements.
All SLRU data of a single bank will be colocated within at most 2 cache line.
I did not come up with idea how to avoid multiplication of bank_number * bank_size in case when user configured 31337 buffers (any number that is radically not a power of 2).
PFA patch implementing this idea.
Best regards, Andrey Borodin.
Attachments:
[application/octet-stream] v17-0002-Divide-SLRU-buffers-into-n-associative-banks.patch (4.2K, ../../[email protected]/2-v17-0002-Divide-SLRU-buffers-into-n-associative-banks.patch)
download | inline diff:
From b418d7eafacfd9aa6d44951deda9690752732197 Mon Sep 17 00:00:00 2001
From: Andrey Borodin <[email protected]>
Date: Sun, 11 Apr 2021 21:18:10 +0300
Subject: [PATCH v17 2/2] Divide SLRU buffers into n-associative banks
---
src/backend/access/transam/slru.c | 36 +++++++++++++++++++++++++++----
src/include/access/slru.h | 2 ++
2 files changed, 34 insertions(+), 4 deletions(-)
diff --git a/src/backend/access/transam/slru.c b/src/backend/access/transam/slru.c
index 82149ad782..84321b7509 100644
--- a/src/backend/access/transam/slru.c
+++ b/src/backend/access/transam/slru.c
@@ -133,7 +133,7 @@ typedef enum
static SlruErrorCause slru_errcause;
static int slru_errno;
-
+static void SlruAdjustNSlots(int* nslots, int* banksize, int* bankoffset);
static void SimpleLruZeroLSNs(SlruCtl ctl, int slotno);
static void SimpleLruWaitIO(SlruCtl ctl, int slotno);
static void SlruInternalWritePage(SlruCtl ctl, int slotno, SlruWriteAll fdata);
@@ -147,6 +147,23 @@ static bool SlruScanDirCbDeleteCutoff(SlruCtl ctl, char *filename,
int segpage, void *data);
static void SlruInternalDeleteSegment(SlruCtl ctl, int segno);
+static void SlruAdjustNSlots(int* nslots, int* banksize, int* bankoffset)
+{
+ *banksize = *nslots;
+ int nbanks = 1;
+ *bankoffset = 0;
+ while (*banksize > 15)
+ {
+ if ((*banksize & 1) != 0)
+ *banksize +=1;
+ *banksize /= 2;
+ nbanks *= 2;
+ *bankoffset += 1;
+ }
+ elog(DEBUG5, "nslots %d banksize %d nbanks %d ", *nslots, *banksize, nbanks);
+ *nslots = *banksize * nbanks;
+}
+
/*
* Initialization of shared memory
*/
@@ -155,6 +172,8 @@ Size
SimpleLruShmemSize(int nslots, int nlsns)
{
Size sz;
+ int bankoffset, banksize;
+ SlruAdjustNSlots(&nslots, &banksize, &bankoffset);
/* we assume nslots isn't so large as to risk overflow */
sz = MAXALIGN(sizeof(SlruSharedData));
@@ -189,6 +208,8 @@ SimpleLruInit(SlruCtl ctl, const char *name, int nslots, int nlsns,
{
SlruShared shared;
bool found;
+ int bankoffset, banksize;
+ SlruAdjustNSlots(&nslots, &banksize, &bankoffset);
shared = (SlruShared) ShmemInitStruct(name,
SimpleLruShmemSize(nslots, nlsns),
@@ -208,6 +229,9 @@ SimpleLruInit(SlruCtl ctl, const char *name, int nslots, int nlsns,
shared->ControlLock = ctllock;
shared->num_slots = nslots;
+ shared->bank_mask = (1 << bankoffset) - 1;
+ shared->bank_size = banksize;
+
shared->lsn_groups_per_page = nlsns;
shared->cur_lru_count = 0;
@@ -500,7 +524,9 @@ SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno, TransactionId xid)
LWLockAcquire(shared->ControlLock, LW_SHARED);
/* See if page is already in a buffer */
- for (slotno = 0; slotno < shared->num_slots; slotno++)
+ int bankstart = (pageno & shared->bank_mask) * shared->bank_size;
+ int bankend = bankstart + shared->bank_size;
+ for (slotno = bankstart; slotno < bankend; slotno++)
{
if (shared->page_number[slotno] == pageno &&
shared->page_status[slotno] != SLRU_PAGE_EMPTY &&
@@ -1029,7 +1055,9 @@ SlruSelectLRUPage(SlruCtl ctl, int pageno)
int best_invalid_page_number = 0; /* keep compiler quiet */
/* See if page already has a buffer assigned */
- for (slotno = 0; slotno < shared->num_slots; slotno++)
+ int bankstart = (pageno & shared->bank_mask) * shared->bank_size;
+ int bankend = bankstart + shared->bank_size;
+ for (slotno = bankstart; slotno < bankend; slotno++)
{
if (shared->page_number[slotno] == pageno &&
shared->page_status[slotno] != SLRU_PAGE_EMPTY)
@@ -1064,7 +1092,7 @@ SlruSelectLRUPage(SlruCtl ctl, int pageno)
* multiple pages with the same lru_count.
*/
cur_count = (shared->cur_lru_count)++;
- for (slotno = 0; slotno < shared->num_slots; slotno++)
+ for (slotno = bankstart; slotno < bankend; slotno++)
{
int this_delta;
int this_page_number;
diff --git a/src/include/access/slru.h b/src/include/access/slru.h
index 793c045f16..f4df54d3c1 100644
--- a/src/include/access/slru.h
+++ b/src/include/access/slru.h
@@ -61,6 +61,8 @@ typedef struct SlruSharedData
/* Number of buffers managed by this SLRU structure */
int num_slots;
+ int bank_size;
+ int bank_mask;
/*
* Arrays holding info for each buffer slot. Page number is undefined
--
2.24.3 (Apple Git-128)
[application/octet-stream] v17-0001-Make-all-SLRU-buffer-sizes-configurable.patch (25.7K, ../../[email protected]/3-v17-0001-Make-all-SLRU-buffer-sizes-configurable.patch)
download | inline diff:
From 46b3aff4ebede9378ea5471d578e8466f0639919 Mon Sep 17 00:00:00 2001
From: Andrey Borodin <[email protected]>
Date: Mon, 15 Feb 2021 21:51:56 +0500
Subject: [PATCH v17 1/2] Make all SLRU buffer sizes configurable.
Provide new GUCs to set the number of buffers, instead of using hard
coded defaults.
Remove the limits on xact_buffers and commit_ts_buffers. The default
sizes for those caches are ~0.2% and ~0.1% of shared_buffers, as before,
but now there is no cap at 128 and 16 buffers respectively (unless
track_commit_timestamp is disabled, in the latter case, then we might as
well keep it tiny). Sizes much larger than the old limits have been
shown to be useful on modern systems, and an earlier commit replaced a
linear search with a hash table to avoid problems with extreme cases.
Author: Andrey M. Borodin <[email protected]>
Reviewed-by: Anastasia Lubennikova <[email protected]>
Reviewed-by: Tomas Vondra <[email protected]>
Reviewed-by: Alexander Korotkov <[email protected]>
Reviewed-by: Gilles Darold <[email protected]>
Reviewed-by: Thomas Munro <[email protected]>
Discussion: https://postgr.es/m/2BEC2B3F-9B61-4C1D-9FB5-5FAB0F05EF86%40yandex-team.ru
---
doc/src/sgml/config.sgml | 135 ++++++++++++++++++
src/backend/access/transam/clog.c | 23 ++-
src/backend/access/transam/commit_ts.c | 10 +-
src/backend/access/transam/multixact.c | 8 +-
src/backend/access/transam/subtrans.c | 5 +-
src/backend/commands/async.c | 8 +-
src/backend/storage/lmgr/predicate.c | 4 +-
src/backend/utils/init/globals.c | 8 ++
src/backend/utils/misc/guc.c | 99 +++++++++++++
src/backend/utils/misc/postgresql.conf.sample | 9 ++
src/include/access/clog.h | 10 ++
src/include/access/commit_ts.h | 1 -
src/include/access/multixact.h | 4 -
src/include/access/slru.h | 5 +
src/include/access/subtrans.h | 2 -
src/include/commands/async.h | 5 -
src/include/miscadmin.h | 7 +
src/include/storage/predicate.h | 4 -
18 files changed, 301 insertions(+), 46 deletions(-)
diff --git a/doc/src/sgml/config.sgml b/doc/src/sgml/config.sgml
index ea5cf3a2dc..43bd93ccf4 100644
--- a/doc/src/sgml/config.sgml
+++ b/doc/src/sgml/config.sgml
@@ -1924,6 +1924,141 @@ include_dir 'conf.d'
</para>
</listitem>
</varlistentry>
+
+ <varlistentry id="guc-multixact-offsets-buffers" xreflabel="multixact_offsets_buffers">
+ <term><varname>multixact_offsets_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>multixact_offsets_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to use to cache the contents
+ of <literal>pg_multixact/offsets</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>8</literal>.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-multixact-members-buffers" xreflabel="multixact_members_buffers">
+ <term><varname>multixact_members_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>multixact_members_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to use to cache the contents
+ of <literal>pg_multixact/members</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>16</literal>.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-subtrans-buffers" xreflabel="subtrans_buffers">
+ <term><varname>subtrans_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>subtrans_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to use to cache the contents
+ of <literal>pg_subtrans</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>8</literal>.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-notify-buffers" xreflabel="notify_buffers">
+ <term><varname>notify_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>notify_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to use to cache the contents
+ of <literal>pg_notify</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>8</literal>.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-serial-buffers" xreflabel="serial_buffers">
+ <term><varname>serial_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>serial_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to use to cache the contents
+ of <literal>pg_serial</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>16</literal>.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-xact-buffers" xreflabel="xact_buffers">
+ <term><varname>xact_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>xact_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to use to cache the contents
+ of <literal>pg_xact</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>0</literal>, which requests
+ <varname>shared_buffers</varname> / 512, but not fewer than 4 blocks.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-commit-ts-buffers" xreflabel="commit_ts_buffers">
+ <term><varname>commit_ts_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>commit_ts_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of memory to use to cache the cotents of
+ <literal>pg_commit_ts</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>0</literal>, which requests
+ <varname>shared_buffers</varname> / 1024, but not fewer than 4 blocks.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
<varlistentry id="guc-max-stack-depth" xreflabel="max_stack_depth">
<term><varname>max_stack_depth</varname> (<type>integer</type>)
diff --git a/src/backend/access/transam/clog.c b/src/backend/access/transam/clog.c
index 6fa4713fb4..dd2d7a5184 100644
--- a/src/backend/access/transam/clog.c
+++ b/src/backend/access/transam/clog.c
@@ -58,8 +58,8 @@
/* We need two bits per xact, so four xacts fit in a byte */
#define CLOG_BITS_PER_XACT 2
-#define CLOG_XACTS_PER_BYTE 4
-#define CLOG_XACTS_PER_PAGE (BLCKSZ * CLOG_XACTS_PER_BYTE)
+StaticAssertDecl((CLOG_BITS_PER_XACT * CLOG_XACTS_PER_BYTE) == BITS_PER_BYTE,
+ "CLOG_BITS_PER_XACT and CLOG_XACTS_PER_BYTE are inconsistent");
#define CLOG_XACT_BITMASK ((1 << CLOG_BITS_PER_XACT) - 1)
#define TransactionIdToPage(xid) ((xid) / (TransactionId) CLOG_XACTS_PER_PAGE)
@@ -659,23 +659,16 @@ TransactionIdGetStatus(TransactionId xid, XLogRecPtr *lsn)
/*
* Number of shared CLOG buffers.
*
- * On larger multi-processor systems, it is possible to have many CLOG page
- * requests in flight at one time which could lead to disk access for CLOG
- * page if the required page is not found in memory. Testing revealed that we
- * can get the best performance by having 128 CLOG buffers, more than that it
- * doesn't improve performance.
- *
- * Unconditionally keeping the number of CLOG buffers to 128 did not seem like
- * a good idea, because it would increase the minimum amount of shared memory
- * required to start, which could be a problem for people running very small
- * configurations. The following formula seems to represent a reasonable
- * compromise: people with very low values for shared_buffers will get fewer
- * CLOG buffers as well, and everyone else will get 128.
+ * By default, we'll use 2MB of for every 1GB of shared buffers, up to the
+ * theoretical maximum useful value, but always at least 4 buffers.
*/
Size
CLOGShmemBuffers(void)
{
- return Min(128, Max(4, NBuffers / 512));
+ /* Use configured value if provided. */
+ if (xact_buffers > 0)
+ return Max(4, xact_buffers);
+ return Min(CLOG_MAX_ALLOWED_BUFFERS, Max(4, NBuffers / 512));
}
/*
diff --git a/src/backend/access/transam/commit_ts.c b/src/backend/access/transam/commit_ts.c
index 268bdba339..729a4b9212 100644
--- a/src/backend/access/transam/commit_ts.c
+++ b/src/backend/access/transam/commit_ts.c
@@ -524,13 +524,17 @@ pg_xact_commit_timestamp_origin(PG_FUNCTION_ARGS)
/*
* Number of shared CommitTS buffers.
*
- * We use a very similar logic as for the number of CLOG buffers; see comments
- * in CLOGShmemBuffers.
+ * By default, we'll use 1MB of for every 1GB of shared buffers, up to the
+ * maximum value that slru.c will allow, but always at least 4 buffers.
*/
Size
CommitTsShmemBuffers(void)
{
- return Min(16, Max(4, NBuffers / 1024));
+ /* Use configured value if provided. */
+ if (commit_ts_buffers > 0)
+ return Max(4, commit_ts_buffers);
+ return Min(track_commit_timestamp ? SLRU_MAX_ALLOWED_BUFFERS : 16,
+ Max(4, NBuffers / 1024));
}
/*
diff --git a/src/backend/access/transam/multixact.c b/src/backend/access/transam/multixact.c
index 1f9f1a1fa1..21787765e2 100644
--- a/src/backend/access/transam/multixact.c
+++ b/src/backend/access/transam/multixact.c
@@ -1831,8 +1831,8 @@ MultiXactShmemSize(void)
mul_size(sizeof(MultiXactId) * 2, MaxOldestSlot))
size = SHARED_MULTIXACT_STATE_SIZE;
- size = add_size(size, SimpleLruShmemSize(NUM_MULTIXACTOFFSET_BUFFERS, 0));
- size = add_size(size, SimpleLruShmemSize(NUM_MULTIXACTMEMBER_BUFFERS, 0));
+ size = add_size(size, SimpleLruShmemSize(multixact_offsets_buffers, 0));
+ size = add_size(size, SimpleLruShmemSize(multixact_members_buffers, 0));
return size;
}
@@ -1848,13 +1848,13 @@ MultiXactShmemInit(void)
MultiXactMemberCtl->PagePrecedes = MultiXactMemberPagePrecedes;
SimpleLruInit(MultiXactOffsetCtl,
- "MultiXactOffset", NUM_MULTIXACTOFFSET_BUFFERS, 0,
+ "MultiXactOffset", multixact_offsets_buffers, 0,
MultiXactOffsetSLRULock, "pg_multixact/offsets",
LWTRANCHE_MULTIXACTOFFSET_BUFFER,
SYNC_HANDLER_MULTIXACT_OFFSET);
SlruPagePrecedesUnitTests(MultiXactOffsetCtl, MULTIXACT_OFFSETS_PER_PAGE);
SimpleLruInit(MultiXactMemberCtl,
- "MultiXactMember", NUM_MULTIXACTMEMBER_BUFFERS, 0,
+ "MultiXactMember", multixact_offsets_buffers, 0,
MultiXactMemberSLRULock, "pg_multixact/members",
LWTRANCHE_MULTIXACTMEMBER_BUFFER,
SYNC_HANDLER_MULTIXACT_MEMBER);
diff --git a/src/backend/access/transam/subtrans.c b/src/backend/access/transam/subtrans.c
index 6a8e521f89..785f2520fd 100644
--- a/src/backend/access/transam/subtrans.c
+++ b/src/backend/access/transam/subtrans.c
@@ -31,6 +31,7 @@
#include "access/slru.h"
#include "access/subtrans.h"
#include "access/transam.h"
+#include "miscadmin.h"
#include "pg_trace.h"
#include "utils/snapmgr.h"
@@ -184,14 +185,14 @@ SubTransGetTopmostTransaction(TransactionId xid)
Size
SUBTRANSShmemSize(void)
{
- return SimpleLruShmemSize(NUM_SUBTRANS_BUFFERS, 0);
+ return SimpleLruShmemSize(subtrans_buffers, 0);
}
void
SUBTRANSShmemInit(void)
{
SubTransCtl->PagePrecedes = SubTransPagePrecedes;
- SimpleLruInit(SubTransCtl, "Subtrans", NUM_SUBTRANS_BUFFERS, 0,
+ SimpleLruInit(SubTransCtl, "Subtrans", subtrans_buffers, 0,
SubtransSLRULock, "pg_subtrans",
LWTRANCHE_SUBTRANS_BUFFER, SYNC_HANDLER_NONE);
SlruPagePrecedesUnitTests(SubTransCtl, SUBTRANS_XACTS_PER_PAGE);
diff --git a/src/backend/commands/async.c b/src/backend/commands/async.c
index 4b16fb5682..de17f52cd7 100644
--- a/src/backend/commands/async.c
+++ b/src/backend/commands/async.c
@@ -107,7 +107,7 @@
* frontend during startup.) The above design guarantees that notifies from
* other backends will never be missed by ignoring self-notifies.
*
- * The amount of shared memory used for notify management (NUM_NOTIFY_BUFFERS)
+ * The amount of shared memory used for notify management (notify_buffers)
* can be varied without affecting anything but performance. The maximum
* amount of notification data that can be queued at one time is determined
* by slru.c's wraparound limit; see QUEUE_MAX_PAGE below.
@@ -225,7 +225,7 @@ typedef struct QueuePosition
*
* Resist the temptation to make this really large. While that would save
* work in some places, it would add cost in others. In particular, this
- * should likely be less than NUM_NOTIFY_BUFFERS, to ensure that backends
+ * should likely be less than notify_buffers, to ensure that backends
* catch up before the pages they'll need to read fall out of SLRU cache.
*/
#define QUEUE_CLEANUP_DELAY 4
@@ -514,7 +514,7 @@ AsyncShmemSize(void)
size = mul_size(MaxBackends + 1, sizeof(QueueBackendStatus));
size = add_size(size, offsetof(AsyncQueueControl, backend));
- size = add_size(size, SimpleLruShmemSize(NUM_NOTIFY_BUFFERS, 0));
+ size = add_size(size, SimpleLruShmemSize(notify_buffers, 0));
return size;
}
@@ -562,7 +562,7 @@ AsyncShmemInit(void)
* Set up SLRU management of the pg_notify data.
*/
NotifyCtl->PagePrecedes = asyncQueuePagePrecedes;
- SimpleLruInit(NotifyCtl, "Notify", NUM_NOTIFY_BUFFERS, 0,
+ SimpleLruInit(NotifyCtl, "Notify", notify_buffers, 0,
NotifySLRULock, "pg_notify", LWTRANCHE_NOTIFY_BUFFER,
SYNC_HANDLER_NONE);
diff --git a/src/backend/storage/lmgr/predicate.c b/src/backend/storage/lmgr/predicate.c
index d493aeef0f..b1f4f1651d 100644
--- a/src/backend/storage/lmgr/predicate.c
+++ b/src/backend/storage/lmgr/predicate.c
@@ -872,7 +872,7 @@ SerialInit(void)
*/
SerialSlruCtl->PagePrecedes = SerialPagePrecedesLogically;
SimpleLruInit(SerialSlruCtl, "Serial",
- NUM_SERIAL_BUFFERS, 0, SerialSLRULock, "pg_serial",
+ serial_buffers, 0, SerialSLRULock, "pg_serial",
LWTRANCHE_SERIAL_BUFFER, SYNC_HANDLER_NONE);
#ifdef USE_ASSERT_CHECKING
SerialPagePrecedesLogicallyUnitTests();
@@ -1395,7 +1395,7 @@ PredicateLockShmemSize(void)
/* Shared memory structures for SLRU tracking of old committed xids. */
size = add_size(size, sizeof(SerialControlData));
- size = add_size(size, SimpleLruShmemSize(NUM_SERIAL_BUFFERS, 0));
+ size = add_size(size, SimpleLruShmemSize(serial_buffers, 0));
return size;
}
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
index 381d9e548d..c83151d5ab 100644
--- a/src/backend/utils/init/globals.c
+++ b/src/backend/utils/init/globals.c
@@ -150,3 +150,11 @@ int64 VacuumPageDirty = 0;
int VacuumCostBalance = 0; /* working state for vacuum */
bool VacuumCostActive = false;
+
+int multixact_offsets_buffers = 8;
+int multixact_members_buffers = 16;
+int subtrans_buffers = 32;
+int notify_buffers = 8;
+int serial_buffers = 16;
+int xact_buffers = 0;
+int commit_ts_buffers = 0;
diff --git a/src/backend/utils/misc/guc.c b/src/backend/utils/misc/guc.c
index bee976bae8..250cef80bd 100644
--- a/src/backend/utils/misc/guc.c
+++ b/src/backend/utils/misc/guc.c
@@ -32,9 +32,11 @@
#endif
#include <unistd.h>
+#include "access/clog.h"
#include "access/commit_ts.h"
#include "access/gin.h"
#include "access/rmgr.h"
+#include "access/slru.h"
#include "access/tableam.h"
#include "access/toast_compression.h"
#include "access/transam.h"
@@ -200,6 +202,8 @@ static const char *show_tcp_keepalives_idle(void);
static const char *show_tcp_keepalives_interval(void);
static const char *show_tcp_keepalives_count(void);
static const char *show_tcp_user_timeout(void);
+static const char *show_xact_buffers(void);
+static const char *show_commit_ts_buffers(void);
static bool check_maxconnections(int *newval, void **extra, GucSource source);
static bool check_max_worker_processes(int *newval, void **extra, GucSource source);
static bool check_autovacuum_max_workers(int *newval, void **extra, GucSource source);
@@ -2340,6 +2344,83 @@ static struct config_int ConfigureNamesInt[] =
NULL, NULL, NULL
},
+ {
+ {"multixact_offsets_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for the MultiXact offset SLRU cache."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &multixact_offsets_buffers,
+ 8, 2, SLRU_MAX_ALLOWED_BUFFERS,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"multixact_members_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for the MultiXact member SLRU cache."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &multixact_members_buffers,
+ 16, 2, SLRU_MAX_ALLOWED_BUFFERS,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"subtrans_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for the sub-transaction SLRU cache."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &subtrans_buffers,
+ 32, 2, SLRU_MAX_ALLOWED_BUFFERS,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"notify_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for the NOTIFY message SLRU cache."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ ¬ify_buffers,
+ 8, 2, SLRU_MAX_ALLOWED_BUFFERS,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"serial_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for the serializable transaction SLRU cache."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &serial_buffers,
+ 16, 2, SLRU_MAX_ALLOWED_BUFFERS,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"xact_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for the transaction status SLRU cache."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &xact_buffers,
+ 0, 0, CLOG_MAX_ALLOWED_BUFFERS,
+ NULL, NULL, show_xact_buffers
+ },
+
+ {
+ {"commit_ts_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the size of the dedicated buffer pool used for the commit timestamp SLRU cache."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &commit_ts_buffers,
+ 0, 0, SLRU_MAX_ALLOWED_BUFFERS,
+ NULL, NULL, show_commit_ts_buffers
+ },
+
{
{"temp_buffers", PGC_USERSET, RESOURCES_MEM,
gettext_noop("Sets the maximum number of temporary buffers used by each session."),
@@ -11959,6 +12040,24 @@ show_tcp_user_timeout(void)
return nbuf;
}
+static const char *
+show_xact_buffers(void)
+{
+ static char nbuf[16];
+
+ snprintf(nbuf, sizeof(nbuf), "%zu", CLOGShmemBuffers());
+ return nbuf;
+}
+
+static const char *
+show_commit_ts_buffers(void)
+{
+ static char nbuf[16];
+
+ snprintf(nbuf, sizeof(nbuf), "%zu", CommitTsShmemBuffers());
+ return nbuf;
+}
+
static bool
check_maxconnections(int *newval, void **extra, GucSource source)
{
diff --git a/src/backend/utils/misc/postgresql.conf.sample b/src/backend/utils/misc/postgresql.conf.sample
index ff9fa006fe..7e14df3b51 100644
--- a/src/backend/utils/misc/postgresql.conf.sample
+++ b/src/backend/utils/misc/postgresql.conf.sample
@@ -190,6 +190,15 @@
# (change requires restart)
#backend_flush_after = 0 # measured in pages, 0 disables
+# - SLRU Buffers (change requires restart) -
+
+#xact_buffers = 0 # memory for pg_xact (0 = auto)
+#subtrans_buffers = 32 # memory for pg_subtrans
+#multixact_offsets_buffers = 8 # memory for pg_multixact/offsets
+#multixact_members_buffers = 16 # memory for pg_multixact/members
+#notify_buffers = 8 # memory for pg_notify
+#serial_buffers = 16 # memory for pg_serial
+#commit_ts_buffers = 0 # memory for pg_commit_ts (0 = auto)
#------------------------------------------------------------------------------
# WRITE-AHEAD LOG
diff --git a/src/include/access/clog.h b/src/include/access/clog.h
index 39b8e4afa8..739a292f7f 100644
--- a/src/include/access/clog.h
+++ b/src/include/access/clog.h
@@ -15,6 +15,16 @@
#include "storage/sync.h"
#include "lib/stringinfo.h"
+/*
+ * Don't allow xact_buffers to be set higher than could possibly be useful or
+ * SLRU would allow.
+ */
+#define CLOG_XACTS_PER_BYTE 4
+#define CLOG_XACTS_PER_PAGE (BLCKSZ * CLOG_XACTS_PER_BYTE)
+#define CLOG_MAX_ALLOWED_BUFFERS \
+ Min(SLRU_MAX_ALLOWED_BUFFERS, \
+ (((MaxTransactionId / 2) + (CLOG_XACTS_PER_PAGE - 1)) / CLOG_XACTS_PER_PAGE))
+
/*
* Possible transaction statuses --- note that all-zeroes is the initial
* state.
diff --git a/src/include/access/commit_ts.h b/src/include/access/commit_ts.h
index 750369104a..e4cf988609 100644
--- a/src/include/access/commit_ts.h
+++ b/src/include/access/commit_ts.h
@@ -17,7 +17,6 @@
#include "storage/sync.h"
#include "utils/guc.h"
-
extern PGDLLIMPORT bool track_commit_timestamp;
extern bool check_track_commit_timestamp(bool *newval, void **extra,
diff --git a/src/include/access/multixact.h b/src/include/access/multixact.h
index 4bbb035eae..97c0a46376 100644
--- a/src/include/access/multixact.h
+++ b/src/include/access/multixact.h
@@ -29,10 +29,6 @@
#define MaxMultiXactOffset ((MultiXactOffset) 0xFFFFFFFF)
-/* Number of SLRU buffers to use for multixact */
-#define NUM_MULTIXACTOFFSET_BUFFERS 8
-#define NUM_MULTIXACTMEMBER_BUFFERS 16
-
/*
* Possible multixact lock modes ("status"). The first four modes are for
* tuple locks (FOR KEY SHARE, FOR SHARE, FOR NO KEY UPDATE, FOR UPDATE); the
diff --git a/src/include/access/slru.h b/src/include/access/slru.h
index dd52e8cec7..793c045f16 100644
--- a/src/include/access/slru.h
+++ b/src/include/access/slru.h
@@ -17,6 +17,11 @@
#include "storage/lwlock.h"
#include "storage/sync.h"
+/*
+ * To avoid overflowing internal arithmetic and the size_t data type, the
+ * number of buffers should not exceed this number.
+ */
+#define SLRU_MAX_ALLOWED_BUFFERS ((1024 * 1024 * 1024) / BLCKSZ)
/*
* Define SLRU segment size. A page is the same BLCKSZ as is used everywhere
diff --git a/src/include/access/subtrans.h b/src/include/access/subtrans.h
index d0ab44ae82..64fa86938e 100644
--- a/src/include/access/subtrans.h
+++ b/src/include/access/subtrans.h
@@ -11,8 +11,6 @@
#ifndef SUBTRANS_H
#define SUBTRANS_H
-/* Number of SLRU buffers to use for subtrans */
-#define NUM_SUBTRANS_BUFFERS 32
extern void SubTransSetParent(TransactionId xid, TransactionId parent);
extern TransactionId SubTransGetParent(TransactionId xid);
diff --git a/src/include/commands/async.h b/src/include/commands/async.h
index 9217f66b91..fa831e3721 100644
--- a/src/include/commands/async.h
+++ b/src/include/commands/async.h
@@ -15,11 +15,6 @@
#include <signal.h>
-/*
- * The number of SLRU page buffers we use for the notification queue.
- */
-#define NUM_NOTIFY_BUFFERS 8
-
extern bool Trace_notify;
extern volatile sig_atomic_t notifyInterruptPending;
diff --git a/src/include/miscadmin.h b/src/include/miscadmin.h
index 95202d37af..495c1bf901 100644
--- a/src/include/miscadmin.h
+++ b/src/include/miscadmin.h
@@ -164,6 +164,13 @@ extern PGDLLIMPORT int MaxBackends;
extern PGDLLIMPORT int MaxConnections;
extern PGDLLIMPORT int max_worker_processes;
extern PGDLLIMPORT int max_parallel_workers;
+extern PGDLLIMPORT int multixact_offsets_buffers;
+extern PGDLLIMPORT int multixact_members_buffers;
+extern PGDLLIMPORT int subtrans_buffers;
+extern PGDLLIMPORT int notify_buffers;
+extern PGDLLIMPORT int serial_buffers;
+extern PGDLLIMPORT int xact_buffers;
+extern PGDLLIMPORT int commit_ts_buffers;
extern PGDLLIMPORT int MyProcPid;
extern PGDLLIMPORT pg_time_t MyStartTime;
diff --git a/src/include/storage/predicate.h b/src/include/storage/predicate.h
index 152b698611..c72779bd88 100644
--- a/src/include/storage/predicate.h
+++ b/src/include/storage/predicate.h
@@ -26,10 +26,6 @@ extern int max_predicate_locks_per_xact;
extern int max_predicate_locks_per_relation;
extern int max_predicate_locks_per_page;
-
-/* Number of SLRU buffers to use for Serial SLRU */
-#define NUM_SERIAL_BUFFERS 16
-
/*
* A handle used for sharing SERIALIZABLEXACT objects between the participants
* in a parallel query.
--
2.24.3 (Apple Git-128)
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2021-06-14 17:40 Васильев Дмитрий <[email protected]>
parent: Andrey Borodin <[email protected]>
1 sibling, 0 replies; 103+ messages in thread
From: Васильев Дмитрий @ 2021-06-14 17:40 UTC (permalink / raw)
To: Andrey Borodin <[email protected]>; +Cc: Thomas Munro <[email protected]>; Gilles Darold <[email protected]>; Tomas Vondra <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
пн, 14 июн. 2021 г. в 15:07, Andrey Borodin <[email protected]>:
> PFA patch implementing this idea.
>
I'm benchmarked v17 patches.
Testing was done on a 96-core machine, with PGDATA completely placed in
tmpfs.
PostgreSQL was built with CFLAGS -O2.
for-update PgBench script:
\set aid random_zipfian(1, 100, 2)
begin;
select :aid from pgbench_accounts where aid = :aid for update;
update pgbench_accounts set abalance = abalance + 1 where aid = :aid;
update pgbench_accounts set abalance = abalance * 2 where aid = :aid;
update pgbench_accounts set abalance = abalance - 2 where aid = :aid;
end;
Before each test sample data was filled with "pgbench -i -s 100", testing
was performed 3 times for 1 hour each test.
The benchmark results are presented with changing
multi_xact_members_buffers and multicast_offsets_buffers (1:2 respectively):
settings tps
multixact_members_buffers_64Kb 693.2
multixact_members_buffers_128Kb 691.4
multixact_members_buffers_192Kb 696.3
multixact_members_buffers_256Kb 694.4
multixact_members_buffers_320Kb 692.3
multixact_members_buffers_448Kb 693.7
multixact_members_buffers_512Kb 693.3
vanilla 676.1
Best regards, Dmitry Vasiliev.
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2021-12-26 10:09 Andrey Borodin <[email protected]>
parent: Andrey Borodin <[email protected]>
1 sibling, 0 replies; 103+ messages in thread
From: Andrey Borodin @ 2021-12-26 10:09 UTC (permalink / raw)
To: Andrey Borodin <[email protected]>; +Cc: Gilles Darold <[email protected]>; Tomas Vondra <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers; Thomas Munro <[email protected]>
>> 8 апр. 2021 г., в 15:22, Thomas Munro <[email protected]> написал(а):
>>
> I have one more idea inspired by CPU caches.
> Let's make SLRU n-associative, where n ~ 8.
> We can divide buffers into "banks", number of banks must be power of 2.
> All banks are of equal size. We choose bank size to approximately satisfy user's configured buffer size.
> Each page can live only within one bank. We use same search and eviction algorithms as we used in SLRU, but we only need to search\evict over 8 elements.
> All SLRU data of a single bank will be colocated within at most 2 cache line.
>
> I did not come up with idea how to avoid multiplication of bank_number * bank_size in case when user configured 31337 buffers (any number that is radically not a power of 2).
We can avoid this multiplication by using gapped memory under SLRU page_statuses, but from my POV here complexity does not worth possible performance gain.
PFA rebase of the patchset. Also I've added a patch to combine page_number, page_status, and page_dirty together to touch less cachelines.
Best regards, Andrey Borodin.
Attachments:
[text/x-diff] v-18-0001-Make-all-SLRU-buffer-sizes-configurable.patch (25.7K, ../../[email protected]/2-v-18-0001-Make-all-SLRU-buffer-sizes-configurable.patch)
download | inline diff:
From ea59d2ebde818ddc2a9111858b3d956cbcc7bff2 Mon Sep 17 00:00:00 2001
From: Andrey Borodin <[email protected]>
Date: Mon, 15 Feb 2021 21:51:56 +0500
Subject: [PATCH v=18 1/3] Make all SLRU buffer sizes configurable.
Provide new GUCs to set the number of buffers, instead of using hard
coded defaults.
Remove the limits on xact_buffers and commit_ts_buffers. The default
sizes for those caches are ~0.2% and ~0.1% of shared_buffers, as before,
but now there is no cap at 128 and 16 buffers respectively (unless
track_commit_timestamp is disabled, in the latter case, then we might as
well keep it tiny). Sizes much larger than the old limits have been
shown to be useful on modern systems, and an earlier commit replaced a
linear search with a hash table to avoid problems with extreme cases.
Author: Andrey M. Borodin <[email protected]>
Reviewed-by: Anastasia Lubennikova <[email protected]>
Reviewed-by: Tomas Vondra <[email protected]>
Reviewed-by: Alexander Korotkov <[email protected]>
Reviewed-by: Gilles Darold <[email protected]>
Reviewed-by: Thomas Munro <[email protected]>
Discussion: https://postgr.es/m/2BEC2B3F-9B61-4C1D-9FB5-5FAB0F05EF86%40yandex-team.ru
---
doc/src/sgml/config.sgml | 135 ++++++++++++++++++
src/backend/access/transam/clog.c | 23 ++-
src/backend/access/transam/commit_ts.c | 5 +
src/backend/access/transam/multixact.c | 8 +-
src/backend/access/transam/subtrans.c | 5 +-
src/backend/commands/async.c | 8 +-
src/backend/storage/lmgr/predicate.c | 4 +-
src/backend/utils/init/globals.c | 8 ++
src/backend/utils/misc/guc.c | 99 +++++++++++++
src/backend/utils/misc/postgresql.conf.sample | 9 ++
src/include/access/clog.h | 10 ++
src/include/access/commit_ts.h | 1 -
src/include/access/multixact.h | 4 -
src/include/access/slru.h | 5 +
src/include/access/subtrans.h | 2 -
src/include/commands/async.h | 5 -
src/include/miscadmin.h | 7 +
src/include/storage/predicate.h | 4 -
18 files changed, 299 insertions(+), 43 deletions(-)
diff --git a/doc/src/sgml/config.sgml b/doc/src/sgml/config.sgml
index afbb6c35e30..57d9696abe8 100644
--- a/doc/src/sgml/config.sgml
+++ b/doc/src/sgml/config.sgml
@@ -1952,6 +1952,141 @@ include_dir 'conf.d'
</para>
</listitem>
</varlistentry>
+
+ <varlistentry id="guc-multixact-offsets-buffers" xreflabel="multixact_offsets_buffers">
+ <term><varname>multixact_offsets_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>multixact_offsets_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to use to cache the contents
+ of <literal>pg_multixact/offsets</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>8</literal>.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-multixact-members-buffers" xreflabel="multixact_members_buffers">
+ <term><varname>multixact_members_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>multixact_members_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to use to cache the contents
+ of <literal>pg_multixact/members</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>16</literal>.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-subtrans-buffers" xreflabel="subtrans_buffers">
+ <term><varname>subtrans_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>subtrans_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to use to cache the contents
+ of <literal>pg_subtrans</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>8</literal>.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-notify-buffers" xreflabel="notify_buffers">
+ <term><varname>notify_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>notify_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to use to cache the contents
+ of <literal>pg_notify</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>8</literal>.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-serial-buffers" xreflabel="serial_buffers">
+ <term><varname>serial_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>serial_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to use to cache the contents
+ of <literal>pg_serial</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>16</literal>.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-xact-buffers" xreflabel="xact_buffers">
+ <term><varname>xact_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>xact_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to use to cache the contents
+ of <literal>pg_xact</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>0</literal>, which requests
+ <varname>shared_buffers</varname> / 512, but not fewer than 4 blocks.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-commit-ts-buffers" xreflabel="commit_ts_buffers">
+ <term><varname>commit_ts_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>commit_ts_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of memory to use to cache the cotents of
+ <literal>pg_commit_ts</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>0</literal>, which requests
+ <varname>shared_buffers</varname> / 1024, but not fewer than 4 blocks.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
<varlistentry id="guc-max-stack-depth" xreflabel="max_stack_depth">
<term><varname>max_stack_depth</varname> (<type>integer</type>)
diff --git a/src/backend/access/transam/clog.c b/src/backend/access/transam/clog.c
index 3ea16a270a8..ca28ada75fa 100644
--- a/src/backend/access/transam/clog.c
+++ b/src/backend/access/transam/clog.c
@@ -58,8 +58,8 @@
/* We need two bits per xact, so four xacts fit in a byte */
#define CLOG_BITS_PER_XACT 2
-#define CLOG_XACTS_PER_BYTE 4
-#define CLOG_XACTS_PER_PAGE (BLCKSZ * CLOG_XACTS_PER_BYTE)
+StaticAssertDecl((CLOG_BITS_PER_XACT * CLOG_XACTS_PER_BYTE) == BITS_PER_BYTE,
+ "CLOG_BITS_PER_XACT and CLOG_XACTS_PER_BYTE are inconsistent");
#define CLOG_XACT_BITMASK ((1 << CLOG_BITS_PER_XACT) - 1)
#define TransactionIdToPage(xid) ((xid) / (TransactionId) CLOG_XACTS_PER_PAGE)
@@ -664,23 +664,16 @@ TransactionIdGetStatus(TransactionId xid, XLogRecPtr *lsn)
/*
* Number of shared CLOG buffers.
*
- * On larger multi-processor systems, it is possible to have many CLOG page
- * requests in flight at one time which could lead to disk access for CLOG
- * page if the required page is not found in memory. Testing revealed that we
- * can get the best performance by having 128 CLOG buffers, more than that it
- * doesn't improve performance.
- *
- * Unconditionally keeping the number of CLOG buffers to 128 did not seem like
- * a good idea, because it would increase the minimum amount of shared memory
- * required to start, which could be a problem for people running very small
- * configurations. The following formula seems to represent a reasonable
- * compromise: people with very low values for shared_buffers will get fewer
- * CLOG buffers as well, and everyone else will get 128.
+ * By default, we'll use 2MB of for every 1GB of shared buffers, up to the
+ * theoretical maximum useful value, but always at least 4 buffers.
*/
Size
CLOGShmemBuffers(void)
{
- return Min(128, Max(4, NBuffers / 512));
+ /* Use configured value if provided. */
+ if (xact_buffers > 0)
+ return Max(4, xact_buffers);
+ return Min(CLOG_MAX_ALLOWED_BUFFERS, Max(4, NBuffers / 512));
}
/*
diff --git a/src/backend/access/transam/commit_ts.c b/src/backend/access/transam/commit_ts.c
index cbbe19fea83..cb2e0ceb1c3 100644
--- a/src/backend/access/transam/commit_ts.c
+++ b/src/backend/access/transam/commit_ts.c
@@ -511,10 +511,15 @@ pg_xact_commit_timestamp_origin(PG_FUNCTION_ARGS)
* We use a very similar logic as for the number of CLOG buffers (except we
* scale up twice as fast with shared buffers, and the maximum is twice as
* high); see comments in CLOGShmemBuffers.
+ * By default, we'll use 1MB of for every 1GB of shared buffers, up to the
+ * maximum value that slru.c will allow, but always at least 4 buffers.
*/
Size
CommitTsShmemBuffers(void)
{
+ /* Use configured value if provided. */
+ if (commit_ts_buffers > 0)
+ return Max(4, commit_ts_buffers);
return Min(256, Max(4, NBuffers / 256));
}
diff --git a/src/backend/access/transam/multixact.c b/src/backend/access/transam/multixact.c
index e6c70ed0bc2..a29ab4769dc 100644
--- a/src/backend/access/transam/multixact.c
+++ b/src/backend/access/transam/multixact.c
@@ -1834,8 +1834,8 @@ MultiXactShmemSize(void)
mul_size(sizeof(MultiXactId) * 2, MaxOldestSlot))
size = SHARED_MULTIXACT_STATE_SIZE;
- size = add_size(size, SimpleLruShmemSize(NUM_MULTIXACTOFFSET_BUFFERS, 0));
- size = add_size(size, SimpleLruShmemSize(NUM_MULTIXACTMEMBER_BUFFERS, 0));
+ size = add_size(size, SimpleLruShmemSize(multixact_offsets_buffers, 0));
+ size = add_size(size, SimpleLruShmemSize(multixact_members_buffers, 0));
return size;
}
@@ -1851,13 +1851,13 @@ MultiXactShmemInit(void)
MultiXactMemberCtl->PagePrecedes = MultiXactMemberPagePrecedes;
SimpleLruInit(MultiXactOffsetCtl,
- "MultiXactOffset", NUM_MULTIXACTOFFSET_BUFFERS, 0,
+ "MultiXactOffset", multixact_offsets_buffers, 0,
MultiXactOffsetSLRULock, "pg_multixact/offsets",
LWTRANCHE_MULTIXACTOFFSET_BUFFER,
SYNC_HANDLER_MULTIXACT_OFFSET);
SlruPagePrecedesUnitTests(MultiXactOffsetCtl, MULTIXACT_OFFSETS_PER_PAGE);
SimpleLruInit(MultiXactMemberCtl,
- "MultiXactMember", NUM_MULTIXACTMEMBER_BUFFERS, 0,
+ "MultiXactMember", multixact_offsets_buffers, 0,
MultiXactMemberSLRULock, "pg_multixact/members",
LWTRANCHE_MULTIXACTMEMBER_BUFFER,
SYNC_HANDLER_MULTIXACT_MEMBER);
diff --git a/src/backend/access/transam/subtrans.c b/src/backend/access/transam/subtrans.c
index 6a8e521f894..785f2520fde 100644
--- a/src/backend/access/transam/subtrans.c
+++ b/src/backend/access/transam/subtrans.c
@@ -31,6 +31,7 @@
#include "access/slru.h"
#include "access/subtrans.h"
#include "access/transam.h"
+#include "miscadmin.h"
#include "pg_trace.h"
#include "utils/snapmgr.h"
@@ -184,14 +185,14 @@ SubTransGetTopmostTransaction(TransactionId xid)
Size
SUBTRANSShmemSize(void)
{
- return SimpleLruShmemSize(NUM_SUBTRANS_BUFFERS, 0);
+ return SimpleLruShmemSize(subtrans_buffers, 0);
}
void
SUBTRANSShmemInit(void)
{
SubTransCtl->PagePrecedes = SubTransPagePrecedes;
- SimpleLruInit(SubTransCtl, "Subtrans", NUM_SUBTRANS_BUFFERS, 0,
+ SimpleLruInit(SubTransCtl, "Subtrans", subtrans_buffers, 0,
SubtransSLRULock, "pg_subtrans",
LWTRANCHE_SUBTRANS_BUFFER, SYNC_HANDLER_NONE);
SlruPagePrecedesUnitTests(SubTransCtl, SUBTRANS_XACTS_PER_PAGE);
diff --git a/src/backend/commands/async.c b/src/backend/commands/async.c
index 85570085450..7f2b7598449 100644
--- a/src/backend/commands/async.c
+++ b/src/backend/commands/async.c
@@ -117,7 +117,7 @@
* frontend during startup.) The above design guarantees that notifies from
* other backends will never be missed by ignoring self-notifies.
*
- * The amount of shared memory used for notify management (NUM_NOTIFY_BUFFERS)
+ * The amount of shared memory used for notify management (notify_buffers)
* can be varied without affecting anything but performance. The maximum
* amount of notification data that can be queued at one time is determined
* by slru.c's wraparound limit; see QUEUE_MAX_PAGE below.
@@ -235,7 +235,7 @@ typedef struct QueuePosition
*
* Resist the temptation to make this really large. While that would save
* work in some places, it would add cost in others. In particular, this
- * should likely be less than NUM_NOTIFY_BUFFERS, to ensure that backends
+ * should likely be less than notify_buffers, to ensure that backends
* catch up before the pages they'll need to read fall out of SLRU cache.
*/
#define QUEUE_CLEANUP_DELAY 4
@@ -521,7 +521,7 @@ AsyncShmemSize(void)
size = mul_size(MaxBackends + 1, sizeof(QueueBackendStatus));
size = add_size(size, offsetof(AsyncQueueControl, backend));
- size = add_size(size, SimpleLruShmemSize(NUM_NOTIFY_BUFFERS, 0));
+ size = add_size(size, SimpleLruShmemSize(notify_buffers, 0));
return size;
}
@@ -569,7 +569,7 @@ AsyncShmemInit(void)
* Set up SLRU management of the pg_notify data.
*/
NotifyCtl->PagePrecedes = asyncQueuePagePrecedes;
- SimpleLruInit(NotifyCtl, "Notify", NUM_NOTIFY_BUFFERS, 0,
+ SimpleLruInit(NotifyCtl, "Notify", notify_buffers, 0,
NotifySLRULock, "pg_notify", LWTRANCHE_NOTIFY_BUFFER,
SYNC_HANDLER_NONE);
diff --git a/src/backend/storage/lmgr/predicate.c b/src/backend/storage/lmgr/predicate.c
index 4f4d5b0d20f..c5e66757643 100644
--- a/src/backend/storage/lmgr/predicate.c
+++ b/src/backend/storage/lmgr/predicate.c
@@ -872,7 +872,7 @@ SerialInit(void)
*/
SerialSlruCtl->PagePrecedes = SerialPagePrecedesLogically;
SimpleLruInit(SerialSlruCtl, "Serial",
- NUM_SERIAL_BUFFERS, 0, SerialSLRULock, "pg_serial",
+ serial_buffers, 0, SerialSLRULock, "pg_serial",
LWTRANCHE_SERIAL_BUFFER, SYNC_HANDLER_NONE);
#ifdef USE_ASSERT_CHECKING
SerialPagePrecedesLogicallyUnitTests();
@@ -1396,7 +1396,7 @@ PredicateLockShmemSize(void)
/* Shared memory structures for SLRU tracking of old committed xids. */
size = add_size(size, sizeof(SerialControlData));
- size = add_size(size, SimpleLruShmemSize(NUM_SERIAL_BUFFERS, 0));
+ size = add_size(size, SimpleLruShmemSize(serial_buffers, 0));
return size;
}
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
index 381d9e548d1..c83151d5ab5 100644
--- a/src/backend/utils/init/globals.c
+++ b/src/backend/utils/init/globals.c
@@ -150,3 +150,11 @@ int64 VacuumPageDirty = 0;
int VacuumCostBalance = 0; /* working state for vacuum */
bool VacuumCostActive = false;
+
+int multixact_offsets_buffers = 8;
+int multixact_members_buffers = 16;
+int subtrans_buffers = 32;
+int notify_buffers = 8;
+int serial_buffers = 16;
+int xact_buffers = 0;
+int commit_ts_buffers = 0;
diff --git a/src/backend/utils/misc/guc.c b/src/backend/utils/misc/guc.c
index bff949a40bc..eb6bf4c0a04 100644
--- a/src/backend/utils/misc/guc.c
+++ b/src/backend/utils/misc/guc.c
@@ -32,9 +32,11 @@
#endif
#include <unistd.h>
+#include "access/clog.h"
#include "access/commit_ts.h"
#include "access/gin.h"
#include "access/rmgr.h"
+#include "access/slru.h"
#include "access/tableam.h"
#include "access/toast_compression.h"
#include "access/transam.h"
@@ -202,6 +204,8 @@ static const char *show_tcp_keepalives_idle(void);
static const char *show_tcp_keepalives_interval(void);
static const char *show_tcp_keepalives_count(void);
static const char *show_tcp_user_timeout(void);
+static const char *show_xact_buffers(void);
+static const char *show_commit_ts_buffers(void);
static bool check_maxconnections(int *newval, void **extra, GucSource source);
static bool check_max_worker_processes(int *newval, void **extra, GucSource source);
static bool check_autovacuum_max_workers(int *newval, void **extra, GucSource source);
@@ -2366,6 +2370,83 @@ static struct config_int ConfigureNamesInt[] =
NULL, NULL, NULL
},
+ {
+ {"multixact_offsets_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for the MultiXact offset SLRU cache."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &multixact_offsets_buffers,
+ 8, 2, SLRU_MAX_ALLOWED_BUFFERS,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"multixact_members_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for the MultiXact member SLRU cache."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &multixact_members_buffers,
+ 16, 2, SLRU_MAX_ALLOWED_BUFFERS,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"subtrans_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for the sub-transaction SLRU cache."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &subtrans_buffers,
+ 32, 2, SLRU_MAX_ALLOWED_BUFFERS,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"notify_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for the NOTIFY message SLRU cache."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ ¬ify_buffers,
+ 8, 2, SLRU_MAX_ALLOWED_BUFFERS,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"serial_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for the serializable transaction SLRU cache."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &serial_buffers,
+ 16, 2, SLRU_MAX_ALLOWED_BUFFERS,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"xact_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for the transaction status SLRU cache."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &xact_buffers,
+ 0, 0, CLOG_MAX_ALLOWED_BUFFERS,
+ NULL, NULL, show_xact_buffers
+ },
+
+ {
+ {"commit_ts_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the size of the dedicated buffer pool used for the commit timestamp SLRU cache."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &commit_ts_buffers,
+ 0, 0, SLRU_MAX_ALLOWED_BUFFERS,
+ NULL, NULL, show_commit_ts_buffers
+ },
+
{
{"temp_buffers", PGC_USERSET, RESOURCES_MEM,
gettext_noop("Sets the maximum number of temporary buffers used by each session."),
@@ -12074,6 +12155,24 @@ show_tcp_user_timeout(void)
return nbuf;
}
+static const char *
+show_xact_buffers(void)
+{
+ static char nbuf[16];
+
+ snprintf(nbuf, sizeof(nbuf), "%zu", CLOGShmemBuffers());
+ return nbuf;
+}
+
+static const char *
+show_commit_ts_buffers(void)
+{
+ static char nbuf[16];
+
+ snprintf(nbuf, sizeof(nbuf), "%zu", CommitTsShmemBuffers());
+ return nbuf;
+}
+
static bool
check_maxconnections(int *newval, void **extra, GucSource source)
{
diff --git a/src/backend/utils/misc/postgresql.conf.sample b/src/backend/utils/misc/postgresql.conf.sample
index a1acd46b611..22bda4383ce 100644
--- a/src/backend/utils/misc/postgresql.conf.sample
+++ b/src/backend/utils/misc/postgresql.conf.sample
@@ -195,6 +195,15 @@
#old_snapshot_threshold = -1 # 1min-60d; -1 disables; 0 is immediate
# (change requires restart)
+# - SLRU Buffers (change requires restart) -
+
+#xact_buffers = 0 # memory for pg_xact (0 = auto)
+#subtrans_buffers = 32 # memory for pg_subtrans
+#multixact_offsets_buffers = 8 # memory for pg_multixact/offsets
+#multixact_members_buffers = 16 # memory for pg_multixact/members
+#notify_buffers = 8 # memory for pg_notify
+#serial_buffers = 16 # memory for pg_serial
+#commit_ts_buffers = 0 # memory for pg_commit_ts (0 = auto)
#------------------------------------------------------------------------------
# WRITE-AHEAD LOG
diff --git a/src/include/access/clog.h b/src/include/access/clog.h
index 39b8e4afa8a..739a292f7f3 100644
--- a/src/include/access/clog.h
+++ b/src/include/access/clog.h
@@ -15,6 +15,16 @@
#include "storage/sync.h"
#include "lib/stringinfo.h"
+/*
+ * Don't allow xact_buffers to be set higher than could possibly be useful or
+ * SLRU would allow.
+ */
+#define CLOG_XACTS_PER_BYTE 4
+#define CLOG_XACTS_PER_PAGE (BLCKSZ * CLOG_XACTS_PER_BYTE)
+#define CLOG_MAX_ALLOWED_BUFFERS \
+ Min(SLRU_MAX_ALLOWED_BUFFERS, \
+ (((MaxTransactionId / 2) + (CLOG_XACTS_PER_PAGE - 1)) / CLOG_XACTS_PER_PAGE))
+
/*
* Possible transaction statuses --- note that all-zeroes is the initial
* state.
diff --git a/src/include/access/commit_ts.h b/src/include/access/commit_ts.h
index a1538978c62..f86760f7240 100644
--- a/src/include/access/commit_ts.h
+++ b/src/include/access/commit_ts.h
@@ -16,7 +16,6 @@
#include "replication/origin.h"
#include "storage/sync.h"
-
extern PGDLLIMPORT bool track_commit_timestamp;
extern void TransactionTreeSetCommitTsData(TransactionId xid, int nsubxids,
diff --git a/src/include/access/multixact.h b/src/include/access/multixact.h
index 4bbb035eaea..97c0a463768 100644
--- a/src/include/access/multixact.h
+++ b/src/include/access/multixact.h
@@ -29,10 +29,6 @@
#define MaxMultiXactOffset ((MultiXactOffset) 0xFFFFFFFF)
-/* Number of SLRU buffers to use for multixact */
-#define NUM_MULTIXACTOFFSET_BUFFERS 8
-#define NUM_MULTIXACTMEMBER_BUFFERS 16
-
/*
* Possible multixact lock modes ("status"). The first four modes are for
* tuple locks (FOR KEY SHARE, FOR SHARE, FOR NO KEY UPDATE, FOR UPDATE); the
diff --git a/src/include/access/slru.h b/src/include/access/slru.h
index dd52e8cec7e..793c045f160 100644
--- a/src/include/access/slru.h
+++ b/src/include/access/slru.h
@@ -17,6 +17,11 @@
#include "storage/lwlock.h"
#include "storage/sync.h"
+/*
+ * To avoid overflowing internal arithmetic and the size_t data type, the
+ * number of buffers should not exceed this number.
+ */
+#define SLRU_MAX_ALLOWED_BUFFERS ((1024 * 1024 * 1024) / BLCKSZ)
/*
* Define SLRU segment size. A page is the same BLCKSZ as is used everywhere
diff --git a/src/include/access/subtrans.h b/src/include/access/subtrans.h
index d0ab44ae828..64fa86938e2 100644
--- a/src/include/access/subtrans.h
+++ b/src/include/access/subtrans.h
@@ -11,8 +11,6 @@
#ifndef SUBTRANS_H
#define SUBTRANS_H
-/* Number of SLRU buffers to use for subtrans */
-#define NUM_SUBTRANS_BUFFERS 32
extern void SubTransSetParent(TransactionId xid, TransactionId parent);
extern TransactionId SubTransGetParent(TransactionId xid);
diff --git a/src/include/commands/async.h b/src/include/commands/async.h
index f371ac896b9..99575974982 100644
--- a/src/include/commands/async.h
+++ b/src/include/commands/async.h
@@ -15,11 +15,6 @@
#include <signal.h>
-/*
- * The number of SLRU page buffers we use for the notification queue.
- */
-#define NUM_NOTIFY_BUFFERS 8
-
extern bool Trace_notify;
extern volatile sig_atomic_t notifyInterruptPending;
diff --git a/src/include/miscadmin.h b/src/include/miscadmin.h
index 90a30160657..22d31546ad5 100644
--- a/src/include/miscadmin.h
+++ b/src/include/miscadmin.h
@@ -176,6 +176,13 @@ extern PGDLLIMPORT int MaxBackends;
extern PGDLLIMPORT int MaxConnections;
extern PGDLLIMPORT int max_worker_processes;
extern PGDLLIMPORT int max_parallel_workers;
+extern PGDLLIMPORT int multixact_offsets_buffers;
+extern PGDLLIMPORT int multixact_members_buffers;
+extern PGDLLIMPORT int subtrans_buffers;
+extern PGDLLIMPORT int notify_buffers;
+extern PGDLLIMPORT int serial_buffers;
+extern PGDLLIMPORT int xact_buffers;
+extern PGDLLIMPORT int commit_ts_buffers;
extern PGDLLIMPORT int MyProcPid;
extern PGDLLIMPORT pg_time_t MyStartTime;
diff --git a/src/include/storage/predicate.h b/src/include/storage/predicate.h
index 152b6986114..c72779bd88d 100644
--- a/src/include/storage/predicate.h
+++ b/src/include/storage/predicate.h
@@ -26,10 +26,6 @@ extern int max_predicate_locks_per_xact;
extern int max_predicate_locks_per_relation;
extern int max_predicate_locks_per_page;
-
-/* Number of SLRU buffers to use for Serial SLRU */
-#define NUM_SERIAL_BUFFERS 16
-
/*
* A handle used for sharing SERIALIZABLEXACT objects between the participants
* in a parallel query.
--
2.33.1
[text/x-diff] v-18-0003-Pack-SLRU-page_number-page_status-and-page_dirt.patch (23.6K, ../../[email protected]/3-v-18-0003-Pack-SLRU-page_number-page_status-and-page_dirt.patch)
download | inline diff:
From 5c1aace8b1f5a2b852bf1476dffbd4a43f196ef3 Mon Sep 17 00:00:00 2001
From: Andrey Borodin <[email protected]>
Date: Sun, 26 Dec 2021 15:03:30 +0500
Subject: [PATCH v=18 3/3] Pack SLRU page_number, page_status and page_dirty
toogether
This allows to test only one cacheline during successfull
SlruSelectLRUPage().
---
src/backend/access/transam/clog.c | 12 +--
src/backend/access/transam/commit_ts.c | 6 +-
src/backend/access/transam/multixact.c | 16 +--
src/backend/access/transam/slru.c | 143 ++++++++++++-------------
src/backend/access/transam/subtrans.c | 4 +-
src/backend/commands/async.c | 2 +-
src/backend/storage/lmgr/predicate.c | 2 +-
src/include/access/slru.h | 13 ++-
8 files changed, 99 insertions(+), 99 deletions(-)
diff --git a/src/backend/access/transam/clog.c b/src/backend/access/transam/clog.c
index ca28ada75fa..7d3a0286a5f 100644
--- a/src/backend/access/transam/clog.c
+++ b/src/backend/access/transam/clog.c
@@ -375,7 +375,7 @@ TransactionIdSetPageStatusInternal(TransactionId xid, int nsubxids,
{
for (i = 0; i < nsubxids; i++)
{
- Assert(XactCtl->shared->page_number[slotno] == TransactionIdToPage(subxids[i]));
+ Assert(XactCtl->shared->page_entries[slotno].page_number == TransactionIdToPage(subxids[i]));
TransactionIdSetStatusBit(subxids[i],
TRANSACTION_STATUS_SUB_COMMITTED,
lsn, slotno);
@@ -389,11 +389,11 @@ TransactionIdSetPageStatusInternal(TransactionId xid, int nsubxids,
/* Set the subtransactions */
for (i = 0; i < nsubxids; i++)
{
- Assert(XactCtl->shared->page_number[slotno] == TransactionIdToPage(subxids[i]));
+ Assert(XactCtl->shared->page_entries[slotno].page_number == TransactionIdToPage(subxids[i]));
TransactionIdSetStatusBit(subxids[i], status, lsn, slotno);
}
- XactCtl->shared->page_dirty[slotno] = true;
+ XactCtl->shared->page_entries[slotno].page_dirty = true;
}
/*
@@ -713,7 +713,7 @@ BootStrapCLOG(void)
/* Make sure it's written out */
SimpleLruWritePage(XactCtl, slotno);
- Assert(!XactCtl->shared->page_dirty[slotno]);
+ Assert(!XactCtl->shared->page_entries[slotno].page_dirty);
LWLockRelease(XactSLRULock);
}
@@ -798,7 +798,7 @@ TrimCLOG(void)
/* Zero the rest of the page */
MemSet(byteptr + 1, 0, BLCKSZ - byteno - 1);
- XactCtl->shared->page_dirty[slotno] = true;
+ XactCtl->shared->page_entries[slotno].page_dirty = true;
}
LWLockRelease(XactSLRULock);
@@ -994,7 +994,7 @@ clog_redo(XLogReaderState *record)
slotno = ZeroCLOGPage(pageno, false);
SimpleLruWritePage(XactCtl, slotno);
- Assert(!XactCtl->shared->page_dirty[slotno]);
+ Assert(!XactCtl->shared->page_entries[slotno].page_dirty);
LWLockRelease(XactSLRULock);
}
diff --git a/src/backend/access/transam/commit_ts.c b/src/backend/access/transam/commit_ts.c
index cb2e0ceb1c3..6879ef3ec71 100644
--- a/src/backend/access/transam/commit_ts.c
+++ b/src/backend/access/transam/commit_ts.c
@@ -227,7 +227,7 @@ SetXidCommitTsInPage(TransactionId xid, int nsubxids,
for (i = 0; i < nsubxids; i++)
TransactionIdSetCommitTs(subxids[i], ts, nodeid, slotno);
- CommitTsCtl->shared->page_dirty[slotno] = true;
+ CommitTsCtl->shared->page_entries[slotno].page_dirty = true;
LWLockRelease(CommitTsSLRULock);
}
@@ -735,7 +735,7 @@ ActivateCommitTs(void)
LWLockAcquire(CommitTsSLRULock, LW_EXCLUSIVE);
slotno = ZeroCommitTsPage(pageno, false);
SimpleLruWritePage(CommitTsCtl, slotno);
- Assert(!CommitTsCtl->shared->page_dirty[slotno]);
+ Assert(!CommitTsCtl->shared->page_entries[slotno].page_dirty);
LWLockRelease(CommitTsSLRULock);
}
@@ -1005,7 +1005,7 @@ commit_ts_redo(XLogReaderState *record)
slotno = ZeroCommitTsPage(pageno, false);
SimpleLruWritePage(CommitTsCtl, slotno);
- Assert(!CommitTsCtl->shared->page_dirty[slotno]);
+ Assert(!CommitTsCtl->shared->page_entries[slotno].page_dirty);
LWLockRelease(CommitTsSLRULock);
}
diff --git a/src/backend/access/transam/multixact.c b/src/backend/access/transam/multixact.c
index a29ab4769dc..eb32821b717 100644
--- a/src/backend/access/transam/multixact.c
+++ b/src/backend/access/transam/multixact.c
@@ -887,7 +887,7 @@ RecordNewMultiXact(MultiXactId multi, MultiXactOffset offset,
*offptr = offset;
- MultiXactOffsetCtl->shared->page_dirty[slotno] = true;
+ MultiXactOffsetCtl->shared->page_entries[slotno].page_dirty = true;
/* Exchange our lock */
LWLockRelease(MultiXactOffsetSLRULock);
@@ -931,7 +931,7 @@ RecordNewMultiXact(MultiXactId multi, MultiXactOffset offset,
flagsval |= (members[i].status << bshift);
*flagsptr = flagsval;
- MultiXactMemberCtl->shared->page_dirty[slotno] = true;
+ MultiXactMemberCtl->shared->page_entries[slotno].page_dirty = true;
}
LWLockRelease(MultiXactMemberSLRULock);
@@ -1902,7 +1902,7 @@ BootStrapMultiXact(void)
/* Make sure it's written out */
SimpleLruWritePage(MultiXactOffsetCtl, slotno);
- Assert(!MultiXactOffsetCtl->shared->page_dirty[slotno]);
+ Assert(!MultiXactOffsetCtl->shared->page_entries[slotno].page_dirty);
LWLockRelease(MultiXactOffsetSLRULock);
@@ -1913,7 +1913,7 @@ BootStrapMultiXact(void)
/* Make sure it's written out */
SimpleLruWritePage(MultiXactMemberCtl, slotno);
- Assert(!MultiXactMemberCtl->shared->page_dirty[slotno]);
+ Assert(!MultiXactMemberCtl->shared->page_entries[slotno].page_dirty);
LWLockRelease(MultiXactMemberSLRULock);
}
@@ -2074,7 +2074,7 @@ TrimMultiXact(void)
MemSet(offptr, 0, BLCKSZ - (entryno * sizeof(MultiXactOffset)));
- MultiXactOffsetCtl->shared->page_dirty[slotno] = true;
+ MultiXactOffsetCtl->shared->page_entries[slotno].page_dirty = true;
}
LWLockRelease(MultiXactOffsetSLRULock);
@@ -2112,7 +2112,7 @@ TrimMultiXact(void)
* writing.
*/
- MultiXactMemberCtl->shared->page_dirty[slotno] = true;
+ MultiXactMemberCtl->shared->page_entries[slotno].page_dirty = true;
}
LWLockRelease(MultiXactMemberSLRULock);
@@ -3251,7 +3251,7 @@ multixact_redo(XLogReaderState *record)
slotno = ZeroMultiXactOffsetPage(pageno, false);
SimpleLruWritePage(MultiXactOffsetCtl, slotno);
- Assert(!MultiXactOffsetCtl->shared->page_dirty[slotno]);
+ Assert(!MultiXactOffsetCtl->shared->page_entries[slotno].page_dirty);
LWLockRelease(MultiXactOffsetSLRULock);
}
@@ -3266,7 +3266,7 @@ multixact_redo(XLogReaderState *record)
slotno = ZeroMultiXactMemberPage(pageno, false);
SimpleLruWritePage(MultiXactMemberCtl, slotno);
- Assert(!MultiXactMemberCtl->shared->page_dirty[slotno]);
+ Assert(!MultiXactMemberCtl->shared->page_entries[slotno].page_dirty);
LWLockRelease(MultiXactMemberSLRULock);
}
diff --git a/src/backend/access/transam/slru.c b/src/backend/access/transam/slru.c
index 33857bffb79..1f87e50f05c 100644
--- a/src/backend/access/transam/slru.c
+++ b/src/backend/access/transam/slru.c
@@ -186,9 +186,7 @@ SimpleLruShmemSize(int nslots, int nlsns)
/* we assume nslots isn't so large as to risk overflow */
sz = MAXALIGN(sizeof(SlruSharedData));
sz += MAXALIGN(nslots * sizeof(char *)); /* page_buffer[] */
- sz += MAXALIGN(nslots * sizeof(SlruPageStatus)); /* page_status[] */
- sz += MAXALIGN(nslots * sizeof(bool)); /* page_dirty[] */
- sz += MAXALIGN(nslots * sizeof(int)); /* page_number[] */
+ sz += MAXALIGN(nslots * sizeof(SlruPageEntry)); /* page_entries[] */
sz += MAXALIGN(nslots * sizeof(int)); /* page_lru_count[] */
sz += MAXALIGN(nslots * sizeof(LWLockPadded)); /* buffer_locks[] */
@@ -248,16 +246,13 @@ SimpleLruInit(SlruCtl ctl, const char *name, int nslots, int nlsns,
shared->slru_stats_idx = pgstat_slru_index(name);
+ Assert(sizeof(SlruPageEntry) == 8);
ptr = (char *) shared;
offset = MAXALIGN(sizeof(SlruSharedData));
shared->page_buffer = (char **) (ptr + offset);
offset += MAXALIGN(nslots * sizeof(char *));
- shared->page_status = (SlruPageStatus *) (ptr + offset);
- offset += MAXALIGN(nslots * sizeof(SlruPageStatus));
- shared->page_dirty = (bool *) (ptr + offset);
- offset += MAXALIGN(nslots * sizeof(bool));
- shared->page_number = (int *) (ptr + offset);
- offset += MAXALIGN(nslots * sizeof(int));
+ shared->page_entries = (SlruPageEntry *) (ptr + offset);
+ offset += MAXALIGN(nslots * sizeof(SlruPageEntry));
shared->page_lru_count = (int *) (ptr + offset);
offset += MAXALIGN(nslots * sizeof(int));
@@ -278,8 +273,8 @@ SimpleLruInit(SlruCtl ctl, const char *name, int nslots, int nlsns,
tranche_id);
shared->page_buffer[slotno] = ptr;
- shared->page_status[slotno] = SLRU_PAGE_EMPTY;
- shared->page_dirty[slotno] = false;
+ shared->page_entries[slotno].page_status = SLRU_PAGE_EMPTY;
+ shared->page_entries[slotno].page_dirty = false;
shared->page_lru_count[slotno] = 0;
ptr += BLCKSZ;
}
@@ -315,15 +310,15 @@ SimpleLruZeroPage(SlruCtl ctl, int pageno)
/* Find a suitable buffer slot for the page */
slotno = SlruSelectLRUPage(ctl, pageno);
- Assert(shared->page_status[slotno] == SLRU_PAGE_EMPTY ||
- (shared->page_status[slotno] == SLRU_PAGE_VALID &&
- !shared->page_dirty[slotno]) ||
- shared->page_number[slotno] == pageno);
+ Assert(shared->page_entries[slotno].page_status == SLRU_PAGE_EMPTY ||
+ (shared->page_entries[slotno].page_status == SLRU_PAGE_VALID &&
+ !shared->page_entries[slotno].page_dirty) ||
+ shared->page_entries[slotno].page_number == pageno);
/* Mark the slot as containing this page */
- shared->page_number[slotno] = pageno;
- shared->page_status[slotno] = SLRU_PAGE_VALID;
- shared->page_dirty[slotno] = true;
+ shared->page_entries[slotno].page_number = pageno;
+ shared->page_entries[slotno].page_status = SLRU_PAGE_VALID;
+ shared->page_entries[slotno].page_dirty = true;
SlruRecentlyUsed(shared, slotno);
/* Set the buffer to zeroes */
@@ -387,18 +382,18 @@ SimpleLruWaitIO(SlruCtl ctl, int slotno)
* cheaply test for failure by seeing if the buffer lock is still held (we
* assume that transaction abort would release the lock).
*/
- if (shared->page_status[slotno] == SLRU_PAGE_READ_IN_PROGRESS ||
- shared->page_status[slotno] == SLRU_PAGE_WRITE_IN_PROGRESS)
+ if (shared->page_entries[slotno].page_status == SLRU_PAGE_READ_IN_PROGRESS ||
+ shared->page_entries[slotno].page_status == SLRU_PAGE_WRITE_IN_PROGRESS)
{
if (LWLockConditionalAcquire(&shared->buffer_locks[slotno].lock, LW_SHARED))
{
/* indeed, the I/O must have failed */
- if (shared->page_status[slotno] == SLRU_PAGE_READ_IN_PROGRESS)
- shared->page_status[slotno] = SLRU_PAGE_EMPTY;
+ if (shared->page_entries[slotno].page_status == SLRU_PAGE_READ_IN_PROGRESS)
+ shared->page_entries[slotno].page_status = SLRU_PAGE_EMPTY;
else /* write_in_progress */
{
- shared->page_status[slotno] = SLRU_PAGE_VALID;
- shared->page_dirty[slotno] = true;
+ shared->page_entries[slotno].page_status = SLRU_PAGE_VALID;
+ shared->page_entries[slotno].page_dirty = true;
}
LWLockRelease(&shared->buffer_locks[slotno].lock);
}
@@ -438,15 +433,15 @@ SimpleLruReadPage(SlruCtl ctl, int pageno, bool write_ok,
slotno = SlruSelectLRUPage(ctl, pageno);
/* Did we find the page in memory? */
- if (shared->page_number[slotno] == pageno &&
- shared->page_status[slotno] != SLRU_PAGE_EMPTY)
+ if (shared->page_entries[slotno].page_number == pageno &&
+ shared->page_entries[slotno].page_status != SLRU_PAGE_EMPTY)
{
/*
* If page is still being read in, we must wait for I/O. Likewise
* if the page is being written and the caller said that's not OK.
*/
- if (shared->page_status[slotno] == SLRU_PAGE_READ_IN_PROGRESS ||
- (shared->page_status[slotno] == SLRU_PAGE_WRITE_IN_PROGRESS &&
+ if (shared->page_entries[slotno].page_status == SLRU_PAGE_READ_IN_PROGRESS ||
+ (shared->page_entries[slotno].page_status == SLRU_PAGE_WRITE_IN_PROGRESS &&
!write_ok))
{
SimpleLruWaitIO(ctl, slotno);
@@ -463,14 +458,14 @@ SimpleLruReadPage(SlruCtl ctl, int pageno, bool write_ok,
}
/* We found no match; assert we selected a freeable slot */
- Assert(shared->page_status[slotno] == SLRU_PAGE_EMPTY ||
- (shared->page_status[slotno] == SLRU_PAGE_VALID &&
- !shared->page_dirty[slotno]));
+ Assert(shared->page_entries[slotno].page_status == SLRU_PAGE_EMPTY ||
+ (shared->page_entries[slotno].page_status == SLRU_PAGE_VALID &&
+ !shared->page_entries[slotno].page_dirty));
/* Mark the slot read-busy */
- shared->page_number[slotno] = pageno;
- shared->page_status[slotno] = SLRU_PAGE_READ_IN_PROGRESS;
- shared->page_dirty[slotno] = false;
+ shared->page_entries[slotno].page_number = pageno;
+ shared->page_entries[slotno].page_status = SLRU_PAGE_READ_IN_PROGRESS;
+ shared->page_entries[slotno].page_dirty = false;
/* Acquire per-buffer lock (cannot deadlock, see notes at top) */
LWLockAcquire(&shared->buffer_locks[slotno].lock, LW_EXCLUSIVE);
@@ -487,11 +482,11 @@ SimpleLruReadPage(SlruCtl ctl, int pageno, bool write_ok,
/* Re-acquire control lock and update page state */
LWLockAcquire(shared->ControlLock, LW_EXCLUSIVE);
- Assert(shared->page_number[slotno] == pageno &&
- shared->page_status[slotno] == SLRU_PAGE_READ_IN_PROGRESS &&
- !shared->page_dirty[slotno]);
+ Assert(shared->page_entries[slotno].page_number == pageno &&
+ shared->page_entries[slotno].page_status == SLRU_PAGE_READ_IN_PROGRESS &&
+ !shared->page_entries[slotno].page_dirty);
- shared->page_status[slotno] = ok ? SLRU_PAGE_VALID : SLRU_PAGE_EMPTY;
+ shared->page_entries[slotno].page_status = ok ? SLRU_PAGE_VALID : SLRU_PAGE_EMPTY;
LWLockRelease(&shared->buffer_locks[slotno].lock);
@@ -536,9 +531,9 @@ SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno, TransactionId xid)
int bankend = bankstart + shared->bank_size;
for (slotno = bankstart; slotno < bankend; slotno++)
{
- if (shared->page_number[slotno] == pageno &&
- shared->page_status[slotno] != SLRU_PAGE_EMPTY &&
- shared->page_status[slotno] != SLRU_PAGE_READ_IN_PROGRESS)
+ if (shared->page_entries[slotno].page_number == pageno &&
+ shared->page_entries[slotno].page_status != SLRU_PAGE_EMPTY &&
+ shared->page_entries[slotno].page_status != SLRU_PAGE_READ_IN_PROGRESS)
{
/* See comments for SlruRecentlyUsed macro */
SlruRecentlyUsed(shared, slotno);
@@ -572,12 +567,12 @@ static void
SlruInternalWritePage(SlruCtl ctl, int slotno, SlruWriteAll fdata)
{
SlruShared shared = ctl->shared;
- int pageno = shared->page_number[slotno];
+ int pageno = shared->page_entries[slotno].page_number;
bool ok;
/* If a write is in progress, wait for it to finish */
- while (shared->page_status[slotno] == SLRU_PAGE_WRITE_IN_PROGRESS &&
- shared->page_number[slotno] == pageno)
+ while (shared->page_entries[slotno].page_status == SLRU_PAGE_WRITE_IN_PROGRESS &&
+ shared->page_entries[slotno].page_number == pageno)
{
SimpleLruWaitIO(ctl, slotno);
}
@@ -586,17 +581,17 @@ SlruInternalWritePage(SlruCtl ctl, int slotno, SlruWriteAll fdata)
* Do nothing if page is not dirty, or if buffer no longer contains the
* same page we were called for.
*/
- if (!shared->page_dirty[slotno] ||
- shared->page_status[slotno] != SLRU_PAGE_VALID ||
- shared->page_number[slotno] != pageno)
+ if (!shared->page_entries[slotno].page_dirty ||
+ shared->page_entries[slotno].page_status != SLRU_PAGE_VALID ||
+ shared->page_entries[slotno].page_number != pageno)
return;
/*
* Mark the slot write-busy, and clear the dirtybit. After this point, a
* transaction status update on this page will mark it dirty again.
*/
- shared->page_status[slotno] = SLRU_PAGE_WRITE_IN_PROGRESS;
- shared->page_dirty[slotno] = false;
+ shared->page_entries[slotno].page_status = SLRU_PAGE_WRITE_IN_PROGRESS;
+ shared->page_entries[slotno].page_dirty = false;
/* Acquire per-buffer lock (cannot deadlock, see notes at top) */
LWLockAcquire(&shared->buffer_locks[slotno].lock, LW_EXCLUSIVE);
@@ -619,14 +614,14 @@ SlruInternalWritePage(SlruCtl ctl, int slotno, SlruWriteAll fdata)
/* Re-acquire control lock and update page state */
LWLockAcquire(shared->ControlLock, LW_EXCLUSIVE);
- Assert(shared->page_number[slotno] == pageno &&
- shared->page_status[slotno] == SLRU_PAGE_WRITE_IN_PROGRESS);
+ Assert(shared->page_entries[slotno].page_number == pageno &&
+ shared->page_entries[slotno].page_status == SLRU_PAGE_WRITE_IN_PROGRESS);
/* If we failed to write, mark the page dirty again */
if (!ok)
- shared->page_dirty[slotno] = true;
+ shared->page_entries[slotno].page_dirty = true;
- shared->page_status[slotno] = SLRU_PAGE_VALID;
+ shared->page_entries[slotno].page_status = SLRU_PAGE_VALID;
LWLockRelease(&shared->buffer_locks[slotno].lock);
@@ -1067,8 +1062,8 @@ SlruSelectLRUPage(SlruCtl ctl, int pageno)
int bankend = bankstart + shared->bank_size;
for (slotno = bankstart; slotno < bankend; slotno++)
{
- if (shared->page_number[slotno] == pageno &&
- shared->page_status[slotno] != SLRU_PAGE_EMPTY)
+ if (shared->page_entries[slotno].page_number == pageno &&
+ shared->page_entries[slotno].page_status != SLRU_PAGE_EMPTY)
return slotno;
}
@@ -1105,7 +1100,7 @@ SlruSelectLRUPage(SlruCtl ctl, int pageno)
int this_delta;
int this_page_number;
- if (shared->page_status[slotno] == SLRU_PAGE_EMPTY)
+ if (shared->page_entries[slotno].page_status == SLRU_PAGE_EMPTY)
return slotno;
this_delta = cur_count - shared->page_lru_count[slotno];
if (this_delta < 0)
@@ -1120,10 +1115,10 @@ SlruSelectLRUPage(SlruCtl ctl, int pageno)
shared->page_lru_count[slotno] = cur_count;
this_delta = 0;
}
- this_page_number = shared->page_number[slotno];
+ this_page_number = shared->page_entries[slotno].page_number;
if (this_page_number == shared->latest_page_number)
continue;
- if (shared->page_status[slotno] == SLRU_PAGE_VALID)
+ if (shared->page_entries[slotno].page_status == SLRU_PAGE_VALID)
{
if (this_delta > best_valid_delta ||
(this_delta == best_valid_delta &&
@@ -1165,7 +1160,7 @@ SlruSelectLRUPage(SlruCtl ctl, int pageno)
/*
* If the selected page is clean, we're set.
*/
- if (!shared->page_dirty[bestvalidslot])
+ if (!shared->page_entries[bestvalidslot].page_dirty)
return bestvalidslot;
/*
@@ -1217,9 +1212,9 @@ SimpleLruWriteAll(SlruCtl ctl, bool allow_redirtied)
* already. That's okay.
*/
Assert(allow_redirtied ||
- shared->page_status[slotno] == SLRU_PAGE_EMPTY ||
- (shared->page_status[slotno] == SLRU_PAGE_VALID &&
- !shared->page_dirty[slotno]));
+ shared->page_entries[slotno].page_status == SLRU_PAGE_EMPTY ||
+ (shared->page_entries[slotno].page_status == SLRU_PAGE_VALID &&
+ !shared->page_entries[slotno].page_dirty));
}
LWLockRelease(shared->ControlLock);
@@ -1291,18 +1286,18 @@ restart:;
for (slotno = 0; slotno < shared->num_slots; slotno++)
{
- if (shared->page_status[slotno] == SLRU_PAGE_EMPTY)
+ if (shared->page_entries[slotno].page_status == SLRU_PAGE_EMPTY)
continue;
- if (!ctl->PagePrecedes(shared->page_number[slotno], cutoffPage))
+ if (!ctl->PagePrecedes(shared->page_entries[slotno].page_number, cutoffPage))
continue;
/*
* If page is clean, just change state to EMPTY (expected case).
*/
- if (shared->page_status[slotno] == SLRU_PAGE_VALID &&
- !shared->page_dirty[slotno])
+ if (shared->page_entries[slotno].page_status == SLRU_PAGE_VALID &&
+ !shared->page_entries[slotno].page_dirty)
{
- shared->page_status[slotno] = SLRU_PAGE_EMPTY;
+ shared->page_entries[slotno].page_status = SLRU_PAGE_EMPTY;
continue;
}
@@ -1316,7 +1311,7 @@ restart:;
* won't have cause to read its data again. For now, keep the logic
* the same as it was.)
*/
- if (shared->page_status[slotno] == SLRU_PAGE_VALID)
+ if (shared->page_entries[slotno].page_status == SLRU_PAGE_VALID)
SlruInternalWritePage(ctl, slotno, NULL);
else
SimpleLruWaitIO(ctl, slotno);
@@ -1371,9 +1366,9 @@ restart:
did_write = false;
for (slotno = 0; slotno < shared->num_slots; slotno++)
{
- int pagesegno = shared->page_number[slotno] / SLRU_PAGES_PER_SEGMENT;
+ int pagesegno = shared->page_entries[slotno].page_number / SLRU_PAGES_PER_SEGMENT;
- if (shared->page_status[slotno] == SLRU_PAGE_EMPTY)
+ if (shared->page_entries[slotno].page_status == SLRU_PAGE_EMPTY)
continue;
/* not the segment we're looking for */
@@ -1381,15 +1376,15 @@ restart:
continue;
/* If page is clean, just change state to EMPTY (expected case). */
- if (shared->page_status[slotno] == SLRU_PAGE_VALID &&
- !shared->page_dirty[slotno])
+ if (shared->page_entries[slotno].page_status == SLRU_PAGE_VALID &&
+ !shared->page_entries[slotno].page_dirty)
{
- shared->page_status[slotno] = SLRU_PAGE_EMPTY;
+ shared->page_entries[slotno].page_status = SLRU_PAGE_EMPTY;
continue;
}
/* Same logic as SimpleLruTruncate() */
- if (shared->page_status[slotno] == SLRU_PAGE_VALID)
+ if (shared->page_entries[slotno].page_status == SLRU_PAGE_VALID)
SlruInternalWritePage(ctl, slotno, NULL);
else
SimpleLruWaitIO(ctl, slotno);
diff --git a/src/backend/access/transam/subtrans.c b/src/backend/access/transam/subtrans.c
index 785f2520fde..11754fdbac6 100644
--- a/src/backend/access/transam/subtrans.c
+++ b/src/backend/access/transam/subtrans.c
@@ -97,7 +97,7 @@ SubTransSetParent(TransactionId xid, TransactionId parent)
{
Assert(*ptr == InvalidTransactionId);
*ptr = parent;
- SubTransCtl->shared->page_dirty[slotno] = true;
+ SubTransCtl->shared->page_entries[slotno].page_dirty = true;
}
LWLockRelease(SubtransSLRULock);
@@ -220,7 +220,7 @@ BootStrapSUBTRANS(void)
/* Make sure it's written out */
SimpleLruWritePage(SubTransCtl, slotno);
- Assert(!SubTransCtl->shared->page_dirty[slotno]);
+ Assert(!SubTransCtl->shared->page_entries[slotno].page_dirty);
LWLockRelease(SubtransSLRULock);
}
diff --git a/src/backend/commands/async.c b/src/backend/commands/async.c
index 7f2b7598449..86a5bc2430c 100644
--- a/src/backend/commands/async.c
+++ b/src/backend/commands/async.c
@@ -1445,7 +1445,7 @@ asyncQueueAddEntries(ListCell *nextNotify)
InvalidTransactionId);
/* Note we mark the page dirty before writing in it */
- NotifyCtl->shared->page_dirty[slotno] = true;
+ NotifyCtl->shared->page_entries[slotno].page_dirty = true;
while (nextNotify != NULL)
{
diff --git a/src/backend/storage/lmgr/predicate.c b/src/backend/storage/lmgr/predicate.c
index c5e66757643..53bd7d957ce 100644
--- a/src/backend/storage/lmgr/predicate.c
+++ b/src/backend/storage/lmgr/predicate.c
@@ -963,7 +963,7 @@ SerialAdd(TransactionId xid, SerCommitSeqNo minConflictCommitSeqNo)
slotno = SimpleLruReadPage(SerialSlruCtl, targetPage, true, xid);
SerialValue(slotno, xid) = minConflictCommitSeqNo;
- SerialSlruCtl->shared->page_dirty[slotno] = true;
+ SerialSlruCtl->shared->page_entries[slotno].page_dirty = true;
LWLockRelease(SerialSLRULock);
}
diff --git a/src/include/access/slru.h b/src/include/access/slru.h
index f4df54d3c12..0a4fae91d7f 100644
--- a/src/include/access/slru.h
+++ b/src/include/access/slru.h
@@ -44,7 +44,7 @@
* in the latter case it implies that the page has been re-dirtied since
* the write started.
*/
-typedef enum
+typedef enum SlruPageStatus:int16_t
{
SLRU_PAGE_EMPTY, /* buffer is not in use */
SLRU_PAGE_READ_IN_PROGRESS, /* page is being read in */
@@ -52,6 +52,13 @@ typedef enum
SLRU_PAGE_WRITE_IN_PROGRESS /* page is being written out */
} SlruPageStatus;
+typedef struct SlruPageEntry
+{
+ int page_number;
+ SlruPageStatus page_status;
+ bool page_dirty;
+} SlruPageEntry;
+
/*
* Shared-memory state
*/
@@ -69,9 +76,7 @@ typedef struct SlruSharedData
* when status is EMPTY, as is page_lru_count.
*/
char **page_buffer;
- SlruPageStatus *page_status;
- bool *page_dirty;
- int *page_number;
+ SlruPageEntry *page_entries;
int *page_lru_count;
LWLockPadded *buffer_locks;
--
2.33.1
[text/x-diff] v-18-0002-Divide-SLRU-buffers-into-8-associative-banks.patch (4.7K, ../../[email protected]/4-v-18-0002-Divide-SLRU-buffers-into-8-associative-banks.patch)
download | inline diff:
From 12a1c31240e338c0e8c3c7211fbdd7b2e9666564 Mon Sep 17 00:00:00 2001
From: Andrey Borodin <[email protected]>
Date: Sun, 11 Apr 2021 21:18:10 +0300
Subject: [PATCH v=18 2/3] Divide SLRU buffers into 8-associative banks
We want to eliminate linear search within SLRU buffers.
To do so we divide SLRU buffers into banks. Each bank holds
approximately 8 buffers. Each SLRU pageno may reside only in one bank.
Adjacent pagenos reside in different banks.
---
src/backend/access/transam/slru.c | 43 ++++++++++++++++++++++++++++---
src/include/access/slru.h | 2 ++
2 files changed, 41 insertions(+), 4 deletions(-)
diff --git a/src/backend/access/transam/slru.c b/src/backend/access/transam/slru.c
index 7585ae24ce9..33857bffb79 100644
--- a/src/backend/access/transam/slru.c
+++ b/src/backend/access/transam/slru.c
@@ -134,7 +134,7 @@ typedef enum
static SlruErrorCause slru_errcause;
static int slru_errno;
-
+static void SlruAdjustNSlots(int* nslots, int* banksize, int* bankoffset);
static void SimpleLruZeroLSNs(SlruCtl ctl, int slotno);
static void SimpleLruWaitIO(SlruCtl ctl, int slotno);
static void SlruInternalWritePage(SlruCtl ctl, int slotno, SlruWriteAll fdata);
@@ -148,6 +148,30 @@ static bool SlruScanDirCbDeleteCutoff(SlruCtl ctl, char *filename,
int segpage, void *data);
static void SlruInternalDeleteSegment(SlruCtl ctl, int segno);
+/*
+ * Pick bank size optimal for N-assiciative SLRU buffers.
+ * We expect bank number to be picked from lowest bits of requested pageno.
+ * Thus we want number of banks to be power of 2. This routine computes number
+ * of banks aiming to make each bank of size 8. So we can pack page number and
+ * statuses of each bank on one cacheline.
+ */
+static void SlruAdjustNSlots(int* nslots, int* banksize, int* bankoffset)
+{
+ *banksize = *nslots;
+ int nbanks = 1;
+ *bankoffset = 0;
+ while (*banksize > 15)
+ {
+ if ((*banksize & 1) != 0)
+ *banksize +=1;
+ *banksize /= 2;
+ nbanks *= 2;
+ *bankoffset += 1;
+ }
+ elog(DEBUG5, "nslots %d banksize %d nbanks %d ", *nslots, *banksize, nbanks);
+ *nslots = *banksize * nbanks;
+}
+
/*
* Initialization of shared memory
*/
@@ -156,6 +180,8 @@ Size
SimpleLruShmemSize(int nslots, int nlsns)
{
Size sz;
+ int bankoffset, banksize;
+ SlruAdjustNSlots(&nslots, &banksize, &bankoffset);
/* we assume nslots isn't so large as to risk overflow */
sz = MAXALIGN(sizeof(SlruSharedData));
@@ -190,6 +216,8 @@ SimpleLruInit(SlruCtl ctl, const char *name, int nslots, int nlsns,
{
SlruShared shared;
bool found;
+ int bankoffset, banksize;
+ SlruAdjustNSlots(&nslots, &banksize, &bankoffset);
shared = (SlruShared) ShmemInitStruct(name,
SimpleLruShmemSize(nslots, nlsns),
@@ -209,6 +237,9 @@ SimpleLruInit(SlruCtl ctl, const char *name, int nslots, int nlsns,
shared->ControlLock = ctllock;
shared->num_slots = nslots;
+ shared->bank_mask = (1 << bankoffset) - 1;
+ shared->bank_size = banksize;
+
shared->lsn_groups_per_page = nlsns;
shared->cur_lru_count = 0;
@@ -501,7 +532,9 @@ SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno, TransactionId xid)
LWLockAcquire(shared->ControlLock, LW_SHARED);
/* See if page is already in a buffer */
- for (slotno = 0; slotno < shared->num_slots; slotno++)
+ int bankstart = (pageno & shared->bank_mask) * shared->bank_size;
+ int bankend = bankstart + shared->bank_size;
+ for (slotno = bankstart; slotno < bankend; slotno++)
{
if (shared->page_number[slotno] == pageno &&
shared->page_status[slotno] != SLRU_PAGE_EMPTY &&
@@ -1030,7 +1063,9 @@ SlruSelectLRUPage(SlruCtl ctl, int pageno)
int best_invalid_page_number = 0; /* keep compiler quiet */
/* See if page already has a buffer assigned */
- for (slotno = 0; slotno < shared->num_slots; slotno++)
+ int bankstart = (pageno & shared->bank_mask) * shared->bank_size;
+ int bankend = bankstart + shared->bank_size;
+ for (slotno = bankstart; slotno < bankend; slotno++)
{
if (shared->page_number[slotno] == pageno &&
shared->page_status[slotno] != SLRU_PAGE_EMPTY)
@@ -1065,7 +1100,7 @@ SlruSelectLRUPage(SlruCtl ctl, int pageno)
* multiple pages with the same lru_count.
*/
cur_count = (shared->cur_lru_count)++;
- for (slotno = 0; slotno < shared->num_slots; slotno++)
+ for (slotno = bankstart; slotno < bankend; slotno++)
{
int this_delta;
int this_page_number;
diff --git a/src/include/access/slru.h b/src/include/access/slru.h
index 793c045f160..f4df54d3c12 100644
--- a/src/include/access/slru.h
+++ b/src/include/access/slru.h
@@ -61,6 +61,8 @@ typedef struct SlruSharedData
/* Number of buffers managed by this SLRU structure */
int num_slots;
+ int bank_size;
+ int bank_mask;
/*
* Arrays holding info for each buffer slot. Page number is undefined
--
2.33.1
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2022-07-21 13:00 Yura Sokolov <[email protected]>
parent: Thomas Munro <[email protected]>
1 sibling, 1 reply; 103+ messages in thread
From: Yura Sokolov @ 2022-07-21 13:00 UTC (permalink / raw)
To: Andrey Borodin <[email protected]>; Andres Freund <[email protected]>; +Cc: Thomas Munro <[email protected]>; Gilles Darold <[email protected]>; Tomas Vondra <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
Good day, all.
I did benchmark of patch on 2 socket Xeon 5220 CPU @ 2.20GHz .
I used "benchmark" used to reproduce problems with SLRU on our
customers setup.
In opposite to Shawn's tests I concentrated on bad case: a lot
of contention.
slru-funcs.sql - function definitions
- functions creates a lot of subtrunsactions to stress subtrans
- and select random rows for share to stress multixacts
slru-call.sql - function call for benchmark
slru-ballast.sql - randomly select 1000 consequent rows
"for update skip locked" to stress multixacts
patch1 - make SLRU buffers configurable
patch2 - make "8-associative banks"
Benchmark done by pgbench.
Inited with scale 1 to induce contention.
pgbench -i -s 1 testdb
Benchmark 1:
- low number of connections (50), 60% slru-call, 40% slru-ballast
pgbench -f slru-call.sql@60 -f slru-ballast.sql@40 -c 50 -j 75 -P 1 -T 30 testdb
version | subtrans | multixact | tps
| buffers | offs/memb | func+ballast
--------+----------+-----------+------
master | 32 | 8/16 | 184+119
patch1 | 32 | 8/16 | 184+119
patch1 | 1024 | 8/16 | 121+77
patch1 | 1024 | 512/1024 | 118+75
patch2 | 32 | 8/16 | 190+122
patch2 | 1024 | 8/16 | 190+125
patch2 | 1024 | 512/1024 | 190+127
As you see, this test case degrades with dumb increase of
SLRU buffers. But use of "hash table" in form of "associative
buckets" makes performance stable.
Benchmark 2:
- high connection number (600), 98% slru-call, 2% slru-ballast
pgbench -f slru-call.sql@98 -f slru-ballast.sql@2 -c 600 -j 75 -P 1 -T 30 testdb
I don't paste "ballast" tps here since 2% make them too small,
and they're very noisy.
version | subtrans | multixact | tps
| buffers | offs/memb | func
--------+----------+-----------+------
master | 32 | 8/16 | 13
patch1 | 32
| 8/16 | 13
patch1 | 1024 | 8/16 | 31
patch1 | 1024 | 512/1024 | 53
patch2 | 32 | 8/16 | 12
patch2 | 1024 | 8/16 | 34
patch2 | 1024 | 512/1024 | 67
In this case simple buffer increase does help. But "buckets"
increase performance gain.
I didn't paste here results third part of patch ("Pack SLRU...")
because I didn't see any major performance gain from it, and
it consumes large part of patch diff.
Rebased versions of first two patch parts are attached.
regards,
Yura Sokolov
Attachments:
[application/sql] slru-ballast.sql (123B, ../../[email protected]/2-slru-ballast.sql)
download
[application/sql] slru-call.sql (25B, ../../[email protected]/3-slru-call.sql)
download
[application/sql] slru-func.sql (1.2K, ../../[email protected]/4-slru-func.sql)
download
[text/x-patch] v21-0002-Divide-SLRU-buffers-into-8-associative-banks.patch (4.9K, ../../[email protected]/5-v21-0002-Divide-SLRU-buffers-into-8-associative-banks.patch)
download | inline diff:
From 41ec9d1c54184c515d53ecc8021c4a998813f2a9 Mon Sep 17 00:00:00 2001
From: Andrey Borodin <[email protected]>
Date: Sun, 11 Apr 2021 21:18:10 +0300
Subject: [PATCH v21 2/2] Divide SLRU buffers into 8-associative banks
We want to eliminate linear search within SLRU buffers.
To do so we divide SLRU buffers into banks. Each bank holds
approximately 8 buffers. Each SLRU pageno may reside only in one bank.
Adjacent pagenos reside in different banks.
---
src/backend/access/transam/slru.c | 43 ++++++++++++++++++++++++++++---
src/include/access/slru.h | 2 ++
2 files changed, 41 insertions(+), 4 deletions(-)
diff --git a/src/backend/access/transam/slru.c b/src/backend/access/transam/slru.c
index b65cb49d7ff..abc534bbd06 100644
--- a/src/backend/access/transam/slru.c
+++ b/src/backend/access/transam/slru.c
@@ -134,7 +134,7 @@ typedef enum
static SlruErrorCause slru_errcause;
static int slru_errno;
-
+static void SlruAdjustNSlots(int* nslots, int* banksize, int* bankoffset);
static void SimpleLruZeroLSNs(SlruCtl ctl, int slotno);
static void SimpleLruWaitIO(SlruCtl ctl, int slotno);
static void SlruInternalWritePage(SlruCtl ctl, int slotno, SlruWriteAll fdata);
@@ -148,6 +148,30 @@ static bool SlruScanDirCbDeleteCutoff(SlruCtl ctl, char *filename,
int segpage, void *data);
static void SlruInternalDeleteSegment(SlruCtl ctl, int segno);
+/*
+ * Pick bank size optimal for N-assiciative SLRU buffers.
+ * We expect bank number to be picked from lowest bits of requested pageno.
+ * Thus we want number of banks to be power of 2. This routine computes number
+ * of banks aiming to make each bank of size 8. So we can pack page number and
+ * statuses of each bank on one cacheline.
+ */
+static void SlruAdjustNSlots(int* nslots, int* banksize, int* bankoffset)
+{
+ int nbanks = 1;
+ *banksize = *nslots;
+ *bankoffset = 0;
+ while (*banksize > 15)
+ {
+ if ((*banksize & 1) != 0)
+ *banksize +=1;
+ *banksize /= 2;
+ nbanks *= 2;
+ *bankoffset += 1;
+ }
+ elog(DEBUG5, "nslots %d banksize %d nbanks %d ", *nslots, *banksize, nbanks);
+ *nslots = *banksize * nbanks;
+}
+
/*
* Initialization of shared memory
*/
@@ -156,6 +180,8 @@ Size
SimpleLruShmemSize(int nslots, int nlsns)
{
Size sz;
+ int bankoffset, banksize;
+ SlruAdjustNSlots(&nslots, &banksize, &bankoffset);
/* we assume nslots isn't so large as to risk overflow */
sz = MAXALIGN(sizeof(SlruSharedData));
@@ -190,6 +216,8 @@ SimpleLruInit(SlruCtl ctl, const char *name, int nslots, int nlsns,
{
SlruShared shared;
bool found;
+ int bankoffset, banksize;
+ SlruAdjustNSlots(&nslots, &banksize, &bankoffset);
shared = (SlruShared) ShmemInitStruct(name,
SimpleLruShmemSize(nslots, nlsns),
@@ -209,6 +237,9 @@ SimpleLruInit(SlruCtl ctl, const char *name, int nslots, int nlsns,
shared->ControlLock = ctllock;
shared->num_slots = nslots;
+ shared->bank_mask = (1 << bankoffset) - 1;
+ shared->bank_size = banksize;
+
shared->lsn_groups_per_page = nlsns;
shared->cur_lru_count = 0;
@@ -496,12 +527,14 @@ SimpleLruReadPage_ReadOnly(SlruCtl ctl, int pageno, TransactionId xid)
{
SlruShared shared = ctl->shared;
int slotno;
+ int bankstart = (pageno & shared->bank_mask) * shared->bank_size;
+ int bankend = bankstart + shared->bank_size;
/* Try to find the page while holding only shared lock */
LWLockAcquire(shared->ControlLock, LW_SHARED);
/* See if page is already in a buffer */
- for (slotno = 0; slotno < shared->num_slots; slotno++)
+ for (slotno = bankstart; slotno < bankend; slotno++)
{
if (shared->page_number[slotno] == pageno &&
shared->page_status[slotno] != SLRU_PAGE_EMPTY &&
@@ -1030,7 +1063,9 @@ SlruSelectLRUPage(SlruCtl ctl, int pageno)
int best_invalid_page_number = 0; /* keep compiler quiet */
/* See if page already has a buffer assigned */
- for (slotno = 0; slotno < shared->num_slots; slotno++)
+ int bankstart = (pageno & shared->bank_mask) * shared->bank_size;
+ int bankend = bankstart + shared->bank_size;
+ for (slotno = bankstart; slotno < bankend; slotno++)
{
if (shared->page_number[slotno] == pageno &&
shared->page_status[slotno] != SLRU_PAGE_EMPTY)
@@ -1065,7 +1100,7 @@ SlruSelectLRUPage(SlruCtl ctl, int pageno)
* multiple pages with the same lru_count.
*/
cur_count = (shared->cur_lru_count)++;
- for (slotno = 0; slotno < shared->num_slots; slotno++)
+ for (slotno = bankstart; slotno < bankend; slotno++)
{
int this_delta;
int this_page_number;
diff --git a/src/include/access/slru.h b/src/include/access/slru.h
index b7e2e2b55e3..cc9cc10d271 100644
--- a/src/include/access/slru.h
+++ b/src/include/access/slru.h
@@ -61,6 +61,8 @@ typedef struct SlruSharedData
/* Number of buffers managed by this SLRU structure */
int num_slots;
+ int bank_size;
+ int bank_mask;
/*
* Arrays holding info for each buffer slot. Page number is undefined
--
2.30.2
[text/x-patch] v21-0001-Make-all-SLRU-buffer-sizes-configurable.patch (25.7K, ../../[email protected]/6-v21-0001-Make-all-SLRU-buffer-sizes-configurable.patch)
download | inline diff:
From e5f03ebac800f06f45c9ecb5f7139c96032e7fa7 Mon Sep 17 00:00:00 2001
From: Andrey Borodin <[email protected]>
Date: Mon, 15 Feb 2021 21:51:56 +0500
Subject: [PATCH v21 1/2] Make all SLRU buffer sizes configurable.
Provide new GUCs to set the number of buffers, instead of using hard
coded defaults.
Remove the limits on xact_buffers and commit_ts_buffers. The default
sizes for those caches are ~0.2% and ~0.1% of shared_buffers, as before,
but now there is no cap at 128 and 16 buffers respectively (unless
track_commit_timestamp is disabled, in the latter case, then we might as
well keep it tiny). Sizes much larger than the old limits have been
shown to be useful on modern systems, and an earlier commit replaced a
linear search with a hash table to avoid problems with extreme cases.
Author: Andrey M. Borodin <[email protected]>
Reviewed-by: Anastasia Lubennikova <[email protected]>
Reviewed-by: Tomas Vondra <[email protected]>
Reviewed-by: Alexander Korotkov <[email protected]>
Reviewed-by: Gilles Darold <[email protected]>
Reviewed-by: Thomas Munro <[email protected]>
Discussion: https://postgr.es/m/2BEC2B3F-9B61-4C1D-9FB5-5FAB0F05EF86%40yandex-team.ru
---
doc/src/sgml/config.sgml | 135 ++++++++++++++++++
src/backend/access/transam/clog.c | 23 ++-
src/backend/access/transam/commit_ts.c | 5 +
src/backend/access/transam/multixact.c | 8 +-
src/backend/access/transam/subtrans.c | 5 +-
src/backend/commands/async.c | 8 +-
src/backend/storage/lmgr/predicate.c | 4 +-
src/backend/utils/init/globals.c | 8 ++
src/backend/utils/misc/guc.c | 99 +++++++++++++
src/backend/utils/misc/postgresql.conf.sample | 9 ++
src/include/access/clog.h | 10 ++
src/include/access/commit_ts.h | 1 -
src/include/access/multixact.h | 4 -
src/include/access/slru.h | 5 +
src/include/access/subtrans.h | 2 -
src/include/commands/async.h | 5 -
src/include/miscadmin.h | 7 +
src/include/storage/predicate.h | 4 -
18 files changed, 299 insertions(+), 43 deletions(-)
diff --git a/doc/src/sgml/config.sgml b/doc/src/sgml/config.sgml
index e2d728e0c4f..991dba69757 100644
--- a/doc/src/sgml/config.sgml
+++ b/doc/src/sgml/config.sgml
@@ -1961,6 +1961,141 @@ include_dir 'conf.d'
</listitem>
</varlistentry>
+ <varlistentry id="guc-multixact-offsets-buffers" xreflabel="multixact_offsets_buffers">
+ <term><varname>multixact_offsets_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>multixact_offsets_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to use to cache the contents
+ of <literal>pg_multixact/offsets</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>8</literal>.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-multixact-members-buffers" xreflabel="multixact_members_buffers">
+ <term><varname>multixact_members_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>multixact_members_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to use to cache the contents
+ of <literal>pg_multixact/members</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>16</literal>.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-subtrans-buffers" xreflabel="subtrans_buffers">
+ <term><varname>subtrans_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>subtrans_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to use to cache the contents
+ of <literal>pg_subtrans</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>8</literal>.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-notify-buffers" xreflabel="notify_buffers">
+ <term><varname>notify_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>notify_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to use to cache the contents
+ of <literal>pg_notify</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>8</literal>.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-serial-buffers" xreflabel="serial_buffers">
+ <term><varname>serial_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>serial_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to use to cache the contents
+ of <literal>pg_serial</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>16</literal>.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-xact-buffers" xreflabel="xact_buffers">
+ <term><varname>xact_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>xact_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of shared memory to use to cache the contents
+ of <literal>pg_xact</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>0</literal>, which requests
+ <varname>shared_buffers</varname> / 512, but not fewer than 4 blocks.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
+ <varlistentry id="guc-commit-ts-buffers" xreflabel="commit_ts_buffers">
+ <term><varname>commit_ts_buffers</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>commit_ts_buffers</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies the amount of memory to use to cache the cotents of
+ <literal>pg_commit_ts</literal> (see
+ <xref linkend="pgdata-contents-table"/>).
+ If this value is specified without units, it is taken as blocks,
+ that is <symbol>BLCKSZ</symbol> bytes, typically 8kB.
+ The default value is <literal>0</literal>, which requests
+ <varname>shared_buffers</varname> / 1024, but not fewer than 4 blocks.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
<varlistentry id="guc-max-stack-depth" xreflabel="max_stack_depth">
<term><varname>max_stack_depth</varname> (<type>integer</type>)
<indexterm>
diff --git a/src/backend/access/transam/clog.c b/src/backend/access/transam/clog.c
index 3d9088a7048..e96a75d6959 100644
--- a/src/backend/access/transam/clog.c
+++ b/src/backend/access/transam/clog.c
@@ -58,8 +58,8 @@
/* We need two bits per xact, so four xacts fit in a byte */
#define CLOG_BITS_PER_XACT 2
-#define CLOG_XACTS_PER_BYTE 4
-#define CLOG_XACTS_PER_PAGE (BLCKSZ * CLOG_XACTS_PER_BYTE)
+StaticAssertDecl((CLOG_BITS_PER_XACT * CLOG_XACTS_PER_BYTE) == BITS_PER_BYTE,
+ "CLOG_BITS_PER_XACT and CLOG_XACTS_PER_BYTE are inconsistent");
#define CLOG_XACT_BITMASK ((1 << CLOG_BITS_PER_XACT) - 1)
#define TransactionIdToPage(xid) ((xid) / (TransactionId) CLOG_XACTS_PER_PAGE)
@@ -665,23 +665,16 @@ TransactionIdGetStatus(TransactionId xid, XLogRecPtr *lsn)
/*
* Number of shared CLOG buffers.
*
- * On larger multi-processor systems, it is possible to have many CLOG page
- * requests in flight at one time which could lead to disk access for CLOG
- * page if the required page is not found in memory. Testing revealed that we
- * can get the best performance by having 128 CLOG buffers, more than that it
- * doesn't improve performance.
- *
- * Unconditionally keeping the number of CLOG buffers to 128 did not seem like
- * a good idea, because it would increase the minimum amount of shared memory
- * required to start, which could be a problem for people running very small
- * configurations. The following formula seems to represent a reasonable
- * compromise: people with very low values for shared_buffers will get fewer
- * CLOG buffers as well, and everyone else will get 128.
+ * By default, we'll use 2MB of for every 1GB of shared buffers, up to the
+ * theoretical maximum useful value, but always at least 4 buffers.
*/
Size
CLOGShmemBuffers(void)
{
- return Min(128, Max(4, NBuffers / 512));
+ /* Use configured value if provided. */
+ if (xact_buffers > 0)
+ return Max(4, xact_buffers);
+ return Min(CLOG_MAX_ALLOWED_BUFFERS, Max(4, NBuffers / 512));
}
/*
diff --git a/src/backend/access/transam/commit_ts.c b/src/backend/access/transam/commit_ts.c
index 4dc8d402bd3..48ca5007475 100644
--- a/src/backend/access/transam/commit_ts.c
+++ b/src/backend/access/transam/commit_ts.c
@@ -514,10 +514,15 @@ pg_xact_commit_timestamp_origin(PG_FUNCTION_ARGS)
* We use a very similar logic as for the number of CLOG buffers (except we
* scale up twice as fast with shared buffers, and the maximum is twice as
* high); see comments in CLOGShmemBuffers.
+ * By default, we'll use 1MB of for every 1GB of shared buffers, up to the
+ * maximum value that slru.c will allow, but always at least 4 buffers.
*/
Size
CommitTsShmemBuffers(void)
{
+ /* Use configured value if provided. */
+ if (commit_ts_buffers > 0)
+ return Max(4, commit_ts_buffers);
return Min(256, Max(4, NBuffers / 256));
}
diff --git a/src/backend/access/transam/multixact.c b/src/backend/access/transam/multixact.c
index 8f7d12950e5..71d17e338e9 100644
--- a/src/backend/access/transam/multixact.c
+++ b/src/backend/access/transam/multixact.c
@@ -1834,8 +1834,8 @@ MultiXactShmemSize(void)
mul_size(sizeof(MultiXactId) * 2, MaxOldestSlot))
size = SHARED_MULTIXACT_STATE_SIZE;
- size = add_size(size, SimpleLruShmemSize(NUM_MULTIXACTOFFSET_BUFFERS, 0));
- size = add_size(size, SimpleLruShmemSize(NUM_MULTIXACTMEMBER_BUFFERS, 0));
+ size = add_size(size, SimpleLruShmemSize(multixact_offsets_buffers, 0));
+ size = add_size(size, SimpleLruShmemSize(multixact_members_buffers, 0));
return size;
}
@@ -1851,13 +1851,13 @@ MultiXactShmemInit(void)
MultiXactMemberCtl->PagePrecedes = MultiXactMemberPagePrecedes;
SimpleLruInit(MultiXactOffsetCtl,
- "MultiXactOffset", NUM_MULTIXACTOFFSET_BUFFERS, 0,
+ "MultiXactOffset", multixact_offsets_buffers, 0,
MultiXactOffsetSLRULock, "pg_multixact/offsets",
LWTRANCHE_MULTIXACTOFFSET_BUFFER,
SYNC_HANDLER_MULTIXACT_OFFSET);
SlruPagePrecedesUnitTests(MultiXactOffsetCtl, MULTIXACT_OFFSETS_PER_PAGE);
SimpleLruInit(MultiXactMemberCtl,
- "MultiXactMember", NUM_MULTIXACTMEMBER_BUFFERS, 0,
+ "MultiXactMember", multixact_offsets_buffers, 0,
MultiXactMemberSLRULock, "pg_multixact/members",
LWTRANCHE_MULTIXACTMEMBER_BUFFER,
SYNC_HANDLER_MULTIXACT_MEMBER);
diff --git a/src/backend/access/transam/subtrans.c b/src/backend/access/transam/subtrans.c
index 66d35481552..862b0eab966 100644
--- a/src/backend/access/transam/subtrans.c
+++ b/src/backend/access/transam/subtrans.c
@@ -31,6 +31,7 @@
#include "access/slru.h"
#include "access/subtrans.h"
#include "access/transam.h"
+#include "miscadmin.h"
#include "pg_trace.h"
#include "utils/snapmgr.h"
@@ -184,14 +185,14 @@ SubTransGetTopmostTransaction(TransactionId xid)
Size
SUBTRANSShmemSize(void)
{
- return SimpleLruShmemSize(NUM_SUBTRANS_BUFFERS, 0);
+ return SimpleLruShmemSize(subtrans_buffers, 0);
}
void
SUBTRANSShmemInit(void)
{
SubTransCtl->PagePrecedes = SubTransPagePrecedes;
- SimpleLruInit(SubTransCtl, "Subtrans", NUM_SUBTRANS_BUFFERS, 0,
+ SimpleLruInit(SubTransCtl, "Subtrans", subtrans_buffers, 0,
SubtransSLRULock, "pg_subtrans",
LWTRANCHE_SUBTRANS_BUFFER, SYNC_HANDLER_NONE);
SlruPagePrecedesUnitTests(SubTransCtl, SUBTRANS_XACTS_PER_PAGE);
diff --git a/src/backend/commands/async.c b/src/backend/commands/async.c
index 3e1b92df030..6b980603302 100644
--- a/src/backend/commands/async.c
+++ b/src/backend/commands/async.c
@@ -117,7 +117,7 @@
* frontend during startup.) The above design guarantees that notifies from
* other backends will never be missed by ignoring self-notifies.
*
- * The amount of shared memory used for notify management (NUM_NOTIFY_BUFFERS)
+ * The amount of shared memory used for notify management (notify_buffers)
* can be varied without affecting anything but performance. The maximum
* amount of notification data that can be queued at one time is determined
* by slru.c's wraparound limit; see QUEUE_MAX_PAGE below.
@@ -235,7 +235,7 @@ typedef struct QueuePosition
*
* Resist the temptation to make this really large. While that would save
* work in some places, it would add cost in others. In particular, this
- * should likely be less than NUM_NOTIFY_BUFFERS, to ensure that backends
+ * should likely be less than notify_buffers, to ensure that backends
* catch up before the pages they'll need to read fall out of SLRU cache.
*/
#define QUEUE_CLEANUP_DELAY 4
@@ -521,7 +521,7 @@ AsyncShmemSize(void)
size = mul_size(MaxBackends + 1, sizeof(QueueBackendStatus));
size = add_size(size, offsetof(AsyncQueueControl, backend));
- size = add_size(size, SimpleLruShmemSize(NUM_NOTIFY_BUFFERS, 0));
+ size = add_size(size, SimpleLruShmemSize(notify_buffers, 0));
return size;
}
@@ -569,7 +569,7 @@ AsyncShmemInit(void)
* Set up SLRU management of the pg_notify data.
*/
NotifyCtl->PagePrecedes = asyncQueuePagePrecedes;
- SimpleLruInit(NotifyCtl, "Notify", NUM_NOTIFY_BUFFERS, 0,
+ SimpleLruInit(NotifyCtl, "Notify", notify_buffers, 0,
NotifySLRULock, "pg_notify", LWTRANCHE_NOTIFY_BUFFER,
SYNC_HANDLER_NONE);
diff --git a/src/backend/storage/lmgr/predicate.c b/src/backend/storage/lmgr/predicate.c
index 5136da6ea36..54a0d66d579 100644
--- a/src/backend/storage/lmgr/predicate.c
+++ b/src/backend/storage/lmgr/predicate.c
@@ -872,7 +872,7 @@ SerialInit(void)
*/
SerialSlruCtl->PagePrecedes = SerialPagePrecedesLogically;
SimpleLruInit(SerialSlruCtl, "Serial",
- NUM_SERIAL_BUFFERS, 0, SerialSLRULock, "pg_serial",
+ serial_buffers, 0, SerialSLRULock, "pg_serial",
LWTRANCHE_SERIAL_BUFFER, SYNC_HANDLER_NONE);
#ifdef USE_ASSERT_CHECKING
SerialPagePrecedesLogicallyUnitTests();
@@ -1396,7 +1396,7 @@ PredicateLockShmemSize(void)
/* Shared memory structures for SLRU tracking of old committed xids. */
size = add_size(size, sizeof(SerialControlData));
- size = add_size(size, SimpleLruShmemSize(NUM_SERIAL_BUFFERS, 0));
+ size = add_size(size, SimpleLruShmemSize(serial_buffers, 0));
return size;
}
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
index 1a5d29ac9ba..cc0ca91a8b3 100644
--- a/src/backend/utils/init/globals.c
+++ b/src/backend/utils/init/globals.c
@@ -151,3 +151,11 @@ int64 VacuumPageDirty = 0;
int VacuumCostBalance = 0; /* working state for vacuum */
bool VacuumCostActive = false;
+
+int multixact_offsets_buffers = 8;
+int multixact_members_buffers = 16;
+int subtrans_buffers = 32;
+int notify_buffers = 8;
+int serial_buffers = 16;
+int xact_buffers = 0;
+int commit_ts_buffers = 0;
diff --git a/src/backend/utils/misc/guc.c b/src/backend/utils/misc/guc.c
index af4a1c30689..2d9121b9265 100644
--- a/src/backend/utils/misc/guc.c
+++ b/src/backend/utils/misc/guc.c
@@ -32,9 +32,11 @@
#endif
#include <unistd.h>
+#include "access/clog.h"
#include "access/commit_ts.h"
#include "access/gin.h"
#include "access/rmgr.h"
+#include "access/slru.h"
#include "access/tableam.h"
#include "access/toast_compression.h"
#include "access/transam.h"
@@ -209,6 +211,8 @@ static const char *show_tcp_keepalives_idle(void);
static const char *show_tcp_keepalives_interval(void);
static const char *show_tcp_keepalives_count(void);
static const char *show_tcp_user_timeout(void);
+static const char *show_xact_buffers(void);
+static const char *show_commit_ts_buffers(void);
static bool check_maxconnections(int *newval, void **extra, GucSource source);
static bool check_max_worker_processes(int *newval, void **extra, GucSource source);
static bool check_autovacuum_max_workers(int *newval, void **extra, GucSource source);
@@ -2426,6 +2430,83 @@ static struct config_int ConfigureNamesInt[] =
NULL, NULL, NULL
},
+ {
+ {"multixact_offsets_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for the MultiXact offset SLRU cache."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &multixact_offsets_buffers,
+ 8, 2, SLRU_MAX_ALLOWED_BUFFERS,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"multixact_members_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for the MultiXact member SLRU cache."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &multixact_members_buffers,
+ 16, 2, SLRU_MAX_ALLOWED_BUFFERS,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"subtrans_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for the sub-transaction SLRU cache."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &subtrans_buffers,
+ 32, 2, SLRU_MAX_ALLOWED_BUFFERS,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"notify_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for the NOTIFY message SLRU cache."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ ¬ify_buffers,
+ 8, 2, SLRU_MAX_ALLOWED_BUFFERS,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"serial_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for the serializable transaction SLRU cache."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &serial_buffers,
+ 16, 2, SLRU_MAX_ALLOWED_BUFFERS,
+ NULL, NULL, NULL
+ },
+
+ {
+ {"xact_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the number of shared memory buffers used for the transaction status SLRU cache."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &xact_buffers,
+ 0, 0, CLOG_MAX_ALLOWED_BUFFERS,
+ NULL, NULL, show_xact_buffers
+ },
+
+ {
+ {"commit_ts_buffers", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("Sets the size of the dedicated buffer pool used for the commit timestamp SLRU cache."),
+ NULL,
+ GUC_UNIT_BLOCKS
+ },
+ &commit_ts_buffers,
+ 0, 0, SLRU_MAX_ALLOWED_BUFFERS,
+ NULL, NULL, show_commit_ts_buffers
+ },
+
{
{"temp_buffers", PGC_USERSET, RESOURCES_MEM,
gettext_noop("Sets the maximum number of temporary buffers used by each session."),
@@ -12447,6 +12528,24 @@ show_tcp_user_timeout(void)
return nbuf;
}
+static const char *
+show_xact_buffers(void)
+{
+ static char nbuf[16];
+
+ snprintf(nbuf, sizeof(nbuf), "%zu", CLOGShmemBuffers());
+ return nbuf;
+}
+
+static const char *
+show_commit_ts_buffers(void)
+{
+ static char nbuf[16];
+
+ snprintf(nbuf, sizeof(nbuf), "%zu", CommitTsShmemBuffers());
+ return nbuf;
+}
+
static bool
check_maxconnections(int *newval, void **extra, GucSource source)
{
diff --git a/src/backend/utils/misc/postgresql.conf.sample b/src/backend/utils/misc/postgresql.conf.sample
index b4bc06e5f5a..2cb827e1e3b 100644
--- a/src/backend/utils/misc/postgresql.conf.sample
+++ b/src/backend/utils/misc/postgresql.conf.sample
@@ -195,6 +195,15 @@
#old_snapshot_threshold = -1 # 1min-60d; -1 disables; 0 is immediate
# (change requires restart)
+# - SLRU Buffers (change requires restart) -
+
+#xact_buffers = 0 # memory for pg_xact (0 = auto)
+#subtrans_buffers = 32 # memory for pg_subtrans
+#multixact_offsets_buffers = 8 # memory for pg_multixact/offsets
+#multixact_members_buffers = 16 # memory for pg_multixact/members
+#notify_buffers = 8 # memory for pg_notify
+#serial_buffers = 16 # memory for pg_serial
+#commit_ts_buffers = 0 # memory for pg_commit_ts (0 = auto)
#------------------------------------------------------------------------------
# WRITE-AHEAD LOG
diff --git a/src/include/access/clog.h b/src/include/access/clog.h
index 543f2e2643a..17d103aa4da 100644
--- a/src/include/access/clog.h
+++ b/src/include/access/clog.h
@@ -15,6 +15,16 @@
#include "storage/sync.h"
#include "lib/stringinfo.h"
+/*
+ * Don't allow xact_buffers to be set higher than could possibly be useful or
+ * SLRU would allow.
+ */
+#define CLOG_XACTS_PER_BYTE 4
+#define CLOG_XACTS_PER_PAGE (BLCKSZ * CLOG_XACTS_PER_BYTE)
+#define CLOG_MAX_ALLOWED_BUFFERS \
+ Min(SLRU_MAX_ALLOWED_BUFFERS, \
+ (((MaxTransactionId / 2) + (CLOG_XACTS_PER_PAGE - 1)) / CLOG_XACTS_PER_PAGE))
+
/*
* Possible transaction statuses --- note that all-zeroes is the initial
* state.
diff --git a/src/include/access/commit_ts.h b/src/include/access/commit_ts.h
index 7662f8e1a9c..d928dcc9352 100644
--- a/src/include/access/commit_ts.h
+++ b/src/include/access/commit_ts.h
@@ -16,7 +16,6 @@
#include "replication/origin.h"
#include "storage/sync.h"
-
extern PGDLLIMPORT bool track_commit_timestamp;
extern void TransactionTreeSetCommitTsData(TransactionId xid, int nsubxids,
diff --git a/src/include/access/multixact.h b/src/include/access/multixact.h
index a5600a320ae..da7b8f0abb9 100644
--- a/src/include/access/multixact.h
+++ b/src/include/access/multixact.h
@@ -29,10 +29,6 @@
#define MaxMultiXactOffset ((MultiXactOffset) 0xFFFFFFFF)
-/* Number of SLRU buffers to use for multixact */
-#define NUM_MULTIXACTOFFSET_BUFFERS 8
-#define NUM_MULTIXACTMEMBER_BUFFERS 16
-
/*
* Possible multixact lock modes ("status"). The first four modes are for
* tuple locks (FOR KEY SHARE, FOR SHARE, FOR NO KEY UPDATE, FOR UPDATE); the
diff --git a/src/include/access/slru.h b/src/include/access/slru.h
index 130c41c8632..b7e2e2b55e3 100644
--- a/src/include/access/slru.h
+++ b/src/include/access/slru.h
@@ -17,6 +17,11 @@
#include "storage/lwlock.h"
#include "storage/sync.h"
+/*
+ * To avoid overflowing internal arithmetic and the size_t data type, the
+ * number of buffers should not exceed this number.
+ */
+#define SLRU_MAX_ALLOWED_BUFFERS ((1024 * 1024 * 1024) / BLCKSZ)
/*
* Define SLRU segment size. A page is the same BLCKSZ as is used everywhere
diff --git a/src/include/access/subtrans.h b/src/include/access/subtrans.h
index f94e116640b..1ddb62883b0 100644
--- a/src/include/access/subtrans.h
+++ b/src/include/access/subtrans.h
@@ -11,8 +11,6 @@
#ifndef SUBTRANS_H
#define SUBTRANS_H
-/* Number of SLRU buffers to use for subtrans */
-#define NUM_SUBTRANS_BUFFERS 32
extern void SubTransSetParent(TransactionId xid, TransactionId parent);
extern TransactionId SubTransGetParent(TransactionId xid);
diff --git a/src/include/commands/async.h b/src/include/commands/async.h
index 926af933d1b..402d184b9b3 100644
--- a/src/include/commands/async.h
+++ b/src/include/commands/async.h
@@ -15,11 +15,6 @@
#include <signal.h>
-/*
- * The number of SLRU page buffers we use for the notification queue.
- */
-#define NUM_NOTIFY_BUFFERS 8
-
extern PGDLLIMPORT bool Trace_notify;
extern PGDLLIMPORT volatile sig_atomic_t notifyInterruptPending;
diff --git a/src/include/miscadmin.h b/src/include/miscadmin.h
index ea9a56d3955..e65990e04e8 100644
--- a/src/include/miscadmin.h
+++ b/src/include/miscadmin.h
@@ -177,6 +177,13 @@ extern PGDLLIMPORT int MaxBackends;
extern PGDLLIMPORT int MaxConnections;
extern PGDLLIMPORT int max_worker_processes;
extern PGDLLIMPORT int max_parallel_workers;
+extern PGDLLIMPORT int multixact_offsets_buffers;
+extern PGDLLIMPORT int multixact_members_buffers;
+extern PGDLLIMPORT int subtrans_buffers;
+extern PGDLLIMPORT int notify_buffers;
+extern PGDLLIMPORT int serial_buffers;
+extern PGDLLIMPORT int xact_buffers;
+extern PGDLLIMPORT int commit_ts_buffers;
extern PGDLLIMPORT int MyProcPid;
extern PGDLLIMPORT pg_time_t MyStartTime;
diff --git a/src/include/storage/predicate.h b/src/include/storage/predicate.h
index 8dfcb3944b4..1e3f757ec30 100644
--- a/src/include/storage/predicate.h
+++ b/src/include/storage/predicate.h
@@ -26,10 +26,6 @@ extern PGDLLIMPORT int max_predicate_locks_per_xact;
extern PGDLLIMPORT int max_predicate_locks_per_relation;
extern PGDLLIMPORT int max_predicate_locks_per_page;
-
-/* Number of SLRU buffers to use for Serial SLRU */
-#define NUM_SERIAL_BUFFERS 16
-
/*
* A handle used for sharing SERIALIZABLEXACT objects between the participants
* in a parallel query.
--
2.30.2
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2022-07-23 08:39 Andrey Borodin <[email protected]>
parent: Yura Sokolov <[email protected]>
0 siblings, 2 replies; 103+ messages in thread
From: Andrey Borodin @ 2022-07-23 08:39 UTC (permalink / raw)
To: Yura Sokolov <[email protected]>; +Cc: Andres Freund <[email protected]>; Thomas Munro <[email protected]>; Gilles Darold <[email protected]>; Tomas Vondra <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
> On 21 Jul 2022, at 18:00, Yura Sokolov <[email protected]> wrote:
>
> In this case simple buffer increase does help. But "buckets"
> increase performance gain.
Yura, thank you for your benchmarks!
We already knew that patch can save the day on pathological workloads, now we have a proof of this.
Also there's the evidence that user can blindly increase size of SLRU if they want (with the second patch). So there's no need for hard explanations on how to tune the buffers size.
Thomas, do you still have any doubts? Or is it certain that SLRU will be replaced by any better subsystem in 16?
Best regards, Andrey Borodin.
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2022-07-23 08:47 Thomas Munro <[email protected]>
parent: Andrey Borodin <[email protected]>
1 sibling, 1 reply; 103+ messages in thread
From: Thomas Munro @ 2022-07-23 08:47 UTC (permalink / raw)
To: Andrey Borodin <[email protected]>; +Cc: Yura Sokolov <[email protected]>; Andres Freund <[email protected]>; Gilles Darold <[email protected]>; Tomas Vondra <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
On Sat, Jul 23, 2022 at 8:41 PM Andrey Borodin <[email protected]> wrote:
> Thomas, do you still have any doubts? Or is it certain that SLRU will be replaced by any better subsystem in 16?
Hi Andrey,
Sorry for my lack of replies on this and the other SLRU thread -- I'm
thinking and experimenting. More soon.
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2022-08-16 19:36 [email protected]
parent: Andrey Borodin <[email protected]>
1 sibling, 1 reply; 103+ messages in thread
From: [email protected] @ 2022-08-16 19:36 UTC (permalink / raw)
To: Andrey Borodin <[email protected]>; +Cc: Yura Sokolov <[email protected]>; Andres Freund <[email protected]>; Thomas Munro <[email protected]>; Gilles Darold <[email protected]>; Tomas Vondra <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
> Andrey Borodin wrote 2022-07-23 11:39:
>
> Yura, thank you for your benchmarks!
> We already knew that patch can save the day on pathological workloads,
> now we have a proof of this.
> Also there's the evidence that user can blindly increase size of SLRU
> if they want (with the second patch). So there's no need for hard
> explanations on how to tune the buffers size.
Hi @Andrey.Borodin, With some considerations and performance checks from
@Yura.Sokolov we simplified your approach by the following:
1. Preamble. We feel free to increase any SLRU's, since there's no
performance degradation on large Buffers count using your SLRU buckets
solution.
2. `slru_buffers_size_scale` is only one config param introduced for all
SLRUs. It scales SLRUs upper cap by power 2.
3. All SLRU buffers count are capped by both `MBuffers (shared_buffers)`
and `slru_buffers_size_scale`. see
4. Magic initial constants `NUM_*_BUFFERS << slru_buffers_size_scale`
are applied for every SLRU.
5. All SLRU buffers are always sized as power of 2, their hash bucket
size is always 8.
There's attached patch for your consideration. It does gather and
simplify both `v21-0001-Make-all-SLRU-buffer-sizes-configurable.patch`
and `v21-0002-Divide-SLRU-buffers-into-8-associative-banks.patch` to
much simpler approach.
Thank you, Yours,
- Ivan
Attachments:
[text/x-diff] v22-0006-bucketed-SLRUs-simplified_patch.patch (18.0K, ../../[email protected]/2-v22-0006-bucketed-SLRUs-simplified_patch.patch)
download | inline diff:
Index: src/include/miscadmin.h
IDEA additional info:
Subsystem: com.intellij.openapi.diff.impl.patch.CharsetEP
<+>UTF-8
===================================================================
diff --git a/src/include/miscadmin.h b/src/include/miscadmin.h
--- a/src/include/miscadmin.h (revision 020258fbd30d37ddd03d0ec68264d1544f8d2838)
+++ b/src/include/miscadmin.h (date 1660678108730)
@@ -176,6 +176,7 @@
extern PGDLLIMPORT int MaxConnections;
extern PGDLLIMPORT int max_worker_processes;
extern PGDLLIMPORT int max_parallel_workers;
+extern PGDLLIMPORT int slru_buffers_size_scale;
extern PGDLLIMPORT int MyProcPid;
extern PGDLLIMPORT pg_time_t MyStartTime;
Index: src/include/access/subtrans.h
IDEA additional info:
Subsystem: com.intellij.openapi.diff.impl.patch.CharsetEP
<+>UTF-8
===================================================================
diff --git a/src/include/access/subtrans.h b/src/include/access/subtrans.h
--- a/src/include/access/subtrans.h (revision 020258fbd30d37ddd03d0ec68264d1544f8d2838)
+++ b/src/include/access/subtrans.h (date 1660678108698)
@@ -12,7 +12,7 @@
#define SUBTRANS_H
/* Number of SLRU buffers to use for subtrans */
-#define NUM_SUBTRANS_BUFFERS 32
+#define NUM_SUBTRANS_BUFFERS (32 << slru_buffers_size_scale)
extern void SubTransSetParent(TransactionId xid, TransactionId parent);
extern TransactionId SubTransGetParent(TransactionId xid);
Index: src/backend/utils/init/globals.c
IDEA additional info:
Subsystem: com.intellij.openapi.diff.impl.patch.CharsetEP
<+>UTF-8
===================================================================
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
--- a/src/backend/utils/init/globals.c (revision 020258fbd30d37ddd03d0ec68264d1544f8d2838)
+++ b/src/backend/utils/init/globals.c (date 1660678064048)
@@ -150,3 +150,5 @@
int VacuumCostBalance = 0; /* working state for vacuum */
bool VacuumCostActive = false;
+
+int slru_buffers_size_scale = 2; /* power 2 scale for SLRU buffers */
Index: src/backend/access/transam/slru.c
IDEA additional info:
Subsystem: com.intellij.openapi.diff.impl.patch.CharsetEP
<+>UTF-8
===================================================================
diff --git a/src/backend/access/transam/slru.c b/src/backend/access/transam/slru.c
--- a/src/backend/access/transam/slru.c (revision 020258fbd30d37ddd03d0ec68264d1544f8d2838)
+++ b/src/backend/access/transam/slru.c (date 1660678063980)
@@ -59,6 +59,7 @@
#include "pgstat.h"
#include "storage/fd.h"
#include "storage/shmem.h"
+#include "port/pg_bitutils.h"
#define SlruFileName(ctl, path, seg) \
snprintf(path, MAXPGPATH, "%s/%04X", (ctl)->Dir, seg)
@@ -71,6 +72,17 @@
*/
#define MAX_WRITEALL_BUFFERS 16
+/*
+ * To avoid overflowing internal arithmetic and the size_t data type, the
+ * number of buffers should not exceed this number.
+ */
+#define SLRU_MAX_ALLOWED_BUFFERS ((1024 * 1024 * 1024) / BLCKSZ)
+
+/*
+ * SLRU bank size for slotno hash banks
+ */
+#define SLRU_BANK_SIZE 8
+
typedef struct SlruWriteAllData
{
int num_files; /* # files actually open */
@@ -134,7 +146,7 @@
static SlruErrorCause slru_errcause;
static int slru_errno;
-
+static void SlruAdjustNSlots(int *nslots, int *bankmask);
static void SimpleLruZeroLSNs(SlruCtl ctl, int slotno);
static void SimpleLruWaitIO(SlruCtl ctl, int slotno);
static void SlruInternalWritePage(SlruCtl ctl, int slotno, SlruWriteAll fdata);
@@ -148,6 +160,25 @@
int segpage, void *data);
static void SlruInternalDeleteSegment(SlruCtl ctl, int segno);
+/*
+ * Pick number of slots and bank size optimal for hashed associative SLRU buffers.
+ * We declare SLRU nslots is always power of 2.
+ * We split SLRU to 8-sized hash banks, after some performance benchmarks.
+ * We hash pageno to banks by pageno masked by 3 upper bits.
+ */
+static void
+SlruAdjustNSlots(int *nslots, int *bankmask)
+{
+ Assert(*nslots > 0);
+ Assert(*nslots <= SLRU_MAX_ALLOWED_BUFFERS);
+
+ *nslots = (int) pg_nextpower2_32(Max(SLRU_BANK_SIZE, Min(*nslots, NBuffers / 256)));
+
+ *bankmask = *nslots / SLRU_BANK_SIZE - 1;
+
+ elog(DEBUG5, "nslots %d banksize %d nbanks %d bankmask %x", *nslots, SLRU_BANK_SIZE, *nslots / SLRU_BANK_SIZE, *bankmask);
+}
+
/*
* Initialization of shared memory
*/
@@ -156,6 +187,9 @@
SimpleLruShmemSize(int nslots, int nlsns)
{
Size sz;
+ int bankmask_ignore;
+
+ SlruAdjustNSlots(&nslots, &bankmask_ignore);
/* we assume nslots isn't so large as to risk overflow */
sz = MAXALIGN(sizeof(SlruSharedData));
@@ -190,6 +224,9 @@
{
SlruShared shared;
bool found;
+ int bankmask;
+
+ SlruAdjustNSlots(&nslots, &bankmask);
shared = (SlruShared) ShmemInitStruct(name,
SimpleLruShmemSize(nslots, nlsns),
@@ -257,7 +294,10 @@
Assert(ptr - (char *) shared <= SimpleLruShmemSize(nslots, nlsns));
}
else
+ {
Assert(found);
+ Assert(shared->num_slots == nslots);
+ }
/*
* Initialize the unshared control struct, including directory path. We
@@ -265,6 +305,7 @@
*/
ctl->shared = shared;
ctl->sync_handler = sync_handler;
+ ctl->bank_mask = bankmask;
strlcpy(ctl->Dir, subdir, sizeof(ctl->Dir));
}
@@ -496,12 +537,14 @@
{
SlruShared shared = ctl->shared;
int slotno;
+ int bankstart = (pageno & ctl->bank_mask) * SLRU_BANK_SIZE;
+ int bankend = bankstart + SLRU_BANK_SIZE;
/* Try to find the page while holding only shared lock */
LWLockAcquire(shared->ControlLock, LW_SHARED);
/* See if page is already in a buffer */
- for (slotno = 0; slotno < shared->num_slots; slotno++)
+ for (slotno = bankstart; slotno < bankend; slotno++)
{
if (shared->page_number[slotno] == pageno &&
shared->page_status[slotno] != SLRU_PAGE_EMPTY &&
@@ -1030,7 +1073,10 @@
int best_invalid_page_number = 0; /* keep compiler quiet */
/* See if page already has a buffer assigned */
- for (slotno = 0; slotno < shared->num_slots; slotno++)
+ int bankstart = (pageno & ctl->bank_mask) * SLRU_BANK_SIZE;
+ int bankend = bankstart + SLRU_BANK_SIZE;
+
+ for (slotno = bankstart; slotno < bankend; slotno++)
{
if (shared->page_number[slotno] == pageno &&
shared->page_status[slotno] != SLRU_PAGE_EMPTY)
@@ -1065,7 +1111,7 @@
* multiple pages with the same lru_count.
*/
cur_count = (shared->cur_lru_count)++;
- for (slotno = 0; slotno < shared->num_slots; slotno++)
+ for (slotno = bankstart; slotno < bankend; slotno++)
{
int this_delta;
int this_page_number;
Index: src/backend/access/transam/clog.c
IDEA additional info:
Subsystem: com.intellij.openapi.diff.impl.patch.CharsetEP
<+>UTF-8
===================================================================
diff --git a/src/backend/access/transam/clog.c b/src/backend/access/transam/clog.c
--- a/src/backend/access/transam/clog.c (revision 020258fbd30d37ddd03d0ec68264d1544f8d2838)
+++ b/src/backend/access/transam/clog.c (date 1660677867877)
@@ -74,6 +74,8 @@
#define GetLSNIndex(slotno, xid) ((slotno) * CLOG_LSNS_PER_PAGE + \
((xid) % (TransactionId) CLOG_XACTS_PER_PAGE) / CLOG_XACTS_PER_LSN_GROUP)
+#define NUM_CLOG_BUFFERS (128 << slru_buffers_size_scale)
+
/*
* The number of subtransactions below which we consider to apply clog group
* update optimization. Testing reveals that the number higher than this can
@@ -661,42 +663,20 @@
return status;
}
-/*
- * Number of shared CLOG buffers.
- *
- * On larger multi-processor systems, it is possible to have many CLOG page
- * requests in flight at one time which could lead to disk access for CLOG
- * page if the required page is not found in memory. Testing revealed that we
- * can get the best performance by having 128 CLOG buffers, more than that it
- * doesn't improve performance.
- *
- * Unconditionally keeping the number of CLOG buffers to 128 did not seem like
- * a good idea, because it would increase the minimum amount of shared memory
- * required to start, which could be a problem for people running very small
- * configurations. The following formula seems to represent a reasonable
- * compromise: people with very low values for shared_buffers will get fewer
- * CLOG buffers as well, and everyone else will get 128.
- */
-Size
-CLOGShmemBuffers(void)
-{
- return Min(128, Max(4, NBuffers / 512));
-}
-
/*
* Initialization of shared memory for CLOG
*/
Size
CLOGShmemSize(void)
{
- return SimpleLruShmemSize(CLOGShmemBuffers(), CLOG_LSNS_PER_PAGE);
+ return SimpleLruShmemSize(NUM_CLOG_BUFFERS, CLOG_LSNS_PER_PAGE);
}
void
CLOGShmemInit(void)
{
XactCtl->PagePrecedes = CLOGPagePrecedes;
- SimpleLruInit(XactCtl, "Xact", CLOGShmemBuffers(), CLOG_LSNS_PER_PAGE,
+ SimpleLruInit(XactCtl, "Xact", NUM_CLOG_BUFFERS, CLOG_LSNS_PER_PAGE,
XactSLRULock, "pg_xact", LWTRANCHE_XACT_BUFFER,
SYNC_HANDLER_CLOG);
SlruPagePrecedesUnitTests(XactCtl, CLOG_XACTS_PER_PAGE);
Index: src/include/storage/predicate.h
IDEA additional info:
Subsystem: com.intellij.openapi.diff.impl.patch.CharsetEP
<+>UTF-8
===================================================================
diff --git a/src/include/storage/predicate.h b/src/include/storage/predicate.h
--- a/src/include/storage/predicate.h (revision 020258fbd30d37ddd03d0ec68264d1544f8d2838)
+++ b/src/include/storage/predicate.h (date 1660678108718)
@@ -28,7 +28,7 @@
/* Number of SLRU buffers to use for Serial SLRU */
-#define NUM_SERIAL_BUFFERS 16
+#define NUM_SERIAL_BUFFERS (16 << slru_buffers_size_scale)
/*
* A handle used for sharing SERIALIZABLEXACT objects between the participants
Index: src/backend/utils/misc/postgresql.conf.sample
IDEA additional info:
Subsystem: com.intellij.openapi.diff.impl.patch.CharsetEP
<+>UTF-8
===================================================================
diff --git a/src/backend/utils/misc/postgresql.conf.sample b/src/backend/utils/misc/postgresql.conf.sample
--- a/src/backend/utils/misc/postgresql.conf.sample (revision 020258fbd30d37ddd03d0ec68264d1544f8d2838)
+++ b/src/backend/utils/misc/postgresql.conf.sample (date 1660678108650)
@@ -155,6 +155,8 @@
# mmap
# (change requires restart)
#min_dynamic_shared_memory = 0MB # (change requires restart)
+#slru_buffers_size_scale = 2 # SLRU buffers size scale of power 2, range 0..7
+ # (change requires restart)
# - Disk -
Index: src/include/access/multixact.h
IDEA additional info:
Subsystem: com.intellij.openapi.diff.impl.patch.CharsetEP
<+>UTF-8
===================================================================
diff --git a/src/include/access/multixact.h b/src/include/access/multixact.h
--- a/src/include/access/multixact.h (revision 020258fbd30d37ddd03d0ec68264d1544f8d2838)
+++ b/src/include/access/multixact.h (date 1660678108678)
@@ -30,8 +30,8 @@
#define MaxMultiXactOffset ((MultiXactOffset) 0xFFFFFFFF)
/* Number of SLRU buffers to use for multixact */
-#define NUM_MULTIXACTOFFSET_BUFFERS 8
-#define NUM_MULTIXACTMEMBER_BUFFERS 16
+#define NUM_MULTIXACTOFFSET_BUFFERS (16 << slru_buffers_size_scale)
+#define NUM_MULTIXACTMEMBER_BUFFERS (32 << slru_buffers_size_scale)
/*
* Possible multixact lock modes ("status"). The first four modes are for
Index: src/backend/utils/misc/guc.c
IDEA additional info:
Subsystem: com.intellij.openapi.diff.impl.patch.CharsetEP
<+>UTF-8
===================================================================
diff --git a/src/backend/utils/misc/guc.c b/src/backend/utils/misc/guc.c
--- a/src/backend/utils/misc/guc.c (revision 020258fbd30d37ddd03d0ec68264d1544f8d2838)
+++ b/src/backend/utils/misc/guc.c (date 1660678064216)
@@ -32,9 +32,11 @@
#endif
#include <unistd.h>
+#include "access/clog.h"
#include "access/commit_ts.h"
#include "access/gin.h"
#include "access/rmgr.h"
+#include "access/slru.h"
#include "access/tableam.h"
#include "access/toast_compression.h"
#include "access/transam.h"
@@ -2375,6 +2377,16 @@
-1, -1, INT_MAX,
NULL, NULL, NULL
},
+
+ {
+ {"slru_buffers_size_scale", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("SLRU buffers size scale of power 2"),
+ NULL
+ },
+ &slru_buffers_size_scale,
+ 2, 0, 7,
+ NULL, NULL, NULL
+ },
{
{"temp_buffers", PGC_USERSET, RESOURCES_MEM,
Index: src/include/commands/async.h
IDEA additional info:
Subsystem: com.intellij.openapi.diff.impl.patch.CharsetEP
<+>UTF-8
===================================================================
diff --git a/src/include/commands/async.h b/src/include/commands/async.h
--- a/src/include/commands/async.h (revision 020258fbd30d37ddd03d0ec68264d1544f8d2838)
+++ b/src/include/commands/async.h (date 1660678108702)
@@ -18,7 +18,7 @@
/*
* The number of SLRU page buffers we use for the notification queue.
*/
-#define NUM_NOTIFY_BUFFERS 8
+#define NUM_NOTIFY_BUFFERS (16 << slru_buffers_size_scale)
extern bool Trace_notify;
extern volatile sig_atomic_t notifyInterruptPending;
Index: src/include/access/slru.h
IDEA additional info:
Subsystem: com.intellij.openapi.diff.impl.patch.CharsetEP
<+>UTF-8
===================================================================
diff --git a/src/include/access/slru.h b/src/include/access/slru.h
--- a/src/include/access/slru.h (revision 020258fbd30d37ddd03d0ec68264d1544f8d2838)
+++ b/src/include/access/slru.h (date 1660678108690)
@@ -134,6 +134,11 @@
* it's always the same, it doesn't need to be in shared memory.
*/
char Dir[64];
+
+ /*
+ * mask for slotno hash bank
+ */
+ Size bank_mask;
} SlruCtlData;
typedef SlruCtlData *SlruCtl;
Index: src/backend/access/transam/subtrans.c
IDEA additional info:
Subsystem: com.intellij.openapi.diff.impl.patch.CharsetEP
<+>UTF-8
===================================================================
diff --git a/src/backend/access/transam/subtrans.c b/src/backend/access/transam/subtrans.c
--- a/src/backend/access/transam/subtrans.c (revision 020258fbd30d37ddd03d0ec68264d1544f8d2838)
+++ b/src/backend/access/transam/subtrans.c (date 1660678064032)
@@ -31,6 +31,7 @@
#include "access/slru.h"
#include "access/subtrans.h"
#include "access/transam.h"
+#include "miscadmin.h"
#include "pg_trace.h"
#include "utils/snapmgr.h"
Index: doc/src/sgml/config.sgml
IDEA additional info:
Subsystem: com.intellij.openapi.diff.impl.patch.CharsetEP
<+>UTF-8
===================================================================
diff --git a/doc/src/sgml/config.sgml b/doc/src/sgml/config.sgml
--- a/doc/src/sgml/config.sgml (revision 020258fbd30d37ddd03d0ec68264d1544f8d2838)
+++ b/doc/src/sgml/config.sgml (date 1660677867861)
@@ -1953,6 +1953,37 @@
</listitem>
</varlistentry>
+ <varlistentry id="guc-slru-buffers-size-scale" xreflabel="slru_buffers_size_scale">
+ <term><varname>slru_buffers_size_scale</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>slru_buffers_size_scale</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies power 2 scale for all SLRU shared memory buffers sizes. Buffers sizes depends on
+ both <literal>guc_slru_buffers_size_scale</literal> and <literal>shared_buffers</literal> params.
+ </para>
+ <para>
+ This affects on buffers in the list below (see also <xref linkend="pgdata-contents-table"/>):
+ <itemizedlist>
+ <listitem><para><literal>NUM_MULTIXACTOFFSET_BUFFERS = Min(32 << slru_buffers_size_scale, shared_buffers/256)</literal></para></listitem>
+ <listitem><para><literal>NUM_MULTIXACTMEMBER_BUFFERS = Min(64 << slru_buffers_size_scale, shared_buffers/256)</literal></para></listitem>
+ <listitem><para><literal>NUM_SUBTRANS_BUFFERS = Min(64 << slru_buffers_size_scale, shared_buffers/256)</literal></para></listitem>
+ <listitem><para><literal>NUM_NOTIFY_BUFFERS = Min(32 << slru_buffers_size_scale, shared_buffers/256)</literal></para></listitem>
+ <listitem><para><literal>NUM_SERIAL_BUFFERS = Min(32 << slru_buffers_size_scale, shared_buffers/256)</literal></para></listitem>
+ <listitem><para><literal>NUM_CLOG_BUFFERS = Min(128 << slru_buffers_size_scale, shared_buffers/256)</literal></para></listitem>
+ <listitem><para><literal>NUM_COMMIT_TS_BUFFERS = Min(128 << slru_buffers_size_scale, shared_buffers/256)</literal></para></listitem>
+ </itemizedlist>
+ </para>
+ <para>
+ Value is in <literal>0..7</literal> bounds.
+ The default value is <literal>2</literal>.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
<varlistentry id="guc-max-stack-depth" xreflabel="max_stack_depth">
<term><varname>max_stack_depth</varname> (<type>integer</type>)
<indexterm>
Index: src/backend/access/transam/commit_ts.c
IDEA additional info:
Subsystem: com.intellij.openapi.diff.impl.patch.CharsetEP
<+>UTF-8
===================================================================
diff --git a/src/backend/access/transam/commit_ts.c b/src/backend/access/transam/commit_ts.c
--- a/src/backend/access/transam/commit_ts.c (revision 020258fbd30d37ddd03d0ec68264d1544f8d2838)
+++ b/src/backend/access/transam/commit_ts.c (date 1660677867889)
@@ -73,6 +73,8 @@
#define TransactionIdToCTsEntry(xid) \
((xid) % (TransactionId) COMMIT_TS_XACTS_PER_PAGE)
+#define NUM_COMMIT_TS_BUFFERS (128 << slru_buffers_size_scale)
+
/*
* Link to shared-memory data structures for CommitTs control
*/
@@ -508,26 +510,13 @@
PG_RETURN_DATUM(HeapTupleGetDatum(htup));
}
-/*
- * Number of shared CommitTS buffers.
- *
- * We use a very similar logic as for the number of CLOG buffers (except we
- * scale up twice as fast with shared buffers, and the maximum is twice as
- * high); see comments in CLOGShmemBuffers.
- */
-Size
-CommitTsShmemBuffers(void)
-{
- return Min(256, Max(4, NBuffers / 256));
-}
-
/*
* Shared memory sizing for CommitTs
*/
Size
CommitTsShmemSize(void)
{
- return SimpleLruShmemSize(CommitTsShmemBuffers(), 0) +
+ return SimpleLruShmemSize(NUM_COMMIT_TS_BUFFERS, 0) +
sizeof(CommitTimestampShared);
}
@@ -541,7 +530,7 @@
bool found;
CommitTsCtl->PagePrecedes = CommitTsPagePrecedes;
- SimpleLruInit(CommitTsCtl, "CommitTs", CommitTsShmemBuffers(), 0,
+ SimpleLruInit(CommitTsCtl, "CommitTs", NUM_COMMIT_TS_BUFFERS, 0,
CommitTsSLRULock, "pg_commit_ts",
LWTRANCHE_COMMITTS_BUFFER,
SYNC_HANDLER_COMMIT_TS);
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2022-08-18 03:35 Andrey Borodin <[email protected]>
parent: [email protected]
0 siblings, 1 reply; 103+ messages in thread
From: Andrey Borodin @ 2022-08-18 03:35 UTC (permalink / raw)
To: [email protected]; +Cc: Yura Sokolov <[email protected]>; Andres Freund <[email protected]>; Thomas Munro <[email protected]>; Gilles Darold <[email protected]>; Tomas Vondra <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
> On 17 Aug 2022, at 00:36, [email protected] wrote:
>
>> Andrey Borodin wrote 2022-07-23 11:39:
>> Yura, thank you for your benchmarks!
>> We already knew that patch can save the day on pathological workloads,
>> now we have a proof of this.
>> Also there's the evidence that user can blindly increase size of SLRU
>> if they want (with the second patch). So there's no need for hard
>> explanations on how to tune the buffers size.
>
> Hi @Andrey.Borodin, With some considerations and performance checks from @Yura.Sokolov we simplified your approach by the following:
>
> 1. Preamble. We feel free to increase any SLRU's, since there's no performance degradation on large Buffers count using your SLRU buckets solution.
> 2. `slru_buffers_size_scale` is only one config param introduced for all SLRUs. It scales SLRUs upper cap by power 2.
> 3. All SLRU buffers count are capped by both `MBuffers (shared_buffers)` and `slru_buffers_size_scale`. see
> 4. Magic initial constants `NUM_*_BUFFERS << slru_buffers_size_scale` are applied for every SLRU.
> 5. All SLRU buffers are always sized as power of 2, their hash bucket size is always 8.
>
> There's attached patch for your consideration. It does gather and simplify both `v21-0001-Make-all-SLRU-buffer-sizes-configurable.patch` and `v21-0002-Divide-SLRU-buffers-into-8-associative-banks.patch` to much simpler approach.
I like the idea of one knob instead of one per each SLRU. Maybe we even could deduce sane value from NBuffers? That would effectively lead to 0 knobs :)
Your patch have a prefix "v22-0006", does it mean there are 5 previous steps of the patchset?
Thank you!
Best regards, Andrey Borodin.
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2022-08-19 15:48 [email protected]
parent: Andrey Borodin <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: [email protected] @ 2022-08-19 15:48 UTC (permalink / raw)
To: Andrey Borodin <[email protected]>; +Cc: Yura Sokolov <[email protected]>; Andres Freund <[email protected]>; Thomas Munro <[email protected]>; Gilles Darold <[email protected]>; Tomas Vondra <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
Andrey Borodin wrote 2022-08-18 06:35:
>
> I like the idea of one knob instead of one per each SLRU. Maybe we
> even could deduce sane value from NBuffers? That would effectively
> lead to 0 knobs :)
>
> Your patch have a prefix "v22-0006", does it mean there are 5 previous
> steps of the patchset?
>
> Thank you!
>
>
> Best regards, Andrey Borodin.
Not sure it's possible to deduce from NBuffers only.
slru_buffers_scale_shift looks like relief valve for systems with ultra
scaled NBuffers.
Regarding v22-0006 I just tried to choose index unique for this thread
so now it's fixed to 0001 indexing.
Attachments:
[text/x-diff] v23-0001-bucketed-SLRUs-simplified.patch (18.0K, ../../[email protected]/2-v23-0001-bucketed-SLRUs-simplified.patch)
download | inline diff:
Index: src/include/miscadmin.h
IDEA additional info:
Subsystem: com.intellij.openapi.diff.impl.patch.CharsetEP
<+>UTF-8
===================================================================
diff --git a/src/include/miscadmin.h b/src/include/miscadmin.h
--- a/src/include/miscadmin.h (revision 020258fbd30d37ddd03d0ec68264d1544f8d2838)
+++ b/src/include/miscadmin.h (date 1660678108730)
@@ -176,6 +176,7 @@
extern PGDLLIMPORT int MaxConnections;
extern PGDLLIMPORT int max_worker_processes;
extern PGDLLIMPORT int max_parallel_workers;
+extern PGDLLIMPORT int slru_buffers_size_scale;
extern PGDLLIMPORT int MyProcPid;
extern PGDLLIMPORT pg_time_t MyStartTime;
Index: src/include/access/subtrans.h
IDEA additional info:
Subsystem: com.intellij.openapi.diff.impl.patch.CharsetEP
<+>UTF-8
===================================================================
diff --git a/src/include/access/subtrans.h b/src/include/access/subtrans.h
--- a/src/include/access/subtrans.h (revision 020258fbd30d37ddd03d0ec68264d1544f8d2838)
+++ b/src/include/access/subtrans.h (date 1660678108698)
@@ -12,7 +12,7 @@
#define SUBTRANS_H
/* Number of SLRU buffers to use for subtrans */
-#define NUM_SUBTRANS_BUFFERS 32
+#define NUM_SUBTRANS_BUFFERS (32 << slru_buffers_size_scale)
extern void SubTransSetParent(TransactionId xid, TransactionId parent);
extern TransactionId SubTransGetParent(TransactionId xid);
Index: src/backend/utils/init/globals.c
IDEA additional info:
Subsystem: com.intellij.openapi.diff.impl.patch.CharsetEP
<+>UTF-8
===================================================================
diff --git a/src/backend/utils/init/globals.c b/src/backend/utils/init/globals.c
--- a/src/backend/utils/init/globals.c (revision 020258fbd30d37ddd03d0ec68264d1544f8d2838)
+++ b/src/backend/utils/init/globals.c (date 1660678064048)
@@ -150,3 +150,5 @@
int VacuumCostBalance = 0; /* working state for vacuum */
bool VacuumCostActive = false;
+
+int slru_buffers_size_scale = 2; /* power 2 scale for SLRU buffers */
Index: src/backend/access/transam/slru.c
IDEA additional info:
Subsystem: com.intellij.openapi.diff.impl.patch.CharsetEP
<+>UTF-8
===================================================================
diff --git a/src/backend/access/transam/slru.c b/src/backend/access/transam/slru.c
--- a/src/backend/access/transam/slru.c (revision 020258fbd30d37ddd03d0ec68264d1544f8d2838)
+++ b/src/backend/access/transam/slru.c (date 1660678063980)
@@ -59,6 +59,7 @@
#include "pgstat.h"
#include "storage/fd.h"
#include "storage/shmem.h"
+#include "port/pg_bitutils.h"
#define SlruFileName(ctl, path, seg) \
snprintf(path, MAXPGPATH, "%s/%04X", (ctl)->Dir, seg)
@@ -71,6 +72,17 @@
*/
#define MAX_WRITEALL_BUFFERS 16
+/*
+ * To avoid overflowing internal arithmetic and the size_t data type, the
+ * number of buffers should not exceed this number.
+ */
+#define SLRU_MAX_ALLOWED_BUFFERS ((1024 * 1024 * 1024) / BLCKSZ)
+
+/*
+ * SLRU bank size for slotno hash banks
+ */
+#define SLRU_BANK_SIZE 8
+
typedef struct SlruWriteAllData
{
int num_files; /* # files actually open */
@@ -134,7 +146,7 @@
static SlruErrorCause slru_errcause;
static int slru_errno;
-
+static void SlruAdjustNSlots(int *nslots, int *bankmask);
static void SimpleLruZeroLSNs(SlruCtl ctl, int slotno);
static void SimpleLruWaitIO(SlruCtl ctl, int slotno);
static void SlruInternalWritePage(SlruCtl ctl, int slotno, SlruWriteAll fdata);
@@ -148,6 +160,25 @@
int segpage, void *data);
static void SlruInternalDeleteSegment(SlruCtl ctl, int segno);
+/*
+ * Pick number of slots and bank size optimal for hashed associative SLRU buffers.
+ * We declare SLRU nslots is always power of 2.
+ * We split SLRU to 8-sized hash banks, after some performance benchmarks.
+ * We hash pageno to banks by pageno masked by 3 upper bits.
+ */
+static void
+SlruAdjustNSlots(int *nslots, int *bankmask)
+{
+ Assert(*nslots > 0);
+ Assert(*nslots <= SLRU_MAX_ALLOWED_BUFFERS);
+
+ *nslots = (int) pg_nextpower2_32(Max(SLRU_BANK_SIZE, Min(*nslots, NBuffers / 256)));
+
+ *bankmask = *nslots / SLRU_BANK_SIZE - 1;
+
+ elog(DEBUG5, "nslots %d banksize %d nbanks %d bankmask %x", *nslots, SLRU_BANK_SIZE, *nslots / SLRU_BANK_SIZE, *bankmask);
+}
+
/*
* Initialization of shared memory
*/
@@ -156,6 +187,9 @@
SimpleLruShmemSize(int nslots, int nlsns)
{
Size sz;
+ int bankmask_ignore;
+
+ SlruAdjustNSlots(&nslots, &bankmask_ignore);
/* we assume nslots isn't so large as to risk overflow */
sz = MAXALIGN(sizeof(SlruSharedData));
@@ -190,6 +224,9 @@
{
SlruShared shared;
bool found;
+ int bankmask;
+
+ SlruAdjustNSlots(&nslots, &bankmask);
shared = (SlruShared) ShmemInitStruct(name,
SimpleLruShmemSize(nslots, nlsns),
@@ -257,7 +294,10 @@
Assert(ptr - (char *) shared <= SimpleLruShmemSize(nslots, nlsns));
}
else
+ {
Assert(found);
+ Assert(shared->num_slots == nslots);
+ }
/*
* Initialize the unshared control struct, including directory path. We
@@ -265,6 +305,7 @@
*/
ctl->shared = shared;
ctl->sync_handler = sync_handler;
+ ctl->bank_mask = bankmask;
strlcpy(ctl->Dir, subdir, sizeof(ctl->Dir));
}
@@ -496,12 +537,14 @@
{
SlruShared shared = ctl->shared;
int slotno;
+ int bankstart = (pageno & ctl->bank_mask) * SLRU_BANK_SIZE;
+ int bankend = bankstart + SLRU_BANK_SIZE;
/* Try to find the page while holding only shared lock */
LWLockAcquire(shared->ControlLock, LW_SHARED);
/* See if page is already in a buffer */
- for (slotno = 0; slotno < shared->num_slots; slotno++)
+ for (slotno = bankstart; slotno < bankend; slotno++)
{
if (shared->page_number[slotno] == pageno &&
shared->page_status[slotno] != SLRU_PAGE_EMPTY &&
@@ -1030,7 +1073,10 @@
int best_invalid_page_number = 0; /* keep compiler quiet */
/* See if page already has a buffer assigned */
- for (slotno = 0; slotno < shared->num_slots; slotno++)
+ int bankstart = (pageno & ctl->bank_mask) * SLRU_BANK_SIZE;
+ int bankend = bankstart + SLRU_BANK_SIZE;
+
+ for (slotno = bankstart; slotno < bankend; slotno++)
{
if (shared->page_number[slotno] == pageno &&
shared->page_status[slotno] != SLRU_PAGE_EMPTY)
@@ -1065,7 +1111,7 @@
* multiple pages with the same lru_count.
*/
cur_count = (shared->cur_lru_count)++;
- for (slotno = 0; slotno < shared->num_slots; slotno++)
+ for (slotno = bankstart; slotno < bankend; slotno++)
{
int this_delta;
int this_page_number;
Index: src/backend/access/transam/clog.c
IDEA additional info:
Subsystem: com.intellij.openapi.diff.impl.patch.CharsetEP
<+>UTF-8
===================================================================
diff --git a/src/backend/access/transam/clog.c b/src/backend/access/transam/clog.c
--- a/src/backend/access/transam/clog.c (revision 020258fbd30d37ddd03d0ec68264d1544f8d2838)
+++ b/src/backend/access/transam/clog.c (date 1660677867877)
@@ -74,6 +74,8 @@
#define GetLSNIndex(slotno, xid) ((slotno) * CLOG_LSNS_PER_PAGE + \
((xid) % (TransactionId) CLOG_XACTS_PER_PAGE) / CLOG_XACTS_PER_LSN_GROUP)
+#define NUM_CLOG_BUFFERS (128 << slru_buffers_size_scale)
+
/*
* The number of subtransactions below which we consider to apply clog group
* update optimization. Testing reveals that the number higher than this can
@@ -661,42 +663,20 @@
return status;
}
-/*
- * Number of shared CLOG buffers.
- *
- * On larger multi-processor systems, it is possible to have many CLOG page
- * requests in flight at one time which could lead to disk access for CLOG
- * page if the required page is not found in memory. Testing revealed that we
- * can get the best performance by having 128 CLOG buffers, more than that it
- * doesn't improve performance.
- *
- * Unconditionally keeping the number of CLOG buffers to 128 did not seem like
- * a good idea, because it would increase the minimum amount of shared memory
- * required to start, which could be a problem for people running very small
- * configurations. The following formula seems to represent a reasonable
- * compromise: people with very low values for shared_buffers will get fewer
- * CLOG buffers as well, and everyone else will get 128.
- */
-Size
-CLOGShmemBuffers(void)
-{
- return Min(128, Max(4, NBuffers / 512));
-}
-
/*
* Initialization of shared memory for CLOG
*/
Size
CLOGShmemSize(void)
{
- return SimpleLruShmemSize(CLOGShmemBuffers(), CLOG_LSNS_PER_PAGE);
+ return SimpleLruShmemSize(NUM_CLOG_BUFFERS, CLOG_LSNS_PER_PAGE);
}
void
CLOGShmemInit(void)
{
XactCtl->PagePrecedes = CLOGPagePrecedes;
- SimpleLruInit(XactCtl, "Xact", CLOGShmemBuffers(), CLOG_LSNS_PER_PAGE,
+ SimpleLruInit(XactCtl, "Xact", NUM_CLOG_BUFFERS, CLOG_LSNS_PER_PAGE,
XactSLRULock, "pg_xact", LWTRANCHE_XACT_BUFFER,
SYNC_HANDLER_CLOG);
SlruPagePrecedesUnitTests(XactCtl, CLOG_XACTS_PER_PAGE);
Index: src/include/storage/predicate.h
IDEA additional info:
Subsystem: com.intellij.openapi.diff.impl.patch.CharsetEP
<+>UTF-8
===================================================================
diff --git a/src/include/storage/predicate.h b/src/include/storage/predicate.h
--- a/src/include/storage/predicate.h (revision 020258fbd30d37ddd03d0ec68264d1544f8d2838)
+++ b/src/include/storage/predicate.h (date 1660678108718)
@@ -28,7 +28,7 @@
/* Number of SLRU buffers to use for Serial SLRU */
-#define NUM_SERIAL_BUFFERS 16
+#define NUM_SERIAL_BUFFERS (16 << slru_buffers_size_scale)
/*
* A handle used for sharing SERIALIZABLEXACT objects between the participants
Index: src/backend/utils/misc/postgresql.conf.sample
IDEA additional info:
Subsystem: com.intellij.openapi.diff.impl.patch.CharsetEP
<+>UTF-8
===================================================================
diff --git a/src/backend/utils/misc/postgresql.conf.sample b/src/backend/utils/misc/postgresql.conf.sample
--- a/src/backend/utils/misc/postgresql.conf.sample (revision 020258fbd30d37ddd03d0ec68264d1544f8d2838)
+++ b/src/backend/utils/misc/postgresql.conf.sample (date 1660678108650)
@@ -155,6 +155,8 @@
# mmap
# (change requires restart)
#min_dynamic_shared_memory = 0MB # (change requires restart)
+#slru_buffers_size_scale = 2 # SLRU buffers size scale of power 2, range 0..7
+ # (change requires restart)
# - Disk -
Index: src/include/access/multixact.h
IDEA additional info:
Subsystem: com.intellij.openapi.diff.impl.patch.CharsetEP
<+>UTF-8
===================================================================
diff --git a/src/include/access/multixact.h b/src/include/access/multixact.h
--- a/src/include/access/multixact.h (revision 020258fbd30d37ddd03d0ec68264d1544f8d2838)
+++ b/src/include/access/multixact.h (date 1660678108678)
@@ -30,8 +30,8 @@
#define MaxMultiXactOffset ((MultiXactOffset) 0xFFFFFFFF)
/* Number of SLRU buffers to use for multixact */
-#define NUM_MULTIXACTOFFSET_BUFFERS 8
-#define NUM_MULTIXACTMEMBER_BUFFERS 16
+#define NUM_MULTIXACTOFFSET_BUFFERS (16 << slru_buffers_size_scale)
+#define NUM_MULTIXACTMEMBER_BUFFERS (32 << slru_buffers_size_scale)
/*
* Possible multixact lock modes ("status"). The first four modes are for
Index: src/backend/utils/misc/guc.c
IDEA additional info:
Subsystem: com.intellij.openapi.diff.impl.patch.CharsetEP
<+>UTF-8
===================================================================
diff --git a/src/backend/utils/misc/guc.c b/src/backend/utils/misc/guc.c
--- a/src/backend/utils/misc/guc.c (revision 020258fbd30d37ddd03d0ec68264d1544f8d2838)
+++ b/src/backend/utils/misc/guc.c (date 1660678064216)
@@ -32,9 +32,11 @@
#endif
#include <unistd.h>
+#include "access/clog.h"
#include "access/commit_ts.h"
#include "access/gin.h"
#include "access/rmgr.h"
+#include "access/slru.h"
#include "access/tableam.h"
#include "access/toast_compression.h"
#include "access/transam.h"
@@ -2375,6 +2377,16 @@
-1, -1, INT_MAX,
NULL, NULL, NULL
},
+
+ {
+ {"slru_buffers_size_scale", PGC_POSTMASTER, RESOURCES_MEM,
+ gettext_noop("SLRU buffers size scale of power 2"),
+ NULL
+ },
+ &slru_buffers_size_scale,
+ 2, 0, 7,
+ NULL, NULL, NULL
+ },
{
{"temp_buffers", PGC_USERSET, RESOURCES_MEM,
Index: src/include/commands/async.h
IDEA additional info:
Subsystem: com.intellij.openapi.diff.impl.patch.CharsetEP
<+>UTF-8
===================================================================
diff --git a/src/include/commands/async.h b/src/include/commands/async.h
--- a/src/include/commands/async.h (revision 020258fbd30d37ddd03d0ec68264d1544f8d2838)
+++ b/src/include/commands/async.h (date 1660678108702)
@@ -18,7 +18,7 @@
/*
* The number of SLRU page buffers we use for the notification queue.
*/
-#define NUM_NOTIFY_BUFFERS 8
+#define NUM_NOTIFY_BUFFERS (16 << slru_buffers_size_scale)
extern bool Trace_notify;
extern volatile sig_atomic_t notifyInterruptPending;
Index: src/include/access/slru.h
IDEA additional info:
Subsystem: com.intellij.openapi.diff.impl.patch.CharsetEP
<+>UTF-8
===================================================================
diff --git a/src/include/access/slru.h b/src/include/access/slru.h
--- a/src/include/access/slru.h (revision 020258fbd30d37ddd03d0ec68264d1544f8d2838)
+++ b/src/include/access/slru.h (date 1660678108690)
@@ -134,6 +134,11 @@
* it's always the same, it doesn't need to be in shared memory.
*/
char Dir[64];
+
+ /*
+ * mask for slotno hash bank
+ */
+ Size bank_mask;
} SlruCtlData;
typedef SlruCtlData *SlruCtl;
Index: src/backend/access/transam/subtrans.c
IDEA additional info:
Subsystem: com.intellij.openapi.diff.impl.patch.CharsetEP
<+>UTF-8
===================================================================
diff --git a/src/backend/access/transam/subtrans.c b/src/backend/access/transam/subtrans.c
--- a/src/backend/access/transam/subtrans.c (revision 020258fbd30d37ddd03d0ec68264d1544f8d2838)
+++ b/src/backend/access/transam/subtrans.c (date 1660678064032)
@@ -31,6 +31,7 @@
#include "access/slru.h"
#include "access/subtrans.h"
#include "access/transam.h"
+#include "miscadmin.h"
#include "pg_trace.h"
#include "utils/snapmgr.h"
Index: doc/src/sgml/config.sgml
IDEA additional info:
Subsystem: com.intellij.openapi.diff.impl.patch.CharsetEP
<+>UTF-8
===================================================================
diff --git a/doc/src/sgml/config.sgml b/doc/src/sgml/config.sgml
--- a/doc/src/sgml/config.sgml (revision 020258fbd30d37ddd03d0ec68264d1544f8d2838)
+++ b/doc/src/sgml/config.sgml (date 1660677867861)
@@ -1953,6 +1953,37 @@
</listitem>
</varlistentry>
+ <varlistentry id="guc-slru-buffers-size-scale" xreflabel="slru_buffers_size_scale">
+ <term><varname>slru_buffers_size_scale</varname> (<type>integer</type>)
+ <indexterm>
+ <primary><varname>slru_buffers_size_scale</varname> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Specifies power 2 scale for all SLRU shared memory buffers sizes. Buffers sizes depends on
+ both <literal>guc_slru_buffers_size_scale</literal> and <literal>shared_buffers</literal> params.
+ </para>
+ <para>
+ This affects on buffers in the list below (see also <xref linkend="pgdata-contents-table"/>):
+ <itemizedlist>
+ <listitem><para><literal>NUM_MULTIXACTOFFSET_BUFFERS = Min(32 << slru_buffers_size_scale, shared_buffers/256)</literal></para></listitem>
+ <listitem><para><literal>NUM_MULTIXACTMEMBER_BUFFERS = Min(64 << slru_buffers_size_scale, shared_buffers/256)</literal></para></listitem>
+ <listitem><para><literal>NUM_SUBTRANS_BUFFERS = Min(64 << slru_buffers_size_scale, shared_buffers/256)</literal></para></listitem>
+ <listitem><para><literal>NUM_NOTIFY_BUFFERS = Min(32 << slru_buffers_size_scale, shared_buffers/256)</literal></para></listitem>
+ <listitem><para><literal>NUM_SERIAL_BUFFERS = Min(32 << slru_buffers_size_scale, shared_buffers/256)</literal></para></listitem>
+ <listitem><para><literal>NUM_CLOG_BUFFERS = Min(128 << slru_buffers_size_scale, shared_buffers/256)</literal></para></listitem>
+ <listitem><para><literal>NUM_COMMIT_TS_BUFFERS = Min(128 << slru_buffers_size_scale, shared_buffers/256)</literal></para></listitem>
+ </itemizedlist>
+ </para>
+ <para>
+ Value is in <literal>0..7</literal> bounds.
+ The default value is <literal>2</literal>.
+ This parameter can only be set at server start.
+ </para>
+ </listitem>
+ </varlistentry>
+
<varlistentry id="guc-max-stack-depth" xreflabel="max_stack_depth">
<term><varname>max_stack_depth</varname> (<type>integer</type>)
<indexterm>
Index: src/backend/access/transam/commit_ts.c
IDEA additional info:
Subsystem: com.intellij.openapi.diff.impl.patch.CharsetEP
<+>UTF-8
===================================================================
diff --git a/src/backend/access/transam/commit_ts.c b/src/backend/access/transam/commit_ts.c
--- a/src/backend/access/transam/commit_ts.c (revision 020258fbd30d37ddd03d0ec68264d1544f8d2838)
+++ b/src/backend/access/transam/commit_ts.c (date 1660677867889)
@@ -73,6 +73,8 @@
#define TransactionIdToCTsEntry(xid) \
((xid) % (TransactionId) COMMIT_TS_XACTS_PER_PAGE)
+#define NUM_COMMIT_TS_BUFFERS (128 << slru_buffers_size_scale)
+
/*
* Link to shared-memory data structures for CommitTs control
*/
@@ -508,26 +510,13 @@
PG_RETURN_DATUM(HeapTupleGetDatum(htup));
}
-/*
- * Number of shared CommitTS buffers.
- *
- * We use a very similar logic as for the number of CLOG buffers (except we
- * scale up twice as fast with shared buffers, and the maximum is twice as
- * high); see comments in CLOGShmemBuffers.
- */
-Size
-CommitTsShmemBuffers(void)
-{
- return Min(256, Max(4, NBuffers / 256));
-}
-
/*
* Shared memory sizing for CommitTs
*/
Size
CommitTsShmemSize(void)
{
- return SimpleLruShmemSize(CommitTsShmemBuffers(), 0) +
+ return SimpleLruShmemSize(NUM_COMMIT_TS_BUFFERS, 0) +
sizeof(CommitTimestampShared);
}
@@ -541,7 +530,7 @@
bool found;
CommitTsCtl->PagePrecedes = CommitTsPagePrecedes;
- SimpleLruInit(CommitTsCtl, "CommitTs", CommitTsShmemBuffers(), 0,
+ SimpleLruInit(CommitTsCtl, "CommitTs", NUM_COMMIT_TS_BUFFERS, 0,
CommitTsSLRULock, "pg_commit_ts",
LWTRANCHE_COMMITTS_BUFFER,
SYNC_HANDLER_COMMIT_TS);
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2022-12-20 18:39 Andrey Borodin <[email protected]>
parent: Thomas Munro <[email protected]>
0 siblings, 0 replies; 103+ messages in thread
From: Andrey Borodin @ 2022-12-20 18:39 UTC (permalink / raw)
To: Thomas Munro <[email protected]>; +Cc: Andrey Borodin <[email protected]>; Yura Sokolov <[email protected]>; Andres Freund <[email protected]>; Gilles Darold <[email protected]>; Tomas Vondra <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
On Sat, Jul 23, 2022 at 1:48 AM Thomas Munro <[email protected]> wrote:
>
> On Sat, Jul 23, 2022 at 8:41 PM Andrey Borodin <[email protected]> wrote:
> > Thomas, do you still have any doubts? Or is it certain that SLRU will be replaced by any better subsystem in 16?
>
> Hi Andrey,
>
> Sorry for my lack of replies on this and the other SLRU thread -- I'm
> thinking and experimenting. More soon.
>
Hi Thomas,
PostgreSQL 16 feature freeze is approaching again. Let's choose
something from possible solutions, even if the chosen one is
temporary.
Thank you!
Best regards, Andrey Borodin.
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2024-01-20 03:31 vignesh C <[email protected]>
parent: [email protected]
0 siblings, 1 reply; 103+ messages in thread
From: vignesh C @ 2024-01-20 03:31 UTC (permalink / raw)
To: Andrey Borodin <[email protected]>; +Cc: [email protected]; Andrey Borodin <[email protected]>; Yura Sokolov <[email protected]>; Andres Freund <[email protected]>; Thomas Munro <[email protected]>; Gilles Darold <[email protected]>; Tomas Vondra <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
On Mon, 9 Jan 2023 at 09:49, Andrey Borodin <[email protected]> wrote:
>
> On Tue, Jan 3, 2023 at 5:02 AM vignesh C <[email protected]> wrote:
> > does not apply on top of HEAD as in [1], please post a rebased patch:
> >
> Thanks! Here's the rebase.
I'm seeing that there has been no activity in this thread for more
than 1 year now, I'm planning to close this in the current commitfest
unless someone is planning to take it forward.
Regards,
Vignesh
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2024-01-27 04:58 Andrey Borodin <[email protected]>
parent: vignesh C <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Andrey Borodin @ 2024-01-27 04:58 UTC (permalink / raw)
To: vignesh C <[email protected]>; +Cc: Andrew Borodin <[email protected]>; [email protected]; Yura Sokolov <[email protected]>; Andres Freund <[email protected]>; Thomas Munro <[email protected]>; Gilles Darold <[email protected]>; Tomas Vondra <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
> On 20 Jan 2024, at 08:31, vignesh C <[email protected]> wrote:
>
> On Mon, 9 Jan 2023 at 09:49, Andrey Borodin <[email protected]> wrote:
>>
>> On Tue, Jan 3, 2023 at 5:02 AM vignesh C <[email protected]> wrote:
>>> does not apply on top of HEAD as in [1], please post a rebased patch:
>>>
>> Thanks! Here's the rebase.
>
> I'm seeing that there has been no activity in this thread for more
> than 1 year now, I'm planning to close this in the current commitfest
> unless someone is planning to take it forward.
Hi Vignesh,
thanks for the ping! Most important parts of this patch set are discussed in [0]. If that patchset will be committed, I'll withdraw entry for this thread from commitfest.
There's a version of Multixact-specific optimizations [1], but I hope they will not be necessary with effective caches developed in [0]. It seems to me that most important part of those optimization is removing sleeps under SLRU lock on standby [2] by Kyotaro Horiguchi. But given that cache optimizations took 4 years to get closer to commit, I'm not sure we will get this optimization any time soon...
Thanks!
Best regards, Andrey Borodin.
[0] https://www.postgresql.org/message-id/flat/CAFiTN-vzDvNz%3DExGXz6gdyjtzGixKSqs0mKHMmaQ8sOSEFZ33A%40m...
[1] https://www.postgresql.org/message-id/flat/2ECE132B-C042-4489-930E-DBC5D0DAB84A%40yandex-team.ru#5f7...
[2] https://www.postgresql.org/message-id/flat/20200515.090333.24867479329066911.horikyota.ntt%40gmail.c...
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2024-01-28 12:49 Alvaro Herrera <[email protected]>
parent: Andrey Borodin <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Alvaro Herrera @ 2024-01-28 12:49 UTC (permalink / raw)
To: Andrey Borodin <[email protected]>; +Cc: vignesh C <[email protected]>; Andrew Borodin <[email protected]>; [email protected]; Yura Sokolov <[email protected]>; Andres Freund <[email protected]>; Thomas Munro <[email protected]>; Gilles Darold <[email protected]>; Tomas Vondra <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
On 2024-Jan-27, Andrey Borodin wrote:
> thanks for the ping! Most important parts of this patch set are discussed in [0]. If that patchset will be committed, I'll withdraw entry for this thread from commitfest.
> There's a version of Multixact-specific optimizations [1], but I hope they will not be necessary with effective caches developed in [0]. It seems to me that most important part of those optimization is removing sleeps under SLRU lock on standby [2] by Kyotaro Horiguchi. But given that cache optimizations took 4 years to get closer to commit, I'm not sure we will get this optimization any time soon...
I'd appreciate it if you or Horiguchi-san can update his patch to remove
use of usleep in favor of a CV in multixact, and keep this CF entry to
cover it.
Perhaps a test to make the code reach the usleep(1000) can be written
using injection points (49cd2b93d7db)?
--
Álvaro Herrera PostgreSQL Developer — https://www.EnterpriseDB.com/
"La experiencia nos dice que el hombre peló millones de veces las patatas,
pero era forzoso admitir la posibilidad de que en un caso entre millones,
las patatas pelarían al hombre" (Ijon Tichy)
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2024-01-28 18:17 Andrey M. Borodin <[email protected]>
parent: Alvaro Herrera <[email protected]>
0 siblings, 0 replies; 103+ messages in thread
From: Andrey M. Borodin @ 2024-01-28 18:17 UTC (permalink / raw)
To: Alvaro Herrera <[email protected]>; +Cc: vignesh C <[email protected]>; Andrew Borodin <[email protected]>; [email protected]; Yura Sokolov <[email protected]>; Andres Freund <[email protected]>; Thomas Munro <[email protected]>; Gilles Darold <[email protected]>; Tomas Vondra <[email protected]>; Tomas Vondra <[email protected]>; Alexander Korotkov <[email protected]>; Anastasia Lubennikova <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
> On 28 Jan 2024, at 17:49, Alvaro Herrera <[email protected]> wrote:
>
> I'd appreciate it if you or Horiguchi-san can update his patch to remove
> use of usleep in favor of a CV in multixact, and keep this CF entry to
> cover it.
Sure! Sounds great!
> Perhaps a test to make the code reach the usleep(1000) can be written
> using injection points (49cd2b93d7db)?
I've tried to prototype something like that. But interesting point between GetNewMultiXactId() and RecordNewMultiXact() is a critical section, and we cannot have injection points in critical sections...
Also, to implement such a test we need "wait" type of injection points, see step 2 in attachment. With this type of injection points I can stop a backend amidst entering information about new MultiXact.
Best regards, Andrey Borodin.
Attachments:
[application/octet-stream] 0001-Add-conditional-variable-to-wait-for-next-MultXact-o.patch (3.3K, ../../[email protected]/2-0001-Add-conditional-variable-to-wait-for-next-MultXact-o.patch)
download | inline diff:
From 975eb3448acbec97c14d48f21261e44aca7b3acc Mon Sep 17 00:00:00 2001
From: Andrey Borodin <[email protected]>
Date: Fri, 22 May 2020 14:13:33 +0500
Subject: [PATCH 1/3] Add conditional variable to wait for next MultXact offset
in edge case
---
src/backend/access/transam/multixact.c | 23 ++++++++++++++++++-
.../utils/activity/wait_event_names.txt | 1 +
2 files changed, 23 insertions(+), 1 deletion(-)
diff --git a/src/backend/access/transam/multixact.c b/src/backend/access/transam/multixact.c
index 59523be901..03fcd25d4c 100644
--- a/src/backend/access/transam/multixact.c
+++ b/src/backend/access/transam/multixact.c
@@ -82,6 +82,7 @@
#include "lib/ilist.h"
#include "miscadmin.h"
#include "pg_trace.h"
+#include "pgstat.h"
#include "postmaster/autovacuum.h"
#include "storage/lmgr.h"
#include "storage/pmsignal.h"
@@ -233,6 +234,7 @@ typedef struct MultiXactStateData
/* support for members anti-wraparound measures */
MultiXactOffset offsetStopLimit; /* known if oldestOffsetKnown */
+ ConditionVariable nextoff_cv;
/*
* Per-backend data starts here. We have two arrays stored in the area
* immediately following the MultiXactStateData struct. Each is indexed by
@@ -894,6 +896,14 @@ RecordNewMultiXact(MultiXactId multi, MultiXactOffset offset,
/* Exchange our lock */
LWLockRelease(MultiXactOffsetSLRULock);
+ /*
+ * Let everybody know the offset of this mxid is recorded now. The waiters
+ * are waiting for the offset of the mxid next of the target to know the
+ * number of members of the target mxid, so we don't need to wait for
+ * members of this mxid are recorded.
+ */
+ ConditionVariableBroadcast(&MultiXactState->nextoff_cv);
+
LWLockAcquire(MultiXactMemberSLRULock, LW_EXCLUSIVE);
prev_pageno = -1;
@@ -1388,9 +1398,19 @@ retry:
if (nextMXOffset == 0)
{
/* Corner case 2: next multixact is still being filled in */
+
+ /*
+ * The recorder of the next mxid is just before writing the offset.
+ * Wait for the offset to be written.
+ */
+ ConditionVariablePrepareToSleep(&MultiXactState->nextoff_cv);
+
LWLockRelease(MultiXactOffsetSLRULock);
CHECK_FOR_INTERRUPTS();
- pg_usleep(1000L);
+
+ ConditionVariableSleep(&MultiXactState->nextoff_cv,
+ WAIT_EVENT_NEXT_MXMEMBERS);
+ ConditionVariableCancelSleep();
goto retry;
}
@@ -1875,6 +1895,7 @@ MultiXactShmemInit(void)
/* Make sure we zero out the per-backend state */
MemSet(MultiXactState, 0, SHARED_MULTIXACT_STATE_SIZE);
+ ConditionVariableInit(&MultiXactState->nextoff_cv);
}
else
Assert(found);
diff --git a/src/backend/utils/activity/wait_event_names.txt b/src/backend/utils/activity/wait_event_names.txt
index a5df835dd4..1bb78d5aad 100644
--- a/src/backend/utils/activity/wait_event_names.txt
+++ b/src/backend/utils/activity/wait_event_names.txt
@@ -144,6 +144,7 @@ SYNC_REP "Waiting for confirmation from a remote server during synchronous repli
WAL_RECEIVER_EXIT "Waiting for the WAL receiver to exit."
WAL_RECEIVER_WAIT_START "Waiting for startup process to send initial data for streaming replication."
WAL_SUMMARY_READY "Waiting for a new WAL summary to be generated."
+NEXT_MXMEMBERS "Waiting for a next multixact member to be filled."
XACT_GROUP_UPDATE "Waiting for the group leader to update transaction status at end of a parallel operation."
--
2.37.1 (Apple Git-137.1)
[application/octet-stream] 0002-Add-wait-type-for-injection-points.patch (3.7K, ../../[email protected]/3-0002-Add-wait-type-for-injection-points.patch)
download | inline diff:
From 93b4c7dac290d6d2d212522b1a10fef03372315e Mon Sep 17 00:00:00 2001
From: "Andrey M. Borodin" <[email protected]>
Date: Sun, 28 Jan 2024 22:22:22 +0500
Subject: [PATCH 2/3] Add wait type for injection points
---
src/backend/utils/misc/injection_point.c | 20 +++++++++++++++++++
src/include/utils/injection_point.h | 1 +
src/test/modules/injection_points/Makefile | 1 +
.../injection_points/injection_points.c | 16 +++++++++++++++
4 files changed, 38 insertions(+)
diff --git a/src/backend/utils/misc/injection_point.c b/src/backend/utils/misc/injection_point.c
index 0cf4d51cac..398ef2cf30 100644
--- a/src/backend/utils/misc/injection_point.c
+++ b/src/backend/utils/misc/injection_point.c
@@ -252,6 +252,26 @@ InjectionPointDetach(const char *name)
#endif
}
+/*
+ * Test if injection point is attached.
+ */
+bool
+InjectionPointIsAttach(const char *name)
+{
+#ifdef USE_INJECTION_POINTS
+ bool found;
+
+ LWLockAcquire(InjectionPointLock, LW_EXCLUSIVE);
+ hash_search(InjectionPointHash, name, HASH_FIND, &found);
+ LWLockRelease(InjectionPointLock);
+
+ return found;
+
+#else
+ elog(ERROR, "Injection points are not supported by this build");
+#endif
+}
+
/*
* Execute an injection point, if defined.
*
diff --git a/src/include/utils/injection_point.h b/src/include/utils/injection_point.h
index 55524b568f..e07f6b7024 100644
--- a/src/include/utils/injection_point.h
+++ b/src/include/utils/injection_point.h
@@ -33,5 +33,6 @@ extern void InjectionPointAttach(const char *name,
const char *function);
extern void InjectionPointRun(const char *name);
extern void InjectionPointDetach(const char *name);
+extern bool InjectionPointIsAttach(const char *name);
#endif /* INJECTION_POINT_H */
diff --git a/src/test/modules/injection_points/Makefile b/src/test/modules/injection_points/Makefile
index 2cbbae4e0a..543d2ab927 100644
--- a/src/test/modules/injection_points/Makefile
+++ b/src/test/modules/injection_points/Makefile
@@ -7,6 +7,7 @@ DATA = injection_points--1.0.sql
PGFILEDESC = "injection_points - facility for injection points"
REGRESS = injection_points
+TAP_TESTS = 1
ifdef USE_PGXS
PG_CONFIG = pg_config
diff --git a/src/test/modules/injection_points/injection_points.c b/src/test/modules/injection_points/injection_points.c
index e843e6594f..fbb30b15ad 100644
--- a/src/test/modules/injection_points/injection_points.c
+++ b/src/test/modules/injection_points/injection_points.c
@@ -18,6 +18,7 @@
#include "postgres.h"
#include "fmgr.h"
+#include "miscadmin.h"
#include "storage/lwlock.h"
#include "storage/shmem.h"
#include "utils/builtins.h"
@@ -28,6 +29,7 @@ PG_MODULE_MAGIC;
extern PGDLLEXPORT void injection_error(const char *name);
extern PGDLLEXPORT void injection_notice(const char *name);
+extern PGDLLEXPORT void injection_wait(const char *name);
/* Set of callbacks available to be attached to an injection point. */
@@ -43,6 +45,18 @@ injection_notice(const char *name)
elog(NOTICE, "notice triggered for injection point %s", name);
}
+void
+injection_wait(const char *name)
+{
+ elog(NOTICE, "waiting triggered for injection point %s", name);
+ do
+ {
+ CHECK_FOR_INTERRUPTS();
+ pg_usleep(1000L);
+ } while (InjectionPointIsAttach(name));
+ elog(NOTICE, "waiting done for injection point %s", name);
+}
+
/*
* SQL function for creating an injection point.
*/
@@ -58,6 +72,8 @@ injection_points_attach(PG_FUNCTION_ARGS)
function = "injection_error";
else if (strcmp(action, "notice") == 0)
function = "injection_notice";
+ else if (strcmp(action, "wait") == 0)
+ function = "injection_wait";
else
elog(ERROR, "incorrect action \"%s\" for injection point creation", action);
--
2.37.1 (Apple Git-137.1)
[application/octet-stream] 0003-Try-to-test-multixact-CV-sleep.patch (5.3K, ../../[email protected]/4-0003-Try-to-test-multixact-CV-sleep.patch)
download | inline diff:
From a61958407770b6cb864e0b245a4341e033824f46 Mon Sep 17 00:00:00 2001
From: "Andrey M. Borodin" <[email protected]>
Date: Sun, 28 Jan 2024 23:10:55 +0500
Subject: [PATCH 3/3] Try to test multixact CV sleep
---
src/backend/access/transam/multixact.c | 5 ++
.../injection_points--1.0.sql | 11 ++++
.../injection_points/injection_points.c | 26 ++++++++
.../modules/injection_points/t/001_wait.pl | 64 +++++++++++++++++++
4 files changed, 106 insertions(+)
create mode 100644 src/test/modules/injection_points/t/001_wait.pl
diff --git a/src/backend/access/transam/multixact.c b/src/backend/access/transam/multixact.c
index 03fcd25d4c..15b2a0010a 100644
--- a/src/backend/access/transam/multixact.c
+++ b/src/backend/access/transam/multixact.c
@@ -89,6 +89,7 @@
#include "storage/proc.h"
#include "storage/procarray.h"
#include "utils/builtins.h"
+#include "utils/injection_point.h"
#include "utils/memutils.h"
#include "utils/snapmgr.h"
@@ -1200,6 +1201,8 @@ GetNewMultiXactId(int nmembers, MultiXactOffset *offset)
LWLockRelease(MultiXactGenLock);
+ INJECTION_POINT("GetNewMultiXactId-done");
+
debug_elog4(DEBUG2, "GetNew: returning %u offset %u", result, *offset);
return result;
}
@@ -1408,6 +1411,8 @@ retry:
LWLockRelease(MultiXactOffsetSLRULock);
CHECK_FOR_INTERRUPTS();
+ INJECTION_POINT("GetMultiXactIdMembers-CV-sleep");
+
ConditionVariableSleep(&MultiXactState->nextoff_cv,
WAIT_EVENT_NEXT_MXMEMBERS);
ConditionVariableCancelSleep();
diff --git a/src/test/modules/injection_points/injection_points--1.0.sql b/src/test/modules/injection_points/injection_points--1.0.sql
index 5944c41716..d3ebda5964 100644
--- a/src/test/modules/injection_points/injection_points--1.0.sql
+++ b/src/test/modules/injection_points/injection_points--1.0.sql
@@ -33,3 +33,14 @@ CREATE FUNCTION injection_points_detach(IN point_name TEXT)
RETURNS void
AS 'MODULE_PATHNAME', 'injection_points_detach'
LANGUAGE C STRICT PARALLEL UNSAFE;
+
+
+CREATE FUNCTION create_test_multixact()
+RETURNS xid
+AS 'MODULE_PATHNAME', 'create_test_multixact'
+LANGUAGE C STRICT PARALLEL UNSAFE;
+
+CREATE FUNCTION read_test_multixact(xid)
+RETURNS void
+AS 'MODULE_PATHNAME', 'read_test_multixact'
+LANGUAGE C STRICT PARALLEL UNSAFE;
\ No newline at end of file
diff --git a/src/test/modules/injection_points/injection_points.c b/src/test/modules/injection_points/injection_points.c
index fbb30b15ad..d3d5faaa25 100644
--- a/src/test/modules/injection_points/injection_points.c
+++ b/src/test/modules/injection_points/injection_points.c
@@ -109,3 +109,29 @@ injection_points_detach(PG_FUNCTION_ARGS)
PG_RETURN_VOID();
}
+
+#include "access/multixact.h"
+#include "access/xact.h"
+
+PG_FUNCTION_INFO_V1(create_test_multixact);
+Datum
+create_test_multixact(PG_FUNCTION_ARGS)
+{
+ MultiXactId id;
+ MultiXactIdSetOldestMember();
+ id = MultiXactIdCreate(GetCurrentTransactionId(), MultiXactStatusUpdate,
+ GetCurrentTransactionId(), MultiXactStatusForShare);
+ PG_RETURN_TRANSACTIONID(id);
+}
+
+PG_FUNCTION_INFO_V1(read_test_multixact);
+Datum
+read_test_multixact(PG_FUNCTION_ARGS)
+{
+ MultiXactId id = PG_GETARG_TRANSACTIONID(0);
+ MultiXactMember *members;
+ INJECTION_POINT("read_test_multixact");
+ if (GetMultiXactIdMembers(id,&members,false, false) == -1)
+ elog(ERROR, "MultiXactId not found");
+ PG_RETURN_VOID();
+}
\ No newline at end of file
diff --git a/src/test/modules/injection_points/t/001_wait.pl b/src/test/modules/injection_points/t/001_wait.pl
new file mode 100644
index 0000000000..6076ffc6ea
--- /dev/null
+++ b/src/test/modules/injection_points/t/001_wait.pl
@@ -0,0 +1,64 @@
+
+# Copyright (c) 2024, PostgreSQL Global Development Group
+
+use strict;
+use warnings FATAL => 'all';
+
+use PostgreSQL::Test::Cluster;
+use PostgreSQL::Test::Utils;
+
+use Test::More;
+
+my ($node, $result);
+
+$node = PostgreSQL::Test::Cluster->new('injection_points');
+$node->init;
+$node->start;
+$node->safe_psql('postgres', q(CREATE EXTENSION injection_points));
+
+$result = $node->psql('postgres', q(select injection_points_attach('FIRST','wait')));
+is($result, '0', 'wait injection point set');
+
+my $bg = $node->background_psql('postgres');
+
+$bg->query_until(
+ qr/start/, q(
+\echo start
+select injection_points_run('FIRST');
+select injection_points_attach('SECOND','wait');
+));
+
+$result = $node->psql('postgres', q(
+select injection_points_run('SECOND');
+select injection_points_detach('FIRST');
+));
+is($result, '0', 'wait injection point set');
+
+$bg->quit;
+
+$node->safe_psql('postgres', q(select injection_points_attach('read_test_multixact','wait')));
+$node->safe_psql('postgres', q(select injection_points_attach('GetMultiXactIdMembers-CV-sleep','notice')));
+
+my $observer = $node->background_psql('postgres');
+
+$observer->query_safe(
+ q(
+select read_test_multixact(create_test_multixact());
+));
+
+$node->safe_psql('postgres', q(select injection_points_attach('GetNewMultiXactId-done','wait')));
+
+my $creator = $node->background_psql('postgres');
+
+$creator->query_safe( q(select create_test_multixact();));
+
+$node->safe_psql('postgres', q(select injection_points_detach('read_test_multixact')));
+
+$node->safe_psql('postgres', q(select injection_points_detach('GetNewMultiXactId-done')));
+
+$observer->quit;
+
+$creator->quit;
+
+$node->stop;
+done_testing();
--
2.37.1 (Apple Git-137.1)
^ permalink raw reply [nested|flat] 103+ messages in thread
* [PATCH v15 2/8] Row pattern recognition patch (parse/analysis).
@ 2024-03-28 10:30 Tatsuo Ishii <[email protected]>
0 siblings, 0 replies; 103+ messages in thread
From: Tatsuo Ishii @ 2024-03-28 10:30 UTC (permalink / raw)
---
src/backend/parser/parse_agg.c | 7 +
src/backend/parser/parse_clause.c | 296 +++++++++++++++++++++++++++++-
src/backend/parser/parse_expr.c | 4 +
src/backend/parser/parse_func.c | 3 +
4 files changed, 309 insertions(+), 1 deletion(-)
diff --git a/src/backend/parser/parse_agg.c b/src/backend/parser/parse_agg.c
index bee7d8346a..9bc22a836a 100644
--- a/src/backend/parser/parse_agg.c
+++ b/src/backend/parser/parse_agg.c
@@ -577,6 +577,10 @@ check_agglevels_and_constraints(ParseState *pstate, Node *expr)
errkind = true;
break;
+ case EXPR_KIND_RPR_DEFINE:
+ errkind = true;
+ break;
+
/*
* There is intentionally no default: case here, so that the
* compiler will warn if we add a new ParseExprKind without
@@ -967,6 +971,9 @@ transformWindowFuncCall(ParseState *pstate, WindowFunc *wfunc,
case EXPR_KIND_CYCLE_MARK:
errkind = true;
break;
+ case EXPR_KIND_RPR_DEFINE:
+ errkind = true;
+ break;
/*
* There is intentionally no default: case here, so that the
diff --git a/src/backend/parser/parse_clause.c b/src/backend/parser/parse_clause.c
index d2ac86777c..de2e5791e3 100644
--- a/src/backend/parser/parse_clause.c
+++ b/src/backend/parser/parse_clause.c
@@ -98,7 +98,14 @@ static WindowClause *findWindowClause(List *wclist, const char *name);
static Node *transformFrameOffset(ParseState *pstate, int frameOptions,
Oid rangeopfamily, Oid rangeopcintype, Oid *inRangeFunc,
Node *clause);
-
+static void transformRPR(ParseState *pstate, WindowClause *wc, WindowDef *windef,
+ List **targetlist);
+static List *transformDefineClause(ParseState *pstate, WindowClause *wc, WindowDef *windef,
+ List **targetlist);
+static void transformPatternClause(ParseState *pstate, WindowClause *wc,
+ WindowDef *windef);
+static List *transformMeasureClause(ParseState *pstate, WindowClause *wc,
+ WindowDef *windef);
/*
* transformFromClause -
@@ -2948,6 +2955,10 @@ transformWindowDefinitions(ParseState *pstate,
rangeopfamily, rangeopcintype,
&wc->endInRangeFunc,
windef->endOffset);
+
+ /* Process Row Pattern Recognition related clauses */
+ transformRPR(pstate, wc, windef, targetlist);
+
wc->runCondition = NIL;
wc->winref = winref;
@@ -3813,3 +3824,286 @@ transformFrameOffset(ParseState *pstate, int frameOptions,
return node;
}
+
+/*
+ * transformRPR
+ * Process Row Pattern Recognition related clauses
+ */
+static void
+transformRPR(ParseState *pstate, WindowClause *wc, WindowDef *windef,
+ List **targetlist)
+{
+ /*
+ * Window definition exists?
+ */
+ if (windef == NULL)
+ return;
+
+ /*
+ * Row Pattern Common Syntax clause exists?
+ */
+ if (windef->rpCommonSyntax == NULL)
+ return;
+
+ /* Check Frame option. Frame must start at current row */
+ if ((wc->frameOptions & FRAMEOPTION_START_CURRENT_ROW) == 0)
+ ereport(ERROR,
+ (errcode(ERRCODE_SYNTAX_ERROR),
+ errmsg("FRAME must start at current row when row patttern recognition is used")));
+
+ /* Transform AFTER MACH SKIP TO clause */
+ wc->rpSkipTo = windef->rpCommonSyntax->rpSkipTo;
+
+ /* Transform AFTER MACH SKIP TO variable */
+ wc->rpSkipVariable = windef->rpCommonSyntax->rpSkipVariable;
+
+ /* Transform SEEK or INITIAL clause */
+ wc->initial = windef->rpCommonSyntax->initial;
+
+ /* Transform DEFINE clause into list of TargetEntry's */
+ wc->defineClause = transformDefineClause(pstate, wc, windef, targetlist);
+
+ /* Check PATTERN clause and copy to patternClause */
+ transformPatternClause(pstate, wc, windef);
+
+ /* Transform MEASURE clause */
+ transformMeasureClause(pstate, wc, windef);
+}
+
+/*
+ * transformDefineClause Process DEFINE clause and transform ResTarget into
+ * list of TargetEntry.
+ *
+ * XXX we only support column reference in row pattern definition search
+ * condition, e.g. "price". <row pattern definition variable name>.<column
+ * reference> is not supported, e.g. "A.price".
+ */
+static List *
+transformDefineClause(ParseState *pstate, WindowClause *wc, WindowDef *windef,
+ List **targetlist)
+{
+ /* DEFINE variable name initials */
+ static char *defineVariableInitials = "abcdefghijklmnopqrstuvwxyz";
+
+ ListCell *lc,
+ *l;
+ ResTarget *restarget,
+ *r;
+ List *restargets;
+ List *defineClause;
+ char *name;
+ int initialLen;
+ int i;
+
+ /*
+ * If Row Definition Common Syntax exists, DEFINE clause must exist. (the
+ * raw parser should have already checked it.)
+ */
+ Assert(windef->rpCommonSyntax->rpDefs != NULL);
+
+ /*
+ * Check and add "A AS A IS TRUE" if pattern variable is missing in DEFINE
+ * per the SQL standard.
+ */
+ restargets = NIL;
+ foreach(lc, windef->rpCommonSyntax->rpPatterns)
+ {
+ A_Expr *a;
+ bool found = false;
+
+ if (!IsA(lfirst(lc), A_Expr))
+ ereport(ERROR,
+ errmsg("node type is not A_Expr"));
+
+ a = (A_Expr *) lfirst(lc);
+ name = strVal(a->lexpr);
+
+ foreach(l, windef->rpCommonSyntax->rpDefs)
+ {
+ restarget = (ResTarget *) lfirst(l);
+
+ if (!strcmp(restarget->name, name))
+ {
+ found = true;
+ break;
+ }
+ }
+
+ if (!found)
+ {
+ /*
+ * "name" is missing. So create "name AS name IS TRUE" ResTarget
+ * node and add it to the temporary list.
+ */
+ A_Const *n;
+
+ restarget = makeNode(ResTarget);
+ n = makeNode(A_Const);
+ n->val.boolval.type = T_Boolean;
+ n->val.boolval.boolval = true;
+ n->location = -1;
+ restarget->name = pstrdup(name);
+ restarget->indirection = NIL;
+ restarget->val = (Node *) n;
+ restarget->location = -1;
+ restargets = lappend((List *) restargets, restarget);
+ }
+ }
+
+ if (list_length(restargets) >= 1)
+ {
+ /* add missing DEFINEs */
+ windef->rpCommonSyntax->rpDefs =
+ list_concat(windef->rpCommonSyntax->rpDefs, restargets);
+ list_free(restargets);
+ }
+
+ /*
+ * Check for duplicate row pattern definition variables. The standard
+ * requires that no two row pattern definition variable names shall be
+ * equivalent.
+ */
+ restargets = NIL;
+ foreach(lc, windef->rpCommonSyntax->rpDefs)
+ {
+ restarget = (ResTarget *) lfirst(lc);
+ name = restarget->name;
+
+ /*
+ * Add DEFINE expression (Restarget->val) to the targetlist as a
+ * TargetEntry if it does not exist yet. Planner will add the column
+ * ref var node to the outer plan's target list later on. This makes
+ * DEFINE expression could access the outer tuple while evaluating
+ * PATTERN.
+ *
+ * XXX: adding whole expressions of DEFINE to the plan.targetlist is
+ * not so good, because it's not necessary to evalute the expression
+ * in the target list while running the plan. We should extract the
+ * var nodes only then add them to the plan.targetlist.
+ */
+ findTargetlistEntrySQL99(pstate, (Node *) restarget->val,
+ targetlist, EXPR_KIND_RPR_DEFINE);
+
+ /*
+ * Make sure that the row pattern definition search condition is a
+ * boolean expression.
+ */
+ transformWhereClause(pstate, restarget->val,
+ EXPR_KIND_RPR_DEFINE, "DEFINE");
+
+ foreach(l, restargets)
+ {
+ char *n;
+
+ r = (ResTarget *) lfirst(l);
+ n = r->name;
+
+ if (!strcmp(n, name))
+ ereport(ERROR,
+ (errcode(ERRCODE_SYNTAX_ERROR),
+ errmsg("row pattern definition variable name \"%s\" appears more than once in DEFINE clause",
+ name),
+ parser_errposition(pstate, exprLocation((Node *) r))));
+ }
+ restargets = lappend(restargets, restarget);
+ }
+ list_free(restargets);
+
+ /*
+ * Create list of row pattern DEFINE variable name's initial. We assign
+ * [a-z] to them (up to 26 variable names are allowed).
+ */
+ restargets = NIL;
+ i = 0;
+ initialLen = strlen(defineVariableInitials);
+
+ foreach(lc, windef->rpCommonSyntax->rpDefs)
+ {
+ char initial[2];
+
+ restarget = (ResTarget *) lfirst(lc);
+ name = restarget->name;
+
+ if (i >= initialLen)
+ {
+ ereport(ERROR,
+ (errcode(ERRCODE_SYNTAX_ERROR),
+ errmsg("number of row pattern definition variable names exceeds %d",
+ initialLen),
+ parser_errposition(pstate,
+ exprLocation((Node *) restarget))));
+ }
+ initial[0] = defineVariableInitials[i++];
+ initial[1] = '\0';
+ wc->defineInitial = lappend(wc->defineInitial,
+ makeString(pstrdup(initial)));
+ }
+
+ defineClause = transformTargetList(pstate, windef->rpCommonSyntax->rpDefs,
+ EXPR_KIND_RPR_DEFINE);
+
+ /* mark column origins */
+ markTargetListOrigins(pstate, defineClause);
+
+ /* mark all nodes in the DEFINE clause tree with collation information */
+ assign_expr_collations(pstate, (Node *) defineClause);
+
+ return defineClause;
+}
+
+/*
+ * transformPatternClause
+ * Process PATTERN clause and return PATTERN clause in the raw parse tree
+ */
+static void
+transformPatternClause(ParseState *pstate, WindowClause *wc,
+ WindowDef *windef)
+{
+ ListCell *lc;
+
+ /*
+ * Row Pattern Common Syntax clause exists?
+ */
+ if (windef->rpCommonSyntax == NULL)
+ return;
+
+ wc->patternVariable = NIL;
+ wc->patternRegexp = NIL;
+ foreach(lc, windef->rpCommonSyntax->rpPatterns)
+ {
+ A_Expr *a;
+ char *name;
+ char *regexp;
+
+ if (!IsA(lfirst(lc), A_Expr))
+ ereport(ERROR,
+ errmsg("node type is not A_Expr"));
+
+ a = (A_Expr *) lfirst(lc);
+ name = strVal(a->lexpr);
+
+ wc->patternVariable = lappend(wc->patternVariable, makeString(pstrdup(name)));
+ regexp = strVal(lfirst(list_head(a->name)));
+
+ wc->patternRegexp = lappend(wc->patternRegexp, makeString(pstrdup(regexp)));
+ }
+}
+
+/*
+ * transformMeasureClause
+ * Process MEASURE clause
+ * XXX MEASURE clause is not supported yet
+ */
+static List *
+transformMeasureClause(ParseState *pstate, WindowClause *wc,
+ WindowDef *windef)
+{
+ if (windef->rowPatternMeasures == NIL)
+ return NIL;
+
+ ereport(ERROR,
+ (errcode(ERRCODE_SYNTAX_ERROR),
+ errmsg("%s", "MEASURE clause is not supported yet"),
+ parser_errposition(pstate, exprLocation((Node *) windef->rowPatternMeasures))));
+ return NIL;
+}
diff --git a/src/backend/parser/parse_expr.c b/src/backend/parser/parse_expr.c
index 73c83cea4a..8b0cc608bc 100644
--- a/src/backend/parser/parse_expr.c
+++ b/src/backend/parser/parse_expr.c
@@ -578,6 +578,7 @@ transformColumnRef(ParseState *pstate, ColumnRef *cref)
case EXPR_KIND_COPY_WHERE:
case EXPR_KIND_GENERATED_COLUMN:
case EXPR_KIND_CYCLE_MARK:
+ case EXPR_KIND_RPR_DEFINE:
/* okay */
break;
@@ -1817,6 +1818,7 @@ transformSubLink(ParseState *pstate, SubLink *sublink)
case EXPR_KIND_VALUES:
case EXPR_KIND_VALUES_SINGLE:
case EXPR_KIND_CYCLE_MARK:
+ case EXPR_KIND_RPR_DEFINE:
/* okay */
break;
case EXPR_KIND_CHECK_CONSTRAINT:
@@ -3197,6 +3199,8 @@ ParseExprKindName(ParseExprKind exprKind)
return "GENERATED AS";
case EXPR_KIND_CYCLE_MARK:
return "CYCLE";
+ case EXPR_KIND_RPR_DEFINE:
+ return "DEFINE";
/*
* There is intentionally no default: case here, so that the
diff --git a/src/backend/parser/parse_func.c b/src/backend/parser/parse_func.c
index 0cbc950c95..ad982a7c17 100644
--- a/src/backend/parser/parse_func.c
+++ b/src/backend/parser/parse_func.c
@@ -2657,6 +2657,9 @@ check_srf_call_placement(ParseState *pstate, Node *last_srf, int location)
case EXPR_KIND_CYCLE_MARK:
errkind = true;
break;
+ case EXPR_KIND_RPR_DEFINE:
+ errkind = true;
+ break;
/*
* There is intentionally no default: case here, so that the
--
2.25.1
----Next_Part(Thu_Mar_28_19_59_25_2024_076)--
Content-Type: Text/X-Patch; charset=us-ascii
Content-Transfer-Encoding: 7bit
Content-Disposition: inline;
filename="v15-0003-Row-pattern-recognition-patch-rewriter.patch"
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: Optimize numeric multiplication for one and two base-NBASE digit multiplicands.
@ 2024-07-03 11:17 Dean Rasheed <[email protected]>
0 siblings, 3 replies; 103+ messages in thread
From: Dean Rasheed @ 2024-07-03 11:17 UTC (permalink / raw)
To: Joel Jacobson <[email protected]>; +Cc: Dagfinn Ilmari Mannsåker <[email protected]>; pgsql-hackers
On Tue, 2 Jul 2024 at 21:10, Joel Jacobson <[email protected]> wrote:
>
> I found the bug in the case 3 code,
> and it turns out the same type of bug also exists in the case 2 code:
>
> case 2:
> newdig = (int) var1digits[1] * var2digits[res_ndigits - 4];
>
> The problem here is that res_ndigits could become less than 4,
Yes. It can't be less than 3 though (per an earlier test), so the case
2 code was correct.
I've been hacking on this a bit and trying to tidy it up. Firstly, I
moved it to a separate function, because it was starting to look messy
having so much extra code in mul_var(). Then I added a bunch more
comments to explain what's going on, and the limits of the various
variables. Note that most of the boundary checks are actually
unnecessary -- in particular all the ones in or after the main loop,
provided you pull out the first 2 result digits from the main loop in
the 3-digit case. That does seem to work very well, but...
I wasn't entirely happy with how messy that code is getting, so I
tried a different approach. Similar to div_var_int(), I tried writing
a mul_var_int() function instead. This can be used for 1 and 2 digit
factors, and we could add a similar mul_var_int64() function on
platforms with 128-bit integers. The code looks quite a lot neater, so
it's probably less likely to contain bugs (though I have just written
it in a hurry,so it might still have bugs). In testing, it seemed to
give a decent speedup, but perhaps a little less than before. But
that's to be balanced against having more maintainable code, and also
a function that might be useful elsewhere in numeric.c.
Anyway, here are both patches for comparison. I'll stop hacking for a
while and let you see what you make of these.
Regards,
Dean
Attachments:
[text/x-patch] v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch (7.0K, ../../CAEZATCUA+GBYvzbqSr7LPDDfmr24jQE1R0iyYfjFk82xeuBCRw@mail.gmail.com/2-v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch)
download | inline diff:
diff --git a/src/backend/utils/adt/numeric.c b/src/backend/utils/adt/numeric.c
new file mode 100644
index 5510a20..81600b3
--- a/src/backend/utils/adt/numeric.c
+++ b/src/backend/utils/adt/numeric.c
@@ -551,6 +551,8 @@ static void sub_var(const NumericVar *va
static void mul_var(const NumericVar *var1, const NumericVar *var2,
NumericVar *result,
int rscale);
+static void mul_var_small(const NumericVar *var1, const NumericVar *var2,
+ NumericVar *result, int rscale);
static void div_var(const NumericVar *var1, const NumericVar *var2,
NumericVar *result,
int rscale, bool round);
@@ -8707,7 +8709,7 @@ mul_var(const NumericVar *var1, const Nu
var1digits = var1->digits;
var2digits = var2->digits;
- if (var1ndigits == 0 || var2ndigits == 0)
+ if (var1ndigits == 0)
{
/* one or both inputs is zero; so is result */
zero_var(result);
@@ -8715,6 +8717,16 @@ mul_var(const NumericVar *var1, const Nu
return;
}
+ /*
+ * If var1 has 3 digits or fewer, delegate to mul_var_small() which uses a
+ * faster short multiplication algorithm.
+ */
+ if (var1ndigits <= 3)
+ {
+ mul_var_small(var1, var2, result, rscale);
+ return;
+ }
+
/* Determine result sign and (maximum possible) weight */
if (var1->sign == var2->sign)
res_sign = NUMERIC_POS;
@@ -8858,6 +8870,188 @@ mul_var(const NumericVar *var1, const Nu
result->sign = res_sign;
/* Round to target rscale (and set result->dscale) */
+ round_var(result, rscale);
+
+ /* Strip leading and trailing zeroes */
+ strip_var(result);
+}
+
+
+/*
+ * mul_var_small() -
+ *
+ * This has the same API as mul_var, but it assumes that var1 has no more
+ * than 3 digits and var2 has at least as many digits as var1. For variables
+ * satisfying these conditions, the product can be computed more quickly than
+ * the general algorithm used in mul_var.
+ */
+static void
+mul_var_small(const NumericVar *var1, const NumericVar *var2,
+ NumericVar *result, int rscale)
+{
+ int var1ndigits = var1->ndigits;
+ int var2ndigits = var2->ndigits;
+ NumericDigit *var1digits = var1->digits;
+ NumericDigit *var2digits = var2->digits;
+ int res_sign;
+ int res_weight;
+ int res_ndigits;
+ int maxdigits;
+ NumericDigit *res_buf;
+ NumericDigit *res_digits;
+ int carry;
+ int term;
+
+ /* Check preconditions */
+ Assert(var1ndigits <= 3);
+ Assert(var2ndigits >= var1ndigits);
+
+ /* Determine result sign and (maximum possible) weight */
+ if (var1->sign == var2->sign)
+ res_sign = NUMERIC_POS;
+ else
+ res_sign = NUMERIC_NEG;
+ res_weight = var1->weight + var2->weight + 2;
+
+ /* Determine the number of result digits to compute - see mul_var() */
+ res_ndigits = var1ndigits + var2ndigits + 1;
+ maxdigits = res_weight + 1 + (rscale + DEC_DIGITS - 1) / DEC_DIGITS +
+ MUL_GUARD_DIGITS;
+ res_ndigits = Min(res_ndigits, maxdigits);
+
+ if (res_ndigits < 3)
+ {
+ /* All input digits will be ignored; so result is zero */
+ zero_var(result);
+ result->dscale = rscale;
+ return;
+ }
+
+ /* Allocate result digit array */
+ res_buf = digitbuf_alloc(res_ndigits);
+ res_buf[0] = 0; /* spare digit for later rounding */
+ res_digits = res_buf + 1;
+
+ /*
+ * Compute the result digits in reverse, in one pass, propagating the
+ * carry up as we go.
+ *
+ * This computes res_digits[res_ndigits - 2], ... res_digits[0] by summing
+ * the products var1digits[i1] * var2digits[i2] for which i1 + i2 + 1 is
+ * the result index.
+ */
+ switch (var1ndigits)
+ {
+ case 1:
+ /* ---------
+ * 1-digit case:
+ * var1ndigits = 1
+ * var2ndigits >= 1
+ * 3 <= res_ndigits <= var2ndigits + 2
+ * ----------
+ */
+ carry = 0;
+ for (int i = res_ndigits - 3; i >= 0; i--)
+ {
+ term = (int) var1digits[0] * var2digits[i] + carry;
+ res_digits[i + 1] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+ }
+ res_digits[0] = (NumericDigit) carry;
+ break;
+
+ case 2:
+ /* ---------
+ * 2-digit case:
+ * var1ndigits = 2
+ * var2ndigits >= 2
+ * 3 <= res_ndigits <= var2ndigits + 3
+ * ----------
+ */
+ /* last result digit and carry */
+ term = 0;
+ if (res_ndigits - 3 < var2ndigits)
+ term += (int) var1digits[0] * var2digits[res_ndigits - 3];
+ if (res_ndigits > 3)
+ term += (int) var1digits[1] * var2digits[res_ndigits - 4];
+ res_digits[res_ndigits - 2] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+
+ /* remaining digits, except for the first two */
+ for (int i = res_ndigits - 4; i >= 1; i--)
+ {
+ term = (int) var1digits[0] * var2digits[i] +
+ (int) var1digits[1] * var2digits[i - 1] + carry;
+ res_digits[i + 1] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+ }
+
+ /* first two digits */
+ term = (int) var1digits[0] * var2digits[0] + carry;
+ res_digits[1] = (NumericDigit) (term % NBASE);
+ res_digits[0] = (NumericDigit) (term / NBASE);
+ break;
+
+ case 3:
+ /* ---------
+ * 3-digit case:
+ * var1ndigits = 3
+ * var2ndigits >= 3
+ * 3 <= res_ndigits <= var2ndigits + 4
+ * ----------
+ */
+ /* last result digit and carry */
+ term = 0;
+ if (res_ndigits - 3 < var2ndigits)
+ term += (int) var1digits[0] * var2digits[res_ndigits - 3];
+ if (res_ndigits > 3 && res_ndigits - 4 < var2ndigits)
+ term += (int) var1digits[1] * var2digits[res_ndigits - 4];
+ if (res_ndigits > 4)
+ term += (int) var1digits[2] * var2digits[res_ndigits - 5];
+ res_digits[res_ndigits - 2] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+
+ /* penultimate result digit */
+ term = carry;
+ if (res_ndigits > 3 && res_ndigits - 4 < var2ndigits)
+ term += (int) var1digits[0] * var2digits[res_ndigits - 4];
+ if (res_ndigits > 4)
+ term += (int) var1digits[1] * var2digits[res_ndigits - 5];
+ if (res_ndigits > 5)
+ term += (int) var1digits[2] * var2digits[res_ndigits - 6];
+ res_digits[res_ndigits - 3] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+
+ /* remaining digits, except for the first three */
+ for (int i = res_ndigits - 5; i >= 2; i--)
+ {
+ term = (int) var1digits[0] * var2digits[i] +
+ (int) var1digits[1] * var2digits[i - 1] +
+ (int) var1digits[2] * var2digits[i - 2] + carry;
+ res_digits[i + 1] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+ }
+
+ /* first three digits */
+ term = (int) var1digits[0] * var2digits[1] +
+ (int) var1digits[1] * var2digits[0] + carry;
+ res_digits[2] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+ term = (int) var1digits[0] * var2digits[0] + carry;
+ res_digits[1] = (NumericDigit) (term % NBASE);
+ res_digits[0] = (NumericDigit) (term / NBASE);
+ break;
+ }
+
+ /* Store the product in result (minus extra rounding digit) */
+ digitbuf_free(result->buf);
+ result->ndigits = res_ndigits - 1;
+ result->buf = res_buf;
+ result->digits = res_digits;
+ result->weight = res_weight - 1;
+ result->sign = res_sign;
+
+ /* Round to target rscale (and set result->dscale) */
round_var(result, rscale);
/* Strip leading and trailing zeroes */
[text/x-patch] v5-add-mul_var_int.patch (5.1K, ../../CAEZATCUA+GBYvzbqSr7LPDDfmr24jQE1R0iyYfjFk82xeuBCRw@mail.gmail.com/3-v5-add-mul_var_int.patch)
download | inline diff:
diff --git a/src/backend/utils/adt/numeric.c b/src/backend/utils/adt/numeric.c
new file mode 100644
index 5510a20..9e50ea7
--- a/src/backend/utils/adt/numeric.c
+++ b/src/backend/utils/adt/numeric.c
@@ -551,6 +551,8 @@ static void sub_var(const NumericVar *va
static void mul_var(const NumericVar *var1, const NumericVar *var2,
NumericVar *result,
int rscale);
+static void mul_var_int(const NumericVar *var, int ival, int ival_weight,
+ NumericVar *result, int rscale);
static void div_var(const NumericVar *var1, const NumericVar *var2,
NumericVar *result,
int rscale, bool round);
@@ -8707,7 +8709,7 @@ mul_var(const NumericVar *var1, const Nu
var1digits = var1->digits;
var2digits = var2->digits;
- if (var1ndigits == 0 || var2ndigits == 0)
+ if (var1ndigits == 0)
{
/* one or both inputs is zero; so is result */
zero_var(result);
@@ -8715,6 +8717,31 @@ mul_var(const NumericVar *var1, const Nu
return;
}
+ /*
+ * If var1 has just one or two digits, delegate to mul_var_int(), which
+ * uses a faster direct multiplication algorithm.
+ *
+ * TODO: Similarly, on platforms with 128-bit integers ...
+ */
+ if (var1ndigits <= 2)
+ {
+ int ifactor;
+ int ifactor_weight;
+
+ ifactor = var1->digits[0];
+ ifactor_weight = var1->weight;
+ if (var1ndigits == 2)
+ {
+ ifactor = ifactor * NBASE + var1->digits[1];
+ ifactor_weight--;
+ }
+ if (var1->sign == NUMERIC_NEG)
+ ifactor = -ifactor;
+
+ mul_var_int(var2, ifactor, ifactor_weight, result, rscale);
+ return;
+ }
+
/* Determine result sign and (maximum possible) weight */
if (var1->sign == var2->sign)
res_sign = NUMERIC_POS;
@@ -8857,6 +8884,123 @@ mul_var(const NumericVar *var1, const Nu
result->weight = res_weight;
result->sign = res_sign;
+ /* Round to target rscale (and set result->dscale) */
+ round_var(result, rscale);
+
+ /* Strip leading and trailing zeroes */
+ strip_var(result);
+}
+
+
+/*
+ * mul_var_int() -
+ *
+ * Multiply a numeric variable by a 32-bit integer with the specified weight.
+ * The product var * ival * NBASE^ival_weight is stored in result.
+ */
+static void
+mul_var_int(const NumericVar *var, int ival, int ival_weight,
+ NumericVar *result, int rscale)
+{
+ NumericDigit *var_digits = var->digits;
+ int var_ndigits = var->ndigits;
+ int res_sign;
+ int res_weight;
+ int res_ndigits;
+ int maxdigits;
+ NumericDigit *res_buf;
+ NumericDigit *res_digits;
+ uint32 factor;
+ uint32 carry;
+
+ if (ival == 0 || var_ndigits == 0)
+ {
+ zero_var(result);
+ result->dscale = rscale;
+ return;
+ }
+
+ /*
+ * Determine the result sign, (maximum possible) weight and number of
+ * digits to calculate. The weight figured here is correct if the emitted
+ * product has no leading zero digits; otherwise strip_var() will fix
+ * things up.
+ */
+ if (var->sign == NUMERIC_POS)
+ res_sign = ival > 0 ? NUMERIC_POS : NUMERIC_NEG;
+ else
+ res_sign = ival > 0 ? NUMERIC_NEG : NUMERIC_POS;
+ res_weight = var->weight + ival_weight + 3;
+ /* The number of accurate result digits we need to produce: */
+ res_ndigits = var_ndigits + 3;
+ maxdigits = res_weight + 1 + (rscale + DEC_DIGITS - 1) / DEC_DIGITS +
+ MUL_GUARD_DIGITS;
+ res_ndigits = Min(res_ndigits, maxdigits);
+
+ if (res_ndigits < 3)
+ {
+ /* All input digits will be ignored; so result is zero */
+ zero_var(result);
+ result->dscale = rscale;
+ return;
+ }
+
+ res_buf = digitbuf_alloc(res_ndigits + 1);
+ res_buf[0] = 0; /* spare digit for later rounding */
+ res_digits = res_buf + 1;
+
+ /*
+ * Now compute the product digits by procssing the input digits in reverse
+ * and propagating the carry up as we go.
+ *
+ * In this algorithm, the carry from one digit to the next is at most
+ * factor - 1, and product is at most factor * NBASE - 1, and so it needs
+ * to be a 64-bit integer if this exceeds UINT_MAX.
+ */
+ factor = abs(ival);
+ carry = 0;
+
+ if (factor <= UINT_MAX / NBASE)
+ {
+ /* product cannot overflow 32 bits */
+ uint32 product;
+
+ for (int i = res_ndigits - 4; i >= 0; i--)
+ {
+ product = factor * var_digits[i] + carry;
+ res_digits[i + 3] = (NumericDigit) (product % NBASE);
+ carry = product / NBASE;
+ }
+ res_digits[2] = (NumericDigit) (carry % NBASE);
+ carry = carry / NBASE;
+ res_digits[1] = (NumericDigit) (carry % NBASE);
+ res_digits[0] = (NumericDigit) (carry / NBASE);
+ }
+ else
+ {
+ /* product may exceed 32 bits */
+ uint64 product;
+
+ for (int i = res_ndigits - 4; i >= 0; i--)
+ {
+ product = (uint64) factor * var_digits[i] + carry;
+ res_digits[i + 3] = (NumericDigit) (product % NBASE);
+ carry = (uint32) (product / NBASE);
+ }
+ res_digits[2] = (NumericDigit) (carry % NBASE);
+ carry = carry / NBASE;
+ res_digits[1] = (NumericDigit) (carry % NBASE);
+ res_digits[0] = (NumericDigit) (carry / NBASE);
+ }
+
+ /* Store the product in result */
+ digitbuf_free(result->buf);
+ result->ndigits = res_ndigits;
+ result->buf = res_buf;
+ result->digits = res_digits;
+ result->weight = res_weight;
+ result->sign = res_sign;
+
/* Round to target rscale (and set result->dscale) */
round_var(result, rscale);
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: Optimize numeric multiplication for one and two base-NBASE digit multiplicands.
@ 2024-07-03 11:43 Ranier Vilela <[email protected]>
parent: Dean Rasheed <[email protected]>
2 siblings, 0 replies; 103+ messages in thread
From: Ranier Vilela @ 2024-07-03 11:43 UTC (permalink / raw)
To: Dean Rasheed <[email protected]>; +Cc: Joel Jacobson <[email protected]>; Dagfinn Ilmari Mannsåker <[email protected]>; pgsql-hackers
Em qua., 3 de jul. de 2024 às 08:18, Dean Rasheed <[email protected]>
escreveu:
> On Tue, 2 Jul 2024 at 21:10, Joel Jacobson <[email protected]> wrote:
> >
> > I found the bug in the case 3 code,
> > and it turns out the same type of bug also exists in the case 2 code:
> >
> > case 2:
> > newdig = (int) var1digits[1] *
> var2digits[res_ndigits - 4];
> >
> > The problem here is that res_ndigits could become less than 4,
>
> Yes. It can't be less than 3 though (per an earlier test), so the case
> 2 code was correct.
>
> I've been hacking on this a bit and trying to tidy it up. Firstly, I
> moved it to a separate function, because it was starting to look messy
> having so much extra code in mul_var(). Then I added a bunch more
> comments to explain what's going on, and the limits of the various
> variables. Note that most of the boundary checks are actually
> unnecessary -- in particular all the ones in or after the main loop,
> provided you pull out the first 2 result digits from the main loop in
> the 3-digit case. That does seem to work very well, but...
>
> I wasn't entirely happy with how messy that code is getting, so I
> tried a different approach. Similar to div_var_int(), I tried writing
> a mul_var_int() function instead. This can be used for 1 and 2 digit
> factors, and we could add a similar mul_var_int64() function on
> platforms with 128-bit integers. The code looks quite a lot neater, so
> it's probably less likely to contain bugs (though I have just written
> it in a hurry,so it might still have bugs). In testing, it seemed to
> give a decent speedup, but perhaps a little less than before. But
> that's to be balanced against having more maintainable code, and also
> a function that might be useful elsewhere in numeric.c.
>
> Anyway, here are both patches for comparison. I'll stop hacking for a
> while and let you see what you make of these.
>
I liked v5-add-mul_var_int.patch better.
I think that *var_digits can be const too.
+ const NumericDigit *var_digits = var->digits;
Typo In the comments:
- by procssing
+ by processing
best regards,
Ranier Vilela
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: Optimize numeric multiplication for one and two base-NBASE digit multiplicands.
@ 2024-07-03 13:48 Joel Jacobson <[email protected]>
parent: Dean Rasheed <[email protected]>
2 siblings, 2 replies; 103+ messages in thread
From: Joel Jacobson @ 2024-07-03 13:48 UTC (permalink / raw)
To: Dean Rasheed <[email protected]>; +Cc: Dagfinn Ilmari Mannsåker <[email protected]>; pgsql-hackers
On Wed, Jul 3, 2024, at 13:17, Dean Rasheed wrote:
> On Tue, 2 Jul 2024 at 21:10, Joel Jacobson <[email protected]> wrote:
>>
>> I found the bug in the case 3 code,
>> and it turns out the same type of bug also exists in the case 2 code:
>>
>> case 2:
>> newdig = (int) var1digits[1] * var2digits[res_ndigits - 4];
>>
>> The problem here is that res_ndigits could become less than 4,
>
> Yes. It can't be less than 3 though (per an earlier test), so the case
> 2 code was correct.
Hmm, I don't see how the case 2 code can be correct?
If, like you say, res_ndigits can't be less than 3, that means it can be 3, right?
And if res_ndigits=3 then `var2digits[res_ndigits - 4]` would try to access `var2digits[-1]`.
> I've been hacking on this a bit and trying to tidy it up. Firstly, I
> moved it to a separate function, because it was starting to look messy
> having so much extra code in mul_var(). Then I added a bunch more
> comments to explain what's going on, and the limits of the various
> variables. Note that most of the boundary checks are actually
> unnecessary -- in particular all the ones in or after the main loop,
> provided you pull out the first 2 result digits from the main loop in
> the 3-digit case. That does seem to work very well, but...
Nice, I was starting to feel a bit uncomfortable with the level of increased complexity.
> I wasn't entirely happy with how messy that code is getting, so I
> tried a different approach. Similar to div_var_int(), I tried writing
> a mul_var_int() function instead. This can be used for 1 and 2 digit
> factors, and we could add a similar mul_var_int64() function on
> platforms with 128-bit integers. The code looks quite a lot neater, so
> it's probably less likely to contain bugs (though I have just written
> it in a hurry,so it might still have bugs). In testing, it seemed to
> give a decent speedup, but perhaps a little less than before. But
> that's to be balanced against having more maintainable code, and also
> a function that might be useful elsewhere in numeric.c.
>
> Anyway, here are both patches for comparison. I'll stop hacking for a
> while and let you see what you make of these.
I've tested both patches, and they produces the same output given the
same input as HEAD, when rscale is unmodified (full precision).
However, for a reduced rscale, there are some differences:
mul_var_small() seems more resilient to rscale reductions than mul_var_int().
The previous version we worked on, I've called "mul_var inlined" in the output below.
```
CREATE TABLE test_numeric_mul_patched (
var1 numeric,
var2 numeric,
rscale_adjustment int,
result numeric
);
DO $$
DECLARE
var1 numeric;
var2 numeric;
BEGIN
FOR i IN 1..1000 LOOP
RAISE NOTICE '%', i;
FOR var1ndigits IN 1..4 LOOP
FOR var2ndigits IN 1..4 LOOP
FOR var1dscale IN 0..(var1ndigits*4) LOOP
FOR var2dscale IN 0..(var2ndigits*4) LOOP
FOR rscale_adjustment IN 0..(var1dscale+var2dscale) LOOP
var1 := round(random(
format('1%s',repeat('0',(var1ndigits-1)*4-1))::numeric,
format('%s',repeat('9',var1ndigits*4))::numeric
) / 10::numeric^var1dscale, var1dscale);
var2 := round(random(
format('1%s',repeat('0',(var2ndigits-1)*4-1))::numeric,
format('%s',repeat('9',var2ndigits*4))::numeric
) / 10::numeric^var2dscale, var2dscale);
INSERT INTO test_numeric_mul_patched
(var1, var2, rscale_adjustment)
VALUES
(var1, var2, -rscale_adjustment);
END LOOP;
END LOOP;
END LOOP;
END LOOP;
END LOOP;
END LOOP;
END $$;
UPDATE test_numeric_mul_patched SET result = numeric_mul_head(var1, var2, rscale_adjustment);
SELECT
rscale_adjustment,
COUNT(*),
COUNT(*) FILTER (WHERE result IS DISTINCT FROM numeric_mul_patch_int(var1,var2,rscale_adjustment)) AS "mul_var_int",
COUNT(*) FILTER (WHERE result IS DISTINCT FROM numeric_mul_patch_small(var1,var2,rscale_adjustment)) AS "mul_var_small",
COUNT(*) FILTER (WHERE result IS DISTINCT FROM numeric_mul_patch_inline(var1,var2,rscale_adjustment)) AS "mul_var inlined"
FROM test_numeric_mul_patched
GROUP BY 1
ORDER BY 1;
rscale_adjustment | count | mul_var_int | mul_var_small | mul_var inlined
-------------------+---------+-------------+---------------+-----------------
-32 | 1000 | 0 | 0 | 0
-31 | 3000 | 0 | 0 | 0
-30 | 6000 | 0 | 0 | 0
-29 | 10000 | 0 | 0 | 0
-28 | 17000 | 0 | 0 | 0
-27 | 27000 | 0 | 0 | 0
-26 | 40000 | 0 | 1 | 0
-25 | 56000 | 1 | 11 | 0
-24 | 78000 | 316 | 119 | 1
-23 | 106000 | 498 | 1696 | 0
-22 | 140000 | 531 | 2480 | 1
-21 | 180000 | 591 | 3145 | 0
-20 | 230000 | 1956 | 5309 | 1
-19 | 290000 | 2189 | 5032 | 0
-18 | 360000 | 2314 | 4868 | 0
-17 | 440000 | 2503 | 4544 | 1
-16 | 533000 | 5201 | 3633 | 0
-15 | 631000 | 5621 | 3006 | 0
-14 | 734000 | 5907 | 2631 | 0
-13 | 842000 | 6268 | 2204 | 0
-12 | 957000 | 9558 | 778 | 0
-11 | 1071000 | 10597 | 489 | 0
-10 | 1184000 | 10765 | 193 | 0
-9 | 1296000 | 9452 | 0 | 0
-8 | 1408000 | 1142 | 0 | 0
-7 | 1512000 | 391 | 0 | 0
-6 | 1608000 | 235 | 0 | 0
-5 | 1696000 | 0 | 0 | 0
-4 | 1776000 | 0 | 0 | 0
-3 | 1840000 | 0 | 0 | 0
-2 | 1888000 | 0 | 0 | 0
-1 | 1920000 | 0 | 0 | 0
0 | 1936000 | 0 | 0 | 0
(33 rows)
SELECT
result - numeric_mul_patch_int(var1,var2,rscale_adjustment),
COUNT(*)
FROM test_numeric_mul_patched
GROUP BY 1
ORDER BY 1;
?column? | count
----------------+----------
0 | 24739964
0.000000000001 | 2170
0.00000000001 | 234
0.0000000001 | 18
0.000000001 | 4
0.00000001 | 8927
0.0000001 | 882
0.000001 | 90
0.00001 | 6
0.0001 | 21963
0.001 | 2174
0.01 | 214
0.1 | 18
1 | 39336
(14 rows)
SELECT
result - numeric_mul_patch_small(var1,var2,rscale_adjustment),
COUNT(*)
FROM test_numeric_mul_patched
GROUP BY 1
ORDER BY 1;
?column? | count
-------------------+----------
-1 | 1233
-0.01 | 9
-0.001 | 73
-0.0001 | 647
-0.000001 | 2
-0.0000001 | 9
-0.00000001 | 116
0.000000000000000 | 24775861
0.00000001 | 1035
0.00000002 | 2
0.0000001 | 96
0.000001 | 9
0.0001 | 8771
0.0002 | 3
0.001 | 952
0.01 | 69
0.1 | 10
1 | 27098
2 | 5
(19 rows)
SELECT
result - numeric_mul_patch_inline(var1,var2,rscale_adjustment),
COUNT(*)
FROM test_numeric_mul_patched
GROUP BY 1
ORDER BY 1;
?column? | count
----------+----------
-1 | 4
0 | 24815996
(2 rows)
```
I found these two interesting to look closer at:
```
0.00000002 | 2
0.0002 | 3
SELECT
*,
numeric_mul_patch_small(var1,var2,rscale_adjustment)
FROM test_numeric_mul_patched
WHERE result - numeric_mul_patch_small(var1,var2,rscale_adjustment) IN (0.00000002, 0.0002);
var1 | var2 | rscale_adjustment | result | numeric_mul_patch_small
-------------------+----------------+-------------------+------------+-------------------------
8952.12658563 | 0.902315486665 | -16 | 8077.6425 | 8077.6423
0.881715409579 | 0.843165739371 | -16 | 0.74343223 | 0.74343221
0.905322758954 | 0.756905996850 | -16 | 0.68524423 | 0.68524421
8464.043170546608 | 0.518100129611 | -20 | 4385.2219 | 4385.2217
5253.006296984449 | 0.989308019355 | -20 | 5196.8413 | 5196.8411
(5 rows)
```
What can be said about mul_var()'s contract with regards to rscale?
It's the number of decimal digits requested by the caller, and if not
requesting full precision, then the decimal digits might not be accurate,
but can something be said about how far off they can be?
The mul_var_int() patch only produces a difference that is exactly
1 less than the exact result, at the last non-zero decimal digit.
Could the difference be more than 1 at the last non-zero digit,
like in the five cases found above?
It would be nice if we could define mul_var()'s contract with regards to
rscale, in terms of what precision can be expected in the result.
Attaching the hacked together version with all the patches, used to do the testing above.
Regards,
Joel
diff --git a/src/backend/utils/adt/numeric.c b/src/backend/utils/adt/numeric.c
index 5510a203b0..63de5cd994 100644
--- a/src/backend/utils/adt/numeric.c
+++ b/src/backend/utils/adt/numeric.c
@@ -551,6 +551,23 @@ static void sub_var(const NumericVar *var1, const NumericVar *var2,
static void mul_var(const NumericVar *var1, const NumericVar *var2,
NumericVar *result,
int rscale);
+static void mul_var_head(const NumericVar *var1, const NumericVar *var2,
+ NumericVar *result,
+ int rscale);
+static void mul_var_patch_inline(const NumericVar *var1, const NumericVar *var2,
+ NumericVar *result,
+ int rscale);
+static void mul_var_patch_small(const NumericVar *var1, const NumericVar *var2,
+ NumericVar *result,
+ int rscale);
+static void mul_var_patch_int(const NumericVar *var1, const NumericVar *var2,
+ NumericVar *result,
+ int rscale);
+static void mul_var_small(const NumericVar *var1, const NumericVar *var2,
+ NumericVar *result, int rscale);
+
+static void mul_var_int(const NumericVar *var, int ival, int ival_weight,
+ NumericVar *result, int rscale);
static void div_var(const NumericVar *var1, const NumericVar *var2,
NumericVar *result,
int rscale, bool round);
@@ -3115,6 +3132,448 @@ numeric_mul_opt_error(Numeric num1, Numeric num2, bool *have_error)
}
+Datum
+numeric_mul_head(PG_FUNCTION_ARGS)
+{
+ Numeric num1 = PG_GETARG_NUMERIC(0);
+ Numeric num2 = PG_GETARG_NUMERIC(1);
+ int32 rscale_adjustment = PG_GETARG_INT32(2);
+ Numeric res;
+
+ res = numeric_mul_head_opt_error(num1, num2, rscale_adjustment, NULL);
+
+ PG_RETURN_NUMERIC(res);
+}
+
+
+Datum
+numeric_mul_patch_inline(PG_FUNCTION_ARGS)
+{
+ Numeric num1 = PG_GETARG_NUMERIC(0);
+ Numeric num2 = PG_GETARG_NUMERIC(1);
+ int32 rscale_adjustment = PG_GETARG_INT32(2);
+ Numeric res;
+
+ res = numeric_mul_patch_inline_opt_error(num1, num2, rscale_adjustment, NULL);
+
+ PG_RETURN_NUMERIC(res);
+}
+
+Datum
+numeric_mul_patch_small(PG_FUNCTION_ARGS)
+{
+ Numeric num1 = PG_GETARG_NUMERIC(0);
+ Numeric num2 = PG_GETARG_NUMERIC(1);
+ int32 rscale_adjustment = PG_GETARG_INT32(2);
+ Numeric res;
+
+ res = numeric_mul_patch_small_opt_error(num1, num2, rscale_adjustment, NULL);
+
+ PG_RETURN_NUMERIC(res);
+}
+
+Datum
+numeric_mul_patch_int(PG_FUNCTION_ARGS)
+{
+ Numeric num1 = PG_GETARG_NUMERIC(0);
+ Numeric num2 = PG_GETARG_NUMERIC(1);
+ int32 rscale_adjustment = PG_GETARG_INT32(2);
+ Numeric res;
+
+ res = numeric_mul_patch_int_opt_error(num1, num2, rscale_adjustment, NULL);
+
+ PG_RETURN_NUMERIC(res);
+}
+
+
+Numeric
+numeric_mul_patch_inline_opt_error(Numeric num1, Numeric num2, int32 rscale_adjustment, bool *have_error)
+{
+ NumericVar arg1;
+ NumericVar arg2;
+ NumericVar result;
+ Numeric res;
+
+ /*
+ * Handle NaN and infinities
+ */
+ if (NUMERIC_IS_SPECIAL(num1) || NUMERIC_IS_SPECIAL(num2))
+ {
+ if (NUMERIC_IS_NAN(num1) || NUMERIC_IS_NAN(num2))
+ return make_result(&const_nan);
+ if (NUMERIC_IS_PINF(num1))
+ {
+ switch (numeric_sign_internal(num2))
+ {
+ case 0:
+ return make_result(&const_nan); /* Inf * 0 */
+ case 1:
+ return make_result(&const_pinf);
+ case -1:
+ return make_result(&const_ninf);
+ }
+ Assert(false);
+ }
+ if (NUMERIC_IS_NINF(num1))
+ {
+ switch (numeric_sign_internal(num2))
+ {
+ case 0:
+ return make_result(&const_nan); /* -Inf * 0 */
+ case 1:
+ return make_result(&const_ninf);
+ case -1:
+ return make_result(&const_pinf);
+ }
+ Assert(false);
+ }
+ /* by here, num1 must be finite, so num2 is not */
+ if (NUMERIC_IS_PINF(num2))
+ {
+ switch (numeric_sign_internal(num1))
+ {
+ case 0:
+ return make_result(&const_nan); /* 0 * Inf */
+ case 1:
+ return make_result(&const_pinf);
+ case -1:
+ return make_result(&const_ninf);
+ }
+ Assert(false);
+ }
+ Assert(NUMERIC_IS_NINF(num2));
+ switch (numeric_sign_internal(num1))
+ {
+ case 0:
+ return make_result(&const_nan); /* 0 * -Inf */
+ case 1:
+ return make_result(&const_ninf);
+ case -1:
+ return make_result(&const_pinf);
+ }
+ Assert(false);
+ }
+
+ /*
+ * Unpack the values, let mul_var() compute the result and return it.
+ * Unlike add_var() and sub_var(), mul_var() will round its result. In the
+ * case of numeric_mul(), which is invoked for the * operator on numerics,
+ * we request exact representation for the product (rscale = sum(dscale of
+ * arg1, dscale of arg2)). If the exact result has more digits after the
+ * decimal point than can be stored in a numeric, we round it. Rounding
+ * after computing the exact result ensures that the final result is
+ * correctly rounded (rounding in mul_var() using a truncated product
+ * would not guarantee this).
+ */
+ init_var_from_num(num1, &arg1);
+ init_var_from_num(num2, &arg2);
+
+ init_var(&result);
+
+ mul_var_patch_inline(&arg1, &arg2, &result, arg1.dscale + arg2.dscale + rscale_adjustment);
+
+ if (result.dscale > NUMERIC_DSCALE_MAX)
+ round_var(&result, NUMERIC_DSCALE_MAX);
+
+ res = make_result_opt_error(&result, have_error);
+
+ free_var(&result);
+
+ return res;
+}
+
+
+Numeric
+numeric_mul_head_opt_error(Numeric num1, Numeric num2, int32 rscale_adjustment, bool *have_error)
+{
+ NumericVar arg1;
+ NumericVar arg2;
+ NumericVar result;
+ Numeric res;
+
+ /*
+ * Handle NaN and infinities
+ */
+ if (NUMERIC_IS_SPECIAL(num1) || NUMERIC_IS_SPECIAL(num2))
+ {
+ if (NUMERIC_IS_NAN(num1) || NUMERIC_IS_NAN(num2))
+ return make_result(&const_nan);
+ if (NUMERIC_IS_PINF(num1))
+ {
+ switch (numeric_sign_internal(num2))
+ {
+ case 0:
+ return make_result(&const_nan); /* Inf * 0 */
+ case 1:
+ return make_result(&const_pinf);
+ case -1:
+ return make_result(&const_ninf);
+ }
+ Assert(false);
+ }
+ if (NUMERIC_IS_NINF(num1))
+ {
+ switch (numeric_sign_internal(num2))
+ {
+ case 0:
+ return make_result(&const_nan); /* -Inf * 0 */
+ case 1:
+ return make_result(&const_ninf);
+ case -1:
+ return make_result(&const_pinf);
+ }
+ Assert(false);
+ }
+ /* by here, num1 must be finite, so num2 is not */
+ if (NUMERIC_IS_PINF(num2))
+ {
+ switch (numeric_sign_internal(num1))
+ {
+ case 0:
+ return make_result(&const_nan); /* 0 * Inf */
+ case 1:
+ return make_result(&const_pinf);
+ case -1:
+ return make_result(&const_ninf);
+ }
+ Assert(false);
+ }
+ Assert(NUMERIC_IS_NINF(num2));
+ switch (numeric_sign_internal(num1))
+ {
+ case 0:
+ return make_result(&const_nan); /* 0 * -Inf */
+ case 1:
+ return make_result(&const_ninf);
+ case -1:
+ return make_result(&const_pinf);
+ }
+ Assert(false);
+ }
+
+ /*
+ * Unpack the values, let mul_var() compute the result and return it.
+ * Unlike add_var() and sub_var(), mul_var() will round its result. In the
+ * case of numeric_mul(), which is invoked for the * operator on numerics,
+ * we request exact representation for the product (rscale = sum(dscale of
+ * arg1, dscale of arg2)). If the exact result has more digits after the
+ * decimal point than can be stored in a numeric, we round it. Rounding
+ * after computing the exact result ensures that the final result is
+ * correctly rounded (rounding in mul_var() using a truncated product
+ * would not guarantee this).
+ */
+ init_var_from_num(num1, &arg1);
+ init_var_from_num(num2, &arg2);
+
+ init_var(&result);
+
+ mul_var_head(&arg1, &arg2, &result, arg1.dscale + arg2.dscale + rscale_adjustment);
+
+ if (result.dscale > NUMERIC_DSCALE_MAX)
+ round_var(&result, NUMERIC_DSCALE_MAX);
+
+ res = make_result_opt_error(&result, have_error);
+
+ free_var(&result);
+
+ return res;
+}
+
+
+Numeric
+numeric_mul_patch_small_opt_error(Numeric num1, Numeric num2, int32 rscale_adjustment, bool *have_error)
+{
+ NumericVar arg1;
+ NumericVar arg2;
+ NumericVar result;
+ Numeric res;
+
+ /*
+ * Handle NaN and infinities
+ */
+ if (NUMERIC_IS_SPECIAL(num1) || NUMERIC_IS_SPECIAL(num2))
+ {
+ if (NUMERIC_IS_NAN(num1) || NUMERIC_IS_NAN(num2))
+ return make_result(&const_nan);
+ if (NUMERIC_IS_PINF(num1))
+ {
+ switch (numeric_sign_internal(num2))
+ {
+ case 0:
+ return make_result(&const_nan); /* Inf * 0 */
+ case 1:
+ return make_result(&const_pinf);
+ case -1:
+ return make_result(&const_ninf);
+ }
+ Assert(false);
+ }
+ if (NUMERIC_IS_NINF(num1))
+ {
+ switch (numeric_sign_internal(num2))
+ {
+ case 0:
+ return make_result(&const_nan); /* -Inf * 0 */
+ case 1:
+ return make_result(&const_ninf);
+ case -1:
+ return make_result(&const_pinf);
+ }
+ Assert(false);
+ }
+ /* by here, num1 must be finite, so num2 is not */
+ if (NUMERIC_IS_PINF(num2))
+ {
+ switch (numeric_sign_internal(num1))
+ {
+ case 0:
+ return make_result(&const_nan); /* 0 * Inf */
+ case 1:
+ return make_result(&const_pinf);
+ case -1:
+ return make_result(&const_ninf);
+ }
+ Assert(false);
+ }
+ Assert(NUMERIC_IS_NINF(num2));
+ switch (numeric_sign_internal(num1))
+ {
+ case 0:
+ return make_result(&const_nan); /* 0 * -Inf */
+ case 1:
+ return make_result(&const_ninf);
+ case -1:
+ return make_result(&const_pinf);
+ }
+ Assert(false);
+ }
+
+ /*
+ * Unpack the values, let mul_var() compute the result and return it.
+ * Unlike add_var() and sub_var(), mul_var() will round its result. In the
+ * case of numeric_mul(), which is invoked for the * operator on numerics,
+ * we request exact representation for the product (rscale = sum(dscale of
+ * arg1, dscale of arg2)). If the exact result has more digits after the
+ * decimal point than can be stored in a numeric, we round it. Rounding
+ * after computing the exact result ensures that the final result is
+ * correctly rounded (rounding in mul_var() using a truncated product
+ * would not guarantee this).
+ */
+ init_var_from_num(num1, &arg1);
+ init_var_from_num(num2, &arg2);
+
+ init_var(&result);
+
+ mul_var_patch_small(&arg1, &arg2, &result, arg1.dscale + arg2.dscale + rscale_adjustment);
+
+ if (result.dscale > NUMERIC_DSCALE_MAX)
+ round_var(&result, NUMERIC_DSCALE_MAX);
+
+ res = make_result_opt_error(&result, have_error);
+
+ free_var(&result);
+
+ return res;
+}
+
+
+Numeric
+numeric_mul_patch_int_opt_error(Numeric num1, Numeric num2, int32 rscale_adjustment, bool *have_error)
+{
+ NumericVar arg1;
+ NumericVar arg2;
+ NumericVar result;
+ Numeric res;
+
+ /*
+ * Handle NaN and infinities
+ */
+ if (NUMERIC_IS_SPECIAL(num1) || NUMERIC_IS_SPECIAL(num2))
+ {
+ if (NUMERIC_IS_NAN(num1) || NUMERIC_IS_NAN(num2))
+ return make_result(&const_nan);
+ if (NUMERIC_IS_PINF(num1))
+ {
+ switch (numeric_sign_internal(num2))
+ {
+ case 0:
+ return make_result(&const_nan); /* Inf * 0 */
+ case 1:
+ return make_result(&const_pinf);
+ case -1:
+ return make_result(&const_ninf);
+ }
+ Assert(false);
+ }
+ if (NUMERIC_IS_NINF(num1))
+ {
+ switch (numeric_sign_internal(num2))
+ {
+ case 0:
+ return make_result(&const_nan); /* -Inf * 0 */
+ case 1:
+ return make_result(&const_ninf);
+ case -1:
+ return make_result(&const_pinf);
+ }
+ Assert(false);
+ }
+ /* by here, num1 must be finite, so num2 is not */
+ if (NUMERIC_IS_PINF(num2))
+ {
+ switch (numeric_sign_internal(num1))
+ {
+ case 0:
+ return make_result(&const_nan); /* 0 * Inf */
+ case 1:
+ return make_result(&const_pinf);
+ case -1:
+ return make_result(&const_ninf);
+ }
+ Assert(false);
+ }
+ Assert(NUMERIC_IS_NINF(num2));
+ switch (numeric_sign_internal(num1))
+ {
+ case 0:
+ return make_result(&const_nan); /* 0 * -Inf */
+ case 1:
+ return make_result(&const_ninf);
+ case -1:
+ return make_result(&const_pinf);
+ }
+ Assert(false);
+ }
+
+ /*
+ * Unpack the values, let mul_var() compute the result and return it.
+ * Unlike add_var() and sub_var(), mul_var() will round its result. In the
+ * case of numeric_mul(), which is invoked for the * operator on numerics,
+ * we request exact representation for the product (rscale = sum(dscale of
+ * arg1, dscale of arg2)). If the exact result has more digits after the
+ * decimal point than can be stored in a numeric, we round it. Rounding
+ * after computing the exact result ensures that the final result is
+ * correctly rounded (rounding in mul_var() using a truncated product
+ * would not guarantee this).
+ */
+ init_var_from_num(num1, &arg1);
+ init_var_from_num(num2, &arg2);
+
+ init_var(&result);
+
+ mul_var_patch_int(&arg1, &arg2, &result, arg1.dscale + arg2.dscale + rscale_adjustment);
+
+ if (result.dscale > NUMERIC_DSCALE_MAX)
+ round_var(&result, NUMERIC_DSCALE_MAX);
+
+ res = make_result_opt_error(&result, have_error);
+
+ free_var(&result);
+
+ return res;
+}
+
+
/*
* numeric_div() -
*
@@ -8864,6 +9323,1245 @@ mul_var(const NumericVar *var1, const NumericVar *var2, NumericVar *result,
strip_var(result);
}
+static void
+mul_var_head(const NumericVar *var1, const NumericVar *var2, NumericVar *result,
+ int rscale)
+{
+ int res_ndigits;
+ int res_sign;
+ int res_weight;
+ int maxdigits;
+ int *dig;
+ int carry;
+ int maxdig;
+ int newdig;
+ int var1ndigits;
+ int var2ndigits;
+ NumericDigit *var1digits;
+ NumericDigit *var2digits;
+ NumericDigit *res_digits;
+ int i,
+ i1,
+ i2;
+
+ /*
+ * Arrange for var1 to be the shorter of the two numbers. This improves
+ * performance because the inner multiplication loop is much simpler than
+ * the outer loop, so it's better to have a smaller number of iterations
+ * of the outer loop. This also reduces the number of times that the
+ * accumulator array needs to be normalized.
+ */
+ if (var1->ndigits > var2->ndigits)
+ {
+ const NumericVar *tmp = var1;
+
+ var1 = var2;
+ var2 = tmp;
+ }
+
+ /* copy these values into local vars for speed in inner loop */
+ var1ndigits = var1->ndigits;
+ var2ndigits = var2->ndigits;
+ var1digits = var1->digits;
+ var2digits = var2->digits;
+
+ if (var1ndigits == 0 || var2ndigits == 0)
+ {
+ /* one or both inputs is zero; so is result */
+ zero_var(result);
+ result->dscale = rscale;
+ return;
+ }
+
+ /* Determine result sign and (maximum possible) weight */
+ if (var1->sign == var2->sign)
+ res_sign = NUMERIC_POS;
+ else
+ res_sign = NUMERIC_NEG;
+ res_weight = var1->weight + var2->weight + 2;
+
+ /*
+ * Determine the number of result digits to compute. If the exact result
+ * would have more than rscale fractional digits, truncate the computation
+ * with MUL_GUARD_DIGITS guard digits, i.e., ignore input digits that
+ * would only contribute to the right of that. (This will give the exact
+ * rounded-to-rscale answer unless carries out of the ignored positions
+ * would have propagated through more than MUL_GUARD_DIGITS digits.)
+ *
+ * Note: an exact computation could not produce more than var1ndigits +
+ * var2ndigits digits, but we allocate one extra output digit in case
+ * rscale-driven rounding produces a carry out of the highest exact digit.
+ */
+ res_ndigits = var1ndigits + var2ndigits + 1;
+ maxdigits = res_weight + 1 + (rscale + DEC_DIGITS - 1) / DEC_DIGITS +
+ MUL_GUARD_DIGITS;
+ res_ndigits = Min(res_ndigits, maxdigits);
+
+ if (res_ndigits < 3)
+ {
+ /* All input digits will be ignored; so result is zero */
+ zero_var(result);
+ result->dscale = rscale;
+ return;
+ }
+
+ /*
+ * We do the arithmetic in an array "dig[]" of signed int's. Since
+ * INT_MAX is noticeably larger than NBASE*NBASE, this gives us headroom
+ * to avoid normalizing carries immediately.
+ *
+ * maxdig tracks the maximum possible value of any dig[] entry; when this
+ * threatens to exceed INT_MAX, we take the time to propagate carries.
+ * Furthermore, we need to ensure that overflow doesn't occur during the
+ * carry propagation passes either. The carry values could be as much as
+ * INT_MAX/NBASE, so really we must normalize when digits threaten to
+ * exceed INT_MAX - INT_MAX/NBASE.
+ *
+ * To avoid overflow in maxdig itself, it actually represents the max
+ * possible value divided by NBASE-1, ie, at the top of the loop it is
+ * known that no dig[] entry exceeds maxdig * (NBASE-1).
+ */
+ dig = (int *) palloc0(res_ndigits * sizeof(int));
+ maxdig = 0;
+
+ /*
+ * The least significant digits of var1 should be ignored if they don't
+ * contribute directly to the first res_ndigits digits of the result that
+ * we are computing.
+ *
+ * Digit i1 of var1 and digit i2 of var2 are multiplied and added to digit
+ * i1+i2+2 of the accumulator array, so we need only consider digits of
+ * var1 for which i1 <= res_ndigits - 3.
+ */
+ for (i1 = Min(var1ndigits - 1, res_ndigits - 3); i1 >= 0; i1--)
+ {
+ NumericDigit var1digit = var1digits[i1];
+
+ if (var1digit == 0)
+ continue;
+
+ /* Time to normalize? */
+ maxdig += var1digit;
+ if (maxdig > (INT_MAX - INT_MAX / NBASE) / (NBASE - 1))
+ {
+ /* Yes, do it */
+ carry = 0;
+ for (i = res_ndigits - 1; i >= 0; i--)
+ {
+ newdig = dig[i] + carry;
+ if (newdig >= NBASE)
+ {
+ carry = newdig / NBASE;
+ newdig -= carry * NBASE;
+ }
+ else
+ carry = 0;
+ dig[i] = newdig;
+ }
+ Assert(carry == 0);
+ /* Reset maxdig to indicate new worst-case */
+ maxdig = 1 + var1digit;
+ }
+
+ /*
+ * Add the appropriate multiple of var2 into the accumulator.
+ *
+ * As above, digits of var2 can be ignored if they don't contribute,
+ * so we only include digits for which i1+i2+2 < res_ndigits.
+ *
+ * This inner loop is the performance bottleneck for multiplication,
+ * so we want to keep it simple enough so that it can be
+ * auto-vectorized. Accordingly, process the digits left-to-right
+ * even though schoolbook multiplication would suggest right-to-left.
+ * Since we aren't propagating carries in this loop, the order does
+ * not matter.
+ */
+ {
+ int i2limit = Min(var2ndigits, res_ndigits - i1 - 2);
+ int *dig_i1_2 = &dig[i1 + 2];
+
+ for (i2 = 0; i2 < i2limit; i2++)
+ dig_i1_2[i2] += var1digit * var2digits[i2];
+ }
+ }
+
+ /*
+ * Now we do a final carry propagation pass to normalize the result, which
+ * we combine with storing the result digits into the output. Note that
+ * this is still done at full precision w/guard digits.
+ */
+ alloc_var(result, res_ndigits);
+ res_digits = result->digits;
+ carry = 0;
+ for (i = res_ndigits - 1; i >= 0; i--)
+ {
+ newdig = dig[i] + carry;
+ if (newdig >= NBASE)
+ {
+ carry = newdig / NBASE;
+ newdig -= carry * NBASE;
+ }
+ else
+ carry = 0;
+ res_digits[i] = newdig;
+ }
+ Assert(carry == 0);
+
+ pfree(dig);
+
+ /*
+ * Finally, round the result to the requested precision.
+ */
+ result->weight = res_weight;
+ result->sign = res_sign;
+
+ /* Round to target rscale (and set result->dscale) */
+ round_var(result, rscale);
+
+ /* Strip leading and trailing zeroes */
+ strip_var(result);
+}
+
+static void
+mul_var_patch_inline(const NumericVar *var1, const NumericVar *var2,
+ NumericVar *result, int rscale)
+{
+ int res_ndigits;
+ int res_sign;
+ int res_weight;
+ int maxdigits;
+ int *dig;
+ int carry;
+ int maxdig;
+ int newdig;
+ int var1ndigits;
+ int var2ndigits;
+ NumericDigit *var1digits;
+ NumericDigit *var2digits;
+ NumericDigit *res_digits;
+ int i,
+ i1,
+ i2;
+
+ /*
+ * Arrange for var1 to be the shorter of the two numbers. This improves
+ * performance because the inner multiplication loop is much simpler than
+ * the outer loop, so it's better to have a smaller number of iterations
+ * of the outer loop. This also reduces the number of times that the
+ * accumulator array needs to be normalized.
+ */
+ if (var1->ndigits > var2->ndigits)
+ {
+ const NumericVar *tmp = var1;
+
+ var1 = var2;
+ var2 = tmp;
+ }
+
+ /* copy these values into local vars for speed in inner loop */
+ var1ndigits = var1->ndigits;
+ var2ndigits = var2->ndigits;
+ var1digits = var1->digits;
+ var2digits = var2->digits;
+
+ if (var1ndigits == 0 || var2ndigits == 0)
+ {
+ /* one or both inputs is zero; so is result */
+ zero_var(result);
+ result->dscale = rscale;
+ return;
+ }
+
+ /* Determine result sign and (maximum possible) weight */
+ if (var1->sign == var2->sign)
+ res_sign = NUMERIC_POS;
+ else
+ res_sign = NUMERIC_NEG;
+ res_weight = var1->weight + var2->weight + 2;
+
+ /*
+ * Determine the number of result digits to compute. If the exact result
+ * would have more than rscale fractional digits, truncate the computation
+ * with MUL_GUARD_DIGITS guard digits, i.e., ignore input digits that
+ * would only contribute to the right of that. (This will give the exact
+ * rounded-to-rscale answer unless carries out of the ignored positions
+ * would have propagated through more than MUL_GUARD_DIGITS digits.)
+ *
+ * Note: an exact computation could not produce more than var1ndigits +
+ * var2ndigits digits, but we allocate one extra output digit in case
+ * rscale-driven rounding produces a carry out of the highest exact digit.
+ */
+ res_ndigits = var1ndigits + var2ndigits + 1;
+ maxdigits = res_weight + 1 + (rscale + DEC_DIGITS - 1) / DEC_DIGITS +
+ MUL_GUARD_DIGITS;
+ res_ndigits = Min(res_ndigits, maxdigits);
+
+ if (res_ndigits < 3)
+ {
+ /* All input digits will be ignored; so result is zero */
+ zero_var(result);
+ result->dscale = rscale;
+ return;
+ }
+
+ /*
+ * Simplified fast-path computation, if var1 has just one or two digits.
+ * This is significantly faster, since it avoids allocating a separate
+ * digit array, making multiple passes over var2, and having separate
+ * carry-propagation passes.
+ */
+ if (var1ndigits <= 3)
+ {
+ NumericDigit *res_buf;
+
+ /* Allocate result digit array */
+ res_buf = digitbuf_alloc(res_ndigits);
+ res_buf[0] = 0; /* spare digit for later rounding */
+ res_digits = res_buf + 1;
+
+ /*
+ * Compute the result digits directly, in one pass, propagating the
+ * carry up as we go.
+ */
+ switch (var1ndigits)
+ {
+ case 1:
+ carry = 0;
+ for (i = res_ndigits - 3; i >= 0; i--)
+ {
+ newdig = (int) var1digits[0] * var2digits[i] + carry;
+ res_digits[i + 1] = (NumericDigit) (newdig % NBASE);
+ carry = newdig / NBASE;
+ }
+ res_digits[0] = (NumericDigit) carry;
+ break;
+
+ case 2:
+ if (res_ndigits - 4 >= 0 && res_ndigits - 4 < var2ndigits)
+ newdig = (int) var1digits[1] * var2digits[res_ndigits - 4];
+ else
+ newdig = 0;
+ if (res_ndigits - 3 >= 0 && res_ndigits - 3 < var2ndigits)
+ newdig += (int) var1digits[0] * var2digits[res_ndigits - 3];
+ res_digits[res_ndigits - 2] = (NumericDigit) (newdig % NBASE);
+ carry = newdig / NBASE;
+ for (i = res_ndigits - 4; i >= 1; i--)
+ {
+ newdig = (int) var1digits[0] * var2digits[i] +
+ (int) var1digits[1] * var2digits[i - 1] + carry;
+ res_digits[i + 1] = (NumericDigit) (newdig % NBASE);
+ carry = newdig / NBASE;
+ }
+ newdig = (int) var1digits[0] * var2digits[0] + carry;
+ res_digits[1] = (NumericDigit) (newdig % NBASE);
+ res_digits[0] = (NumericDigit) (newdig / NBASE);
+ break;
+
+ case 3:
+ if (res_ndigits - 5 >= 0 && res_ndigits - 5 < var2ndigits)
+ newdig = (int) var1digits[2] * var2digits[res_ndigits - 5];
+ else
+ newdig = 0;
+ if (res_ndigits - 4 >= 0 && res_ndigits - 4 < var2ndigits)
+ newdig += (int) var1digits[1] * var2digits[res_ndigits - 4];
+ if (res_ndigits - 3 >= 0 && res_ndigits - 3 < var2ndigits)
+ newdig += (int) var1digits[0] * var2digits[res_ndigits - 3];
+ res_digits[res_ndigits - 2] = (NumericDigit) (newdig % NBASE);
+ carry = newdig / NBASE;
+ for (i = res_ndigits - 4; i >= 2; i--)
+ {
+ newdig = carry;
+ if (i < var2ndigits)
+ newdig += (int) var1digits[0] * var2digits[i];
+ if (i - 1 >= 0 && i - 1 < var2ndigits)
+ newdig += (int) var1digits[1] * var2digits[i - 1];
+ if (i - 2 >= 0 && i - 2 < var2ndigits)
+ newdig += (int) var1digits[2] * var2digits[i - 2];
+ res_digits[i + 1] = (NumericDigit) (newdig % NBASE);
+ carry = newdig / NBASE;
+ }
+ newdig = carry;
+ if (var2ndigits > 1)
+ newdig += (int) var1digits[0] * var2digits[1];
+ if (var2ndigits > 0)
+ newdig += (int) var1digits[1] * var2digits[0];
+ res_digits[2] = (NumericDigit) (newdig % NBASE);
+ carry = newdig / NBASE;
+ newdig = (int) var1digits[0] * var2digits[0] + carry;
+ res_digits[1] = (NumericDigit) (newdig % NBASE);
+ res_digits[0] = (NumericDigit) (newdig / NBASE);
+ break;
+ }
+
+ /* Store the product in result (minus extra rounding digit) */
+ digitbuf_free(result->buf);
+ result->ndigits = res_ndigits - 1;
+ result->buf = res_buf;
+ result->digits = res_digits;
+ result->weight = res_weight - 1;
+ result->sign = res_sign;
+
+ /* Round to target rscale (and set result->dscale) */
+ round_var(result, rscale);
+
+ /* Strip leading and trailing zeroes */
+ strip_var(result);
+
+ return;
+ }
+
+ /*
+ * We do the arithmetic in an array "dig[]" of signed int's. Since
+ * INT_MAX is noticeably larger than NBASE*NBASE, this gives us headroom
+ * to avoid normalizing carries immediately.
+ *
+ * maxdig tracks the maximum possible value of any dig[] entry; when this
+ * threatens to exceed INT_MAX, we take the time to propagate carries.
+ * Furthermore, we need to ensure that overflow doesn't occur during the
+ * carry propagation passes either. The carry values could be as much as
+ * INT_MAX/NBASE, so really we must normalize when digits threaten to
+ * exceed INT_MAX - INT_MAX/NBASE.
+ *
+ * To avoid overflow in maxdig itself, it actually represents the max
+ * possible value divided by NBASE-1, ie, at the top of the loop it is
+ * known that no dig[] entry exceeds maxdig * (NBASE-1).
+ */
+ dig = (int *) palloc0(res_ndigits * sizeof(int));
+ maxdig = 0;
+
+ /*
+ * The least significant digits of var1 should be ignored if they don't
+ * contribute directly to the first res_ndigits digits of the result that
+ * we are computing.
+ *
+ * Digit i1 of var1 and digit i2 of var2 are multiplied and added to digit
+ * i1+i2+2 of the accumulator array, so we need only consider digits of
+ * var1 for which i1 <= res_ndigits - 3.
+ */
+ for (i1 = Min(var1ndigits - 1, res_ndigits - 3); i1 >= 0; i1--)
+ {
+ NumericDigit var1digit = var1digits[i1];
+
+ if (var1digit == 0)
+ continue;
+
+ /* Time to normalize? */
+ maxdig += var1digit;
+ if (maxdig > (INT_MAX - INT_MAX / NBASE) / (NBASE - 1))
+ {
+ /* Yes, do it */
+ carry = 0;
+ for (i = res_ndigits - 1; i >= 0; i--)
+ {
+ newdig = dig[i] + carry;
+ if (newdig >= NBASE)
+ {
+ carry = newdig / NBASE;
+ newdig -= carry * NBASE;
+ }
+ else
+ carry = 0;
+ dig[i] = newdig;
+ }
+ Assert(carry == 0);
+ /* Reset maxdig to indicate new worst-case */
+ maxdig = 1 + var1digit;
+ }
+
+ /*
+ * Add the appropriate multiple of var2 into the accumulator.
+ *
+ * As above, digits of var2 can be ignored if they don't contribute,
+ * so we only include digits for which i1+i2+2 < res_ndigits.
+ *
+ * This inner loop is the performance bottleneck for multiplication,
+ * so we want to keep it simple enough so that it can be
+ * auto-vectorized. Accordingly, process the digits left-to-right
+ * even though schoolbook multiplication would suggest right-to-left.
+ * Since we aren't propagating carries in this loop, the order does
+ * not matter.
+ */
+ {
+ int i2limit = Min(var2ndigits, res_ndigits - i1 - 2);
+ int *dig_i1_2 = &dig[i1 + 2];
+
+ for (i2 = 0; i2 < i2limit; i2++)
+ {
+ dig_i1_2[i2] += var1digit * var2digits[i2];
+ }
+ }
+ }
+
+ /*
+ * Now we do a final carry propagation pass to normalize the result, which
+ * we combine with storing the result digits into the output. Note that
+ * this is still done at full precision w/guard digits.
+ */
+ alloc_var(result, res_ndigits);
+ res_digits = result->digits;
+ carry = 0;
+ for (i = res_ndigits - 1; i >= 0; i--)
+ {
+ newdig = dig[i] + carry;
+ if (newdig >= NBASE)
+ {
+ carry = newdig / NBASE;
+ newdig -= carry * NBASE;
+ }
+ else
+ carry = 0;
+ res_digits[i] = newdig;
+ }
+ Assert(carry == 0);
+
+ pfree(dig);
+
+ /*
+ * Finally, round the result to the requested precision.
+ */
+ result->weight = res_weight;
+ result->sign = res_sign;
+
+ /* Round to target rscale (and set result->dscale) */
+ round_var(result, rscale);
+
+ /* Strip leading and trailing zeroes */
+ strip_var(result);
+
+}
+
+
+static void
+mul_var_patch_int(const NumericVar *var1, const NumericVar *var2, NumericVar *result,
+ int rscale)
+{
+ int res_ndigits;
+ int res_sign;
+ int res_weight;
+ int maxdigits;
+ int *dig;
+ int carry;
+ int maxdig;
+ int newdig;
+ int var1ndigits;
+ int var2ndigits;
+ NumericDigit *var1digits;
+ NumericDigit *var2digits;
+ NumericDigit *res_digits;
+ int i,
+ i1,
+ i2;
+
+ /*
+ * Arrange for var1 to be the shorter of the two numbers. This improves
+ * performance because the inner multiplication loop is much simpler than
+ * the outer loop, so it's better to have a smaller number of iterations
+ * of the outer loop. This also reduces the number of times that the
+ * accumulator array needs to be normalized.
+ */
+ if (var1->ndigits > var2->ndigits)
+ {
+ const NumericVar *tmp = var1;
+
+ var1 = var2;
+ var2 = tmp;
+ }
+
+ /* copy these values into local vars for speed in inner loop */
+ var1ndigits = var1->ndigits;
+ var2ndigits = var2->ndigits;
+ var1digits = var1->digits;
+ var2digits = var2->digits;
+
+ if (var1ndigits == 0)
+ {
+ /* one or both inputs is zero; so is result */
+ zero_var(result);
+ result->dscale = rscale;
+ return;
+ }
+
+ /*
+ * If var1 has just one or two digits, delegate to mul_var_int(), which
+ * uses a faster direct multiplication algorithm.
+ *
+ * TODO: Similarly, on platforms with 128-bit integers ...
+ */
+ if (var1ndigits <= 2)
+ {
+ int ifactor;
+ int ifactor_weight;
+
+ ifactor = var1->digits[0];
+ ifactor_weight = var1->weight;
+ if (var1ndigits == 2)
+ {
+ ifactor = ifactor * NBASE + var1->digits[1];
+ ifactor_weight--;
+ }
+ if (var1->sign == NUMERIC_NEG)
+ ifactor = -ifactor;
+
+ mul_var_int(var2, ifactor, ifactor_weight, result, rscale);
+ return;
+ }
+
+ /* Determine result sign and (maximum possible) weight */
+ if (var1->sign == var2->sign)
+ res_sign = NUMERIC_POS;
+ else
+ res_sign = NUMERIC_NEG;
+ res_weight = var1->weight + var2->weight + 2;
+
+ /*
+ * Determine the number of result digits to compute. If the exact result
+ * would have more than rscale fractional digits, truncate the computation
+ * with MUL_GUARD_DIGITS guard digits, i.e., ignore input digits that
+ * would only contribute to the right of that. (This will give the exact
+ * rounded-to-rscale answer unless carries out of the ignored positions
+ * would have propagated through more than MUL_GUARD_DIGITS digits.)
+ *
+ * Note: an exact computation could not produce more than var1ndigits +
+ * var2ndigits digits, but we allocate one extra output digit in case
+ * rscale-driven rounding produces a carry out of the highest exact digit.
+ */
+ res_ndigits = var1ndigits + var2ndigits + 1;
+ maxdigits = res_weight + 1 + (rscale + DEC_DIGITS - 1) / DEC_DIGITS +
+ MUL_GUARD_DIGITS;
+ res_ndigits = Min(res_ndigits, maxdigits);
+
+ if (res_ndigits < 3)
+ {
+ /* All input digits will be ignored; so result is zero */
+ zero_var(result);
+ result->dscale = rscale;
+ return;
+ }
+
+ /*
+ * We do the arithmetic in an array "dig[]" of signed int's. Since
+ * INT_MAX is noticeably larger than NBASE*NBASE, this gives us headroom
+ * to avoid normalizing carries immediately.
+ *
+ * maxdig tracks the maximum possible value of any dig[] entry; when this
+ * threatens to exceed INT_MAX, we take the time to propagate carries.
+ * Furthermore, we need to ensure that overflow doesn't occur during the
+ * carry propagation passes either. The carry values could be as much as
+ * INT_MAX/NBASE, so really we must normalize when digits threaten to
+ * exceed INT_MAX - INT_MAX/NBASE.
+ *
+ * To avoid overflow in maxdig itself, it actually represents the max
+ * possible value divided by NBASE-1, ie, at the top of the loop it is
+ * known that no dig[] entry exceeds maxdig * (NBASE-1).
+ */
+ dig = (int *) palloc0(res_ndigits * sizeof(int));
+ maxdig = 0;
+
+ /*
+ * The least significant digits of var1 should be ignored if they don't
+ * contribute directly to the first res_ndigits digits of the result that
+ * we are computing.
+ *
+ * Digit i1 of var1 and digit i2 of var2 are multiplied and added to digit
+ * i1+i2+2 of the accumulator array, so we need only consider digits of
+ * var1 for which i1 <= res_ndigits - 3.
+ */
+ for (i1 = Min(var1ndigits - 1, res_ndigits - 3); i1 >= 0; i1--)
+ {
+ NumericDigit var1digit = var1digits[i1];
+
+ if (var1digit == 0)
+ continue;
+
+ /* Time to normalize? */
+ maxdig += var1digit;
+ if (maxdig > (INT_MAX - INT_MAX / NBASE) / (NBASE - 1))
+ {
+ /* Yes, do it */
+ carry = 0;
+ for (i = res_ndigits - 1; i >= 0; i--)
+ {
+ newdig = dig[i] + carry;
+ if (newdig >= NBASE)
+ {
+ carry = newdig / NBASE;
+ newdig -= carry * NBASE;
+ }
+ else
+ carry = 0;
+ dig[i] = newdig;
+ }
+ Assert(carry == 0);
+ /* Reset maxdig to indicate new worst-case */
+ maxdig = 1 + var1digit;
+ }
+
+ /*
+ * Add the appropriate multiple of var2 into the accumulator.
+ *
+ * As above, digits of var2 can be ignored if they don't contribute,
+ * so we only include digits for which i1+i2+2 < res_ndigits.
+ *
+ * This inner loop is the performance bottleneck for multiplication,
+ * so we want to keep it simple enough so that it can be
+ * auto-vectorized. Accordingly, process the digits left-to-right
+ * even though schoolbook multiplication would suggest right-to-left.
+ * Since we aren't propagating carries in this loop, the order does
+ * not matter.
+ */
+ {
+ int i2limit = Min(var2ndigits, res_ndigits - i1 - 2);
+ int *dig_i1_2 = &dig[i1 + 2];
+
+ for (i2 = 0; i2 < i2limit; i2++)
+ dig_i1_2[i2] += var1digit * var2digits[i2];
+ }
+ }
+
+ /*
+ * Now we do a final carry propagation pass to normalize the result, which
+ * we combine with storing the result digits into the output. Note that
+ * this is still done at full precision w/guard digits.
+ */
+ alloc_var(result, res_ndigits);
+ res_digits = result->digits;
+ carry = 0;
+ for (i = res_ndigits - 1; i >= 0; i--)
+ {
+ newdig = dig[i] + carry;
+ if (newdig >= NBASE)
+ {
+ carry = newdig / NBASE;
+ newdig -= carry * NBASE;
+ }
+ else
+ carry = 0;
+ res_digits[i] = newdig;
+ }
+ Assert(carry == 0);
+
+ pfree(dig);
+
+ /*
+ * Finally, round the result to the requested precision.
+ */
+ result->weight = res_weight;
+ result->sign = res_sign;
+
+ /* Round to target rscale (and set result->dscale) */
+ round_var(result, rscale);
+
+ /* Strip leading and trailing zeroes */
+ strip_var(result);
+}
+
+/*
+ * mul_var_int() -
+ *
+ * Multiply a numeric variable by a 32-bit integer with the specified weight.
+ * The product var * ival * NBASE^ival_weight is stored in result.
+ */
+static void
+mul_var_int(const NumericVar *var, int ival, int ival_weight,
+ NumericVar *result, int rscale)
+{
+ NumericDigit *var_digits = var->digits;
+ int var_ndigits = var->ndigits;
+ int res_sign;
+ int res_weight;
+ int res_ndigits;
+ int maxdigits;
+ NumericDigit *res_buf;
+ NumericDigit *res_digits;
+ uint32 factor;
+ uint32 carry;
+
+ if (ival == 0 || var_ndigits == 0)
+ {
+ zero_var(result);
+ result->dscale = rscale;
+ return;
+ }
+
+ /*
+ * Determine the result sign, (maximum possible) weight and number of
+ * digits to calculate. The weight figured here is correct if the emitted
+ * product has no leading zero digits; otherwise strip_var() will fix
+ * things up.
+ */
+ if (var->sign == NUMERIC_POS)
+ res_sign = ival > 0 ? NUMERIC_POS : NUMERIC_NEG;
+ else
+ res_sign = ival > 0 ? NUMERIC_NEG : NUMERIC_POS;
+ res_weight = var->weight + ival_weight + 3;
+ /* The number of accurate result digits we need to produce: */
+ res_ndigits = var_ndigits + 3;
+ maxdigits = res_weight + 1 + (rscale + DEC_DIGITS - 1) / DEC_DIGITS +
+ MUL_GUARD_DIGITS;
+ res_ndigits = Min(res_ndigits, maxdigits);
+
+ if (res_ndigits < 3)
+ {
+ /* All input digits will be ignored; so result is zero */
+ zero_var(result);
+ result->dscale = rscale;
+ return;
+ }
+
+ res_buf = digitbuf_alloc(res_ndigits + 1);
+ res_buf[0] = 0; /* spare digit for later rounding */
+ res_digits = res_buf + 1;
+
+ /*
+ * Now compute the product digits by procssing the input digits in reverse
+ * and propagating the carry up as we go.
+ *
+ * In this algorithm, the carry from one digit to the next is at most
+ * factor - 1, and product is at most factor * NBASE - 1, and so it needs
+ * to be a 64-bit integer if this exceeds UINT_MAX.
+ */
+ factor = abs(ival);
+ carry = 0;
+
+ if (factor <= UINT_MAX / NBASE)
+ {
+ /* product cannot overflow 32 bits */
+ uint32 product;
+
+ for (int i = res_ndigits - 4; i >= 0; i--)
+ {
+ product = factor * var_digits[i] + carry;
+ res_digits[i + 3] = (NumericDigit) (product % NBASE);
+ carry = product / NBASE;
+ }
+ res_digits[2] = (NumericDigit) (carry % NBASE);
+ carry = carry / NBASE;
+ res_digits[1] = (NumericDigit) (carry % NBASE);
+ res_digits[0] = (NumericDigit) (carry / NBASE);
+ }
+ else
+ {
+ /* product may exceed 32 bits */
+ uint64 product;
+
+ for (int i = res_ndigits - 4; i >= 0; i--)
+ {
+ product = (uint64) factor * var_digits[i] + carry;
+ res_digits[i + 3] = (NumericDigit) (product % NBASE);
+ carry = (uint32) (product / NBASE);
+ }
+ res_digits[2] = (NumericDigit) (carry % NBASE);
+ carry = carry / NBASE;
+ res_digits[1] = (NumericDigit) (carry % NBASE);
+ res_digits[0] = (NumericDigit) (carry / NBASE);
+ }
+
+ /* Store the product in result */
+ digitbuf_free(result->buf);
+ result->ndigits = res_ndigits;
+ result->buf = res_buf;
+ result->digits = res_digits;
+ result->weight = res_weight;
+ result->sign = res_sign;
+
+ /* Round to target rscale (and set result->dscale) */
+ round_var(result, rscale);
+
+ /* Strip leading and trailing zeroes */
+ strip_var(result);
+}
+
+
+static void
+mul_var_patch_small(const NumericVar *var1, const NumericVar *var2, NumericVar *result,
+ int rscale)
+{
+ int res_ndigits;
+ int res_sign;
+ int res_weight;
+ int maxdigits;
+ int *dig;
+ int carry;
+ int maxdig;
+ int newdig;
+ int var1ndigits;
+ int var2ndigits;
+ NumericDigit *var1digits;
+ NumericDigit *var2digits;
+ NumericDigit *res_digits;
+ int i,
+ i1,
+ i2;
+
+ /*
+ * Arrange for var1 to be the shorter of the two numbers. This improves
+ * performance because the inner multiplication loop is much simpler than
+ * the outer loop, so it's better to have a smaller number of iterations
+ * of the outer loop. This also reduces the number of times that the
+ * accumulator array needs to be normalized.
+ */
+ if (var1->ndigits > var2->ndigits)
+ {
+ const NumericVar *tmp = var1;
+
+ var1 = var2;
+ var2 = tmp;
+ }
+
+ /* copy these values into local vars for speed in inner loop */
+ var1ndigits = var1->ndigits;
+ var2ndigits = var2->ndigits;
+ var1digits = var1->digits;
+ var2digits = var2->digits;
+
+ if (var1ndigits == 0)
+ {
+ /* one or both inputs is zero; so is result */
+ zero_var(result);
+ result->dscale = rscale;
+ return;
+ }
+
+ /*
+ * If var1 has 3 digits or fewer, delegate to mul_var_small() which uses a
+ * faster short multiplication algorithm.
+ */
+ if (var1ndigits <= 3)
+ {
+ mul_var_small(var1, var2, result, rscale);
+ return;
+ }
+
+ /* Determine result sign and (maximum possible) weight */
+ if (var1->sign == var2->sign)
+ res_sign = NUMERIC_POS;
+ else
+ res_sign = NUMERIC_NEG;
+ res_weight = var1->weight + var2->weight + 2;
+
+ /*
+ * Determine the number of result digits to compute. If the exact result
+ * would have more than rscale fractional digits, truncate the computation
+ * with MUL_GUARD_DIGITS guard digits, i.e., ignore input digits that
+ * would only contribute to the right of that. (This will give the exact
+ * rounded-to-rscale answer unless carries out of the ignored positions
+ * would have propagated through more than MUL_GUARD_DIGITS digits.)
+ *
+ * Note: an exact computation could not produce more than var1ndigits +
+ * var2ndigits digits, but we allocate one extra output digit in case
+ * rscale-driven rounding produces a carry out of the highest exact digit.
+ */
+ res_ndigits = var1ndigits + var2ndigits + 1;
+ maxdigits = res_weight + 1 + (rscale + DEC_DIGITS - 1) / DEC_DIGITS +
+ MUL_GUARD_DIGITS;
+ res_ndigits = Min(res_ndigits, maxdigits);
+
+ if (res_ndigits < 3)
+ {
+ /* All input digits will be ignored; so result is zero */
+ zero_var(result);
+ result->dscale = rscale;
+ return;
+ }
+
+ /*
+ * We do the arithmetic in an array "dig[]" of signed int's. Since
+ * INT_MAX is noticeably larger than NBASE*NBASE, this gives us headroom
+ * to avoid normalizing carries immediately.
+ *
+ * maxdig tracks the maximum possible value of any dig[] entry; when this
+ * threatens to exceed INT_MAX, we take the time to propagate carries.
+ * Furthermore, we need to ensure that overflow doesn't occur during the
+ * carry propagation passes either. The carry values could be as much as
+ * INT_MAX/NBASE, so really we must normalize when digits threaten to
+ * exceed INT_MAX - INT_MAX/NBASE.
+ *
+ * To avoid overflow in maxdig itself, it actually represents the max
+ * possible value divided by NBASE-1, ie, at the top of the loop it is
+ * known that no dig[] entry exceeds maxdig * (NBASE-1).
+ */
+ dig = (int *) palloc0(res_ndigits * sizeof(int));
+ maxdig = 0;
+
+ /*
+ * The least significant digits of var1 should be ignored if they don't
+ * contribute directly to the first res_ndigits digits of the result that
+ * we are computing.
+ *
+ * Digit i1 of var1 and digit i2 of var2 are multiplied and added to digit
+ * i1+i2+2 of the accumulator array, so we need only consider digits of
+ * var1 for which i1 <= res_ndigits - 3.
+ */
+ for (i1 = Min(var1ndigits - 1, res_ndigits - 3); i1 >= 0; i1--)
+ {
+ NumericDigit var1digit = var1digits[i1];
+
+ if (var1digit == 0)
+ continue;
+
+ /* Time to normalize? */
+ maxdig += var1digit;
+ if (maxdig > (INT_MAX - INT_MAX / NBASE) / (NBASE - 1))
+ {
+ /* Yes, do it */
+ carry = 0;
+ for (i = res_ndigits - 1; i >= 0; i--)
+ {
+ newdig = dig[i] + carry;
+ if (newdig >= NBASE)
+ {
+ carry = newdig / NBASE;
+ newdig -= carry * NBASE;
+ }
+ else
+ carry = 0;
+ dig[i] = newdig;
+ }
+ Assert(carry == 0);
+ /* Reset maxdig to indicate new worst-case */
+ maxdig = 1 + var1digit;
+ }
+
+ /*
+ * Add the appropriate multiple of var2 into the accumulator.
+ *
+ * As above, digits of var2 can be ignored if they don't contribute,
+ * so we only include digits for which i1+i2+2 < res_ndigits.
+ *
+ * This inner loop is the performance bottleneck for multiplication,
+ * so we want to keep it simple enough so that it can be
+ * auto-vectorized. Accordingly, process the digits left-to-right
+ * even though schoolbook multiplication would suggest right-to-left.
+ * Since we aren't propagating carries in this loop, the order does
+ * not matter.
+ */
+ {
+ int i2limit = Min(var2ndigits, res_ndigits - i1 - 2);
+ int *dig_i1_2 = &dig[i1 + 2];
+
+ for (i2 = 0; i2 < i2limit; i2++)
+ dig_i1_2[i2] += var1digit * var2digits[i2];
+ }
+ }
+
+ /*
+ * Now we do a final carry propagation pass to normalize the result, which
+ * we combine with storing the result digits into the output. Note that
+ * this is still done at full precision w/guard digits.
+ */
+ alloc_var(result, res_ndigits);
+ res_digits = result->digits;
+ carry = 0;
+ for (i = res_ndigits - 1; i >= 0; i--)
+ {
+ newdig = dig[i] + carry;
+ if (newdig >= NBASE)
+ {
+ carry = newdig / NBASE;
+ newdig -= carry * NBASE;
+ }
+ else
+ carry = 0;
+ res_digits[i] = newdig;
+ }
+ Assert(carry == 0);
+
+ pfree(dig);
+
+ /*
+ * Finally, round the result to the requested precision.
+ */
+ result->weight = res_weight;
+ result->sign = res_sign;
+
+ /* Round to target rscale (and set result->dscale) */
+ round_var(result, rscale);
+
+ /* Strip leading and trailing zeroes */
+ strip_var(result);
+}
+
+
+/*
+ * mul_var_small() -
+ *
+ * This has the same API as mul_var, but it assumes that var1 has no more
+ * than 3 digits and var2 has at least as many digits as var1. For variables
+ * satisfying these conditions, the product can be computed more quickly than
+ * the general algorithm used in mul_var.
+ */
+static void
+mul_var_small(const NumericVar *var1, const NumericVar *var2,
+ NumericVar *result, int rscale)
+{
+ int var1ndigits = var1->ndigits;
+ int var2ndigits = var2->ndigits;
+ NumericDigit *var1digits = var1->digits;
+ NumericDigit *var2digits = var2->digits;
+ int res_sign;
+ int res_weight;
+ int res_ndigits;
+ int maxdigits;
+ NumericDigit *res_buf;
+ NumericDigit *res_digits;
+ int carry;
+ int term;
+
+ /* Check preconditions */
+ Assert(var1ndigits <= 3);
+ Assert(var2ndigits >= var1ndigits);
+
+ /* Determine result sign and (maximum possible) weight */
+ if (var1->sign == var2->sign)
+ res_sign = NUMERIC_POS;
+ else
+ res_sign = NUMERIC_NEG;
+ res_weight = var1->weight + var2->weight + 2;
+
+ /* Determine the number of result digits to compute - see mul_var() */
+ res_ndigits = var1ndigits + var2ndigits + 1;
+ maxdigits = res_weight + 1 + (rscale + DEC_DIGITS - 1) / DEC_DIGITS +
+ MUL_GUARD_DIGITS;
+ res_ndigits = Min(res_ndigits, maxdigits);
+
+ if (res_ndigits < 3)
+ {
+ /* All input digits will be ignored; so result is zero */
+ zero_var(result);
+ result->dscale = rscale;
+ return;
+ }
+
+ /* Allocate result digit array */
+ res_buf = digitbuf_alloc(res_ndigits);
+ res_buf[0] = 0; /* spare digit for later rounding */
+ res_digits = res_buf + 1;
+
+ /*
+ * Compute the result digits in reverse, in one pass, propagating the
+ * carry up as we go.
+ *
+ * This computes res_digits[res_ndigits - 2], ... res_digits[0] by summing
+ * the products var1digits[i1] * var2digits[i2] for which i1 + i2 + 1 is
+ * the result index.
+ */
+ switch (var1ndigits)
+ {
+ case 1:
+ /* ---------
+ * 1-digit case:
+ * var1ndigits = 1
+ * var2ndigits >= 1
+ * 3 <= res_ndigits <= var2ndigits + 2
+ * ----------
+ */
+ carry = 0;
+ for (int i = res_ndigits - 3; i >= 0; i--)
+ {
+ term = (int) var1digits[0] * var2digits[i] + carry;
+ res_digits[i + 1] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+ }
+ res_digits[0] = (NumericDigit) carry;
+ break;
+
+ case 2:
+ /* ---------
+ * 2-digit case:
+ * var1ndigits = 2
+ * var2ndigits >= 2
+ * 3 <= res_ndigits <= var2ndigits + 3
+ * ----------
+ */
+ /* last result digit and carry */
+ term = 0;
+ if (res_ndigits - 3 < var2ndigits)
+ term += (int) var1digits[0] * var2digits[res_ndigits - 3];
+ if (res_ndigits > 3)
+ term += (int) var1digits[1] * var2digits[res_ndigits - 4];
+ res_digits[res_ndigits - 2] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+
+ /* remaining digits, except for the first two */
+ for (int i = res_ndigits - 4; i >= 1; i--)
+ {
+ term = (int) var1digits[0] * var2digits[i] +
+ (int) var1digits[1] * var2digits[i - 1] + carry;
+ res_digits[i + 1] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+ }
+
+ /* first two digits */
+ term = (int) var1digits[0] * var2digits[0] + carry;
+ res_digits[1] = (NumericDigit) (term % NBASE);
+ res_digits[0] = (NumericDigit) (term / NBASE);
+ break;
+
+ case 3:
+ /* ---------
+ * 3-digit case:
+ * var1ndigits = 3
+ * var2ndigits >= 3
+ * 3 <= res_ndigits <= var2ndigits + 4
+ * ----------
+ */
+ /* last result digit and carry */
+ term = 0;
+ if (res_ndigits - 3 < var2ndigits)
+ term += (int) var1digits[0] * var2digits[res_ndigits - 3];
+ if (res_ndigits > 3 && res_ndigits - 4 < var2ndigits)
+ term += (int) var1digits[1] * var2digits[res_ndigits - 4];
+ if (res_ndigits > 4)
+ term += (int) var1digits[2] * var2digits[res_ndigits - 5];
+ res_digits[res_ndigits - 2] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+
+ /* penultimate result digit */
+ term = carry;
+ if (res_ndigits > 3 && res_ndigits - 4 < var2ndigits)
+ term += (int) var1digits[0] * var2digits[res_ndigits - 4];
+ if (res_ndigits > 4)
+ term += (int) var1digits[1] * var2digits[res_ndigits - 5];
+ if (res_ndigits > 5)
+ term += (int) var1digits[2] * var2digits[res_ndigits - 6];
+ res_digits[res_ndigits - 3] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+
+ /* remaining digits, except for the first three */
+ for (int i = res_ndigits - 5; i >= 2; i--)
+ {
+ term = (int) var1digits[0] * var2digits[i] +
+ (int) var1digits[1] * var2digits[i - 1] +
+ (int) var1digits[2] * var2digits[i - 2] + carry;
+ res_digits[i + 1] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+ }
+
+ /* first three digits */
+ term = (int) var1digits[0] * var2digits[1] +
+ (int) var1digits[1] * var2digits[0] + carry;
+ res_digits[2] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+ term = (int) var1digits[0] * var2digits[0] + carry;
+ res_digits[1] = (NumericDigit) (term % NBASE);
+ res_digits[0] = (NumericDigit) (term / NBASE);
+ break;
+ }
+
+ /* Store the product in result (minus extra rounding digit) */
+ digitbuf_free(result->buf);
+ result->ndigits = res_ndigits - 1;
+ result->buf = res_buf;
+ result->digits = res_digits;
+ result->weight = res_weight - 1;
+ result->sign = res_sign;
+
+ /* Round to target rscale (and set result->dscale) */
+ round_var(result, rscale);
+
+ /* Strip leading and trailing zeroes */
+ strip_var(result);
+}
/*
* div_var() -
diff --git a/src/include/catalog/pg_proc.dat b/src/include/catalog/pg_proc.dat
index d4ac578ae6..c85ed20a99 100644
--- a/src/include/catalog/pg_proc.dat
+++ b/src/include/catalog/pg_proc.dat
@@ -4465,6 +4465,18 @@
{ oid => '1726',
proname => 'numeric_mul', prorettype => 'numeric',
proargtypes => 'numeric numeric', prosrc => 'numeric_mul' },
+{ oid => '6347',
+ proname => 'numeric_mul_head', prorettype => 'numeric',
+ proargtypes => 'numeric numeric int4', prosrc => 'numeric_mul_head' },
+{ oid => '6348',
+ proname => 'numeric_mul_patch_inline', prorettype => 'numeric',
+ proargtypes => 'numeric numeric int4', prosrc => 'numeric_mul_patch_inline' },
+{ oid => '6349',
+ proname => 'numeric_mul_patch_int', prorettype => 'numeric',
+ proargtypes => 'numeric numeric int4', prosrc => 'numeric_mul_patch_int' },
+{ oid => '6350',
+ proname => 'numeric_mul_patch_small', prorettype => 'numeric',
+ proargtypes => 'numeric numeric int4', prosrc => 'numeric_mul_patch_small' },
{ oid => '1727',
proname => 'numeric_div', prorettype => 'numeric',
proargtypes => 'numeric numeric', prosrc => 'numeric_div' },
diff --git a/src/include/utils/numeric.h b/src/include/utils/numeric.h
index 43c75c436f..2bc400c741 100644
--- a/src/include/utils/numeric.h
+++ b/src/include/utils/numeric.h
@@ -97,6 +97,14 @@ extern Numeric numeric_sub_opt_error(Numeric num1, Numeric num2,
bool *have_error);
extern Numeric numeric_mul_opt_error(Numeric num1, Numeric num2,
bool *have_error);
+extern Numeric numeric_mul_head_opt_error(Numeric num1, Numeric num2,
+ int32 rscale_adjustment, bool *have_error);
+extern Numeric numeric_mul_patch_inline_opt_error(Numeric num1, Numeric num2,
+ int32 rscale_adjustment, bool *have_error);
+extern Numeric numeric_mul_patch_int_opt_error(Numeric num1, Numeric num2,
+ int32 rscale_adjustment, bool *have_error);
+extern Numeric numeric_mul_patch_small_opt_error(Numeric num1, Numeric num2,
+ int32 rscale_adjustment, bool *have_error);
extern Numeric numeric_div_opt_error(Numeric num1, Numeric num2,
bool *have_error);
extern Numeric numeric_mod_opt_error(Numeric num1, Numeric num2,
diff --git a/test-mul-var.sql b/test-mul-var.sql
new file mode 100644
index 0000000000..cf7c55a089
--- /dev/null
+++ b/test-mul-var.sql
@@ -0,0 +1,50 @@
+CREATE TABLE test_numeric_mul_patched (
+ var1 numeric,
+ var2 numeric,
+ rscale_adjustment int,
+ result numeric
+);
+
+DO $$
+DECLARE
+var1 numeric;
+var2 numeric;
+BEGIN
+ FOR i IN 1..1000 LOOP
+ RAISE NOTICE '%', i;
+ FOR var1ndigits IN 1..4 LOOP
+ FOR var2ndigits IN 1..4 LOOP
+ FOR var1dscale IN 0..(var1ndigits*4) LOOP
+ FOR var2dscale IN 0..(var2ndigits*4) LOOP
+ FOR rscale_adjustment IN 0..(var1dscale+var2dscale) LOOP
+ var1 := round(random(
+ format('1%s',repeat('0',(var1ndigits-1)*4-1))::numeric,
+ format('%s',repeat('9',var1ndigits*4))::numeric
+ ) / 10::numeric^var1dscale, var1dscale);
+ var2 := round(random(
+ format('1%s',repeat('0',(var2ndigits-1)*4-1))::numeric,
+ format('%s',repeat('9',var2ndigits*4))::numeric
+ ) / 10::numeric^var2dscale, var2dscale);
+ INSERT INTO test_numeric_mul_patched
+ (var1, var2, rscale_adjustment)
+ VALUES
+ (var1, var2, -rscale_adjustment);
+ END LOOP;
+ END LOOP;
+ END LOOP;
+ END LOOP;
+ END LOOP;
+ END LOOP;
+END $$;
+
+UPDATE test_numeric_mul_patched SET result = numeric_mul_head(var1, var2, rscale_adjustment);
+
+SELECT
+ rscale_adjustment,
+ COUNT(*),
+ COUNT(*) FILTER (WHERE result IS DISTINCT FROM numeric_mul_patch_int(var1,var2,rscale_adjustment)) AS "mul_var_int",
+ COUNT(*) FILTER (WHERE result IS DISTINCT FROM numeric_mul_patch_small(var1,var2,rscale_adjustment)) AS "mul_var_small",
+ COUNT(*) FILTER (WHERE result IS DISTINCT FROM numeric_mul_patch_inline(var1,var2,rscale_adjustment)) AS "mul_var inlined"
+FROM test_numeric_mul_patched
+GROUP BY 1
+ORDER BY 1;
Attachments:
[text/plain] test-mul-var-versions.patch.txt (53.3K, ../../[email protected]/2-test-mul-var-versions.patch.txt)
download | inline diff:
diff --git a/src/backend/utils/adt/numeric.c b/src/backend/utils/adt/numeric.c
index 5510a203b0..63de5cd994 100644
--- a/src/backend/utils/adt/numeric.c
+++ b/src/backend/utils/adt/numeric.c
@@ -551,6 +551,23 @@ static void sub_var(const NumericVar *var1, const NumericVar *var2,
static void mul_var(const NumericVar *var1, const NumericVar *var2,
NumericVar *result,
int rscale);
+static void mul_var_head(const NumericVar *var1, const NumericVar *var2,
+ NumericVar *result,
+ int rscale);
+static void mul_var_patch_inline(const NumericVar *var1, const NumericVar *var2,
+ NumericVar *result,
+ int rscale);
+static void mul_var_patch_small(const NumericVar *var1, const NumericVar *var2,
+ NumericVar *result,
+ int rscale);
+static void mul_var_patch_int(const NumericVar *var1, const NumericVar *var2,
+ NumericVar *result,
+ int rscale);
+static void mul_var_small(const NumericVar *var1, const NumericVar *var2,
+ NumericVar *result, int rscale);
+
+static void mul_var_int(const NumericVar *var, int ival, int ival_weight,
+ NumericVar *result, int rscale);
static void div_var(const NumericVar *var1, const NumericVar *var2,
NumericVar *result,
int rscale, bool round);
@@ -3115,6 +3132,448 @@ numeric_mul_opt_error(Numeric num1, Numeric num2, bool *have_error)
}
+Datum
+numeric_mul_head(PG_FUNCTION_ARGS)
+{
+ Numeric num1 = PG_GETARG_NUMERIC(0);
+ Numeric num2 = PG_GETARG_NUMERIC(1);
+ int32 rscale_adjustment = PG_GETARG_INT32(2);
+ Numeric res;
+
+ res = numeric_mul_head_opt_error(num1, num2, rscale_adjustment, NULL);
+
+ PG_RETURN_NUMERIC(res);
+}
+
+
+Datum
+numeric_mul_patch_inline(PG_FUNCTION_ARGS)
+{
+ Numeric num1 = PG_GETARG_NUMERIC(0);
+ Numeric num2 = PG_GETARG_NUMERIC(1);
+ int32 rscale_adjustment = PG_GETARG_INT32(2);
+ Numeric res;
+
+ res = numeric_mul_patch_inline_opt_error(num1, num2, rscale_adjustment, NULL);
+
+ PG_RETURN_NUMERIC(res);
+}
+
+Datum
+numeric_mul_patch_small(PG_FUNCTION_ARGS)
+{
+ Numeric num1 = PG_GETARG_NUMERIC(0);
+ Numeric num2 = PG_GETARG_NUMERIC(1);
+ int32 rscale_adjustment = PG_GETARG_INT32(2);
+ Numeric res;
+
+ res = numeric_mul_patch_small_opt_error(num1, num2, rscale_adjustment, NULL);
+
+ PG_RETURN_NUMERIC(res);
+}
+
+Datum
+numeric_mul_patch_int(PG_FUNCTION_ARGS)
+{
+ Numeric num1 = PG_GETARG_NUMERIC(0);
+ Numeric num2 = PG_GETARG_NUMERIC(1);
+ int32 rscale_adjustment = PG_GETARG_INT32(2);
+ Numeric res;
+
+ res = numeric_mul_patch_int_opt_error(num1, num2, rscale_adjustment, NULL);
+
+ PG_RETURN_NUMERIC(res);
+}
+
+
+Numeric
+numeric_mul_patch_inline_opt_error(Numeric num1, Numeric num2, int32 rscale_adjustment, bool *have_error)
+{
+ NumericVar arg1;
+ NumericVar arg2;
+ NumericVar result;
+ Numeric res;
+
+ /*
+ * Handle NaN and infinities
+ */
+ if (NUMERIC_IS_SPECIAL(num1) || NUMERIC_IS_SPECIAL(num2))
+ {
+ if (NUMERIC_IS_NAN(num1) || NUMERIC_IS_NAN(num2))
+ return make_result(&const_nan);
+ if (NUMERIC_IS_PINF(num1))
+ {
+ switch (numeric_sign_internal(num2))
+ {
+ case 0:
+ return make_result(&const_nan); /* Inf * 0 */
+ case 1:
+ return make_result(&const_pinf);
+ case -1:
+ return make_result(&const_ninf);
+ }
+ Assert(false);
+ }
+ if (NUMERIC_IS_NINF(num1))
+ {
+ switch (numeric_sign_internal(num2))
+ {
+ case 0:
+ return make_result(&const_nan); /* -Inf * 0 */
+ case 1:
+ return make_result(&const_ninf);
+ case -1:
+ return make_result(&const_pinf);
+ }
+ Assert(false);
+ }
+ /* by here, num1 must be finite, so num2 is not */
+ if (NUMERIC_IS_PINF(num2))
+ {
+ switch (numeric_sign_internal(num1))
+ {
+ case 0:
+ return make_result(&const_nan); /* 0 * Inf */
+ case 1:
+ return make_result(&const_pinf);
+ case -1:
+ return make_result(&const_ninf);
+ }
+ Assert(false);
+ }
+ Assert(NUMERIC_IS_NINF(num2));
+ switch (numeric_sign_internal(num1))
+ {
+ case 0:
+ return make_result(&const_nan); /* 0 * -Inf */
+ case 1:
+ return make_result(&const_ninf);
+ case -1:
+ return make_result(&const_pinf);
+ }
+ Assert(false);
+ }
+
+ /*
+ * Unpack the values, let mul_var() compute the result and return it.
+ * Unlike add_var() and sub_var(), mul_var() will round its result. In the
+ * case of numeric_mul(), which is invoked for the * operator on numerics,
+ * we request exact representation for the product (rscale = sum(dscale of
+ * arg1, dscale of arg2)). If the exact result has more digits after the
+ * decimal point than can be stored in a numeric, we round it. Rounding
+ * after computing the exact result ensures that the final result is
+ * correctly rounded (rounding in mul_var() using a truncated product
+ * would not guarantee this).
+ */
+ init_var_from_num(num1, &arg1);
+ init_var_from_num(num2, &arg2);
+
+ init_var(&result);
+
+ mul_var_patch_inline(&arg1, &arg2, &result, arg1.dscale + arg2.dscale + rscale_adjustment);
+
+ if (result.dscale > NUMERIC_DSCALE_MAX)
+ round_var(&result, NUMERIC_DSCALE_MAX);
+
+ res = make_result_opt_error(&result, have_error);
+
+ free_var(&result);
+
+ return res;
+}
+
+
+Numeric
+numeric_mul_head_opt_error(Numeric num1, Numeric num2, int32 rscale_adjustment, bool *have_error)
+{
+ NumericVar arg1;
+ NumericVar arg2;
+ NumericVar result;
+ Numeric res;
+
+ /*
+ * Handle NaN and infinities
+ */
+ if (NUMERIC_IS_SPECIAL(num1) || NUMERIC_IS_SPECIAL(num2))
+ {
+ if (NUMERIC_IS_NAN(num1) || NUMERIC_IS_NAN(num2))
+ return make_result(&const_nan);
+ if (NUMERIC_IS_PINF(num1))
+ {
+ switch (numeric_sign_internal(num2))
+ {
+ case 0:
+ return make_result(&const_nan); /* Inf * 0 */
+ case 1:
+ return make_result(&const_pinf);
+ case -1:
+ return make_result(&const_ninf);
+ }
+ Assert(false);
+ }
+ if (NUMERIC_IS_NINF(num1))
+ {
+ switch (numeric_sign_internal(num2))
+ {
+ case 0:
+ return make_result(&const_nan); /* -Inf * 0 */
+ case 1:
+ return make_result(&const_ninf);
+ case -1:
+ return make_result(&const_pinf);
+ }
+ Assert(false);
+ }
+ /* by here, num1 must be finite, so num2 is not */
+ if (NUMERIC_IS_PINF(num2))
+ {
+ switch (numeric_sign_internal(num1))
+ {
+ case 0:
+ return make_result(&const_nan); /* 0 * Inf */
+ case 1:
+ return make_result(&const_pinf);
+ case -1:
+ return make_result(&const_ninf);
+ }
+ Assert(false);
+ }
+ Assert(NUMERIC_IS_NINF(num2));
+ switch (numeric_sign_internal(num1))
+ {
+ case 0:
+ return make_result(&const_nan); /* 0 * -Inf */
+ case 1:
+ return make_result(&const_ninf);
+ case -1:
+ return make_result(&const_pinf);
+ }
+ Assert(false);
+ }
+
+ /*
+ * Unpack the values, let mul_var() compute the result and return it.
+ * Unlike add_var() and sub_var(), mul_var() will round its result. In the
+ * case of numeric_mul(), which is invoked for the * operator on numerics,
+ * we request exact representation for the product (rscale = sum(dscale of
+ * arg1, dscale of arg2)). If the exact result has more digits after the
+ * decimal point than can be stored in a numeric, we round it. Rounding
+ * after computing the exact result ensures that the final result is
+ * correctly rounded (rounding in mul_var() using a truncated product
+ * would not guarantee this).
+ */
+ init_var_from_num(num1, &arg1);
+ init_var_from_num(num2, &arg2);
+
+ init_var(&result);
+
+ mul_var_head(&arg1, &arg2, &result, arg1.dscale + arg2.dscale + rscale_adjustment);
+
+ if (result.dscale > NUMERIC_DSCALE_MAX)
+ round_var(&result, NUMERIC_DSCALE_MAX);
+
+ res = make_result_opt_error(&result, have_error);
+
+ free_var(&result);
+
+ return res;
+}
+
+
+Numeric
+numeric_mul_patch_small_opt_error(Numeric num1, Numeric num2, int32 rscale_adjustment, bool *have_error)
+{
+ NumericVar arg1;
+ NumericVar arg2;
+ NumericVar result;
+ Numeric res;
+
+ /*
+ * Handle NaN and infinities
+ */
+ if (NUMERIC_IS_SPECIAL(num1) || NUMERIC_IS_SPECIAL(num2))
+ {
+ if (NUMERIC_IS_NAN(num1) || NUMERIC_IS_NAN(num2))
+ return make_result(&const_nan);
+ if (NUMERIC_IS_PINF(num1))
+ {
+ switch (numeric_sign_internal(num2))
+ {
+ case 0:
+ return make_result(&const_nan); /* Inf * 0 */
+ case 1:
+ return make_result(&const_pinf);
+ case -1:
+ return make_result(&const_ninf);
+ }
+ Assert(false);
+ }
+ if (NUMERIC_IS_NINF(num1))
+ {
+ switch (numeric_sign_internal(num2))
+ {
+ case 0:
+ return make_result(&const_nan); /* -Inf * 0 */
+ case 1:
+ return make_result(&const_ninf);
+ case -1:
+ return make_result(&const_pinf);
+ }
+ Assert(false);
+ }
+ /* by here, num1 must be finite, so num2 is not */
+ if (NUMERIC_IS_PINF(num2))
+ {
+ switch (numeric_sign_internal(num1))
+ {
+ case 0:
+ return make_result(&const_nan); /* 0 * Inf */
+ case 1:
+ return make_result(&const_pinf);
+ case -1:
+ return make_result(&const_ninf);
+ }
+ Assert(false);
+ }
+ Assert(NUMERIC_IS_NINF(num2));
+ switch (numeric_sign_internal(num1))
+ {
+ case 0:
+ return make_result(&const_nan); /* 0 * -Inf */
+ case 1:
+ return make_result(&const_ninf);
+ case -1:
+ return make_result(&const_pinf);
+ }
+ Assert(false);
+ }
+
+ /*
+ * Unpack the values, let mul_var() compute the result and return it.
+ * Unlike add_var() and sub_var(), mul_var() will round its result. In the
+ * case of numeric_mul(), which is invoked for the * operator on numerics,
+ * we request exact representation for the product (rscale = sum(dscale of
+ * arg1, dscale of arg2)). If the exact result has more digits after the
+ * decimal point than can be stored in a numeric, we round it. Rounding
+ * after computing the exact result ensures that the final result is
+ * correctly rounded (rounding in mul_var() using a truncated product
+ * would not guarantee this).
+ */
+ init_var_from_num(num1, &arg1);
+ init_var_from_num(num2, &arg2);
+
+ init_var(&result);
+
+ mul_var_patch_small(&arg1, &arg2, &result, arg1.dscale + arg2.dscale + rscale_adjustment);
+
+ if (result.dscale > NUMERIC_DSCALE_MAX)
+ round_var(&result, NUMERIC_DSCALE_MAX);
+
+ res = make_result_opt_error(&result, have_error);
+
+ free_var(&result);
+
+ return res;
+}
+
+
+Numeric
+numeric_mul_patch_int_opt_error(Numeric num1, Numeric num2, int32 rscale_adjustment, bool *have_error)
+{
+ NumericVar arg1;
+ NumericVar arg2;
+ NumericVar result;
+ Numeric res;
+
+ /*
+ * Handle NaN and infinities
+ */
+ if (NUMERIC_IS_SPECIAL(num1) || NUMERIC_IS_SPECIAL(num2))
+ {
+ if (NUMERIC_IS_NAN(num1) || NUMERIC_IS_NAN(num2))
+ return make_result(&const_nan);
+ if (NUMERIC_IS_PINF(num1))
+ {
+ switch (numeric_sign_internal(num2))
+ {
+ case 0:
+ return make_result(&const_nan); /* Inf * 0 */
+ case 1:
+ return make_result(&const_pinf);
+ case -1:
+ return make_result(&const_ninf);
+ }
+ Assert(false);
+ }
+ if (NUMERIC_IS_NINF(num1))
+ {
+ switch (numeric_sign_internal(num2))
+ {
+ case 0:
+ return make_result(&const_nan); /* -Inf * 0 */
+ case 1:
+ return make_result(&const_ninf);
+ case -1:
+ return make_result(&const_pinf);
+ }
+ Assert(false);
+ }
+ /* by here, num1 must be finite, so num2 is not */
+ if (NUMERIC_IS_PINF(num2))
+ {
+ switch (numeric_sign_internal(num1))
+ {
+ case 0:
+ return make_result(&const_nan); /* 0 * Inf */
+ case 1:
+ return make_result(&const_pinf);
+ case -1:
+ return make_result(&const_ninf);
+ }
+ Assert(false);
+ }
+ Assert(NUMERIC_IS_NINF(num2));
+ switch (numeric_sign_internal(num1))
+ {
+ case 0:
+ return make_result(&const_nan); /* 0 * -Inf */
+ case 1:
+ return make_result(&const_ninf);
+ case -1:
+ return make_result(&const_pinf);
+ }
+ Assert(false);
+ }
+
+ /*
+ * Unpack the values, let mul_var() compute the result and return it.
+ * Unlike add_var() and sub_var(), mul_var() will round its result. In the
+ * case of numeric_mul(), which is invoked for the * operator on numerics,
+ * we request exact representation for the product (rscale = sum(dscale of
+ * arg1, dscale of arg2)). If the exact result has more digits after the
+ * decimal point than can be stored in a numeric, we round it. Rounding
+ * after computing the exact result ensures that the final result is
+ * correctly rounded (rounding in mul_var() using a truncated product
+ * would not guarantee this).
+ */
+ init_var_from_num(num1, &arg1);
+ init_var_from_num(num2, &arg2);
+
+ init_var(&result);
+
+ mul_var_patch_int(&arg1, &arg2, &result, arg1.dscale + arg2.dscale + rscale_adjustment);
+
+ if (result.dscale > NUMERIC_DSCALE_MAX)
+ round_var(&result, NUMERIC_DSCALE_MAX);
+
+ res = make_result_opt_error(&result, have_error);
+
+ free_var(&result);
+
+ return res;
+}
+
+
/*
* numeric_div() -
*
@@ -8864,6 +9323,1245 @@ mul_var(const NumericVar *var1, const NumericVar *var2, NumericVar *result,
strip_var(result);
}
+static void
+mul_var_head(const NumericVar *var1, const NumericVar *var2, NumericVar *result,
+ int rscale)
+{
+ int res_ndigits;
+ int res_sign;
+ int res_weight;
+ int maxdigits;
+ int *dig;
+ int carry;
+ int maxdig;
+ int newdig;
+ int var1ndigits;
+ int var2ndigits;
+ NumericDigit *var1digits;
+ NumericDigit *var2digits;
+ NumericDigit *res_digits;
+ int i,
+ i1,
+ i2;
+
+ /*
+ * Arrange for var1 to be the shorter of the two numbers. This improves
+ * performance because the inner multiplication loop is much simpler than
+ * the outer loop, so it's better to have a smaller number of iterations
+ * of the outer loop. This also reduces the number of times that the
+ * accumulator array needs to be normalized.
+ */
+ if (var1->ndigits > var2->ndigits)
+ {
+ const NumericVar *tmp = var1;
+
+ var1 = var2;
+ var2 = tmp;
+ }
+
+ /* copy these values into local vars for speed in inner loop */
+ var1ndigits = var1->ndigits;
+ var2ndigits = var2->ndigits;
+ var1digits = var1->digits;
+ var2digits = var2->digits;
+
+ if (var1ndigits == 0 || var2ndigits == 0)
+ {
+ /* one or both inputs is zero; so is result */
+ zero_var(result);
+ result->dscale = rscale;
+ return;
+ }
+
+ /* Determine result sign and (maximum possible) weight */
+ if (var1->sign == var2->sign)
+ res_sign = NUMERIC_POS;
+ else
+ res_sign = NUMERIC_NEG;
+ res_weight = var1->weight + var2->weight + 2;
+
+ /*
+ * Determine the number of result digits to compute. If the exact result
+ * would have more than rscale fractional digits, truncate the computation
+ * with MUL_GUARD_DIGITS guard digits, i.e., ignore input digits that
+ * would only contribute to the right of that. (This will give the exact
+ * rounded-to-rscale answer unless carries out of the ignored positions
+ * would have propagated through more than MUL_GUARD_DIGITS digits.)
+ *
+ * Note: an exact computation could not produce more than var1ndigits +
+ * var2ndigits digits, but we allocate one extra output digit in case
+ * rscale-driven rounding produces a carry out of the highest exact digit.
+ */
+ res_ndigits = var1ndigits + var2ndigits + 1;
+ maxdigits = res_weight + 1 + (rscale + DEC_DIGITS - 1) / DEC_DIGITS +
+ MUL_GUARD_DIGITS;
+ res_ndigits = Min(res_ndigits, maxdigits);
+
+ if (res_ndigits < 3)
+ {
+ /* All input digits will be ignored; so result is zero */
+ zero_var(result);
+ result->dscale = rscale;
+ return;
+ }
+
+ /*
+ * We do the arithmetic in an array "dig[]" of signed int's. Since
+ * INT_MAX is noticeably larger than NBASE*NBASE, this gives us headroom
+ * to avoid normalizing carries immediately.
+ *
+ * maxdig tracks the maximum possible value of any dig[] entry; when this
+ * threatens to exceed INT_MAX, we take the time to propagate carries.
+ * Furthermore, we need to ensure that overflow doesn't occur during the
+ * carry propagation passes either. The carry values could be as much as
+ * INT_MAX/NBASE, so really we must normalize when digits threaten to
+ * exceed INT_MAX - INT_MAX/NBASE.
+ *
+ * To avoid overflow in maxdig itself, it actually represents the max
+ * possible value divided by NBASE-1, ie, at the top of the loop it is
+ * known that no dig[] entry exceeds maxdig * (NBASE-1).
+ */
+ dig = (int *) palloc0(res_ndigits * sizeof(int));
+ maxdig = 0;
+
+ /*
+ * The least significant digits of var1 should be ignored if they don't
+ * contribute directly to the first res_ndigits digits of the result that
+ * we are computing.
+ *
+ * Digit i1 of var1 and digit i2 of var2 are multiplied and added to digit
+ * i1+i2+2 of the accumulator array, so we need only consider digits of
+ * var1 for which i1 <= res_ndigits - 3.
+ */
+ for (i1 = Min(var1ndigits - 1, res_ndigits - 3); i1 >= 0; i1--)
+ {
+ NumericDigit var1digit = var1digits[i1];
+
+ if (var1digit == 0)
+ continue;
+
+ /* Time to normalize? */
+ maxdig += var1digit;
+ if (maxdig > (INT_MAX - INT_MAX / NBASE) / (NBASE - 1))
+ {
+ /* Yes, do it */
+ carry = 0;
+ for (i = res_ndigits - 1; i >= 0; i--)
+ {
+ newdig = dig[i] + carry;
+ if (newdig >= NBASE)
+ {
+ carry = newdig / NBASE;
+ newdig -= carry * NBASE;
+ }
+ else
+ carry = 0;
+ dig[i] = newdig;
+ }
+ Assert(carry == 0);
+ /* Reset maxdig to indicate new worst-case */
+ maxdig = 1 + var1digit;
+ }
+
+ /*
+ * Add the appropriate multiple of var2 into the accumulator.
+ *
+ * As above, digits of var2 can be ignored if they don't contribute,
+ * so we only include digits for which i1+i2+2 < res_ndigits.
+ *
+ * This inner loop is the performance bottleneck for multiplication,
+ * so we want to keep it simple enough so that it can be
+ * auto-vectorized. Accordingly, process the digits left-to-right
+ * even though schoolbook multiplication would suggest right-to-left.
+ * Since we aren't propagating carries in this loop, the order does
+ * not matter.
+ */
+ {
+ int i2limit = Min(var2ndigits, res_ndigits - i1 - 2);
+ int *dig_i1_2 = &dig[i1 + 2];
+
+ for (i2 = 0; i2 < i2limit; i2++)
+ dig_i1_2[i2] += var1digit * var2digits[i2];
+ }
+ }
+
+ /*
+ * Now we do a final carry propagation pass to normalize the result, which
+ * we combine with storing the result digits into the output. Note that
+ * this is still done at full precision w/guard digits.
+ */
+ alloc_var(result, res_ndigits);
+ res_digits = result->digits;
+ carry = 0;
+ for (i = res_ndigits - 1; i >= 0; i--)
+ {
+ newdig = dig[i] + carry;
+ if (newdig >= NBASE)
+ {
+ carry = newdig / NBASE;
+ newdig -= carry * NBASE;
+ }
+ else
+ carry = 0;
+ res_digits[i] = newdig;
+ }
+ Assert(carry == 0);
+
+ pfree(dig);
+
+ /*
+ * Finally, round the result to the requested precision.
+ */
+ result->weight = res_weight;
+ result->sign = res_sign;
+
+ /* Round to target rscale (and set result->dscale) */
+ round_var(result, rscale);
+
+ /* Strip leading and trailing zeroes */
+ strip_var(result);
+}
+
+static void
+mul_var_patch_inline(const NumericVar *var1, const NumericVar *var2,
+ NumericVar *result, int rscale)
+{
+ int res_ndigits;
+ int res_sign;
+ int res_weight;
+ int maxdigits;
+ int *dig;
+ int carry;
+ int maxdig;
+ int newdig;
+ int var1ndigits;
+ int var2ndigits;
+ NumericDigit *var1digits;
+ NumericDigit *var2digits;
+ NumericDigit *res_digits;
+ int i,
+ i1,
+ i2;
+
+ /*
+ * Arrange for var1 to be the shorter of the two numbers. This improves
+ * performance because the inner multiplication loop is much simpler than
+ * the outer loop, so it's better to have a smaller number of iterations
+ * of the outer loop. This also reduces the number of times that the
+ * accumulator array needs to be normalized.
+ */
+ if (var1->ndigits > var2->ndigits)
+ {
+ const NumericVar *tmp = var1;
+
+ var1 = var2;
+ var2 = tmp;
+ }
+
+ /* copy these values into local vars for speed in inner loop */
+ var1ndigits = var1->ndigits;
+ var2ndigits = var2->ndigits;
+ var1digits = var1->digits;
+ var2digits = var2->digits;
+
+ if (var1ndigits == 0 || var2ndigits == 0)
+ {
+ /* one or both inputs is zero; so is result */
+ zero_var(result);
+ result->dscale = rscale;
+ return;
+ }
+
+ /* Determine result sign and (maximum possible) weight */
+ if (var1->sign == var2->sign)
+ res_sign = NUMERIC_POS;
+ else
+ res_sign = NUMERIC_NEG;
+ res_weight = var1->weight + var2->weight + 2;
+
+ /*
+ * Determine the number of result digits to compute. If the exact result
+ * would have more than rscale fractional digits, truncate the computation
+ * with MUL_GUARD_DIGITS guard digits, i.e., ignore input digits that
+ * would only contribute to the right of that. (This will give the exact
+ * rounded-to-rscale answer unless carries out of the ignored positions
+ * would have propagated through more than MUL_GUARD_DIGITS digits.)
+ *
+ * Note: an exact computation could not produce more than var1ndigits +
+ * var2ndigits digits, but we allocate one extra output digit in case
+ * rscale-driven rounding produces a carry out of the highest exact digit.
+ */
+ res_ndigits = var1ndigits + var2ndigits + 1;
+ maxdigits = res_weight + 1 + (rscale + DEC_DIGITS - 1) / DEC_DIGITS +
+ MUL_GUARD_DIGITS;
+ res_ndigits = Min(res_ndigits, maxdigits);
+
+ if (res_ndigits < 3)
+ {
+ /* All input digits will be ignored; so result is zero */
+ zero_var(result);
+ result->dscale = rscale;
+ return;
+ }
+
+ /*
+ * Simplified fast-path computation, if var1 has just one or two digits.
+ * This is significantly faster, since it avoids allocating a separate
+ * digit array, making multiple passes over var2, and having separate
+ * carry-propagation passes.
+ */
+ if (var1ndigits <= 3)
+ {
+ NumericDigit *res_buf;
+
+ /* Allocate result digit array */
+ res_buf = digitbuf_alloc(res_ndigits);
+ res_buf[0] = 0; /* spare digit for later rounding */
+ res_digits = res_buf + 1;
+
+ /*
+ * Compute the result digits directly, in one pass, propagating the
+ * carry up as we go.
+ */
+ switch (var1ndigits)
+ {
+ case 1:
+ carry = 0;
+ for (i = res_ndigits - 3; i >= 0; i--)
+ {
+ newdig = (int) var1digits[0] * var2digits[i] + carry;
+ res_digits[i + 1] = (NumericDigit) (newdig % NBASE);
+ carry = newdig / NBASE;
+ }
+ res_digits[0] = (NumericDigit) carry;
+ break;
+
+ case 2:
+ if (res_ndigits - 4 >= 0 && res_ndigits - 4 < var2ndigits)
+ newdig = (int) var1digits[1] * var2digits[res_ndigits - 4];
+ else
+ newdig = 0;
+ if (res_ndigits - 3 >= 0 && res_ndigits - 3 < var2ndigits)
+ newdig += (int) var1digits[0] * var2digits[res_ndigits - 3];
+ res_digits[res_ndigits - 2] = (NumericDigit) (newdig % NBASE);
+ carry = newdig / NBASE;
+ for (i = res_ndigits - 4; i >= 1; i--)
+ {
+ newdig = (int) var1digits[0] * var2digits[i] +
+ (int) var1digits[1] * var2digits[i - 1] + carry;
+ res_digits[i + 1] = (NumericDigit) (newdig % NBASE);
+ carry = newdig / NBASE;
+ }
+ newdig = (int) var1digits[0] * var2digits[0] + carry;
+ res_digits[1] = (NumericDigit) (newdig % NBASE);
+ res_digits[0] = (NumericDigit) (newdig / NBASE);
+ break;
+
+ case 3:
+ if (res_ndigits - 5 >= 0 && res_ndigits - 5 < var2ndigits)
+ newdig = (int) var1digits[2] * var2digits[res_ndigits - 5];
+ else
+ newdig = 0;
+ if (res_ndigits - 4 >= 0 && res_ndigits - 4 < var2ndigits)
+ newdig += (int) var1digits[1] * var2digits[res_ndigits - 4];
+ if (res_ndigits - 3 >= 0 && res_ndigits - 3 < var2ndigits)
+ newdig += (int) var1digits[0] * var2digits[res_ndigits - 3];
+ res_digits[res_ndigits - 2] = (NumericDigit) (newdig % NBASE);
+ carry = newdig / NBASE;
+ for (i = res_ndigits - 4; i >= 2; i--)
+ {
+ newdig = carry;
+ if (i < var2ndigits)
+ newdig += (int) var1digits[0] * var2digits[i];
+ if (i - 1 >= 0 && i - 1 < var2ndigits)
+ newdig += (int) var1digits[1] * var2digits[i - 1];
+ if (i - 2 >= 0 && i - 2 < var2ndigits)
+ newdig += (int) var1digits[2] * var2digits[i - 2];
+ res_digits[i + 1] = (NumericDigit) (newdig % NBASE);
+ carry = newdig / NBASE;
+ }
+ newdig = carry;
+ if (var2ndigits > 1)
+ newdig += (int) var1digits[0] * var2digits[1];
+ if (var2ndigits > 0)
+ newdig += (int) var1digits[1] * var2digits[0];
+ res_digits[2] = (NumericDigit) (newdig % NBASE);
+ carry = newdig / NBASE;
+ newdig = (int) var1digits[0] * var2digits[0] + carry;
+ res_digits[1] = (NumericDigit) (newdig % NBASE);
+ res_digits[0] = (NumericDigit) (newdig / NBASE);
+ break;
+ }
+
+ /* Store the product in result (minus extra rounding digit) */
+ digitbuf_free(result->buf);
+ result->ndigits = res_ndigits - 1;
+ result->buf = res_buf;
+ result->digits = res_digits;
+ result->weight = res_weight - 1;
+ result->sign = res_sign;
+
+ /* Round to target rscale (and set result->dscale) */
+ round_var(result, rscale);
+
+ /* Strip leading and trailing zeroes */
+ strip_var(result);
+
+ return;
+ }
+
+ /*
+ * We do the arithmetic in an array "dig[]" of signed int's. Since
+ * INT_MAX is noticeably larger than NBASE*NBASE, this gives us headroom
+ * to avoid normalizing carries immediately.
+ *
+ * maxdig tracks the maximum possible value of any dig[] entry; when this
+ * threatens to exceed INT_MAX, we take the time to propagate carries.
+ * Furthermore, we need to ensure that overflow doesn't occur during the
+ * carry propagation passes either. The carry values could be as much as
+ * INT_MAX/NBASE, so really we must normalize when digits threaten to
+ * exceed INT_MAX - INT_MAX/NBASE.
+ *
+ * To avoid overflow in maxdig itself, it actually represents the max
+ * possible value divided by NBASE-1, ie, at the top of the loop it is
+ * known that no dig[] entry exceeds maxdig * (NBASE-1).
+ */
+ dig = (int *) palloc0(res_ndigits * sizeof(int));
+ maxdig = 0;
+
+ /*
+ * The least significant digits of var1 should be ignored if they don't
+ * contribute directly to the first res_ndigits digits of the result that
+ * we are computing.
+ *
+ * Digit i1 of var1 and digit i2 of var2 are multiplied and added to digit
+ * i1+i2+2 of the accumulator array, so we need only consider digits of
+ * var1 for which i1 <= res_ndigits - 3.
+ */
+ for (i1 = Min(var1ndigits - 1, res_ndigits - 3); i1 >= 0; i1--)
+ {
+ NumericDigit var1digit = var1digits[i1];
+
+ if (var1digit == 0)
+ continue;
+
+ /* Time to normalize? */
+ maxdig += var1digit;
+ if (maxdig > (INT_MAX - INT_MAX / NBASE) / (NBASE - 1))
+ {
+ /* Yes, do it */
+ carry = 0;
+ for (i = res_ndigits - 1; i >= 0; i--)
+ {
+ newdig = dig[i] + carry;
+ if (newdig >= NBASE)
+ {
+ carry = newdig / NBASE;
+ newdig -= carry * NBASE;
+ }
+ else
+ carry = 0;
+ dig[i] = newdig;
+ }
+ Assert(carry == 0);
+ /* Reset maxdig to indicate new worst-case */
+ maxdig = 1 + var1digit;
+ }
+
+ /*
+ * Add the appropriate multiple of var2 into the accumulator.
+ *
+ * As above, digits of var2 can be ignored if they don't contribute,
+ * so we only include digits for which i1+i2+2 < res_ndigits.
+ *
+ * This inner loop is the performance bottleneck for multiplication,
+ * so we want to keep it simple enough so that it can be
+ * auto-vectorized. Accordingly, process the digits left-to-right
+ * even though schoolbook multiplication would suggest right-to-left.
+ * Since we aren't propagating carries in this loop, the order does
+ * not matter.
+ */
+ {
+ int i2limit = Min(var2ndigits, res_ndigits - i1 - 2);
+ int *dig_i1_2 = &dig[i1 + 2];
+
+ for (i2 = 0; i2 < i2limit; i2++)
+ {
+ dig_i1_2[i2] += var1digit * var2digits[i2];
+ }
+ }
+ }
+
+ /*
+ * Now we do a final carry propagation pass to normalize the result, which
+ * we combine with storing the result digits into the output. Note that
+ * this is still done at full precision w/guard digits.
+ */
+ alloc_var(result, res_ndigits);
+ res_digits = result->digits;
+ carry = 0;
+ for (i = res_ndigits - 1; i >= 0; i--)
+ {
+ newdig = dig[i] + carry;
+ if (newdig >= NBASE)
+ {
+ carry = newdig / NBASE;
+ newdig -= carry * NBASE;
+ }
+ else
+ carry = 0;
+ res_digits[i] = newdig;
+ }
+ Assert(carry == 0);
+
+ pfree(dig);
+
+ /*
+ * Finally, round the result to the requested precision.
+ */
+ result->weight = res_weight;
+ result->sign = res_sign;
+
+ /* Round to target rscale (and set result->dscale) */
+ round_var(result, rscale);
+
+ /* Strip leading and trailing zeroes */
+ strip_var(result);
+
+}
+
+
+static void
+mul_var_patch_int(const NumericVar *var1, const NumericVar *var2, NumericVar *result,
+ int rscale)
+{
+ int res_ndigits;
+ int res_sign;
+ int res_weight;
+ int maxdigits;
+ int *dig;
+ int carry;
+ int maxdig;
+ int newdig;
+ int var1ndigits;
+ int var2ndigits;
+ NumericDigit *var1digits;
+ NumericDigit *var2digits;
+ NumericDigit *res_digits;
+ int i,
+ i1,
+ i2;
+
+ /*
+ * Arrange for var1 to be the shorter of the two numbers. This improves
+ * performance because the inner multiplication loop is much simpler than
+ * the outer loop, so it's better to have a smaller number of iterations
+ * of the outer loop. This also reduces the number of times that the
+ * accumulator array needs to be normalized.
+ */
+ if (var1->ndigits > var2->ndigits)
+ {
+ const NumericVar *tmp = var1;
+
+ var1 = var2;
+ var2 = tmp;
+ }
+
+ /* copy these values into local vars for speed in inner loop */
+ var1ndigits = var1->ndigits;
+ var2ndigits = var2->ndigits;
+ var1digits = var1->digits;
+ var2digits = var2->digits;
+
+ if (var1ndigits == 0)
+ {
+ /* one or both inputs is zero; so is result */
+ zero_var(result);
+ result->dscale = rscale;
+ return;
+ }
+
+ /*
+ * If var1 has just one or two digits, delegate to mul_var_int(), which
+ * uses a faster direct multiplication algorithm.
+ *
+ * TODO: Similarly, on platforms with 128-bit integers ...
+ */
+ if (var1ndigits <= 2)
+ {
+ int ifactor;
+ int ifactor_weight;
+
+ ifactor = var1->digits[0];
+ ifactor_weight = var1->weight;
+ if (var1ndigits == 2)
+ {
+ ifactor = ifactor * NBASE + var1->digits[1];
+ ifactor_weight--;
+ }
+ if (var1->sign == NUMERIC_NEG)
+ ifactor = -ifactor;
+
+ mul_var_int(var2, ifactor, ifactor_weight, result, rscale);
+ return;
+ }
+
+ /* Determine result sign and (maximum possible) weight */
+ if (var1->sign == var2->sign)
+ res_sign = NUMERIC_POS;
+ else
+ res_sign = NUMERIC_NEG;
+ res_weight = var1->weight + var2->weight + 2;
+
+ /*
+ * Determine the number of result digits to compute. If the exact result
+ * would have more than rscale fractional digits, truncate the computation
+ * with MUL_GUARD_DIGITS guard digits, i.e., ignore input digits that
+ * would only contribute to the right of that. (This will give the exact
+ * rounded-to-rscale answer unless carries out of the ignored positions
+ * would have propagated through more than MUL_GUARD_DIGITS digits.)
+ *
+ * Note: an exact computation could not produce more than var1ndigits +
+ * var2ndigits digits, but we allocate one extra output digit in case
+ * rscale-driven rounding produces a carry out of the highest exact digit.
+ */
+ res_ndigits = var1ndigits + var2ndigits + 1;
+ maxdigits = res_weight + 1 + (rscale + DEC_DIGITS - 1) / DEC_DIGITS +
+ MUL_GUARD_DIGITS;
+ res_ndigits = Min(res_ndigits, maxdigits);
+
+ if (res_ndigits < 3)
+ {
+ /* All input digits will be ignored; so result is zero */
+ zero_var(result);
+ result->dscale = rscale;
+ return;
+ }
+
+ /*
+ * We do the arithmetic in an array "dig[]" of signed int's. Since
+ * INT_MAX is noticeably larger than NBASE*NBASE, this gives us headroom
+ * to avoid normalizing carries immediately.
+ *
+ * maxdig tracks the maximum possible value of any dig[] entry; when this
+ * threatens to exceed INT_MAX, we take the time to propagate carries.
+ * Furthermore, we need to ensure that overflow doesn't occur during the
+ * carry propagation passes either. The carry values could be as much as
+ * INT_MAX/NBASE, so really we must normalize when digits threaten to
+ * exceed INT_MAX - INT_MAX/NBASE.
+ *
+ * To avoid overflow in maxdig itself, it actually represents the max
+ * possible value divided by NBASE-1, ie, at the top of the loop it is
+ * known that no dig[] entry exceeds maxdig * (NBASE-1).
+ */
+ dig = (int *) palloc0(res_ndigits * sizeof(int));
+ maxdig = 0;
+
+ /*
+ * The least significant digits of var1 should be ignored if they don't
+ * contribute directly to the first res_ndigits digits of the result that
+ * we are computing.
+ *
+ * Digit i1 of var1 and digit i2 of var2 are multiplied and added to digit
+ * i1+i2+2 of the accumulator array, so we need only consider digits of
+ * var1 for which i1 <= res_ndigits - 3.
+ */
+ for (i1 = Min(var1ndigits - 1, res_ndigits - 3); i1 >= 0; i1--)
+ {
+ NumericDigit var1digit = var1digits[i1];
+
+ if (var1digit == 0)
+ continue;
+
+ /* Time to normalize? */
+ maxdig += var1digit;
+ if (maxdig > (INT_MAX - INT_MAX / NBASE) / (NBASE - 1))
+ {
+ /* Yes, do it */
+ carry = 0;
+ for (i = res_ndigits - 1; i >= 0; i--)
+ {
+ newdig = dig[i] + carry;
+ if (newdig >= NBASE)
+ {
+ carry = newdig / NBASE;
+ newdig -= carry * NBASE;
+ }
+ else
+ carry = 0;
+ dig[i] = newdig;
+ }
+ Assert(carry == 0);
+ /* Reset maxdig to indicate new worst-case */
+ maxdig = 1 + var1digit;
+ }
+
+ /*
+ * Add the appropriate multiple of var2 into the accumulator.
+ *
+ * As above, digits of var2 can be ignored if they don't contribute,
+ * so we only include digits for which i1+i2+2 < res_ndigits.
+ *
+ * This inner loop is the performance bottleneck for multiplication,
+ * so we want to keep it simple enough so that it can be
+ * auto-vectorized. Accordingly, process the digits left-to-right
+ * even though schoolbook multiplication would suggest right-to-left.
+ * Since we aren't propagating carries in this loop, the order does
+ * not matter.
+ */
+ {
+ int i2limit = Min(var2ndigits, res_ndigits - i1 - 2);
+ int *dig_i1_2 = &dig[i1 + 2];
+
+ for (i2 = 0; i2 < i2limit; i2++)
+ dig_i1_2[i2] += var1digit * var2digits[i2];
+ }
+ }
+
+ /*
+ * Now we do a final carry propagation pass to normalize the result, which
+ * we combine with storing the result digits into the output. Note that
+ * this is still done at full precision w/guard digits.
+ */
+ alloc_var(result, res_ndigits);
+ res_digits = result->digits;
+ carry = 0;
+ for (i = res_ndigits - 1; i >= 0; i--)
+ {
+ newdig = dig[i] + carry;
+ if (newdig >= NBASE)
+ {
+ carry = newdig / NBASE;
+ newdig -= carry * NBASE;
+ }
+ else
+ carry = 0;
+ res_digits[i] = newdig;
+ }
+ Assert(carry == 0);
+
+ pfree(dig);
+
+ /*
+ * Finally, round the result to the requested precision.
+ */
+ result->weight = res_weight;
+ result->sign = res_sign;
+
+ /* Round to target rscale (and set result->dscale) */
+ round_var(result, rscale);
+
+ /* Strip leading and trailing zeroes */
+ strip_var(result);
+}
+
+/*
+ * mul_var_int() -
+ *
+ * Multiply a numeric variable by a 32-bit integer with the specified weight.
+ * The product var * ival * NBASE^ival_weight is stored in result.
+ */
+static void
+mul_var_int(const NumericVar *var, int ival, int ival_weight,
+ NumericVar *result, int rscale)
+{
+ NumericDigit *var_digits = var->digits;
+ int var_ndigits = var->ndigits;
+ int res_sign;
+ int res_weight;
+ int res_ndigits;
+ int maxdigits;
+ NumericDigit *res_buf;
+ NumericDigit *res_digits;
+ uint32 factor;
+ uint32 carry;
+
+ if (ival == 0 || var_ndigits == 0)
+ {
+ zero_var(result);
+ result->dscale = rscale;
+ return;
+ }
+
+ /*
+ * Determine the result sign, (maximum possible) weight and number of
+ * digits to calculate. The weight figured here is correct if the emitted
+ * product has no leading zero digits; otherwise strip_var() will fix
+ * things up.
+ */
+ if (var->sign == NUMERIC_POS)
+ res_sign = ival > 0 ? NUMERIC_POS : NUMERIC_NEG;
+ else
+ res_sign = ival > 0 ? NUMERIC_NEG : NUMERIC_POS;
+ res_weight = var->weight + ival_weight + 3;
+ /* The number of accurate result digits we need to produce: */
+ res_ndigits = var_ndigits + 3;
+ maxdigits = res_weight + 1 + (rscale + DEC_DIGITS - 1) / DEC_DIGITS +
+ MUL_GUARD_DIGITS;
+ res_ndigits = Min(res_ndigits, maxdigits);
+
+ if (res_ndigits < 3)
+ {
+ /* All input digits will be ignored; so result is zero */
+ zero_var(result);
+ result->dscale = rscale;
+ return;
+ }
+
+ res_buf = digitbuf_alloc(res_ndigits + 1);
+ res_buf[0] = 0; /* spare digit for later rounding */
+ res_digits = res_buf + 1;
+
+ /*
+ * Now compute the product digits by procssing the input digits in reverse
+ * and propagating the carry up as we go.
+ *
+ * In this algorithm, the carry from one digit to the next is at most
+ * factor - 1, and product is at most factor * NBASE - 1, and so it needs
+ * to be a 64-bit integer if this exceeds UINT_MAX.
+ */
+ factor = abs(ival);
+ carry = 0;
+
+ if (factor <= UINT_MAX / NBASE)
+ {
+ /* product cannot overflow 32 bits */
+ uint32 product;
+
+ for (int i = res_ndigits - 4; i >= 0; i--)
+ {
+ product = factor * var_digits[i] + carry;
+ res_digits[i + 3] = (NumericDigit) (product % NBASE);
+ carry = product / NBASE;
+ }
+ res_digits[2] = (NumericDigit) (carry % NBASE);
+ carry = carry / NBASE;
+ res_digits[1] = (NumericDigit) (carry % NBASE);
+ res_digits[0] = (NumericDigit) (carry / NBASE);
+ }
+ else
+ {
+ /* product may exceed 32 bits */
+ uint64 product;
+
+ for (int i = res_ndigits - 4; i >= 0; i--)
+ {
+ product = (uint64) factor * var_digits[i] + carry;
+ res_digits[i + 3] = (NumericDigit) (product % NBASE);
+ carry = (uint32) (product / NBASE);
+ }
+ res_digits[2] = (NumericDigit) (carry % NBASE);
+ carry = carry / NBASE;
+ res_digits[1] = (NumericDigit) (carry % NBASE);
+ res_digits[0] = (NumericDigit) (carry / NBASE);
+ }
+
+ /* Store the product in result */
+ digitbuf_free(result->buf);
+ result->ndigits = res_ndigits;
+ result->buf = res_buf;
+ result->digits = res_digits;
+ result->weight = res_weight;
+ result->sign = res_sign;
+
+ /* Round to target rscale (and set result->dscale) */
+ round_var(result, rscale);
+
+ /* Strip leading and trailing zeroes */
+ strip_var(result);
+}
+
+
+static void
+mul_var_patch_small(const NumericVar *var1, const NumericVar *var2, NumericVar *result,
+ int rscale)
+{
+ int res_ndigits;
+ int res_sign;
+ int res_weight;
+ int maxdigits;
+ int *dig;
+ int carry;
+ int maxdig;
+ int newdig;
+ int var1ndigits;
+ int var2ndigits;
+ NumericDigit *var1digits;
+ NumericDigit *var2digits;
+ NumericDigit *res_digits;
+ int i,
+ i1,
+ i2;
+
+ /*
+ * Arrange for var1 to be the shorter of the two numbers. This improves
+ * performance because the inner multiplication loop is much simpler than
+ * the outer loop, so it's better to have a smaller number of iterations
+ * of the outer loop. This also reduces the number of times that the
+ * accumulator array needs to be normalized.
+ */
+ if (var1->ndigits > var2->ndigits)
+ {
+ const NumericVar *tmp = var1;
+
+ var1 = var2;
+ var2 = tmp;
+ }
+
+ /* copy these values into local vars for speed in inner loop */
+ var1ndigits = var1->ndigits;
+ var2ndigits = var2->ndigits;
+ var1digits = var1->digits;
+ var2digits = var2->digits;
+
+ if (var1ndigits == 0)
+ {
+ /* one or both inputs is zero; so is result */
+ zero_var(result);
+ result->dscale = rscale;
+ return;
+ }
+
+ /*
+ * If var1 has 3 digits or fewer, delegate to mul_var_small() which uses a
+ * faster short multiplication algorithm.
+ */
+ if (var1ndigits <= 3)
+ {
+ mul_var_small(var1, var2, result, rscale);
+ return;
+ }
+
+ /* Determine result sign and (maximum possible) weight */
+ if (var1->sign == var2->sign)
+ res_sign = NUMERIC_POS;
+ else
+ res_sign = NUMERIC_NEG;
+ res_weight = var1->weight + var2->weight + 2;
+
+ /*
+ * Determine the number of result digits to compute. If the exact result
+ * would have more than rscale fractional digits, truncate the computation
+ * with MUL_GUARD_DIGITS guard digits, i.e., ignore input digits that
+ * would only contribute to the right of that. (This will give the exact
+ * rounded-to-rscale answer unless carries out of the ignored positions
+ * would have propagated through more than MUL_GUARD_DIGITS digits.)
+ *
+ * Note: an exact computation could not produce more than var1ndigits +
+ * var2ndigits digits, but we allocate one extra output digit in case
+ * rscale-driven rounding produces a carry out of the highest exact digit.
+ */
+ res_ndigits = var1ndigits + var2ndigits + 1;
+ maxdigits = res_weight + 1 + (rscale + DEC_DIGITS - 1) / DEC_DIGITS +
+ MUL_GUARD_DIGITS;
+ res_ndigits = Min(res_ndigits, maxdigits);
+
+ if (res_ndigits < 3)
+ {
+ /* All input digits will be ignored; so result is zero */
+ zero_var(result);
+ result->dscale = rscale;
+ return;
+ }
+
+ /*
+ * We do the arithmetic in an array "dig[]" of signed int's. Since
+ * INT_MAX is noticeably larger than NBASE*NBASE, this gives us headroom
+ * to avoid normalizing carries immediately.
+ *
+ * maxdig tracks the maximum possible value of any dig[] entry; when this
+ * threatens to exceed INT_MAX, we take the time to propagate carries.
+ * Furthermore, we need to ensure that overflow doesn't occur during the
+ * carry propagation passes either. The carry values could be as much as
+ * INT_MAX/NBASE, so really we must normalize when digits threaten to
+ * exceed INT_MAX - INT_MAX/NBASE.
+ *
+ * To avoid overflow in maxdig itself, it actually represents the max
+ * possible value divided by NBASE-1, ie, at the top of the loop it is
+ * known that no dig[] entry exceeds maxdig * (NBASE-1).
+ */
+ dig = (int *) palloc0(res_ndigits * sizeof(int));
+ maxdig = 0;
+
+ /*
+ * The least significant digits of var1 should be ignored if they don't
+ * contribute directly to the first res_ndigits digits of the result that
+ * we are computing.
+ *
+ * Digit i1 of var1 and digit i2 of var2 are multiplied and added to digit
+ * i1+i2+2 of the accumulator array, so we need only consider digits of
+ * var1 for which i1 <= res_ndigits - 3.
+ */
+ for (i1 = Min(var1ndigits - 1, res_ndigits - 3); i1 >= 0; i1--)
+ {
+ NumericDigit var1digit = var1digits[i1];
+
+ if (var1digit == 0)
+ continue;
+
+ /* Time to normalize? */
+ maxdig += var1digit;
+ if (maxdig > (INT_MAX - INT_MAX / NBASE) / (NBASE - 1))
+ {
+ /* Yes, do it */
+ carry = 0;
+ for (i = res_ndigits - 1; i >= 0; i--)
+ {
+ newdig = dig[i] + carry;
+ if (newdig >= NBASE)
+ {
+ carry = newdig / NBASE;
+ newdig -= carry * NBASE;
+ }
+ else
+ carry = 0;
+ dig[i] = newdig;
+ }
+ Assert(carry == 0);
+ /* Reset maxdig to indicate new worst-case */
+ maxdig = 1 + var1digit;
+ }
+
+ /*
+ * Add the appropriate multiple of var2 into the accumulator.
+ *
+ * As above, digits of var2 can be ignored if they don't contribute,
+ * so we only include digits for which i1+i2+2 < res_ndigits.
+ *
+ * This inner loop is the performance bottleneck for multiplication,
+ * so we want to keep it simple enough so that it can be
+ * auto-vectorized. Accordingly, process the digits left-to-right
+ * even though schoolbook multiplication would suggest right-to-left.
+ * Since we aren't propagating carries in this loop, the order does
+ * not matter.
+ */
+ {
+ int i2limit = Min(var2ndigits, res_ndigits - i1 - 2);
+ int *dig_i1_2 = &dig[i1 + 2];
+
+ for (i2 = 0; i2 < i2limit; i2++)
+ dig_i1_2[i2] += var1digit * var2digits[i2];
+ }
+ }
+
+ /*
+ * Now we do a final carry propagation pass to normalize the result, which
+ * we combine with storing the result digits into the output. Note that
+ * this is still done at full precision w/guard digits.
+ */
+ alloc_var(result, res_ndigits);
+ res_digits = result->digits;
+ carry = 0;
+ for (i = res_ndigits - 1; i >= 0; i--)
+ {
+ newdig = dig[i] + carry;
+ if (newdig >= NBASE)
+ {
+ carry = newdig / NBASE;
+ newdig -= carry * NBASE;
+ }
+ else
+ carry = 0;
+ res_digits[i] = newdig;
+ }
+ Assert(carry == 0);
+
+ pfree(dig);
+
+ /*
+ * Finally, round the result to the requested precision.
+ */
+ result->weight = res_weight;
+ result->sign = res_sign;
+
+ /* Round to target rscale (and set result->dscale) */
+ round_var(result, rscale);
+
+ /* Strip leading and trailing zeroes */
+ strip_var(result);
+}
+
+
+/*
+ * mul_var_small() -
+ *
+ * This has the same API as mul_var, but it assumes that var1 has no more
+ * than 3 digits and var2 has at least as many digits as var1. For variables
+ * satisfying these conditions, the product can be computed more quickly than
+ * the general algorithm used in mul_var.
+ */
+static void
+mul_var_small(const NumericVar *var1, const NumericVar *var2,
+ NumericVar *result, int rscale)
+{
+ int var1ndigits = var1->ndigits;
+ int var2ndigits = var2->ndigits;
+ NumericDigit *var1digits = var1->digits;
+ NumericDigit *var2digits = var2->digits;
+ int res_sign;
+ int res_weight;
+ int res_ndigits;
+ int maxdigits;
+ NumericDigit *res_buf;
+ NumericDigit *res_digits;
+ int carry;
+ int term;
+
+ /* Check preconditions */
+ Assert(var1ndigits <= 3);
+ Assert(var2ndigits >= var1ndigits);
+
+ /* Determine result sign and (maximum possible) weight */
+ if (var1->sign == var2->sign)
+ res_sign = NUMERIC_POS;
+ else
+ res_sign = NUMERIC_NEG;
+ res_weight = var1->weight + var2->weight + 2;
+
+ /* Determine the number of result digits to compute - see mul_var() */
+ res_ndigits = var1ndigits + var2ndigits + 1;
+ maxdigits = res_weight + 1 + (rscale + DEC_DIGITS - 1) / DEC_DIGITS +
+ MUL_GUARD_DIGITS;
+ res_ndigits = Min(res_ndigits, maxdigits);
+
+ if (res_ndigits < 3)
+ {
+ /* All input digits will be ignored; so result is zero */
+ zero_var(result);
+ result->dscale = rscale;
+ return;
+ }
+
+ /* Allocate result digit array */
+ res_buf = digitbuf_alloc(res_ndigits);
+ res_buf[0] = 0; /* spare digit for later rounding */
+ res_digits = res_buf + 1;
+
+ /*
+ * Compute the result digits in reverse, in one pass, propagating the
+ * carry up as we go.
+ *
+ * This computes res_digits[res_ndigits - 2], ... res_digits[0] by summing
+ * the products var1digits[i1] * var2digits[i2] for which i1 + i2 + 1 is
+ * the result index.
+ */
+ switch (var1ndigits)
+ {
+ case 1:
+ /* ---------
+ * 1-digit case:
+ * var1ndigits = 1
+ * var2ndigits >= 1
+ * 3 <= res_ndigits <= var2ndigits + 2
+ * ----------
+ */
+ carry = 0;
+ for (int i = res_ndigits - 3; i >= 0; i--)
+ {
+ term = (int) var1digits[0] * var2digits[i] + carry;
+ res_digits[i + 1] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+ }
+ res_digits[0] = (NumericDigit) carry;
+ break;
+
+ case 2:
+ /* ---------
+ * 2-digit case:
+ * var1ndigits = 2
+ * var2ndigits >= 2
+ * 3 <= res_ndigits <= var2ndigits + 3
+ * ----------
+ */
+ /* last result digit and carry */
+ term = 0;
+ if (res_ndigits - 3 < var2ndigits)
+ term += (int) var1digits[0] * var2digits[res_ndigits - 3];
+ if (res_ndigits > 3)
+ term += (int) var1digits[1] * var2digits[res_ndigits - 4];
+ res_digits[res_ndigits - 2] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+
+ /* remaining digits, except for the first two */
+ for (int i = res_ndigits - 4; i >= 1; i--)
+ {
+ term = (int) var1digits[0] * var2digits[i] +
+ (int) var1digits[1] * var2digits[i - 1] + carry;
+ res_digits[i + 1] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+ }
+
+ /* first two digits */
+ term = (int) var1digits[0] * var2digits[0] + carry;
+ res_digits[1] = (NumericDigit) (term % NBASE);
+ res_digits[0] = (NumericDigit) (term / NBASE);
+ break;
+
+ case 3:
+ /* ---------
+ * 3-digit case:
+ * var1ndigits = 3
+ * var2ndigits >= 3
+ * 3 <= res_ndigits <= var2ndigits + 4
+ * ----------
+ */
+ /* last result digit and carry */
+ term = 0;
+ if (res_ndigits - 3 < var2ndigits)
+ term += (int) var1digits[0] * var2digits[res_ndigits - 3];
+ if (res_ndigits > 3 && res_ndigits - 4 < var2ndigits)
+ term += (int) var1digits[1] * var2digits[res_ndigits - 4];
+ if (res_ndigits > 4)
+ term += (int) var1digits[2] * var2digits[res_ndigits - 5];
+ res_digits[res_ndigits - 2] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+
+ /* penultimate result digit */
+ term = carry;
+ if (res_ndigits > 3 && res_ndigits - 4 < var2ndigits)
+ term += (int) var1digits[0] * var2digits[res_ndigits - 4];
+ if (res_ndigits > 4)
+ term += (int) var1digits[1] * var2digits[res_ndigits - 5];
+ if (res_ndigits > 5)
+ term += (int) var1digits[2] * var2digits[res_ndigits - 6];
+ res_digits[res_ndigits - 3] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+
+ /* remaining digits, except for the first three */
+ for (int i = res_ndigits - 5; i >= 2; i--)
+ {
+ term = (int) var1digits[0] * var2digits[i] +
+ (int) var1digits[1] * var2digits[i - 1] +
+ (int) var1digits[2] * var2digits[i - 2] + carry;
+ res_digits[i + 1] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+ }
+
+ /* first three digits */
+ term = (int) var1digits[0] * var2digits[1] +
+ (int) var1digits[1] * var2digits[0] + carry;
+ res_digits[2] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+ term = (int) var1digits[0] * var2digits[0] + carry;
+ res_digits[1] = (NumericDigit) (term % NBASE);
+ res_digits[0] = (NumericDigit) (term / NBASE);
+ break;
+ }
+
+ /* Store the product in result (minus extra rounding digit) */
+ digitbuf_free(result->buf);
+ result->ndigits = res_ndigits - 1;
+ result->buf = res_buf;
+ result->digits = res_digits;
+ result->weight = res_weight - 1;
+ result->sign = res_sign;
+
+ /* Round to target rscale (and set result->dscale) */
+ round_var(result, rscale);
+
+ /* Strip leading and trailing zeroes */
+ strip_var(result);
+}
/*
* div_var() -
diff --git a/src/include/catalog/pg_proc.dat b/src/include/catalog/pg_proc.dat
index d4ac578ae6..c85ed20a99 100644
--- a/src/include/catalog/pg_proc.dat
+++ b/src/include/catalog/pg_proc.dat
@@ -4465,6 +4465,18 @@
{ oid => '1726',
proname => 'numeric_mul', prorettype => 'numeric',
proargtypes => 'numeric numeric', prosrc => 'numeric_mul' },
+{ oid => '6347',
+ proname => 'numeric_mul_head', prorettype => 'numeric',
+ proargtypes => 'numeric numeric int4', prosrc => 'numeric_mul_head' },
+{ oid => '6348',
+ proname => 'numeric_mul_patch_inline', prorettype => 'numeric',
+ proargtypes => 'numeric numeric int4', prosrc => 'numeric_mul_patch_inline' },
+{ oid => '6349',
+ proname => 'numeric_mul_patch_int', prorettype => 'numeric',
+ proargtypes => 'numeric numeric int4', prosrc => 'numeric_mul_patch_int' },
+{ oid => '6350',
+ proname => 'numeric_mul_patch_small', prorettype => 'numeric',
+ proargtypes => 'numeric numeric int4', prosrc => 'numeric_mul_patch_small' },
{ oid => '1727',
proname => 'numeric_div', prorettype => 'numeric',
proargtypes => 'numeric numeric', prosrc => 'numeric_div' },
diff --git a/src/include/utils/numeric.h b/src/include/utils/numeric.h
index 43c75c436f..2bc400c741 100644
--- a/src/include/utils/numeric.h
+++ b/src/include/utils/numeric.h
@@ -97,6 +97,14 @@ extern Numeric numeric_sub_opt_error(Numeric num1, Numeric num2,
bool *have_error);
extern Numeric numeric_mul_opt_error(Numeric num1, Numeric num2,
bool *have_error);
+extern Numeric numeric_mul_head_opt_error(Numeric num1, Numeric num2,
+ int32 rscale_adjustment, bool *have_error);
+extern Numeric numeric_mul_patch_inline_opt_error(Numeric num1, Numeric num2,
+ int32 rscale_adjustment, bool *have_error);
+extern Numeric numeric_mul_patch_int_opt_error(Numeric num1, Numeric num2,
+ int32 rscale_adjustment, bool *have_error);
+extern Numeric numeric_mul_patch_small_opt_error(Numeric num1, Numeric num2,
+ int32 rscale_adjustment, bool *have_error);
extern Numeric numeric_div_opt_error(Numeric num1, Numeric num2,
bool *have_error);
extern Numeric numeric_mod_opt_error(Numeric num1, Numeric num2,
diff --git a/test-mul-var.sql b/test-mul-var.sql
new file mode 100644
index 0000000000..cf7c55a089
--- /dev/null
+++ b/test-mul-var.sql
@@ -0,0 +1,50 @@
+CREATE TABLE test_numeric_mul_patched (
+ var1 numeric,
+ var2 numeric,
+ rscale_adjustment int,
+ result numeric
+);
+
+DO $$
+DECLARE
+var1 numeric;
+var2 numeric;
+BEGIN
+ FOR i IN 1..1000 LOOP
+ RAISE NOTICE '%', i;
+ FOR var1ndigits IN 1..4 LOOP
+ FOR var2ndigits IN 1..4 LOOP
+ FOR var1dscale IN 0..(var1ndigits*4) LOOP
+ FOR var2dscale IN 0..(var2ndigits*4) LOOP
+ FOR rscale_adjustment IN 0..(var1dscale+var2dscale) LOOP
+ var1 := round(random(
+ format('1%s',repeat('0',(var1ndigits-1)*4-1))::numeric,
+ format('%s',repeat('9',var1ndigits*4))::numeric
+ ) / 10::numeric^var1dscale, var1dscale);
+ var2 := round(random(
+ format('1%s',repeat('0',(var2ndigits-1)*4-1))::numeric,
+ format('%s',repeat('9',var2ndigits*4))::numeric
+ ) / 10::numeric^var2dscale, var2dscale);
+ INSERT INTO test_numeric_mul_patched
+ (var1, var2, rscale_adjustment)
+ VALUES
+ (var1, var2, -rscale_adjustment);
+ END LOOP;
+ END LOOP;
+ END LOOP;
+ END LOOP;
+ END LOOP;
+ END LOOP;
+END $$;
+
+UPDATE test_numeric_mul_patched SET result = numeric_mul_head(var1, var2, rscale_adjustment);
+
+SELECT
+ rscale_adjustment,
+ COUNT(*),
+ COUNT(*) FILTER (WHERE result IS DISTINCT FROM numeric_mul_patch_int(var1,var2,rscale_adjustment)) AS "mul_var_int",
+ COUNT(*) FILTER (WHERE result IS DISTINCT FROM numeric_mul_patch_small(var1,var2,rscale_adjustment)) AS "mul_var_small",
+ COUNT(*) FILTER (WHERE result IS DISTINCT FROM numeric_mul_patch_inline(var1,var2,rscale_adjustment)) AS "mul_var inlined"
+FROM test_numeric_mul_patched
+GROUP BY 1
+ORDER BY 1;
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: Optimize numeric multiplication for one and two base-NBASE digit multiplicands.
@ 2024-07-03 18:57 Dean Rasheed <[email protected]>
parent: Joel Jacobson <[email protected]>
1 sibling, 1 reply; 103+ messages in thread
From: Dean Rasheed @ 2024-07-03 18:57 UTC (permalink / raw)
To: Joel Jacobson <[email protected]>; +Cc: Dagfinn Ilmari Mannsåker <[email protected]>; pgsql-hackers
On Wed, 3 Jul 2024 at 14:49, Joel Jacobson <[email protected]> wrote:
>
> Hmm, I don't see how the case 2 code can be correct?
> If, like you say, res_ndigits can't be less than 3, that means it can be 3, right?
> And if res_ndigits=3 then `var2digits[res_ndigits - 4]` would try to access `var2digits[-1]`.
>
Ah yes, I think I was looking at a newer version of the code where I'd
already fixed that bug. Unless you think there are still bugs in any
of the boundary checks, which is entirely possible.
> I've tested both patches, and they produces the same output given the
> same input as HEAD, when rscale is unmodified (full precision).
>
> However, for a reduced rscale, there are some differences:
>
> mul_var_small() seems more resilient to rscale reductions than mul_var_int().
>
Ah, I can see what's going on. It's perhaps best illustrated with a
simple example. Suppose you are multiplying a 4-digit integer x by a
2-digit integer y (both with dscale=0). Then the terms of the full
product computed by mul_var() or mul_var_small() would look something
like this:
x0 x1 x2 x3
* y0 y1
---------------------------------
x0*y0 x1*y0 x2*y0 x3*y0
x0*y1 x1*y1 x2*y1 x3*y1
In the reduced-rscale case, it might perform a truncated computation,
computing just the first 3 columns (say), and discarding the last two
columns. Therefore it would skip the 3 rightmost digit products.
However, in mul_var_int(), y0 and y1 have been combined into a single
integer equal to y0*NBASE+y1, and the terms of full product are
computed as follows:
x0*(y0*NBASE+y1) x1*(y0*NBASE+y1) x2*(y0*NBASE+y1) x3*(y0*NBASE+y1)
In the full product, that gives the same result, but if you follow the
same rule in the reduced-rscale case, skipping the last two terms, it
would actually discard 4 digit products, making it less accurate.
That could be avoided by increasing maxdigits by 1 in mul_var_int() so
it would always be at least as accurate as it was before, but that
might not really be necessary. However, if we implemented
mul_var_int64() in the same way, it would be discarding much
higher-order digit products, and so we probably would have to increase
maxdigits to get sufficiently accurate results. But there's an even
bigger problem: the results would be different between platforms that
did and didn't have 128-bit integers, which I really don't like. We
could avoid that by not using it in reduced-rscale cases, but that
would involve another test condition.
By contrast, mul_var_small() is intended to replicate the arithmetic
in mul_var() exactly (but in a different order) for all rscales. So if
you've found any cases where they give different results, that's a
bug.
In light of that, it might be that mul_var_small() is the better
option, rather than mul_var_int(), but it'd be interesting to see how
they compare in terms of performance first.
> What can be said about mul_var()'s contract with regards to rscale?
> It's the number of decimal digits requested by the caller, and if not
> requesting full precision, then the decimal digits might not be accurate,
> but can something be said about how far off they can be?
>
I wouldn't expect it to ever be off by more than 1, given that
MUL_GUARD_DIGITS = 2, which corresponds to 8 decimal digits, and the
number of digits in the smaller input (and hence the number of digit
products in each column) is limited to something like 16,000 NBASE
digits.
Regards,
Dean
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: Optimize numeric multiplication for one and two base-NBASE digit multiplicands.
@ 2024-07-03 19:05 Joel Jacobson <[email protected]>
parent: Joel Jacobson <[email protected]>
1 sibling, 0 replies; 103+ messages in thread
From: Joel Jacobson @ 2024-07-03 19:05 UTC (permalink / raw)
To: Dean Rasheed <[email protected]>; +Cc: Dagfinn Ilmari Mannsåker <[email protected]>; pgsql-hackers
On Wed, Jul 3, 2024, at 15:48, Joel Jacobson wrote:
> On Wed, Jul 3, 2024, at 13:17, Dean Rasheed wrote:
>> On Tue, 2 Jul 2024 at 21:10, Joel Jacobson <[email protected]> wrote:
>>>
>>> I found the bug in the case 3 code,
>>> and it turns out the same type of bug also exists in the case 2 code:
>>>
>>> case 2:
>>> newdig = (int) var1digits[1] * var2digits[res_ndigits - 4];
>>>
>>> The problem here is that res_ndigits could become less than 4,
>>
>> Yes. It can't be less than 3 though (per an earlier test), so the case
>> 2 code was correct.
>
> Hmm, I don't see how the case 2 code can be correct?
> If, like you say, res_ndigits can't be less than 3, that means it can
> be 3, right?
> And if res_ndigits=3 then `var2digits[res_ndigits - 4]` would try to
> access `var2digits[-1]`.
Here is an example on how to trigger the bug:
```
case 2:
if (res_ndigits - 4 < 0)
{
printf("var1=%s\n",get_str_from_var(var1));
printf("var2=%s\n",get_str_from_var(var2));
printf("rscale=%d\n", rscale);
printf("res_ndigits - 4 < 0 => var2digits[%d]=%d\n", res_ndigits - 4, var2digits[res_ndigits - 4]);
}
```
Running through my tests, I hit lots of cases, including:
var1=0.10968501
var2=0.903728177113
rscale=0
res_ndigits - 4 < 0 => var2digits[-1]=-31105
All of the spotted cases had rscale=0.
If we know that mul_var() will never be called with rscale=0 when dealing with decimal inputs, perhaps we should enforce this with an Assert(), to prevent the otherwise possible out-of-bounds access (negative indexing) and provide early detection?
Regards,
Joel
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: Optimize numeric multiplication for one and two base-NBASE digit multiplicands.
@ 2024-07-03 20:27 Joel Jacobson <[email protected]>
parent: Dean Rasheed <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Joel Jacobson @ 2024-07-03 20:27 UTC (permalink / raw)
To: Dean Rasheed <[email protected]>; +Cc: Dagfinn Ilmari Mannsåker <[email protected]>; pgsql-hackers
On Wed, Jul 3, 2024, at 20:57, Dean Rasheed wrote:
> Ah yes, I think I was looking at a newer version of the code where I'd
> already fixed that bug. Unless you think there are still bugs in any
> of the boundary checks, which is entirely possible.
Ah, that explains it.
And no, I can't find any other bugs in the boundary checks.
> Ah, I can see what's going on. It's perhaps best illustrated with a
> simple example. Suppose you are multiplying a 4-digit integer x by a
> 2-digit integer y (both with dscale=0). Then the terms of the full
> product computed by mul_var() or mul_var_small() would look something
> like this:
>
> x0 x1 x2 x3
> * y0 y1
> ---------------------------------
> x0*y0 x1*y0 x2*y0 x3*y0
> x0*y1 x1*y1 x2*y1 x3*y1
>
> In the reduced-rscale case, it might perform a truncated computation,
> computing just the first 3 columns (say), and discarding the last two
> columns. Therefore it would skip the 3 rightmost digit products.
>
> However, in mul_var_int(), y0 and y1 have been combined into a single
> integer equal to y0*NBASE+y1, and the terms of full product are
> computed as follows:
>
> x0*(y0*NBASE+y1) x1*(y0*NBASE+y1) x2*(y0*NBASE+y1) x3*(y0*NBASE+y1)
>
> In the full product, that gives the same result, but if you follow the
> same rule in the reduced-rscale case, skipping the last two terms, it
> would actually discard 4 digit products, making it less accurate.
>
> That could be avoided by increasing maxdigits by 1 in mul_var_int() so
> it would always be at least as accurate as it was before, but that
> might not really be necessary. However, if we implemented
> mul_var_int64() in the same way, it would be discarding much
> higher-order digit products, and so we probably would have to increase
> maxdigits to get sufficiently accurate results. But there's an even
> bigger problem: the results would be different between platforms that
> did and didn't have 128-bit integers, which I really don't like. We
> could avoid that by not using it in reduced-rscale cases, but that
> would involve another test condition.
>
> By contrast, mul_var_small() is intended to replicate the arithmetic
> in mul_var() exactly (but in a different order) for all rscales. So if
> you've found any cases where they give different results, that's a
> bug.
>
> In light of that, it might be that mul_var_small() is the better
> option, rather than mul_var_int(), but it'd be interesting to see how
> they compare in terms of performance first.
Thanks for explaining, very helpful.
I agree on your reasoning about the pros and cons.
Not sure yet which version I prefer. Let's see how it evolves.
I've done some benchmarks.
Haven't tested Intel and AMD yet, but this is what I get on my Apple M3 Max:
--
-- varndigits=1
--
-- HEAD
SELECT SUM(numeric_mul_head(var1,var2,0)) FROM bench_mul_var_var1ndigits_1;
Time: 2976.896 ms (00:02.977)
Time: 2984.759 ms (00:02.985)
Time: 2970.364 ms (00:02.970)
-- mul_var_int() patch
SELECT SUM(numeric_mul_patch_int(var1,var2,0)) FROM bench_mul_var_var1ndigits_1;
Time: 2790.227 ms (00:02.790)
Time: 2786.338 ms (00:02.786)
Time: 2784.957 ms (00:02.785)
-- mul_var_small() patch
SELECT SUM(numeric_mul_patch_small(var1,var2,0)) FROM bench_mul_var_var1ndigits_1;
Time: 2770.211 ms (00:02.770)
Time: 2760.685 ms (00:02.761)
Time: 2773.221 ms (00:02.773)
--
-- varndigits=2
--
-- HEAD
SELECT SUM(numeric_mul_head(var1,var2,0)) FROM bench_mul_var_var1ndigits_2;
Time: 3353.258 ms (00:03.353)
Time: 3273.055 ms (00:03.273)
Time: 3266.392 ms (00:03.266)
-- mul_var_int() patch
SELECT SUM(numeric_mul_patch_int(var1,var2,0)) FROM bench_mul_var_var1ndigits_2;
Time: 2694.169 ms (00:02.694)
Time: 2687.935 ms (00:02.688)
Time: 2692.398 ms (00:02.692)
-- mul_var_small() patch
SELECT SUM(numeric_mul_patch_small(var1,var2,0)) FROM bench_mul_var_var1ndigits_2;
Time: 2997.685 ms (00:02.998)
Time: 2984.418 ms (00:02.984)
Time: 2986.976 ms (00:02.987)
--
-- varndigits=3
--
-- HEAD
SELECT SUM(numeric_mul_head(var1,var2,0)) FROM bench_mul_var_var1ndigits_3;
Time: 3471.391 ms (00:03.471)
Time: 3384.114 ms (00:03.384)
Time: 3387.031 ms (00:03.387)
-- mul_var_int() patch
SELECT SUM(numeric_mul_patch_int(var1,var2,0)) FROM bench_mul_var_var1ndigits_3;
Time: 3384.428 ms (00:03.384)
Time: 3398.044 ms (00:03.398)
Time: 3393.727 ms (00:03.394)
-- mul_var_small() patch
SELECT SUM(numeric_mul_patch_small(var1,var2,0)) FROM bench_mul_var_var1ndigits_3;
Time: 3100.567 ms (00:03.101)
Time: 3114.225 ms (00:03.114)
Time: 3116.137 ms (00:03.116)
Interesting, mul_var_small() seems to be the winner for var1ndigits=3
and mul_var_int() to be the winner for var1ndigits=2,
and about the same for var1ndigits=1.
>> What can be said about mul_var()'s contract with regards to rscale?
>> It's the number of decimal digits requested by the caller, and if not
>> requesting full precision, then the decimal digits might not be accurate,
>> but can something be said about how far off they can be?
>>
>
> I wouldn't expect it to ever be off by more than 1, given that
> MUL_GUARD_DIGITS = 2, which corresponds to 8 decimal digits, and the
> number of digits in the smaller input (and hence the number of digit
> products in each column) is limited to something like 16,000 NBASE
> digits.
OK, so then the cases I found where it was off by 2 for the mul_var_int() patch
are unexpected?
Regards,
Joel
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: Optimize numeric multiplication for one and two base-NBASE digit multiplicands.
@ 2024-07-03 20:45 Joel Jacobson <[email protected]>
parent: Joel Jacobson <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Joel Jacobson @ 2024-07-03 20:45 UTC (permalink / raw)
To: Dean Rasheed <[email protected]>; +Cc: Dagfinn Ilmari Mannsåker <[email protected]>; pgsql-hackers
On Wed, Jul 3, 2024, at 22:27, Joel Jacobson wrote:
> On Wed, Jul 3, 2024, at 20:57, Dean Rasheed wrote:
>> I wouldn't expect it to ever be off by more than 1, given that
>> MUL_GUARD_DIGITS = 2, which corresponds to 8 decimal digits, and the
>> number of digits in the smaller input (and hence the number of digit
>> products in each column) is limited to something like 16,000 NBASE
>> digits.
>
> OK, so then the cases I found where it was off by 2 for the mul_var_int() patch
> are unexpected?
Sorry, I meant off by 2 for the mul_var_small() patch, these cases that I found:
var1 | var2 | rscale_adjustment | result | numeric_mul_patch_small
-------------------+----------------+-------------------+------------+-------------------------
8952.12658563 | 0.902315486665 | -16 | 8077.6425 | 8077.6423
0.881715409579 | 0.843165739371 | -16 | 0.74343223 | 0.74343221
0.905322758954 | 0.756905996850 | -16 | 0.68524423 | 0.68524421
8464.043170546608 | 0.518100129611 | -20 | 4385.2219 | 4385.2217
5253.006296984449 | 0.989308019355 | -20 | 5196.8413 | 5196.8411
(5 rows)
Regards,
Joel
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: Optimize numeric multiplication for one and two base-NBASE digit multiplicands.
@ 2024-07-04 07:38 Joel Jacobson <[email protected]>
parent: Dean Rasheed <[email protected]>
2 siblings, 1 reply; 103+ messages in thread
From: Joel Jacobson @ 2024-07-04 07:38 UTC (permalink / raw)
To: Dean Rasheed <[email protected]>; +Cc: Dagfinn Ilmari Mannsåker <[email protected]>; pgsql-hackers
On Wed, Jul 3, 2024, at 13:17, Dean Rasheed wrote:
> Anyway, here are both patches for comparison. I'll stop hacking for a
> while and let you see what you make of these.
>
> Regards,
> Dean
>
> Attachments:
> * v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch
> * v5-add-mul_var_int.patch
I've now benchmarked the patches on all my machines,
see bench_mul_var.sql for details.
Summary of benchmark results:
cpu | var1ndigits | winner
----------------------+-------------+-------------------------------------------------------------
AMD Ryzen 9 7950X3D | 1 | v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch
AMD Ryzen 9 7950X3D | 2 | v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch
AMD Ryzen 9 7950X3D | 3 | v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch
Apple M3 Max | 1 | v5-add-mul_var_int.patch
Apple M3 Max | 2 | v5-add-mul_var_int.patch
Apple M3 Max | 3 | v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch
Intel Core i9-14900K | 1 | v5-add-mul_var_int.patch
Intel Core i9-14900K | 2 | v5-add-mul_var_int.patch
Intel Core i9-14900K | 3 | v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch
(9 rows)
Performance ratio against HEAD per CPU and var1ndigits:
cpu | var1ndigits | version | performance_ratio
----------------------+-------------+-------------------------------------------------------------+-------------------
AMD Ryzen 9 7950X3D | 1 | HEAD | 1.00
AMD Ryzen 9 7950X3D | 1 | v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch | 1.11
AMD Ryzen 9 7950X3D | 1 | v5-add-mul_var_int.patch | 1.07
AMD Ryzen 9 7950X3D | 1 | v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch | 1.12
AMD Ryzen 9 7950X3D | 2 | HEAD | 1.00
AMD Ryzen 9 7950X3D | 2 | v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch | 1.10
AMD Ryzen 9 7950X3D | 2 | v5-add-mul_var_int.patch | 1.11
AMD Ryzen 9 7950X3D | 2 | v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch | 1.13
AMD Ryzen 9 7950X3D | 3 | HEAD | 1.00
AMD Ryzen 9 7950X3D | 3 | v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch | 1.10
AMD Ryzen 9 7950X3D | 3 | v5-add-mul_var_int.patch | 0.98
AMD Ryzen 9 7950X3D | 3 | v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch | 1.15
Apple M3 Max | 1 | HEAD | 1.00
Apple M3 Max | 1 | v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch | 1.07
Apple M3 Max | 1 | v5-add-mul_var_int.patch | 1.08
Apple M3 Max | 1 | v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch | 1.07
Apple M3 Max | 2 | HEAD | 1.00
Apple M3 Max | 2 | v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch | 1.09
Apple M3 Max | 2 | v5-add-mul_var_int.patch | 1.21
Apple M3 Max | 2 | v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch | 1.10
Apple M3 Max | 3 | HEAD | 1.00
Apple M3 Max | 3 | v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch | 1.09
Apple M3 Max | 3 | v5-add-mul_var_int.patch | 0.99
Apple M3 Max | 3 | v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch | 1.09
Intel Core i9-14900K | 1 | HEAD | 1.00
Intel Core i9-14900K | 1 | v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch | 1.05
Intel Core i9-14900K | 1 | v5-add-mul_var_int.patch | 1.07
Intel Core i9-14900K | 1 | v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch | 1.06
Intel Core i9-14900K | 2 | HEAD | 1.00
Intel Core i9-14900K | 2 | v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch | 1.06
Intel Core i9-14900K | 2 | v5-add-mul_var_int.patch | 1.08
Intel Core i9-14900K | 2 | v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch | 1.06
Intel Core i9-14900K | 3 | HEAD | 1.00
Intel Core i9-14900K | 3 | v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch | 1.04
Intel Core i9-14900K | 3 | v5-add-mul_var_int.patch | 1.00
Intel Core i9-14900K | 3 | v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch | 1.04
(36 rows)
The queries to produce the above are in bench_csv_queries.txt
/Joel
CREATE TABLE bench (cpu text, var1ndigits int, version text, time numeric);
\COPY bench FROM bench.csv WITH CSV HEADER;
WITH ranked_bench AS (
SELECT
cpu,
var1ndigits,
version,
ROW_NUMBER() OVER (PARTITION BY cpu, var1ndigits ORDER BY AVG(time)) AS rn
FROM bench
GROUP BY cpu, var1ndigits, version
)
SELECT
cpu,
var1ndigits,
version AS winner
FROM ranked_bench
WHERE rn = 1
ORDER BY cpu, var1ndigits;
WITH avg_times AS (
SELECT
cpu,
var1ndigits,
version,
AVG(time) AS avg_time
FROM bench
GROUP BY cpu, var1ndigits, version
),
head_times AS (
SELECT
cpu,
var1ndigits,
avg_time AS head_avg_time
FROM avg_times
WHERE version = 'HEAD'
)
SELECT
a.cpu,
a.var1ndigits,
a.version,
ROUND(h.head_avg_time / a.avg_time,2) AS performance_ratio
FROM avg_times a
JOIN head_times h
ON a.cpu = h.cpu AND a.var1ndigits = h.var1ndigits
ORDER BY a.cpu, a.var1ndigits, a.version;
Attachments:
[text/plain] bench_csv_queries.txt (1016B, ../../[email protected]/2-bench_csv_queries.txt)
download | inline:
CREATE TABLE bench (cpu text, var1ndigits int, version text, time numeric);
\COPY bench FROM bench.csv WITH CSV HEADER;
WITH ranked_bench AS (
SELECT
cpu,
var1ndigits,
version,
ROW_NUMBER() OVER (PARTITION BY cpu, var1ndigits ORDER BY AVG(time)) AS rn
FROM bench
GROUP BY cpu, var1ndigits, version
)
SELECT
cpu,
var1ndigits,
version AS winner
FROM ranked_bench
WHERE rn = 1
ORDER BY cpu, var1ndigits;
WITH avg_times AS (
SELECT
cpu,
var1ndigits,
version,
AVG(time) AS avg_time
FROM bench
GROUP BY cpu, var1ndigits, version
),
head_times AS (
SELECT
cpu,
var1ndigits,
avg_time AS head_avg_time
FROM avg_times
WHERE version = 'HEAD'
)
SELECT
a.cpu,
a.var1ndigits,
a.version,
ROUND(h.head_avg_time / a.avg_time,2) AS performance_ratio
FROM avg_times a
JOIN head_times h
ON a.cpu = h.cpu AND a.var1ndigits = h.var1ndigits
ORDER BY a.cpu, a.var1ndigits, a.version;
[text/csv] bench.csv (11.9K, ../../[email protected]/3-bench.csv)
download | inline:
cpu,var1ndigits,version,time
Apple M3 Max,1,HEAD,3090.147
Apple M3 Max,1,HEAD,3095.153
Apple M3 Max,1,HEAD,3096.725
Apple M3 Max,1,HEAD,3070.083
Apple M3 Max,1,HEAD,3081.267
Apple M3 Max,1,v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,2883.311
Apple M3 Max,1,v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,2882.971
Apple M3 Max,1,v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,2884.639
Apple M3 Max,1,v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,2884.728
Apple M3 Max,1,v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,2887.346
Apple M3 Max,1,v5-add-mul_var_int.patch,2859.045
Apple M3 Max,1,v5-add-mul_var_int.patch,2854.941
Apple M3 Max,1,v5-add-mul_var_int.patch,2851.976
Apple M3 Max,1,v5-add-mul_var_int.patch,2863.930
Apple M3 Max,1,v5-add-mul_var_int.patch,2864.494
Apple M3 Max,1,v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,2869.741
Apple M3 Max,1,v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,2870.023
Apple M3 Max,1,v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,2870.653
Apple M3 Max,1,v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,2869.711
Apple M3 Max,1,v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,2878.178
Apple M3 Max,2,HEAD,3397.181
Apple M3 Max,2,HEAD,3371.557
Apple M3 Max,2,HEAD,3356.081
Apple M3 Max,2,HEAD,3371.946
Apple M3 Max,2,HEAD,3385.859
Apple M3 Max,2,v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3084.038
Apple M3 Max,2,v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3082.308
Apple M3 Max,2,v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3089.160
Apple M3 Max,2,v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3083.793
Apple M3 Max,2,v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3081.382
Apple M3 Max,2,v5-add-mul_var_int.patch,2782.388
Apple M3 Max,2,v5-add-mul_var_int.patch,2780.564
Apple M3 Max,2,v5-add-mul_var_int.patch,2781.664
Apple M3 Max,2,v5-add-mul_var_int.patch,2776.481
Apple M3 Max,2,v5-add-mul_var_int.patch,2781.443
Apple M3 Max,2,v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3074.596
Apple M3 Max,2,v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3073.305
Apple M3 Max,2,v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3078.297
Apple M3 Max,2,v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3084.720
Apple M3 Max,2,v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3076.612
Apple M3 Max,3,HEAD,3500.698
Apple M3 Max,3,HEAD,3490.170
Apple M3 Max,3,HEAD,3481.300
Apple M3 Max,3,HEAD,3486.962
Apple M3 Max,3,HEAD,3473.165
Apple M3 Max,3,v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3196.107
Apple M3 Max,3,v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3203.074
Apple M3 Max,3,v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3215.363
Apple M3 Max,3,v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3204.599
Apple M3 Max,3,v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3203.819
Apple M3 Max,3,v5-add-mul_var_int.patch,3505.878
Apple M3 Max,3,v5-add-mul_var_int.patch,3544.366
Apple M3 Max,3,v5-add-mul_var_int.patch,3521.562
Apple M3 Max,3,v5-add-mul_var_int.patch,3510.695
Apple M3 Max,3,v5-add-mul_var_int.patch,3523.758
Apple M3 Max,3,v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3203.702
Apple M3 Max,3,v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3206.802
Apple M3 Max,3,v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3208.966
Apple M3 Max,3,v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3198.790
Apple M3 Max,3,v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3202.307
Intel Core i9-14900K,1,HEAD,3294.044
Intel Core i9-14900K,1,HEAD,3296.176
Intel Core i9-14900K,1,HEAD,3263.968
Intel Core i9-14900K,1,HEAD,3262.892
Intel Core i9-14900K,1,HEAD,3263.531
Intel Core i9-14900K,1,v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3125.577
Intel Core i9-14900K,1,v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3126.274
Intel Core i9-14900K,1,v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3122.623
Intel Core i9-14900K,1,v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3123.057
Intel Core i9-14900K,1,v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3124.143
Intel Core i9-14900K,1,v5-add-mul_var_int.patch,3056.697
Intel Core i9-14900K,1,v5-add-mul_var_int.patch,3053.200
Intel Core i9-14900K,1,v5-add-mul_var_int.patch,3053.484
Intel Core i9-14900K,1,v5-add-mul_var_int.patch,3053.548
Intel Core i9-14900K,1,v5-add-mul_var_int.patch,3052.770
Intel Core i9-14900K,1,v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3076.845
Intel Core i9-14900K,1,v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3078.734
Intel Core i9-14900K,1,v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3078.344
Intel Core i9-14900K,1,v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3079.325
Intel Core i9-14900K,1,v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3077.803
Intel Core i9-14900K,2,HEAD,3398.412
Intel Core i9-14900K,2,HEAD,3380.465
Intel Core i9-14900K,2,HEAD,3349.829
Intel Core i9-14900K,2,HEAD,3351.368
Intel Core i9-14900K,2,HEAD,3346.712
Intel Core i9-14900K,2,v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3180.399
Intel Core i9-14900K,2,v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3179.782
Intel Core i9-14900K,2,v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3181.158
Intel Core i9-14900K,2,v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3180.896
Intel Core i9-14900K,2,v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3184.998
Intel Core i9-14900K,2,v5-add-mul_var_int.patch,3116.356
Intel Core i9-14900K,2,v5-add-mul_var_int.patch,3111.412
Intel Core i9-14900K,2,v5-add-mul_var_int.patch,3116.011
Intel Core i9-14900K,2,v5-add-mul_var_int.patch,3117.422
Intel Core i9-14900K,2,v5-add-mul_var_int.patch,3118.254
Intel Core i9-14900K,2,v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3189.580
Intel Core i9-14900K,2,v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3188.140
Intel Core i9-14900K,2,v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3187.830
Intel Core i9-14900K,2,v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3189.053
Intel Core i9-14900K,2,v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3189.717
Intel Core i9-14900K,3,HEAD,4171.285
Intel Core i9-14900K,3,HEAD,4162.330
Intel Core i9-14900K,3,HEAD,4142.480
Intel Core i9-14900K,3,HEAD,4137.238
Intel Core i9-14900K,3,HEAD,4137.180
Intel Core i9-14900K,3,v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3983.753
Intel Core i9-14900K,3,v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3983.630
Intel Core i9-14900K,3,v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3983.684
Intel Core i9-14900K,3,v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3992.266
Intel Core i9-14900K,3,v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3984.959
Intel Core i9-14900K,3,v5-add-mul_var_int.patch,4159.903
Intel Core i9-14900K,3,v5-add-mul_var_int.patch,4155.236
Intel Core i9-14900K,3,v5-add-mul_var_int.patch,4156.482
Intel Core i9-14900K,3,v5-add-mul_var_int.patch,4154.075
Intel Core i9-14900K,3,v5-add-mul_var_int.patch,4152.739
Intel Core i9-14900K,3,v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3982.057
Intel Core i9-14900K,3,v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3983.134
Intel Core i9-14900K,3,v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3984.192
Intel Core i9-14900K,3,v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3982.792
Intel Core i9-14900K,3,v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3980.789
AMD Ryzen 9 7950X3D,1,HEAD,3306.653
AMD Ryzen 9 7950X3D,1,HEAD,3307.260
AMD Ryzen 9 7950X3D,1,HEAD,3268.231
AMD Ryzen 9 7950X3D,1,HEAD,3276.653
AMD Ryzen 9 7950X3D,1,HEAD,3267.375
AMD Ryzen 9 7950X3D,1,v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,2954.318
AMD Ryzen 9 7950X3D,1,v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,2961.550
AMD Ryzen 9 7950X3D,1,v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,2963.000
AMD Ryzen 9 7950X3D,1,v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,2958.896
AMD Ryzen 9 7950X3D,1,v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,2957.998
AMD Ryzen 9 7950X3D,1,v5-add-mul_var_int.patch,3080.628
AMD Ryzen 9 7950X3D,1,v5-add-mul_var_int.patch,3064.396
AMD Ryzen 9 7950X3D,1,v5-add-mul_var_int.patch,3073.567
AMD Ryzen 9 7950X3D,1,v5-add-mul_var_int.patch,3074.281
AMD Ryzen 9 7950X3D,1,v5-add-mul_var_int.patch,3072.741
AMD Ryzen 9 7950X3D,1,v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,2932.485
AMD Ryzen 9 7950X3D,1,v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,2931.049
AMD Ryzen 9 7950X3D,1,v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,2932.589
AMD Ryzen 9 7950X3D,1,v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,2927.167
AMD Ryzen 9 7950X3D,1,v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,2932.993
AMD Ryzen 9 7950X3D,2,HEAD,3505.663
AMD Ryzen 9 7950X3D,2,HEAD,3510.422
AMD Ryzen 9 7950X3D,2,HEAD,3470.290
AMD Ryzen 9 7950X3D,2,HEAD,3490.446
AMD Ryzen 9 7950X3D,2,HEAD,3470.264
AMD Ryzen 9 7950X3D,2,v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3166.622
AMD Ryzen 9 7950X3D,2,v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3162.453
AMD Ryzen 9 7950X3D,2,v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3168.570
AMD Ryzen 9 7950X3D,2,v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3173.971
AMD Ryzen 9 7950X3D,2,v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3170.914
AMD Ryzen 9 7950X3D,2,v5-add-mul_var_int.patch,3135.849
AMD Ryzen 9 7950X3D,2,v5-add-mul_var_int.patch,3131.497
AMD Ryzen 9 7950X3D,2,v5-add-mul_var_int.patch,3138.993
AMD Ryzen 9 7950X3D,2,v5-add-mul_var_int.patch,3135.383
AMD Ryzen 9 7950X3D,2,v5-add-mul_var_int.patch,3143.074
AMD Ryzen 9 7950X3D,2,v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3075.388
AMD Ryzen 9 7950X3D,2,v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3078.407
AMD Ryzen 9 7950X3D,2,v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3075.033
AMD Ryzen 9 7950X3D,2,v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3078.055
AMD Ryzen 9 7950X3D,2,v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3075.422
AMD Ryzen 9 7950X3D,3,HEAD,3809.949
AMD Ryzen 9 7950X3D,3,HEAD,3824.012
AMD Ryzen 9 7950X3D,3,HEAD,3782.624
AMD Ryzen 9 7950X3D,3,HEAD,3783.997
AMD Ryzen 9 7950X3D,3,HEAD,3761.864
AMD Ryzen 9 7950X3D,3,v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3459.177
AMD Ryzen 9 7950X3D,3,v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3474.086
AMD Ryzen 9 7950X3D,3,v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3456.303
AMD Ryzen 9 7950X3D,3,v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3475.646
AMD Ryzen 9 7950X3D,3,v4-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3450.919
AMD Ryzen 9 7950X3D,3,v5-add-mul_var_int.patch,3882.610
AMD Ryzen 9 7950X3D,3,v5-add-mul_var_int.patch,3885.023
AMD Ryzen 9 7950X3D,3,v5-add-mul_var_int.patch,3884.721
AMD Ryzen 9 7950X3D,3,v5-add-mul_var_int.patch,3894.463
AMD Ryzen 9 7950X3D,3,v5-add-mul_var_int.patch,3878.118
AMD Ryzen 9 7950X3D,3,v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3302.385
AMD Ryzen 9 7950X3D,3,v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3320.810
AMD Ryzen 9 7950X3D,3,v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3306.740
AMD Ryzen 9 7950X3D,3,v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3307.842
AMD Ryzen 9 7950X3D,3,v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch,3302.253
[application/octet-stream] bench_mul_var.sql (9.8K, ../../[email protected]/4-bench_mul_var.sql)
download
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: Optimize numeric multiplication for one and two base-NBASE digit multiplicands.
@ 2024-07-04 10:38 Joel Jacobson <[email protected]>
parent: Joel Jacobson <[email protected]>
0 siblings, 0 replies; 103+ messages in thread
From: Joel Jacobson @ 2024-07-04 10:38 UTC (permalink / raw)
To: Dean Rasheed <[email protected]>; +Cc: Dagfinn Ilmari Mannsåker <[email protected]>; pgsql-hackers
On Thu, Jul 4, 2024, at 09:38, Joel Jacobson wrote:
> Summary of benchmark results:
>
> cpu | var1ndigits | winner
> ----------------------+-------------+-------------------------------------------------------------
..
> v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch
> AMD Ryzen 9 7950X3D | 3 |
> v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch
...
> Apple M3 Max | 3 |
> v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch
...
> Intel Core i9-14900K | 3 |
> v5-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch
Since v5-add-mul_var_int.patch only implements (var1ndigits <= 2)
it can't possibly win the var1ndigits=3 competition.
/Joel
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: Optimize numeric multiplication for one and two base-NBASE digit multiplicands.
@ 2024-07-04 18:43 Dean Rasheed <[email protected]>
parent: Joel Jacobson <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Dean Rasheed @ 2024-07-04 18:43 UTC (permalink / raw)
To: Joel Jacobson <[email protected]>; +Cc: Dagfinn Ilmari Mannsåker <[email protected]>; pgsql-hackers
On Wed, 3 Jul 2024 at 21:45, Joel Jacobson <[email protected]> wrote:
>
> > On Wed, Jul 3, 2024, at 20:57, Dean Rasheed wrote:
> >> I wouldn't expect it to ever be off by more than 1
> >
> > OK, so then the cases I found where it was off by 2 for the mul_var_int() patch
> > are unexpected?
>
> Sorry, I meant off by 2 for the mul_var_small() patch, these cases that I found:
>
Yeah, so that was another bug in mul_var_small(). If rscale is made
small enough, the result index for the digits computed before the main
loop overlaps the ones after, so it would overwrite digits already
computed.
Of course, that's fairly easy to fix, but at this point I think the
better solution is to only use mul_var_small() when an exact product
is requested. We would have to do that for mul_var_int() anyway,
because of its accuracy issues discussed earlier. I think this is a
reasonable thing to do because only functions like ln_var() and
exp_var() will ask mul_var() for a reduced-rscale result, and those
functions are likely to be dominated by computations involving larger
numbers, for which this patch wouldn't help anyway. Also those
functions are probably less widely used.
If we make that decision, a lot of the complexity in mul_var_small()
goes away, including all the conditional array accesses, making it
simpler and more efficient. v6 patch attached.
I also updated the mul_var_int() patch so that it is also only invoked
when an exact product is requested, and I noticed a couple of other
minor optimisations that could be made. Then I decided to try
implementing mul_var_int64(). This gives a pretty decent speedup for
3-digit inputs, but unfortunately it is much slower for 4-digit inputs
(for which most values will go through the 128-bit code path). I'm
attaching that too, just for information, but it's clearly not going
to be acceptable as-is.
Running your benchmark queries, I got these results:
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_1;
Time: 4520.874 ms (00:04.521) -- HEAD
Time: 3937.536 ms (00:03.938) -- v5-mul_var_int.patch
Time: 3919.495 ms (00:03.919) -- v5-mul_var_small.patch
Time: 3916.964 ms (00:03.917) -- v6-mul_var_int64.patch
Time: 3811.118 ms (00:03.811) -- v6-mul_var_small.patch
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_2;
Time: 4762.528 ms (00:04.763) -- HEAD
Time: 4075.546 ms (00:04.076) -- v5-mul_var_int.patch
Time: 4055.180 ms (00:04.055) -- v5-mul_var_small.patch
Time: 4037.866 ms (00:04.038) -- v6-mul_var_int64.patch
Time: 4018.488 ms (00:04.018) -- v6-mul_var_small.patch
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_3;
Time: 5387.514 ms (00:05.388) -- HEAD
Time: 5350.736 ms (00:05.351) -- v5-mul_var_int.patch
Time: 4648.449 ms (00:04.648) -- v5-mul_var_small.patch
Time: 4655.204 ms (00:04.655) -- v6-mul_var_int64.patch
Time: 4645.962 ms (00:04.646) -- v6-mul_var_small.patch
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_4;
Time: 5617.150 ms (00:05.617) -- HEAD
Time: 5505.913 ms (00:05.506) -- v5-mul_var_int.patch
Time: 5486.441 ms (00:05.486) -- v5-mul_var_small.patch
Time: 8203.081 ms (00:08.203) -- v6-mul_var_int64.patch
Time: 5598.909 ms (00:05.599) -- v6-mul_var_small.patch
So v6-mul_var_int64 improves on v5-mul_var_int in the 3-digit case,
but is terrible in the 4-digit case. None of the other patches touch
the 4-digit case, but it might be interesting to try mul_var_small()
with 4 digits.
Regards,
Dean
Attachments:
[text/x-patch] v6-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch (6.3K, ../../CAEZATCXVB15_RfzROFiO_7Fh3voOuOe7FotbvON-6a+ManuGJg@mail.gmail.com/2-v6-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch)
download | inline diff:
diff --git a/src/backend/utils/adt/numeric.c b/src/backend/utils/adt/numeric.c
new file mode 100644
index 5510a20..bae07f2
--- a/src/backend/utils/adt/numeric.c
+++ b/src/backend/utils/adt/numeric.c
@@ -551,6 +551,8 @@ static void sub_var(const NumericVar *va
static void mul_var(const NumericVar *var1, const NumericVar *var2,
NumericVar *result,
int rscale);
+static void mul_var_small(const NumericVar *var1, const NumericVar *var2,
+ NumericVar *result, int rscale);
static void div_var(const NumericVar *var1, const NumericVar *var2,
NumericVar *result,
int rscale, bool round);
@@ -8707,7 +8709,7 @@ mul_var(const NumericVar *var1, const Nu
var1digits = var1->digits;
var2digits = var2->digits;
- if (var1ndigits == 0 || var2ndigits == 0)
+ if (var1ndigits == 0)
{
/* one or both inputs is zero; so is result */
zero_var(result);
@@ -8715,6 +8717,17 @@ mul_var(const NumericVar *var1, const Nu
return;
}
+ /*
+ * If var1 has 3 digits or fewer, and we are computing the exact result,
+ * with no rounding, delegate to mul_var_small() which uses a faster short
+ * multiplication algorithm.
+ */
+ if (var1ndigits <= 3 && rscale == var1->dscale + var2->dscale)
+ {
+ mul_var_small(var1, var2, result, rscale);
+ return;
+ }
+
/* Determine result sign and (maximum possible) weight */
if (var1->sign == var2->sign)
res_sign = NUMERIC_POS;
@@ -8858,6 +8871,168 @@ mul_var(const NumericVar *var1, const Nu
result->sign = res_sign;
/* Round to target rscale (and set result->dscale) */
+ round_var(result, rscale);
+
+ /* Strip leading and trailing zeroes */
+ strip_var(result);
+}
+
+
+/*
+ * mul_var_small() -
+ *
+ * This has the same API as mul_var, but it assumes that var1 has no more
+ * than 3 digits and var2 has at least as many digits as var1. For variables
+ * satisfying these conditions, the product can be computed more quickly than
+ * the general algorithm used in mul_var.
+ */
+static void
+mul_var_small(const NumericVar *var1, const NumericVar *var2,
+ NumericVar *result, int rscale)
+{
+ int var1ndigits = var1->ndigits;
+ int var2ndigits = var2->ndigits;
+ NumericDigit *var1digits = var1->digits;
+ NumericDigit *var2digits = var2->digits;
+ int res_sign;
+ int res_weight;
+ int res_ndigits;
+ NumericDigit *res_buf;
+ NumericDigit *res_digits;
+ uint32 carry;
+ uint32 term;
+
+ /* Check preconditions */
+ Assert(var1ndigits <= 3);
+ Assert(var2ndigits >= var1ndigits);
+
+ /* Determine result sign and (maximum possible) weight */
+ if (var1->sign == var2->sign)
+ res_sign = NUMERIC_POS;
+ else
+ res_sign = NUMERIC_NEG;
+ res_weight = var1->weight + var2->weight + 2;
+
+ /* Determine the number of result digits to compute - cf. mul_var() */
+ res_ndigits = var1ndigits + var2ndigits + 1;
+
+ if (res_ndigits < 3)
+ {
+ /* All input digits will be ignored; so result is zero */
+ zero_var(result);
+ result->dscale = rscale;
+ return;
+ }
+
+ /* Allocate result digit array */
+ res_buf = digitbuf_alloc(res_ndigits);
+ res_buf[0] = 0; /* spare digit for later rounding */
+ res_digits = res_buf + 1;
+
+ /*
+ * Compute the result digits in reverse, in one pass, propagating the
+ * carry up as we go.
+ *
+ * This computes res_digits[res_ndigits - 2], ... res_digits[0] by summing
+ * the products var1digits[i1] * var2digits[i2] for which i1 + i2 + 1 is
+ * the result index.
+ */
+ switch (var1ndigits)
+ {
+ case 1:
+ /* ---------
+ * 1-digit case:
+ * var1ndigits = 1
+ * var2ndigits >= 1
+ * res_ndigits = var2ndigits + 2
+ * ----------
+ */
+ carry = 0;
+ for (int i = res_ndigits - 3; i >= 0; i--)
+ {
+ term = (uint32) var1digits[0] * var2digits[i] + carry;
+ res_digits[i + 1] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+ }
+ res_digits[0] = (NumericDigit) carry;
+ break;
+
+ case 2:
+ /* ---------
+ * 2-digit case:
+ * var1ndigits = 2
+ * var2ndigits >= 2
+ * res_ndigits = var2ndigits + 3
+ * ----------
+ */
+ /* last result digit and carry */
+ term = (uint32) var1digits[1] * var2digits[res_ndigits - 4];
+ res_digits[res_ndigits - 2] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+
+ /* remaining digits, except for the first two */
+ for (int i = res_ndigits - 4; i >= 1; i--)
+ {
+ term = (uint32) var1digits[0] * var2digits[i] +
+ (uint32) var1digits[1] * var2digits[i - 1] + carry;
+ res_digits[i + 1] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+ }
+
+ /* first two digits */
+ term = (uint32) var1digits[0] * var2digits[0] + carry;
+ res_digits[1] = (NumericDigit) (term % NBASE);
+ res_digits[0] = (NumericDigit) (term / NBASE);
+ break;
+
+ case 3:
+ /* ---------
+ * 3-digit case:
+ * var1ndigits = 3
+ * var2ndigits >= 3
+ * res_ndigits = var2ndigits + 4
+ * ----------
+ */
+ /* last two result digits */
+ term = (uint32) var1digits[2] * var2digits[res_ndigits - 5];
+ res_digits[res_ndigits - 2] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+
+ term = (uint32) var1digits[1] * var2digits[res_ndigits - 5] +
+ (uint32) var1digits[2] * var2digits[res_ndigits - 6] + carry;
+ res_digits[res_ndigits - 3] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+
+ /* remaining digits, except for the first three */
+ for (int i = res_ndigits - 5; i >= 2; i--)
+ {
+ term = (uint32) var1digits[0] * var2digits[i] +
+ (uint32) var1digits[1] * var2digits[i - 1] +
+ (uint32) var1digits[2] * var2digits[i - 2] + carry;
+ res_digits[i + 1] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+ }
+
+ /* first three digits */
+ term = (uint32) var1digits[0] * var2digits[1] +
+ (uint32) var1digits[1] * var2digits[0] + carry;
+ res_digits[2] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+ term = (uint32) var1digits[0] * var2digits[0] + carry;
+ res_digits[1] = (NumericDigit) (term % NBASE);
+ res_digits[0] = (NumericDigit) (term / NBASE);
+ break;
+ }
+
+ /* Store the product in result (minus extra rounding digit) */
+ digitbuf_free(result->buf);
+ result->ndigits = res_ndigits - 1;
+ result->buf = res_buf;
+ result->digits = res_digits;
+ result->weight = res_weight - 1;
+ result->sign = res_sign;
+
+ /* Round to target rscale (and set result->dscale) */
round_var(result, rscale);
/* Strip leading and trailing zeroes */
[text/x-patch] v6-add-mul_var_int64.patch (9.0K, ../../CAEZATCXVB15_RfzROFiO_7Fh3voOuOe7FotbvON-6a+ManuGJg@mail.gmail.com/3-v6-add-mul_var_int64.patch)
download | inline diff:
diff --git a/src/backend/utils/adt/numeric.c b/src/backend/utils/adt/numeric.c
new file mode 100644
index 5510a20..96456d8
--- a/src/backend/utils/adt/numeric.c
+++ b/src/backend/utils/adt/numeric.c
@@ -551,6 +551,12 @@ static void sub_var(const NumericVar *va
static void mul_var(const NumericVar *var1, const NumericVar *var2,
NumericVar *result,
int rscale);
+static void mul_var_int(const NumericVar *var, int ival, int ival_weight,
+ NumericVar *result, int rscale);
+#ifdef HAVE_INT128
+static void mul_var_int64(const NumericVar *var, int64 ival, int ival_weight,
+ NumericVar *result, int rscale);
+#endif
static void div_var(const NumericVar *var1, const NumericVar *var2,
NumericVar *result,
int rscale, bool round);
@@ -8707,7 +8713,7 @@ mul_var(const NumericVar *var1, const Nu
var1digits = var1->digits;
var2digits = var2->digits;
- if (var1ndigits == 0 || var2ndigits == 0)
+ if (var1ndigits == 0)
{
/* one or both inputs is zero; so is result */
zero_var(result);
@@ -8715,6 +8721,54 @@ mul_var(const NumericVar *var1, const Nu
return;
}
+ /*
+ * If var1 has just one or two digits, and we are computing the exact
+ * result, with no rounding, delegate to mul_var_int(), which uses a
+ * faster direct multiplication algorithm.
+ *
+ * Similarly, on platforms with 128-bit integer support, delegate to
+ * mul_var_int64() if var1 has three or four digits, and we are computing
+ * the exact result, with no rounding.
+ */
+ if (var1ndigits <= 2 && rscale == var1->dscale + var2->dscale)
+ {
+ int ifactor;
+ int ifactor_weight;
+
+ ifactor = var1->digits[0];
+ ifactor_weight = var1->weight;
+ if (var1ndigits == 2)
+ {
+ ifactor = ifactor * NBASE + var1->digits[1];
+ ifactor_weight--;
+ }
+ if (var1->sign == NUMERIC_NEG)
+ ifactor = -ifactor;
+
+ mul_var_int(var2, ifactor, ifactor_weight, result, rscale);
+ return;
+ }
+#ifdef HAVE_INT128
+ if (var1ndigits <= 4 && rscale == var1->dscale + var2->dscale)
+ {
+ int64 ifactor;
+ int ifactor_weight;
+
+ ifactor = var1->digits[0];
+ ifactor_weight = var1->weight;
+ for (i = 1; i < var1ndigits; i++)
+ {
+ ifactor = ifactor * NBASE + var1->digits[i];
+ ifactor_weight--;
+ }
+ if (var1->sign == NUMERIC_NEG)
+ ifactor = -ifactor;
+
+ mul_var_int64(var2, ifactor, ifactor_weight, result, rscale);
+ return;
+ }
+#endif
+
/* Determine result sign and (maximum possible) weight */
if (var1->sign == var2->sign)
res_sign = NUMERIC_POS;
@@ -8866,6 +8920,233 @@ mul_var(const NumericVar *var1, const Nu
/*
+ * mul_var_int() -
+ *
+ * Multiply a numeric variable by a 32-bit integer with the specified weight.
+ * The product var * ival * NBASE^ival_weight is stored in result.
+ */
+static void
+mul_var_int(const NumericVar *var, int ival, int ival_weight,
+ NumericVar *result, int rscale)
+{
+ NumericDigit *var_digits = var->digits;
+ int var_ndigits = var->ndigits;
+ int res_sign;
+ int res_weight;
+ int res_ndigits;
+ NumericDigit *res_buf;
+ NumericDigit *res_digits;
+ uint32 factor;
+ uint32 carry;
+
+ if (ival == 0 || var_ndigits == 0)
+ {
+ zero_var(result);
+ result->dscale = rscale;
+ return;
+ }
+
+ /*
+ * Determine the result sign, (maximum possible) weight and number of
+ * digits to calculate. The weight figured here is correct if the emitted
+ * product has no leading zero digits; otherwise strip_var() will fix
+ * things up.
+ */
+ if (var->sign == NUMERIC_POS)
+ res_sign = ival > 0 ? NUMERIC_POS : NUMERIC_NEG;
+ else
+ res_sign = ival > 0 ? NUMERIC_NEG : NUMERIC_POS;
+ res_weight = var->weight + ival_weight + 3;
+ /* The number of accurate result digits we need to produce: */
+ res_ndigits = var_ndigits + 3;
+
+ res_buf = digitbuf_alloc(res_ndigits + 1);
+ res_buf[0] = 0; /* spare digit for later rounding */
+ res_digits = res_buf + 1;
+
+ /*
+ * Now compute the product digits by processing the input digits in
+ * reverse and propagating the carry up as we go.
+ *
+ * In this algorithm, the carry from one digit to the next is at most
+ * factor - 1, and product is at most factor * NBASE - 1, and so it needs
+ * to be a 64-bit integer if this exceeds UINT_MAX.
+ */
+ factor = abs(ival);
+ carry = 0;
+
+ if (factor <= UINT_MAX / NBASE)
+ {
+ /* product cannot overflow 32 bits */
+ uint32 product;
+
+ for (int i = res_ndigits - 4; i >= 0; i--)
+ {
+ product = factor * var_digits[i] + carry;
+ res_digits[i + 3] = (NumericDigit) (product % NBASE);
+ carry = product / NBASE;
+ }
+ /* note: carry < UINT_MAX / NBASE in this branch */
+ res_digits[2] = (NumericDigit) (carry % NBASE);
+ carry = carry / NBASE;
+ Assert(carry < NBASE);
+ res_digits[1] = (NumericDigit) carry;
+ res_digits[0] = 0;
+ }
+ else
+ {
+ /* product may exceed 32 bits */
+ uint64 product;
+
+ for (int i = res_ndigits - 4; i >= 0; i--)
+ {
+ product = (uint64) factor * var_digits[i] + carry;
+ res_digits[i + 3] = (NumericDigit) (product % NBASE);
+ carry = (uint32) (product / NBASE);
+ }
+ res_digits[2] = (NumericDigit) (carry % NBASE);
+ carry = carry / NBASE;
+ res_digits[1] = (NumericDigit) (carry % NBASE);
+ res_digits[0] = (NumericDigit) (carry / NBASE);
+ }
+
+ /* Store the product in result */
+ digitbuf_free(result->buf);
+ result->ndigits = res_ndigits;
+ result->buf = res_buf;
+ result->digits = res_digits;
+ result->weight = res_weight;
+ result->sign = res_sign;
+
+ /* Round to target rscale (and set result->dscale) */
+ round_var(result, rscale);
+
+ /* Strip leading and trailing zeroes */
+ strip_var(result);
+}
+
+
+#ifdef HAVE_INT128
+/*
+ * mul_var_int64() -
+ *
+ * Multiply a numeric variable by a 64-bit integer with the specified weight.
+ * The product var * ival * NBASE^ival_weight is stored in result.
+ *
+ * This duplicates the logic in mul_var_int(), so any changes made there
+ * should be made here too.
+ */
+static void
+mul_var_int64(const NumericVar *var, int64 ival, int ival_weight,
+ NumericVar *result, int rscale)
+{
+ NumericDigit *var_digits = var->digits;
+ int var_ndigits = var->ndigits;
+ int res_sign;
+ int res_weight;
+ int res_ndigits;
+ NumericDigit *res_buf;
+ NumericDigit *res_digits;
+ uint64 factor;
+ uint64 carry;
+
+ if (ival == 0 || var_ndigits == 0)
+ {
+ zero_var(result);
+ result->dscale = rscale;
+ return;
+ }
+
+ /*
+ * Determine the result sign, (maximum possible) weight and number of
+ * digits to calculate. The weight figured here is correct if the emitted
+ * product has no leading zero digits; otherwise strip_var() will fix
+ * things up.
+ */
+ if (var->sign == NUMERIC_POS)
+ res_sign = ival > 0 ? NUMERIC_POS : NUMERIC_NEG;
+ else
+ res_sign = ival > 0 ? NUMERIC_NEG : NUMERIC_POS;
+ res_weight = var->weight + ival_weight + 5;
+ /* The number of accurate result digits we need to produce: */
+ res_ndigits = var_ndigits + 5;
+
+ res_buf = digitbuf_alloc(res_ndigits + 1);
+ res_buf[0] = 0; /* spare digit for later rounding */
+ res_digits = res_buf + 1;
+
+ /*
+ * Now compute the product digits by processing the input digits in
+ * reverse and propagating the carry up as we go.
+ *
+ * In this algorithm, the carry from one digit to the next is at most
+ * factor - 1, and product is at most factor * NBASE - 1, and so it needs
+ * to be a 128-bit integer if this exceeds PG_UINT64_MAX.
+ */
+ factor = i64abs(ival);
+ carry = 0;
+
+ if (factor <= PG_UINT64_MAX / NBASE)
+ {
+ /* product cannot overflow 64 bits */
+ uint64 product;
+
+ for (int i = res_ndigits - 6; i >= 0; i--)
+ {
+ product = factor * var_digits[i] + carry;
+ res_digits[i + 5] = (NumericDigit) (product % NBASE);
+ carry = product / NBASE;
+ }
+ /* note: carry < PG_UINT64_MAX / NBASE in this branch */
+ res_digits[4] = (NumericDigit) (carry % NBASE);
+ carry = carry / NBASE;
+ res_digits[3] = (NumericDigit) (carry % NBASE);
+ carry = carry / NBASE;
+ res_digits[2] = (NumericDigit) (carry % NBASE);
+ carry = carry / NBASE;
+ Assert(carry < NBASE);
+ res_digits[1] = (NumericDigit) carry;
+ res_digits[0] = 0;
+ }
+ else
+ {
+ /* product may exceed 64 bits */
+ uint128 product;
+
+ for (int i = res_ndigits - 6; i >= 0; i--)
+ {
+ product = (uint128) factor * var_digits[i] + carry;
+ res_digits[i + 5] = (NumericDigit) (product % NBASE);
+ carry = (uint64) (product / NBASE);
+ }
+ res_digits[4] = (NumericDigit) (carry % NBASE);
+ carry = carry / NBASE;
+ res_digits[3] = (NumericDigit) (carry % NBASE);
+ carry = carry / NBASE;
+ res_digits[2] = (NumericDigit) (carry % NBASE);
+ carry = carry / NBASE;
+ res_digits[1] = (NumericDigit) (carry % NBASE);
+ res_digits[0] = (NumericDigit) (carry / NBASE);
+ }
+
+ /* Store the product in result */
+ digitbuf_free(result->buf);
+ result->ndigits = res_ndigits;
+ result->buf = res_buf;
+ result->digits = res_digits;
+ result->weight = res_weight;
+ result->sign = res_sign;
+
+ /* Round to target rscale (and set result->dscale) */
+ round_var(result, rscale);
+
+ /* Strip leading and trailing zeroes */
+ strip_var(result);
+}
+#endif
+
+
+/*
* div_var() -
*
* Division on variable level. Quotient of var1 / var2 is stored in result.
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: Optimize numeric multiplication for one and two base-NBASE digit multiplicands.
@ 2024-07-05 11:56 Joel Jacobson <[email protected]>
parent: Dean Rasheed <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Joel Jacobson @ 2024-07-05 11:56 UTC (permalink / raw)
To: Dean Rasheed <[email protected]>; +Cc: Dagfinn Ilmari Mannsåker <[email protected]>; pgsql-hackers
On Thu, Jul 4, 2024, at 20:43, Dean Rasheed wrote:
> On Wed, 3 Jul 2024 at 21:45, Joel Jacobson <[email protected]> wrote:
>>
>> > On Wed, Jul 3, 2024, at 20:57, Dean Rasheed wrote:
>> >> I wouldn't expect it to ever be off by more than 1
>> >
>> > OK, so then the cases I found where it was off by 2 for the mul_var_int() patch
>> > are unexpected?
>>
>> Sorry, I meant off by 2 for the mul_var_small() patch, these cases that I found:
>>
>
> Yeah, so that was another bug in mul_var_small(). If rscale is made
> small enough, the result index for the digits computed before the main
> loop overlaps the ones after, so it would overwrite digits already
> computed.
>
> Of course, that's fairly easy to fix, but at this point I think the
> better solution is to only use mul_var_small() when an exact product
> is requested. We would have to do that for mul_var_int() anyway,
> because of its accuracy issues discussed earlier. I think this is a
> reasonable thing to do because only functions like ln_var() and
> exp_var() will ask mul_var() for a reduced-rscale result, and those
> functions are likely to be dominated by computations involving larger
> numbers, for which this patch wouldn't help anyway. Also those
> functions are probably less widely used.
>
> If we make that decision, a lot of the complexity in mul_var_small()
> goes away, including all the conditional array accesses, making it
> simpler and more efficient. v6 patch attached.
Nice, I think that looks like the better trade-off.
> I also updated the mul_var_int() patch so that it is also only invoked
> when an exact product is requested, and I noticed a couple of other
> minor optimisations that could be made.
Looks nice.
> Then I decided to try
> implementing mul_var_int64(). This gives a pretty decent speedup for
> 3-digit inputs, but unfortunately it is much slower for 4-digit inputs
> (for which most values will go through the 128-bit code path). I'm
> attaching that too, just for information, but it's clearly not going
> to be acceptable as-is.
>
> Running your benchmark queries, I got these results:
>
> SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_1;
> Time: 4520.874 ms (00:04.521) -- HEAD
> Time: 3937.536 ms (00:03.938) -- v5-mul_var_int.patch
> Time: 3919.495 ms (00:03.919) -- v5-mul_var_small.patch
> Time: 3916.964 ms (00:03.917) -- v6-mul_var_int64.patch
> Time: 3811.118 ms (00:03.811) -- v6-mul_var_small.patch
My benchmarks also indicate v6-mul_var_small.patch (v6-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch)
is the winner on all my three CPUs, for var1ndigits=1:
-- Apple M3 Max
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_1; -- HEAD
Time: 3046.668 ms (00:03.047)
Time: 3053.327 ms (00:03.053)
Time: 3045.517 ms (00:03.046)
Time: 3049.626 ms (00:03.050)
Time: 3075.101 ms (00:03.075)
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_1; -- v6-add-mul_var_int64.patch
Time: 2781.536 ms (00:02.782)
Time: 2781.324 ms (00:02.781)
Time: 2781.301 ms (00:02.781)
Time: 2786.524 ms (00:02.787)
Time: 2784.494 ms (00:02.784)
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_1; -- v6-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch
Time: 2711.167 ms (00:02.711)
Time: 2723.647 ms (00:02.724)
Time: 2706.372 ms (00:02.706)
Time: 2708.883 ms (00:02.709)
Time: 2704.621 ms (00:02.705)
-- Intel Core i9-14900K
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_1; -- HEAD
Time: 3496.714 ms (00:03.497)
Time: 3278.909 ms (00:03.279)
Time: 3278.631 ms (00:03.279)
Time: 3277.658 ms (00:03.278)
Time: 3276.121 ms (00:03.276)
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_1; -- v6-add-mul_var_int64.patch
Time: 3080.751 ms (00:03.081)
Time: 3078.118 ms (00:03.078)
Time: 3079.932 ms (00:03.080)
Time: 3080.668 ms (00:03.081)
Time: 3080.697 ms (00:03.081)
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_1; -- v6-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch
Time: 3043.635 ms (00:03.044)
Time: 3043.606 ms (00:03.044)
Time: 3041.391 ms (00:03.041)
Time: 3041.997 ms (00:03.042)
Time: 3045.464 ms (00:03.045)
-- AMD Ryzen 9 7950X3D
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_1; -- HEAD
Time: 3421.307 ms (00:03.421)
Time: 3400.935 ms (00:03.401)
Time: 3359.839 ms (00:03.360)
Time: 3374.246 ms (00:03.374)
Time: 3375.085 ms (00:03.375)
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_1; -- v6-add-mul_var_int64.patch
Time: 3002.214 ms (00:03.002)
Time: 3016.042 ms (00:03.016)
Time: 3010.541 ms (00:03.011)
Time: 3009.204 ms (00:03.009)
Time: 3002.088 ms (00:03.002)
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_1; -- v6-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch
Time: 2959.319 ms (00:02.959)
Time: 2957.694 ms (00:02.958)
Time: 2971.559 ms (00:02.972)
Time: 2958.788 ms (00:02.959)
Time: 2958.812 ms (00:02.959)
> SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_2;
> Time: 4762.528 ms (00:04.763) -- HEAD
> Time: 4075.546 ms (00:04.076) -- v5-mul_var_int.patch
> Time: 4055.180 ms (00:04.055) -- v5-mul_var_small.patch
> Time: 4037.866 ms (00:04.038) -- v6-mul_var_int64.patch
> Time: 4018.488 ms (00:04.018) -- v6-mul_var_small.patch
I get mixed results for var1ndigits=2:
Winner on Apple M3 Max and AMD Ryzen 9 7950X3D is v6-add-mul_var_int64.patch.
Winner on Intel Core i9-14900K is v6-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch
-- Apple M3 Max
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_2; -- HEAD
Time: 3369.724 ms (00:03.370)
Time: 3340.977 ms (00:03.341)
Time: 3331.407 ms (00:03.331)
Time: 3333.304 ms (00:03.333)
Time: 3332.136 ms (00:03.332)
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_2; -- v6-add-mul_var_int64.patch
Time: 2768.108 ms (00:02.768)
Time: 2736.996 ms (00:02.737)
Time: 2730.217 ms (00:02.730)
Time: 2743.556 ms (00:02.744)
Time: 2746.725 ms (00:02.747)
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_2; -- v6-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch
Time: 2895.200 ms (00:02.895)
Time: 2894.823 ms (00:02.895)
Time: 2899.475 ms (00:02.899)
Time: 2895.385 ms (00:02.895)
Time: 2898.647 ms (00:02.899)
-- Intel Core i9-14900K
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_2; -- HEAD
Time: 3385.836 ms (00:03.386)
Time: 3367.739 ms (00:03.368)
Time: 3363.321 ms (00:03.363)
Time: 3365.433 ms (00:03.365)
Time: 3365.301 ms (00:03.365)
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_2; -- v6-add-mul_var_int64.patch
Time: 3086.253 ms (00:03.086)
Time: 3085.272 ms (00:03.085)
Time: 3085.769 ms (00:03.086)
Time: 3089.128 ms (00:03.089)
Time: 3086.735 ms (00:03.087)
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_2; -- v6-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch
Time: 3053.775 ms (00:03.054)
Time: 3058.392 ms (00:03.058)
Time: 3068.113 ms (00:03.068)
Time: 3057.333 ms (00:03.057)
Time: 3057.218 ms (00:03.057)
-- AMD Ryzen 9 7950X3D
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_2; -- HEAD
Time: 3619.441 ms (00:03.619)
Time: 3609.553 ms (00:03.610)
Time: 3574.277 ms (00:03.574)
Time: 3578.031 ms (00:03.578)
Time: 3558.545 ms (00:03.559)
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_2; -- v6-add-mul_var_int64.patch
Time: 3061.858 ms (00:03.062)
Time: 3082.174 ms (00:03.082)
Time: 3081.093 ms (00:03.081)
Time: 3093.610 ms (00:03.094)
Time: 3064.507 ms (00:03.065)
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_2; -- v6-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch
Time: 3100.287 ms (00:03.100)
Time: 3100.264 ms (00:03.100)
Time: 3097.207 ms (00:03.097)
Time: 3101.477 ms (00:03.101)
Time: 3098.035 ms (00:03.098)
> SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_3;
> Time: 5387.514 ms (00:05.388) -- HEAD
> Time: 5350.736 ms (00:05.351) -- v5-mul_var_int.patch
> Time: 4648.449 ms (00:04.648) -- v5-mul_var_small.patch
> Time: 4655.204 ms (00:04.655) -- v6-mul_var_int64.patch
> Time: 4645.962 ms (00:04.646) -- v6-mul_var_small.patch
I get mixed results for var1ndigits=3:
Winner on Apple M3 Max and AMD Ryzen 9 7950X3D is v6-add-mul_var_int64.patch.
Winner on Intel Core i9-14900K is v6-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch
Same winners as for var1ndigits=2.
-- Apple M3 Max
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_3; -- HEAD
Time: 3466.932 ms (00:03.467)
Time: 3447.001 ms (00:03.447)
Time: 3457.259 ms (00:03.457)
Time: 3445.938 ms (00:03.446)
Time: 3443.310 ms (00:03.443)
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_3; -- v6-add-mul_var_int64.patch
Time: 2988.444 ms (00:02.988)
Time: 2976.036 ms (00:02.976)
Time: 2982.756 ms (00:02.983)
Time: 2986.436 ms (00:02.986)
Time: 2973.457 ms (00:02.973)
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_3; -- v6-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch
Time: 3026.666 ms (00:03.027)
Time: 3024.458 ms (00:03.024)
Time: 3022.976 ms (00:03.023)
Time: 3029.526 ms (00:03.030)
Time: 3021.543 ms (00:03.022)
-- Intel Core i9-14900K
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_3; -- HEAD
Time: 4510.078 ms (00:04.510)
Time: 4222.501 ms (00:04.223)
Time: 4179.509 ms (00:04.180)
Time: 4179.307 ms (00:04.179)
Time: 4183.026 ms (00:04.183)
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_3; -- v6-add-mul_var_int64.patch
Time: 3811.866 ms (00:03.812)
Time: 3816.664 ms (00:03.817)
Time: 3811.695 ms (00:03.812)
Time: 3811.674 ms (00:03.812)
Time: 3812.265 ms (00:03.812)
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_3; -- v6-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch
Time: 4095.053 ms (00:04.095)
Time: 3896.002 ms (00:03.896)
Time: 3888.999 ms (00:03.889)
Time: 3889.346 ms (00:03.889)
Time: 3889.017 ms (00:03.889)
-- AMD Ryzen 9 7950X3D
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_3; -- HEAD
Time: 3946.255 ms (00:03.946)
Time: 3896.110 ms (00:03.896)
Time: 3877.470 ms (00:03.877)
Time: 3854.402 ms (00:03.854)
Time: 3901.218 ms (00:03.901)
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_3; -- v6-add-mul_var_int64.patch
Time: 3393.572 ms (00:03.394)
Time: 3395.401 ms (00:03.395)
Time: 3395.199 ms (00:03.395)
Time: 3418.555 ms (00:03.419)
Time: 3399.619 ms (00:03.400)
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_3; -- v6-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch
Time: 3308.791 ms (00:03.309)
Time: 3309.316 ms (00:03.309)
Time: 3318.238 ms (00:03.318)
Time: 3296.061 ms (00:03.296)
Time: 3320.282 ms (00:03.320)
> SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_4;
> Time: 5617.150 ms (00:05.617) -- HEAD
> Time: 5505.913 ms (00:05.506) -- v5-mul_var_int.patch
> Time: 5486.441 ms (00:05.486) -- v5-mul_var_small.patch
> Time: 8203.081 ms (00:08.203) -- v6-mul_var_int64.patch
> Time: 5598.909 ms (00:05.599) -- v6-mul_var_small.patch
> So v6-mul_var_int64 improves on v5-mul_var_int in the 3-digit case,
> but is terrible in the 4-digit case. None of the other patches touch
> the 4-digit case, but it might be interesting to try mul_var_small()
> with 4 digits.
Interesting you got so bad bench results for v6-mul_var_int64.patch
for var1ndigits=4, that patch is actually the winner on AMD Ryzen 9 7950X3D.
On Intel Core i9-14900K the winner is v6-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch.
On Apple M3 Max, HEAD is the winner.
-- Apple M3 Max
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_4; -- HEAD
Time: 3618.530 ms (00:03.619)
Time: 3595.239 ms (00:03.595)
Time: 3600.412 ms (00:03.600)
Time: 3607.500 ms (00:03.607)
Time: 3604.122 ms (00:03.604)
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_4; -- v6-add-mul_var_int64.patch
Time: 4498.993 ms (00:04.499)
Time: 4527.302 ms (00:04.527)
Time: 4493.613 ms (00:04.494)
Time: 4482.194 ms (00:04.482)
Time: 4493.660 ms (00:04.494)
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_4; -- v6-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch
Time: 3628.177 ms (00:03.628)
Time: 3613.975 ms (00:03.614)
Time: 3612.213 ms (00:03.612)
Time: 3614.026 ms (00:03.614)
Time: 3622.959 ms (00:03.623)
-- Intel Core i9-14900K
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_4; -- HEAD
Time: 4130.810 ms (00:04.131)
Time: 3836.462 ms (00:03.836)
Time: 3810.604 ms (00:03.811)
Time: 3805.443 ms (00:03.805)
Time: 3803.055 ms (00:03.803)
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_4; -- v6-add-mul_var_int64.patch
Time: 3952.862 ms (00:03.953)
Time: 3792.272 ms (00:03.792)
Time: 3793.995 ms (00:03.794)
Time: 3790.531 ms (00:03.791)
Time: 3793.647 ms (00:03.794)
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_4; -- v6-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch
Time: 3762.828 ms (00:03.763)
Time: 3754.869 ms (00:03.755)
Time: 3756.041 ms (00:03.756)
Time: 3758.719 ms (00:03.759)
Time: 3754.894 ms (00:03.755)
-- AMD Ryzen 9 7950X3D
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_4; -- HEAD
Time: 4075.988 ms (00:04.076)
Time: 4067.702 ms (00:04.068)
Time: 4035.191 ms (00:04.035)
Time: 4022.411 ms (00:04.022)
Time: 4016.475 ms (00:04.016)
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_4; -- v6-add-mul_var_int64.patch
Time: 3830.021 ms (00:03.830)
Time: 3837.947 ms (00:03.838)
Time: 3834.894 ms (00:03.835)
Time: 3832.755 ms (00:03.833)
Time: 3834.512 ms (00:03.835)
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_4; -- v6-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch
Time: 4031.128 ms (00:04.031)
Time: 4001.590 ms (00:04.002)
Time: 4031.212 ms (00:04.031)
Time: 4035.941 ms (00:04.036)
Time: 4031.218 ms (00:04.031)
Regards,
Joel
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: Optimize numeric multiplication for one and two base-NBASE digit multiplicands.
@ 2024-07-05 15:41 Dean Rasheed <[email protected]>
parent: Joel Jacobson <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Dean Rasheed @ 2024-07-05 15:41 UTC (permalink / raw)
To: Joel Jacobson <[email protected]>; +Cc: Dagfinn Ilmari Mannsåker <[email protected]>; pgsql-hackers
On Fri, 5 Jul 2024 at 12:56, Joel Jacobson <[email protected]> wrote:
>
> Interesting you got so bad bench results for v6-mul_var_int64.patch
> for var1ndigits=4, that patch is actually the winner on AMD Ryzen 9 7950X3D.
Interesting.
> On Intel Core i9-14900K the winner is v6-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch.
That must be random noise, since
v6-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch doesn't
invoke mul_var_small() for 4-digit inputs.
> On Apple M3 Max, HEAD is the winner.
Importantly, mul_var_int64() is around 1.25x slower there, and it was
even worse on my machine.
Attached is a v7 mul_var_small() patch adding 4-digit support. For me,
this gives a nice speedup:
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_4;
Time: 5617.150 ms (00:05.617) -- HEAD
Time: 8203.081 ms (00:08.203) -- v6-mul_var_int64.patch
Time: 4750.212 ms (00:04.750) -- v7-mul_var_small.patch
The other advantage, of course, is that it doesn't require 128-bit
integer support.
Regards,
Dean
Attachments:
[text/x-patch] v7-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch (8.2K, ../../CAEZATCWY_h7jTzsQnZY3bChNF85W4KLYV-rRgy=cc4QAUXEUdg@mail.gmail.com/2-v7-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch)
download | inline diff:
diff --git a/src/backend/utils/adt/numeric.c b/src/backend/utils/adt/numeric.c
new file mode 100644
index 5510a20..b9497e1
--- a/src/backend/utils/adt/numeric.c
+++ b/src/backend/utils/adt/numeric.c
@@ -551,6 +551,8 @@ static void sub_var(const NumericVar *va
static void mul_var(const NumericVar *var1, const NumericVar *var2,
NumericVar *result,
int rscale);
+static void mul_var_small(const NumericVar *var1, const NumericVar *var2,
+ NumericVar *result, int rscale);
static void div_var(const NumericVar *var1, const NumericVar *var2,
NumericVar *result,
int rscale, bool round);
@@ -8707,7 +8709,7 @@ mul_var(const NumericVar *var1, const Nu
var1digits = var1->digits;
var2digits = var2->digits;
- if (var1ndigits == 0 || var2ndigits == 0)
+ if (var1ndigits == 0)
{
/* one or both inputs is zero; so is result */
zero_var(result);
@@ -8715,6 +8717,17 @@ mul_var(const NumericVar *var1, const Nu
return;
}
+ /*
+ * If var1 has 4 digits or fewer, and we are computing the exact result,
+ * with no rounding, delegate to mul_var_small() which uses a faster short
+ * multiplication algorithm.
+ */
+ if (var1ndigits <= 4 && rscale == var1->dscale + var2->dscale)
+ {
+ mul_var_small(var1, var2, result, rscale);
+ return;
+ }
+
/* Determine result sign and (maximum possible) weight */
if (var1->sign == var2->sign)
res_sign = NUMERIC_POS;
@@ -8858,6 +8871,221 @@ mul_var(const NumericVar *var1, const Nu
result->sign = res_sign;
/* Round to target rscale (and set result->dscale) */
+ round_var(result, rscale);
+
+ /* Strip leading and trailing zeroes */
+ strip_var(result);
+}
+
+
+/*
+ * mul_var_small() -
+ *
+ * This has the same API as mul_var, but it assumes that var1 has no more
+ * than 4 digits and var2 has at least as many digits as var1. For variables
+ * satisfying these conditions, the product can be computed more quickly than
+ * the general algorithm used in mul_var.
+ */
+static void
+mul_var_small(const NumericVar *var1, const NumericVar *var2,
+ NumericVar *result, int rscale)
+{
+ int var1ndigits = var1->ndigits;
+ int var2ndigits = var2->ndigits;
+ NumericDigit *var1digits = var1->digits;
+ NumericDigit *var2digits = var2->digits;
+ int res_sign;
+ int res_weight;
+ int res_ndigits;
+ NumericDigit *res_buf;
+ NumericDigit *res_digits;
+ uint32 carry;
+ uint32 term;
+
+ /* Check preconditions */
+ Assert(var1ndigits <= 4);
+ Assert(var2ndigits >= var1ndigits);
+
+ /* Determine result sign and (maximum possible) weight */
+ if (var1->sign == var2->sign)
+ res_sign = NUMERIC_POS;
+ else
+ res_sign = NUMERIC_NEG;
+ res_weight = var1->weight + var2->weight + 2;
+
+ /* Determine the number of result digits to compute - cf. mul_var() */
+ res_ndigits = var1ndigits + var2ndigits + 1;
+
+ if (res_ndigits < 3)
+ {
+ /* All input digits will be ignored; so result is zero */
+ zero_var(result);
+ result->dscale = rscale;
+ return;
+ }
+
+ /* Allocate result digit array */
+ res_buf = digitbuf_alloc(res_ndigits);
+ res_buf[0] = 0; /* spare digit for later rounding */
+ res_digits = res_buf + 1;
+
+ /*
+ * Compute the result digits in reverse, in one pass, propagating the
+ * carry up as we go.
+ *
+ * This computes res_digits[res_ndigits - 2], ... res_digits[0] by summing
+ * the products var1digits[i1] * var2digits[i2] for which i1 + i2 + 1 is
+ * the result index.
+ */
+ switch (var1ndigits)
+ {
+ case 1:
+ /* ---------
+ * 1-digit case:
+ * var1ndigits = 1
+ * var2ndigits >= 1
+ * res_ndigits = var2ndigits + 2
+ * ----------
+ */
+ carry = 0;
+ for (int i = res_ndigits - 3; i >= 0; i--)
+ {
+ term = (uint32) var1digits[0] * var2digits[i] + carry;
+ res_digits[i + 1] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+ }
+ res_digits[0] = (NumericDigit) carry;
+ break;
+
+ case 2:
+ /* ---------
+ * 2-digit case:
+ * var1ndigits = 2
+ * var2ndigits >= 2
+ * res_ndigits = var2ndigits + 3
+ * ----------
+ */
+ /* last result digit and carry */
+ term = (uint32) var1digits[1] * var2digits[res_ndigits - 4];
+ res_digits[res_ndigits - 2] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+
+ /* remaining digits, except for the first two */
+ for (int i = res_ndigits - 4; i >= 1; i--)
+ {
+ term = (uint32) var1digits[0] * var2digits[i] +
+ (uint32) var1digits[1] * var2digits[i - 1] + carry;
+ res_digits[i + 1] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+ }
+
+ /* first two digits */
+ term = (uint32) var1digits[0] * var2digits[0] + carry;
+ res_digits[1] = (NumericDigit) (term % NBASE);
+ res_digits[0] = (NumericDigit) (term / NBASE);
+ break;
+
+ case 3:
+ /* ---------
+ * 3-digit case:
+ * var1ndigits = 3
+ * var2ndigits >= 3
+ * res_ndigits = var2ndigits + 4
+ * ----------
+ */
+ /* last two result digits */
+ term = (uint32) var1digits[2] * var2digits[res_ndigits - 5];
+ res_digits[res_ndigits - 2] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+
+ term = (uint32) var1digits[1] * var2digits[res_ndigits - 5] +
+ (uint32) var1digits[2] * var2digits[res_ndigits - 6] + carry;
+ res_digits[res_ndigits - 3] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+
+ /* remaining digits, except for the first three */
+ for (int i = res_ndigits - 5; i >= 2; i--)
+ {
+ term = (uint32) var1digits[0] * var2digits[i] +
+ (uint32) var1digits[1] * var2digits[i - 1] +
+ (uint32) var1digits[2] * var2digits[i - 2] + carry;
+ res_digits[i + 1] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+ }
+
+ /* first three digits */
+ term = (uint32) var1digits[0] * var2digits[1] +
+ (uint32) var1digits[1] * var2digits[0] + carry;
+ res_digits[2] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+
+ term = (uint32) var1digits[0] * var2digits[0] + carry;
+ res_digits[1] = (NumericDigit) (term % NBASE);
+ res_digits[0] = (NumericDigit) (term / NBASE);
+ break;
+
+ case 4:
+ /* ---------
+ * 4-digit case:
+ * var1ndigits = 4
+ * var2ndigits >= 4
+ * res_ndigits = var2ndigits + 5
+ * ----------
+ */
+ /* last three result digits */
+ term = (uint32) var1digits[3] * var2digits[res_ndigits - 6];
+ res_digits[res_ndigits - 2] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+
+ term = (uint32) var1digits[2] * var2digits[res_ndigits - 6] +
+ (uint32) var1digits[3] * var2digits[res_ndigits - 7] + carry;
+ res_digits[res_ndigits - 3] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+
+ term = (uint32) var1digits[1] * var2digits[res_ndigits - 6] +
+ (uint32) var1digits[2] * var2digits[res_ndigits - 7] +
+ (uint32) var1digits[3] * var2digits[res_ndigits - 8] + carry;
+ res_digits[res_ndigits - 4] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+
+ /* remaining digits, except for the first four */
+ for (int i = res_ndigits - 6; i >= 3; i--)
+ {
+ term = (uint32) var1digits[0] * var2digits[i] +
+ (uint32) var1digits[1] * var2digits[i - 1] +
+ (uint32) var1digits[2] * var2digits[i - 2] +
+ (uint32) var1digits[3] * var2digits[i - 3] + carry;
+ res_digits[i + 1] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+ }
+
+ /* first four digits */
+ term = (uint32) var1digits[0] * var2digits[2] +
+ (uint32) var1digits[1] * var2digits[1] +
+ (uint32) var1digits[2] * var2digits[0] + carry;
+ res_digits[3] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+
+ term = (uint32) var1digits[0] * var2digits[1] +
+ (uint32) var1digits[1] * var2digits[0] + carry;
+ res_digits[2] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+
+ term = (uint32) var1digits[0] * var2digits[0] + carry;
+ res_digits[1] = (NumericDigit) (term % NBASE);
+ res_digits[0] = (NumericDigit) (term / NBASE);
+ break;
+ }
+
+ /* Store the product in result (minus extra rounding digit) */
+ digitbuf_free(result->buf);
+ result->ndigits = res_ndigits - 1;
+ result->buf = res_buf;
+ result->digits = res_digits;
+ result->weight = res_weight - 1;
+ result->sign = res_sign;
+
+ /* Round to target rscale (and set result->dscale) */
round_var(result, rscale);
/* Strip leading and trailing zeroes */
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: Optimize numeric multiplication for one and two base-NBASE digit multiplicands.
@ 2024-07-05 16:42 Joel Jacobson <[email protected]>
parent: Dean Rasheed <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Joel Jacobson @ 2024-07-05 16:42 UTC (permalink / raw)
To: Dean Rasheed <[email protected]>; +Cc: Dagfinn Ilmari Mannsåker <[email protected]>; pgsql-hackers
On Fri, Jul 5, 2024, at 17:41, Dean Rasheed wrote:
> On Fri, 5 Jul 2024 at 12:56, Joel Jacobson <[email protected]> wrote:
>>
>> Interesting you got so bad bench results for v6-mul_var_int64.patch
>> for var1ndigits=4, that patch is actually the winner on AMD Ryzen 9 7950X3D.
>
> Interesting.
I remeasured just to be sure, and yes, it was the winner among the previous patches,
but the new v7 beats it.
>> On Intel Core i9-14900K the winner is v6-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch.
>
> That must be random noise, since
> v6-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch doesn't
> invoke mul_var_small() for 4-digit inputs.
Yes, something was off with the HEAD measurements for that one,
I remeasured and then got almost identical results (as expected)
between HEAD and v6-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch
for 4-digit inputs.
>> On Apple M3 Max, HEAD is the winner.
>
> Importantly, mul_var_int64() is around 1.25x slower there, and it was
> even worse on my machine.
>
> Attached is a v7 mul_var_small() patch adding 4-digit support. For me,
> this gives a nice speedup:
>
> SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_4;
> Time: 5617.150 ms (00:05.617) -- HEAD
> Time: 8203.081 ms (00:08.203) -- v6-mul_var_int64.patch
> Time: 4750.212 ms (00:04.750) -- v7-mul_var_small.patch
>
> The other advantage, of course, is that it doesn't require 128-bit
> integer support.
Very nice, v7-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch
is now the winner on all my CPUs:
-- Apple M3 Max
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_4; -- HEAD
Time: 3574.865 ms (00:03.575)
Time: 3573.678 ms (00:03.574)
Time: 3576.953 ms (00:03.577)
Time: 3580.536 ms (00:03.581)
Time: 3589.007 ms (00:03.589)
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_4; -- v7-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch
Time: 3110.171 ms (00:03.110)
Time: 3098.558 ms (00:03.099)
Time: 3105.873 ms (00:03.106)
Time: 3104.484 ms (00:03.104)
Time: 3109.035 ms (00:03.109)
-- Intel Core i9-14900K
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_4; -- HEAD
Time: 3751.767 ms (00:03.752)
Time: 3745.916 ms (00:03.746)
Time: 3742.542 ms (00:03.743)
Time: 3746.139 ms (00:03.746)
Time: 3745.493 ms (00:03.745)
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_4; -- v6-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch
Time: 3747.640 ms (00:03.748)
Time: 3747.231 ms (00:03.747)
Time: 3747.965 ms (00:03.748)
Time: 3748.309 ms (00:03.748)
Time: 3746.498 ms (00:03.746)
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_4; -- v7-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch
Time: 3417.924 ms (00:03.418)
Time: 3417.088 ms (00:03.417)
Time: 3415.708 ms (00:03.416)
Time: 3415.453 ms (00:03.415)
Time: 3419.566 ms (00:03.420)
-- AMD Ryzen 9 7950X3D
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_4; -- HEAD
Time: 3970.131 ms (00:03.970)
Time: 3924.335 ms (00:03.924)
Time: 3927.863 ms (00:03.928)
Time: 3924.761 ms (00:03.925)
Time: 3926.290 ms (00:03.926)
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_4; -- v6-add-mul_var_int64.patch
Time: 3874.769 ms (00:03.875)
Time: 3858.071 ms (00:03.858)
Time: 3836.698 ms (00:03.837)
Time: 3871.388 ms (00:03.871)
Time: 3844.907 ms (00:03.845)
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_4; -- v7-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch
Time: 3397.846 ms (00:03.398)
Time: 3398.050 ms (00:03.398)
Time: 3395.279 ms (00:03.395)
Time: 3393.285 ms (00:03.393)
Time: 3402.570 ms (00:03.403)
Code wise I think it's now very nice and clear, with just enough comments.
Also nice to see that the var1ndigits=4 case isn't much more complex
than var1ndigits=3, since it follows the same pattern.
Regards,
Joel
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: Optimize numeric multiplication for one and two base-NBASE digit multiplicands.
@ 2024-07-05 17:37 Joel Jacobson <[email protected]>
parent: Joel Jacobson <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Joel Jacobson @ 2024-07-05 17:37 UTC (permalink / raw)
To: Dean Rasheed <[email protected]>; +Cc: Dagfinn Ilmari Mannsåker <[email protected]>; pgsql-hackers
On Fri, Jul 5, 2024, at 18:42, Joel Jacobson wrote:
> Very nice, v7-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch
> is now the winner on all my CPUs:
I thought it would be interesting to also measure the isolated effect
on just numeric_mul() without the query overhead.
Included var1ndigits=5 var2ndigits=5, that should be unaffected,
just to get a sense of the noise level.
SELECT timeit.h('numeric_mul',array['9999','9999'],2,min_time:='1 s'::interval);
SELECT timeit.h('numeric_mul',array['9999_9999','9999_9999'],2,min_time:='1 s'::interval);
SELECT timeit.h('numeric_mul',array['9999_9999_9999','9999_9999_9999'],2,min_time:='1 s'::interval);
SELECT timeit.h('numeric_mul',array['9999_9999_9999_9999','9999_9999_9999_9999'],2,min_time:='1 s'::interval);
SELECT timeit.h('numeric_mul',array['9999_9999_9999_9999_9999','9999_9999_9999_9999_9999'],2,min_time:='1 s'::interval);
CPU | var1ndigits | var2ndigits | HEAD | v7 | HEAD/v7
---------------------+-------------+-------------+-------+-------+---------
Apple M3 Max | 1 | 1 | 28 ns | 18 ns | 1.56
Apple M3 Max | 2 | 2 | 32 ns | 18 ns | 1.78
Apple M3 Max | 3 | 3 | 38 ns | 21 ns | 1.81
Apple M3 Max | 4 | 4 | 42 ns | 24 ns | 1.75
Intel Core i9-14900K | 1 | 1 | 25 ns | 20 ns | 1.25
Intel Core i9-14900K | 2 | 2 | 28 ns | 20 ns | 1.40
Intel Core i9-14900K | 3 | 3 | 33 ns | 24 ns | 1.38
Intel Core i9-14900K | 4 | 4 | 37 ns | 25 ns | 1.48
AMD Ryzen 9 7950X3D | 1 | 1 | 37 ns | 29 ns | 1.28
AMD Ryzen 9 7950X3D | 2 | 2 | 43 ns | 31 ns | 1.39
AMD Ryzen 9 7950X3D | 3 | 3 | 50 ns | 37 ns | 1.35
AMD Ryzen 9 7950X3D | 4 | 4 | 55 ns | 39 ns | 1.41
Impressive speed-up, between 25% - 81%.
Regards,
Joel
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: Optimize numeric multiplication for one and two base-NBASE digit multiplicands.
@ 2024-07-06 09:34 Dean Rasheed <[email protected]>
parent: Joel Jacobson <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Dean Rasheed @ 2024-07-06 09:34 UTC (permalink / raw)
To: Joel Jacobson <[email protected]>; +Cc: Dagfinn Ilmari Mannsåker <[email protected]>; pgsql-hackers
On Fri, 5 Jul 2024 at 18:37, Joel Jacobson <[email protected]> wrote:
>
> On Fri, Jul 5, 2024, at 18:42, Joel Jacobson wrote:
> > Very nice, v7-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch
> > is now the winner on all my CPUs:
>
> I thought it would be interesting to also measure the isolated effect
> on just numeric_mul() without the query overhead.
>
> Impressive speed-up, between 25% - 81%.
>
Cool. I think we should go with the mul_var_small() patch then, since
it's more generally applicable.
I also did some testing with much larger var2 values, and saw similar
speed-ups. One high-level function that benefits from that is
factorial(), which accepts inputs up to 32177, and so uses both the
1-digit and 2-digit code with very large var2 values. I doubt anyone
actually uses it with such large inputs, but it's interesting
nonetheless:
SELECT factorial(32177);
Time: 923.117 ms -- HEAD
Time: 534.375 ms -- mul_var_small() patch
I did one more round of (mostly cosmetic) copy-editing. Aside from
improving some of the comments, it occurred to me that there's no need
to pass rscale to mul_var_small(), or for it to call round_var(),
since it's always computing the exact result. That shaves off a few
more cycles.
Additionally, I didn't like how res_weight and res_ndigits were being
set 1 higher than they needed to be. That makes sense in mul_var()
because it may round the result, causing a non-zero carry to propagate
into the next digit up, but it's just confusing in mul_var_small(). So
I've reduced those by 1, which makes the look much more logical. To be
clear, this doesn't change how many digits we're calculating. But now
res_ndigits is actually the number of digits being calculated, whereas
before, res_ndigits was 1 larger and we were calculating res_ndigits -
1 digits, which was confusing.
I think this is good to go, so unless there are any further comments,
I plan to commit it soon.
Possible future work would be to try extending it to larger var1
values. I have a feeling that might work quite well for 5 or 6 digits,
but at some point, we'll start seeing diminishing returns, and the
code bloat won't be worth it.
Regards,
Dean
Attachments:
[text/x-patch] v8-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch (7.9K, ../../CAEZATCUmrq4pgJdQNQendiPkmyM=qdnZv1g-vtoMyrOyCi6xdA@mail.gmail.com/2-v8-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch)
download | inline diff:
diff --git a/src/backend/utils/adt/numeric.c b/src/backend/utils/adt/numeric.c
new file mode 100644
index 5510a20..b556861
--- a/src/backend/utils/adt/numeric.c
+++ b/src/backend/utils/adt/numeric.c
@@ -551,6 +551,8 @@ static void sub_var(const NumericVar *va
static void mul_var(const NumericVar *var1, const NumericVar *var2,
NumericVar *result,
int rscale);
+static void mul_var_small(const NumericVar *var1, const NumericVar *var2,
+ NumericVar *result);
static void div_var(const NumericVar *var1, const NumericVar *var2,
NumericVar *result,
int rscale, bool round);
@@ -8707,7 +8709,7 @@ mul_var(const NumericVar *var1, const Nu
var1digits = var1->digits;
var2digits = var2->digits;
- if (var1ndigits == 0 || var2ndigits == 0)
+ if (var1ndigits == 0)
{
/* one or both inputs is zero; so is result */
zero_var(result);
@@ -8715,6 +8717,16 @@ mul_var(const NumericVar *var1, const Nu
return;
}
+ /*
+ * If var1 has 1-4 digits and the exact result was requested, delegate to
+ * mul_var_small() which uses a faster direct multiplication algorithm.
+ */
+ if (var1ndigits <= 4 && rscale == var1->dscale + var2->dscale)
+ {
+ mul_var_small(var1, var2, result);
+ return;
+ }
+
/* Determine result sign and (maximum possible) weight */
if (var1->sign == var2->sign)
res_sign = NUMERIC_POS;
@@ -8862,6 +8874,212 @@ mul_var(const NumericVar *var1, const Nu
/* Strip leading and trailing zeroes */
strip_var(result);
+}
+
+
+/*
+ * mul_var_small() -
+ *
+ * Special-case multiplication function used when var1 has 1-4 digits, var2
+ * has at least as many digits as var1, and the exact product var1 * var2 is
+ * requested.
+ */
+static void
+mul_var_small(const NumericVar *var1, const NumericVar *var2,
+ NumericVar *result)
+{
+ int var1ndigits = var1->ndigits;
+ int var2ndigits = var2->ndigits;
+ NumericDigit *var1digits = var1->digits;
+ NumericDigit *var2digits = var2->digits;
+ int res_sign;
+ int res_weight;
+ int res_ndigits;
+ NumericDigit *res_buf;
+ NumericDigit *res_digits;
+ uint32 carry;
+ uint32 term;
+
+ /* Check preconditions */
+ Assert(var1ndigits >= 1);
+ Assert(var1ndigits <= 4);
+ Assert(var2ndigits >= var1ndigits);
+
+ /*
+ * Determine the result sign, weight, and number of digits to calculate.
+ * The weight figured here is correct if the product has no leading zero
+ * digits; otherwise strip_var() will fix things up. Note that, unlike
+ * mul_var(), we do not need to allocate an extra output digit, because we
+ * are not rounding here.
+ */
+ if (var1->sign == var2->sign)
+ res_sign = NUMERIC_POS;
+ else
+ res_sign = NUMERIC_NEG;
+ res_weight = var1->weight + var2->weight + 1;
+ res_ndigits = var1ndigits + var2ndigits;
+
+ /* Allocate result digit array */
+ res_buf = digitbuf_alloc(res_ndigits + 1);
+ res_buf[0] = 0; /* spare digit for later rounding */
+ res_digits = res_buf + 1;
+
+ /*
+ * Compute the result digits in reverse, in one pass, propagating the
+ * carry up as we go. The i'th result digit consists of the sum of the
+ * products var1digits[i1] * var2digits[i2] for which i = i1 + i2 + 1.
+ */
+ switch (var1ndigits)
+ {
+ case 1:
+ /* ---------
+ * 1-digit case:
+ * var1ndigits = 1
+ * var2ndigits >= 1
+ * res_ndigits = var2ndigits + 1
+ * ----------
+ */
+ carry = 0;
+ for (int i = res_ndigits - 2; i >= 0; i--)
+ {
+ term = (uint32) var1digits[0] * var2digits[i] + carry;
+ res_digits[i + 1] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+ }
+ res_digits[0] = (NumericDigit) carry;
+ break;
+
+ case 2:
+ /* ---------
+ * 2-digit case:
+ * var1ndigits = 2
+ * var2ndigits >= 2
+ * res_ndigits = var2ndigits + 2
+ * ----------
+ */
+ /* last result digit and carry */
+ term = (uint32) var1digits[1] * var2digits[res_ndigits - 3];
+ res_digits[res_ndigits - 1] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+
+ /* remaining digits, except for the first two */
+ for (int i = res_ndigits - 3; i >= 1; i--)
+ {
+ term = (uint32) var1digits[0] * var2digits[i] +
+ (uint32) var1digits[1] * var2digits[i - 1] + carry;
+ res_digits[i + 1] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+ }
+
+ /* first two digits */
+ term = (uint32) var1digits[0] * var2digits[0] + carry;
+ res_digits[1] = (NumericDigit) (term % NBASE);
+ res_digits[0] = (NumericDigit) (term / NBASE);
+ break;
+
+ case 3:
+ /* ---------
+ * 3-digit case:
+ * var1ndigits = 3
+ * var2ndigits >= 3
+ * res_ndigits = var2ndigits + 3
+ * ----------
+ */
+ /* last two result digits */
+ term = (uint32) var1digits[2] * var2digits[res_ndigits - 4];
+ res_digits[res_ndigits - 1] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+
+ term = (uint32) var1digits[1] * var2digits[res_ndigits - 4] +
+ (uint32) var1digits[2] * var2digits[res_ndigits - 5] + carry;
+ res_digits[res_ndigits - 2] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+
+ /* remaining digits, except for the first three */
+ for (int i = res_ndigits - 4; i >= 2; i--)
+ {
+ term = (uint32) var1digits[0] * var2digits[i] +
+ (uint32) var1digits[1] * var2digits[i - 1] +
+ (uint32) var1digits[2] * var2digits[i - 2] + carry;
+ res_digits[i + 1] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+ }
+
+ /* first three digits */
+ term = (uint32) var1digits[0] * var2digits[1] +
+ (uint32) var1digits[1] * var2digits[0] + carry;
+ res_digits[2] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+
+ term = (uint32) var1digits[0] * var2digits[0] + carry;
+ res_digits[1] = (NumericDigit) (term % NBASE);
+ res_digits[0] = (NumericDigit) (term / NBASE);
+ break;
+
+ case 4:
+ /* ---------
+ * 4-digit case:
+ * var1ndigits = 4
+ * var2ndigits >= 4
+ * res_ndigits = var2ndigits + 4
+ * ----------
+ */
+ /* last three result digits */
+ term = (uint32) var1digits[3] * var2digits[res_ndigits - 5];
+ res_digits[res_ndigits - 1] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+
+ term = (uint32) var1digits[2] * var2digits[res_ndigits - 5] +
+ (uint32) var1digits[3] * var2digits[res_ndigits - 6] + carry;
+ res_digits[res_ndigits - 2] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+
+ term = (uint32) var1digits[1] * var2digits[res_ndigits - 5] +
+ (uint32) var1digits[2] * var2digits[res_ndigits - 6] +
+ (uint32) var1digits[3] * var2digits[res_ndigits - 7] + carry;
+ res_digits[res_ndigits - 3] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+
+ /* remaining digits, except for the first four */
+ for (int i = res_ndigits - 5; i >= 3; i--)
+ {
+ term = (uint32) var1digits[0] * var2digits[i] +
+ (uint32) var1digits[1] * var2digits[i - 1] +
+ (uint32) var1digits[2] * var2digits[i - 2] +
+ (uint32) var1digits[3] * var2digits[i - 3] + carry;
+ res_digits[i + 1] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+ }
+
+ /* first four digits */
+ term = (uint32) var1digits[0] * var2digits[2] +
+ (uint32) var1digits[1] * var2digits[1] +
+ (uint32) var1digits[2] * var2digits[0] + carry;
+ res_digits[3] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+
+ term = (uint32) var1digits[0] * var2digits[1] +
+ (uint32) var1digits[1] * var2digits[0] + carry;
+ res_digits[2] = (NumericDigit) (term % NBASE);
+ carry = term / NBASE;
+
+ term = (uint32) var1digits[0] * var2digits[0] + carry;
+ res_digits[1] = (NumericDigit) (term % NBASE);
+ res_digits[0] = (NumericDigit) (term / NBASE);
+ break;
+ }
+
+ /* Store the product in result */
+ digitbuf_free(result->buf);
+ result->ndigits = res_ndigits;
+ result->buf = res_buf;
+ result->digits = res_digits;
+ result->weight = res_weight;
+ result->sign = res_sign;
+ result->dscale = var1->dscale + var2->dscale;
+
+ /* Strip leading and trailing zeroes */
+ strip_var(result);
}
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: Optimize numeric multiplication for one and two base-NBASE digit multiplicands.
@ 2024-07-06 11:16 Joel Jacobson <[email protected]>
parent: Dean Rasheed <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Joel Jacobson @ 2024-07-06 11:16 UTC (permalink / raw)
To: Dean Rasheed <[email protected]>; +Cc: Dagfinn Ilmari Mannsåker <[email protected]>; pgsql-hackers
On Sat, Jul 6, 2024, at 11:34, Dean Rasheed wrote:
> On Fri, 5 Jul 2024 at 18:37, Joel Jacobson <[email protected]> wrote:
>>
>> On Fri, Jul 5, 2024, at 18:42, Joel Jacobson wrote:
>> > Very nice, v7-optimize-numeric-mul_var-small-var1-arbitrary-var2.patch
>> > is now the winner on all my CPUs:
>>
>> I thought it would be interesting to also measure the isolated effect
>> on just numeric_mul() without the query overhead.
>>
>> Impressive speed-up, between 25% - 81%.
>>
>
> Cool. I think we should go with the mul_var_small() patch then, since
> it's more generally applicable.
I agree.
> I also did some testing with much larger var2 values, and saw similar
> speed-ups. One high-level function that benefits from that is
> factorial(), which accepts inputs up to 32177, and so uses both the
> 1-digit and 2-digit code with very large var2 values. I doubt anyone
> actually uses it with such large inputs, but it's interesting
> nonetheless:
>
> SELECT factorial(32177);
> Time: 923.117 ms -- HEAD
> Time: 534.375 ms -- mul_var_small() patch
Nice!
> I did one more round of (mostly cosmetic) copy-editing. Aside from
> improving some of the comments, it occurred to me that there's no need
> to pass rscale to mul_var_small(), or for it to call round_var(),
> since it's always computing the exact result. That shaves off a few
> more cycles.
Nice, also cleaner.
> Additionally, I didn't like how res_weight and res_ndigits were being
> set 1 higher than they needed to be. That makes sense in mul_var()
> because it may round the result, causing a non-zero carry to propagate
> into the next digit up, but it's just confusing in mul_var_small(). So
> I've reduced those by 1, which makes the look much more logical. To be
> clear, this doesn't change how many digits we're calculating. But now
> res_ndigits is actually the number of digits being calculated, whereas
> before, res_ndigits was 1 larger and we were calculating res_ndigits -
> 1 digits, which was confusing.
Nice, much cleaner.
> I think this is good to go, so unless there are any further comments,
> I plan to commit it soon.
LGTM.
Benchmark, only on Apple M3 Max:
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_1; -- HEAD
Time: 3042.157 ms (00:03.042)
Time: 3027.711 ms (00:03.028)
Time: 3078.215 ms (00:03.078)
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_1; -- v8
Time: 2700.676 ms (00:02.701)
Time: 2713.594 ms (00:02.714)
Time: 2704.139 ms (00:02.704)
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_2; -- HEAD
Time: 4506.064 ms (00:04.506)
Time: 3316.204 ms (00:03.316)
Time: 3321.086 ms (00:03.321)
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_2; -- v8
Time: 2904.786 ms (00:02.905)
Time: 2921.996 ms (00:02.922)
Time: 2919.269 ms (00:02.919)
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_3; -- HEAD
Time: 4636.051 ms (00:04.636)
Time: 3439.951 ms (00:03.440)
Time: 3471.245 ms (00:03.471)
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_3; -- v8
Time: 3034.364 ms (00:03.034)
Time: 3025.351 ms (00:03.025)
Time: 3075.024 ms (00:03.075)
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_4; -- HEAD
Time: 4978.086 ms (00:04.978)
Time: 3580.283 ms (00:03.580)
Time: 3582.719 ms (00:03.583)
SELECT SUM(var1*var2) FROM bench_mul_var_var1ndigits_4; -- v8
Time: 3147.352 ms (00:03.147)
Time: 3135.903 ms (00:03.136)
Time: 3172.491 ms (00:03.172)
Regards,
Joel
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: Optimize numeric multiplication for one and two base-NBASE digit multiplicands.
@ 2024-07-09 09:11 Dean Rasheed <[email protected]>
parent: Joel Jacobson <[email protected]>
0 siblings, 1 reply; 103+ messages in thread
From: Dean Rasheed @ 2024-07-09 09:11 UTC (permalink / raw)
To: Joel Jacobson <[email protected]>; +Cc: Dagfinn Ilmari Mannsåker <[email protected]>; pgsql-hackers
On Sat, 6 Jul 2024 at 12:17, Joel Jacobson <[email protected]> wrote:
>
> > I think this is good to go, so unless there are any further comments,
> > I plan to commit it soon.
>
> LGTM.
>
OK, I have committed this.
At the last minute, I changed the name of the new function to
mul_var_short() because "short" is probably a better term to use in
this context (we already use it in a preceding comment). "Small" is
potentially misleading, because the numbers themselves could be
numerically very large.
Regards,
Dean
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: Optimize numeric multiplication for one and two base-NBASE digit multiplicands.
@ 2024-07-09 12:01 Dean Rasheed <[email protected]>
parent: Dean Rasheed <[email protected]>
0 siblings, 2 replies; 103+ messages in thread
From: Dean Rasheed @ 2024-07-09 12:01 UTC (permalink / raw)
To: Joel Jacobson <[email protected]>; +Cc: Dagfinn Ilmari Mannsåker <[email protected]>; pgsql-hackers
On Tue, 9 Jul 2024 at 10:11, Dean Rasheed <[email protected]> wrote:
>
> OK, I have committed this.
>
Before considering the other patches to optimise for larger inputs, I
think it's worth optimising the existing mul_var() code as much as
possible.
One thing I noticed while testing the earlier patches on this thread
was that they were significantly faster if they used unsigned integers
rather than signed integers. I think the reason is that operations
like "x / 10000" and "x % 10000" use fewer CPU instructions (on every
platform, according to godbolt.org) if x is unsigned.
In addition, this reduces the number of times the digit array needs to
be renormalised, which seems to be the biggest factor.
Another small optimisation that seems to be just about worthwhile is
to pull the first digit of var1 out of the main loop, so that its
contributions can be set directly in dig[], rather than being added to
it. This allows palloc() to be used to allocate dig[], rather than
palloc0(), and only requires the upper part of dig[] to be initialised
to zeros, rather than all of it.
Together, these seem to give a decent speed-up:
NBASE digits | HEAD rate | patch rate
--------------+---------------+---------------
5 | 5.8797105e+06 | 6.047134e+06
12 | 4.140017e+06 | 4.3429845e+06
25 | 2.5711072e+06 | 2.7530615e+06
50 | 1.0367389e+06 | 1.3370771e+06
100 | 367924.1 | 462437.38
250 | 77231.32 | 104593.95
2500 | 881.48694 | 1197.4739
15000 | 25.064987 | 32.78391
The largest gains are above around 50 NBASE digits, where the time
spent renormalising dig[] becomes significant.
Regards,
Dean
Attachments:
[text/x-patch] optimise-mul_var.patch (4.2K, ../../CAEZATCXoemvhECHiL8Ug1MQxxtU0WqZfYqE853fDr_PvUpFerA@mail.gmail.com/2-optimise-mul_var.patch)
download | inline diff:
diff --git a/src/backend/utils/adt/numeric.c b/src/backend/utils/adt/numeric.c
new file mode 100644
index f6e20cf..2a2e22d
--- a/src/backend/utils/adt/numeric.c
+++ b/src/backend/utils/adt/numeric.c
@@ -8690,18 +8690,21 @@ mul_var(const NumericVar *var1, const Nu
int res_sign;
int res_weight;
int maxdigits;
- int *dig;
- int carry;
+ uint32 *dig,
+ *dig_i1_2;
+ uint32 carry;
int maxdig;
- int newdig;
+ uint32 newdig;
int var1ndigits;
int var2ndigits;
NumericDigit *var1digits;
NumericDigit *var2digits;
NumericDigit *res_digits;
+ NumericDigit var1digit;
int i,
i1,
- i2;
+ i2,
+ i2limit;
/*
* Arrange for var1 to be the shorter of the two numbers. This improves
@@ -8775,23 +8778,22 @@ mul_var(const NumericVar *var1, const Nu
}
/*
- * We do the arithmetic in an array "dig[]" of signed int's. Since
- * INT_MAX is noticeably larger than NBASE*NBASE, this gives us headroom
- * to avoid normalizing carries immediately.
+ * We do the arithmetic in an array "dig[]" of unsigned 32-bit integers.
+ * Since PG_UINT32_MAX is noticeably larger than NBASE*NBASE, this gives
+ * us headroom to avoid normalizing carries immediately.
*
* maxdig tracks the maximum possible value of any dig[] entry; when this
- * threatens to exceed INT_MAX, we take the time to propagate carries.
- * Furthermore, we need to ensure that overflow doesn't occur during the
- * carry propagation passes either. The carry values could be as much as
- * INT_MAX/NBASE, so really we must normalize when digits threaten to
- * exceed INT_MAX - INT_MAX/NBASE.
+ * threatens to exceed PG_UINT32_MAX, we take the time to propagate
+ * carries. Furthermore, we need to ensure that overflow doesn't occur
+ * during the carry propagation passes either. The carry values could be
+ * as much as PG_UINT32_MAX/NBASE, so really we must normalize when digits
+ * threaten to exceed PG_UINT32_MAX - PG_UINT32_MAX/NBASE.
*
* To avoid overflow in maxdig itself, it actually represents the max
* possible value divided by NBASE-1, ie, at the top of the loop it is
* known that no dig[] entry exceeds maxdig * (NBASE-1).
*/
- dig = (int *) palloc0(res_ndigits * sizeof(int));
- maxdig = 0;
+ dig = (uint32 *) palloc(res_ndigits * sizeof(uint32));
/*
* The least significant digits of var1 should be ignored if they don't
@@ -8801,17 +8803,37 @@ mul_var(const NumericVar *var1, const Nu
* Digit i1 of var1 and digit i2 of var2 are multiplied and added to digit
* i1+i2+2 of the accumulator array, so we need only consider digits of
* var1 for which i1 <= res_ndigits - 3.
+ *
+ * We process the least significant digit of var1 outside the main loop,
+ * since the results can be applied directly to dig[], rather than being
+ * added to it.
*/
- for (i1 = Min(var1ndigits - 1, res_ndigits - 3); i1 >= 0; i1--)
+ i1 = Min(var1ndigits - 1, res_ndigits - 3);
+ var1digit = var1digits[i1];
+ maxdig = var1digit;
+
+ /*
+ * Multiply by the least significant applicable digit of var1 -- see
+ * comments below. Any changes made there should be reflected here.
+ */
+ i2limit = Min(var2ndigits, res_ndigits - i1 - 2);
+ dig_i1_2 = &dig[i1 + 2];
+
+ memset(dig, 0, sizeof(uint32) * (i1 + 2));
+ for (i2 = 0; i2 < i2limit; i2++)
+ dig_i1_2[i2] = var1digit * var2digits[i2];
+
+ /* Add contributions from remaining digits of var1 */
+ for (i1 = i1 - 1; i1 >= 0; i1--)
{
- NumericDigit var1digit = var1digits[i1];
+ var1digit = var1digits[i1];
if (var1digit == 0)
continue;
/* Time to normalize? */
maxdig += var1digit;
- if (maxdig > (INT_MAX - INT_MAX / NBASE) / (NBASE - 1))
+ if (maxdig > (PG_UINT32_MAX - PG_UINT32_MAX / NBASE) / (NBASE - 1))
{
/* Yes, do it */
carry = 0;
@@ -8845,13 +8867,11 @@ mul_var(const NumericVar *var1, const Nu
* Since we aren't propagating carries in this loop, the order does
* not matter.
*/
- {
- int i2limit = Min(var2ndigits, res_ndigits - i1 - 2);
- int *dig_i1_2 = &dig[i1 + 2];
+ i2limit = Min(var2ndigits, res_ndigits - i1 - 2);
+ dig_i1_2 = &dig[i1 + 2];
- for (i2 = 0; i2 < i2limit; i2++)
- dig_i1_2[i2] += var1digit * var2digits[i2];
- }
+ for (i2 = 0; i2 < i2limit; i2++)
+ dig_i1_2[i2] += var1digit * var2digits[i2];
}
/*
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: Optimize numeric multiplication for one and two base-NBASE digit multiplicands.
@ 2024-07-09 14:11 Joel Jacobson <[email protected]>
parent: Dean Rasheed <[email protected]>
1 sibling, 1 reply; 103+ messages in thread
From: Joel Jacobson @ 2024-07-09 14:11 UTC (permalink / raw)
To: Dean Rasheed <[email protected]>; +Cc: Dagfinn Ilmari Mannsåker <[email protected]>; pgsql-hackers
On Tue, Jul 9, 2024, at 14:01, Dean Rasheed wrote:
> Before considering the other patches to optimise for larger inputs, I
> think it's worth optimising the existing mul_var() code as much as
> possible.
>
> One thing I noticed while testing the earlier patches on this thread
> was that they were significantly faster if they used unsigned integers
> rather than signed integers. I think the reason is that operations
> like "x / 10000" and "x % 10000" use fewer CPU instructions (on every
> platform, according to godbolt.org) if x is unsigned.
>
> In addition, this reduces the number of times the digit array needs to
> be renormalised, which seems to be the biggest factor.
>
> Another small optimisation that seems to be just about worthwhile is
> to pull the first digit of var1 out of the main loop, so that its
> contributions can be set directly in dig[], rather than being added to
> it. This allows palloc() to be used to allocate dig[], rather than
> palloc0(), and only requires the upper part of dig[] to be initialised
> to zeros, rather than all of it.
Nice, really smart!
> Together, these seem to give a decent speed-up:
>
> NBASE digits | HEAD rate | patch rate
> --------------+---------------+---------------
> 5 | 5.8797105e+06 | 6.047134e+06
> 12 | 4.140017e+06 | 4.3429845e+06
> 25 | 2.5711072e+06 | 2.7530615e+06
> 50 | 1.0367389e+06 | 1.3370771e+06
> 100 | 367924.1 | 462437.38
> 250 | 77231.32 | 104593.95
> 2500 | 881.48694 | 1197.4739
> 15000 | 25.064987 | 32.78391
>
> The largest gains are above around 50 NBASE digits, where the time
> spent renormalising dig[] becomes significant.
I added some more ndigits test cases:
/*
* Intel Core i9-14900K
*/
NBASE digits | HEAD rate | patch rate | relative difference
--------------+----------------+----------------+---------------------
1 | 4.7846890e+07 | 4.7846890e+07 | 0.00%
2 | 4.9504950e+07 | 4.7393365e+07 | -4.27%
3 | 4.0816327e+07 | 4.0983607e+07 | 0.41%
4 | 4.1152263e+07 | 3.9370079e+07 | -4.33%
5 | 2.2573363e+07 | 2.1978022e+07 | -2.64%
6 | 2.1739130e+07 | 1.9646365e+07 | -9.63%
7 | 1.6393443e+07 | 1.6339869e+07 | -0.33%
8 | 1.6863406e+07 | 1.6778523e+07 | -0.50%
9 | 1.5105740e+07 | 1.6420361e+07 | 8.70%
10 | 1.3315579e+07 | 1.5527950e+07 | 16.61%
11 | 1.2360939e+07 | 1.4124294e+07 | 14.27%
12 | 1.1764706e+07 | 1.2836970e+07 | 9.11%
13 | 1.0060362e+07 | 1.1820331e+07 | 17.49%
14 | 9.0909091e+06 | 1.0000000e+07 | 10.00%
15 | 7.6923077e+06 | 8.0000000e+06 | 4.00%
16 | 9.0909091e+06 | 9.4339623e+06 | 3.77%
17 | 7.2992701e+06 | 9.0909091e+06 | 24.55%
18 | 7.0921986e+06 | 7.8125000e+06 | 10.16%
19 | 6.5789474e+06 | 6.6666667e+06 | 1.33%
20 | 6.2500000e+06 | 6.5789474e+06 | 5.26%
21 | 5.8479532e+06 | 6.1728395e+06 | 5.56%
22 | 5.5555556e+06 | 5.9880240e+06 | 7.78%
24 | 5.2631579e+06 | 5.8823529e+06 | 11.76%
25 | 5.2083333e+06 | 5.5555556e+06 | 6.67%
26 | 4.7619048e+06 | 5.2631579e+06 | 10.53%
27 | 4.5045045e+06 | 5.2083333e+06 | 15.63%
28 | 4.4247788e+06 | 4.7619048e+06 | 7.62%
29 | 4.1666667e+06 | 4.5454545e+06 | 9.09%
30 | 4.0000000e+06 | 4.3478261e+06 | 8.70%
31 | 3.4482759e+06 | 4.0000000e+06 | 16.00%
32 | 3.9840637e+06 | 4.2016807e+06 | 5.46%
50 | 2.0964361e+06 | 2.6595745e+06 | 26.86%
100 | 666666.67 | 869565.22 | 30.43%
250 | 142653.35 | 171526.59 | 20.24%
2500 | 1642.04 | 2197.80 | 33.85%
15000 | 41.67 | 52.63 | 26.32%
(36 rows)
/*
* AMD Ryzen 9 7950X3D
*/
NBASE digits | HEAD rate | patch rate | relative difference
--------------+----------------+----------------+---------------------
1 | 3.6900369e+07 | 3.8022814e+07 | 3.04%
2 | 3.4602076e+07 | 3.5714286e+07 | 3.21%
3 | 2.8011204e+07 | 2.7777778e+07 | -0.83%
4 | 2.7932961e+07 | 2.8328612e+07 | 1.42%
5 | 1.6420361e+07 | 1.7123288e+07 | 4.28%
6 | 1.4705882e+07 | 1.5313936e+07 | 4.13%
7 | 1.3192612e+07 | 1.3888889e+07 | 5.28%
8 | 1.2121212e+07 | 1.2919897e+07 | 6.59%
9 | 1.1235955e+07 | 1.2135922e+07 | 8.01%
10 | 1.0000000e+07 | 1.1312217e+07 | 13.12%
11 | 9.0909091e+06 | 1.0000000e+07 | 10.00%
12 | 8.1967213e+06 | 8.4033613e+06 | 2.52%
13 | 7.2463768e+06 | 7.7519380e+06 | 6.98%
14 | 6.7567568e+06 | 7.1428571e+06 | 5.71%
15 | 5.5555556e+06 | 5.8823529e+06 | 5.88%
16 | 6.3291139e+06 | 5.7803468e+06 | -8.67%
17 | 5.8823529e+06 | 5.9880240e+06 | 1.80%
18 | 5.5555556e+06 | 5.7142857e+06 | 2.86%
19 | 5.2356021e+06 | 5.6179775e+06 | 7.30%
20 | 4.9019608e+06 | 5.1020408e+06 | 4.08%
21 | 4.5454545e+06 | 4.8543689e+06 | 6.80%
22 | 4.1841004e+06 | 4.5871560e+06 | 9.63%
24 | 4.4642857e+06 | 4.4052863e+06 | -1.32%
25 | 4.1666667e+06 | 4.2194093e+06 | 1.27%
26 | 4.0000000e+06 | 3.9525692e+06 | -1.19%
27 | 3.8461538e+06 | 3.8022814e+06 | -1.14%
28 | 3.9062500e+06 | 3.8759690e+06 | -0.78%
29 | 3.7878788e+06 | 3.8022814e+06 | 0.38%
30 | 3.3898305e+06 | 3.7174721e+06 | 9.67%
31 | 2.7472527e+06 | 2.8571429e+06 | 4.00%
32 | 3.0395137e+06 | 3.1446541e+06 | 3.46%
50 | 1.7094017e+06 | 2.0576132e+06 | 20.37%
100 | 518134.72 | 609756.10 | 17.68%
250 | 108577.63 | 136612.02 | 25.82%
2500 | 1264.22 | 1610.31 | 27.38%
15000 | 34.48 | 43.48 | 26.09%
(36 rows)
/*
* Apple M3 Max
*/
NBASE digits | HEAD rate | patch rate | relative difference
--------------+----------------+----------------+---------------------
1 | 4.9504950e+07 | 4.9504950e+07 | 0.00%
2 | 6.1349693e+07 | 5.9171598e+07 | -3.55%
3 | 5.2631579e+07 | 5.2083333e+07 | -1.04%
4 | 4.5248869e+07 | 4.5248869e+07 | 0.00%
5 | 2.1978022e+07 | 2.2727273e+07 | 3.41%
6 | 1.9920319e+07 | 2.0876827e+07 | 4.80%
7 | 1.7182131e+07 | 1.8018018e+07 | 4.86%
8 | 1.5822785e+07 | 1.6051364e+07 | 1.44%
9 | 1.3368984e+07 | 1.3333333e+07 | -0.27%
10 | 1.1709602e+07 | 1.1627907e+07 | -0.70%
11 | 1.0020040e+07 | 1.0526316e+07 | 5.05%
12 | 9.0909091e+06 | 9.0909091e+06 | 0.00%
13 | 8.2644628e+06 | 8.2644628e+06 | 0.00%
14 | 7.6923077e+06 | 7.6335878e+06 | -0.76%
15 | 7.1428571e+06 | 7.0921986e+06 | -0.71%
16 | 6.6225166e+06 | 6.5789474e+06 | -0.66%
17 | 5.9880240e+06 | 6.2111801e+06 | 3.73%
18 | 5.7803468e+06 | 5.5865922e+06 | -3.35%
19 | 5.2631579e+06 | 5.2356021e+06 | -0.52%
20 | 4.6296296e+06 | 4.8543689e+06 | 4.85%
21 | 4.4444444e+06 | 4.3859649e+06 | -1.32%
22 | 4.2016807e+06 | 4.0485830e+06 | -3.64%
24 | 3.7453184e+06 | 3.5714286e+06 | -4.64%
25 | 3.4843206e+06 | 3.4013605e+06 | -2.38%
26 | 3.2786885e+06 | 3.2786885e+06 | 0.00%
27 | 3.0674847e+06 | 3.1055901e+06 | 1.24%
28 | 2.8818444e+06 | 2.9069767e+06 | 0.87%
29 | 2.7322404e+06 | 2.7700831e+06 | 1.39%
30 | 2.5839793e+06 | 2.6246719e+06 | 1.57%
31 | 2.5062657e+06 | 2.4630542e+06 | -1.72%
32 | 4.5871560e+06 | 4.6082949e+06 | 0.46%
50 | 1.7513135e+06 | 1.9880716e+06 | 13.52%
100 | 714285.71 | 833333.33 | 16.67%
250 | 124223.60 | 149925.04 | 20.69%
2500 | 1440.92 | 1760.56 | 22.18%
15000 | 39.53 | 48.08 | 21.63%
(36 rows)
Regards,
Joel
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: Optimize numeric multiplication for one and two base-NBASE digit multiplicands.
@ 2024-07-09 19:48 Joel Jacobson <[email protected]>
parent: Joel Jacobson <[email protected]>
0 siblings, 0 replies; 103+ messages in thread
From: Joel Jacobson @ 2024-07-09 19:48 UTC (permalink / raw)
To: Dean Rasheed <[email protected]>; +Cc: Dagfinn Ilmari Mannsåker <[email protected]>; pgsql-hackers
On Tue, Jul 9, 2024, at 16:11, Joel Jacobson wrote:
> I added some more ndigits test cases:
Ops, please ignore previous benchmark;
I had forgot to commit in between the measurements,
so they all ran in the same db txn,
which caused a lot of noise on few ndigits.
New benchmark:
> /*
> * Intel Core i9-14900K
> */
NBASE digits | HEAD rate | patch rate | relative difference
--------------+----------------+----------------+---------------------
1 | 5.0251256e+07 | 5.2631579e+07 | 4.74%
2 | 4.8543689e+07 | 4.9751244e+07 | 2.49%
3 | 4.1493776e+07 | 4.3478261e+07 | 4.78%
4 | 4.1493776e+07 | 4.0816327e+07 | -1.63%
5 | 2.2371365e+07 | 2.3364486e+07 | 4.44%
6 | 2.1008403e+07 | 2.1186441e+07 | 0.85%
7 | 1.7152659e+07 | 1.6233766e+07 | -5.36%
8 | 1.7123288e+07 | 1.8450185e+07 | 7.75%
9 | 1.5290520e+07 | 1.7271157e+07 | 12.95%
10 | 1.3351135e+07 | 1.5384615e+07 | 15.23%
11 | 1.2453300e+07 | 1.4164306e+07 | 13.74%
12 | 1.1655012e+07 | 1.2936611e+07 | 11.00%
13 | 1.0373444e+07 | 1.1904762e+07 | 14.76%
14 | 9.0909091e+06 | 1.0162602e+07 | 11.79%
15 | 7.7519380e+06 | 8.1300813e+06 | 4.88%
16 | 9.0909091e+06 | 9.8039216e+06 | 7.84%
17 | 7.5757576e+06 | 9.0909091e+06 | 20.00%
18 | 7.2463768e+06 | 8.2644628e+06 | 14.05%
19 | 6.6225166e+06 | 7.5757576e+06 | 14.39%
20 | 6.4516129e+06 | 7.0422535e+06 | 9.15%
21 | 6.0606061e+06 | 6.5789474e+06 | 8.55%
22 | 5.7142857e+06 | 6.2500000e+06 | 9.38%
24 | 5.4054054e+06 | 6.0240964e+06 | 11.45%
25 | 5.2356021e+06 | 5.8139535e+06 | 11.05%
26 | 5.0251256e+06 | 5.8139535e+06 | 15.70%
27 | 4.7393365e+06 | 5.7142857e+06 | 20.57%
28 | 4.6082949e+06 | 5.2083333e+06 | 13.02%
29 | 4.3478261e+06 | 4.9504950e+06 | 13.86%
30 | 4.0816327e+06 | 4.6728972e+06 | 14.49%
31 | 3.4843206e+06 | 3.9682540e+06 | 13.89%
32 | 4.0000000e+06 | 4.1666667e+06 | 4.17%
50 | 2.1097046e+06 | 2.8571429e+06 | 35.43%
100 | 680272.11 | 909090.91 | 33.64%
250 | 141643.06 | 174216.03 | 23.00%
2500 | 1626.02 | 2188.18 | 34.57%
15000 | 41.67 | 52.63 | 26.32%
(36 rows)
> /*
> * AMD Ryzen 9 7950X3D
> */
NBASE digits | HEAD rate | patch rate | relative difference
--------------+----------------+----------------+---------------------
1 | 3.7037037e+07 | 3.8910506e+07 | 5.06%
2 | 3.5587189e+07 | 3.5971223e+07 | 1.08%
3 | 3.0581040e+07 | 2.9239766e+07 | -4.39%
4 | 2.7322404e+07 | 3.0303030e+07 | 10.91%
5 | 1.8050542e+07 | 1.9011407e+07 | 5.32%
6 | 1.5974441e+07 | 1.6233766e+07 | 1.62%
7 | 1.3106160e+07 | 1.3071895e+07 | -0.26%
8 | 1.2285012e+07 | 1.3106160e+07 | 6.68%
9 | 1.1534025e+07 | 1.2269939e+07 | 6.38%
10 | 1.1135857e+07 | 1.1507480e+07 | 3.34%
11 | 9.7943193e+06 | 1.0976948e+07 | 12.07%
12 | 9.5238095e+06 | 1.0256410e+07 | 7.69%
13 | 8.6206897e+06 | 8.7719298e+06 | 1.75%
14 | 7.3529412e+06 | 8.1967213e+06 | 11.48%
15 | 6.2893082e+06 | 6.7114094e+06 | 6.71%
16 | 7.2463768e+06 | 7.0422535e+06 | -2.82%
17 | 6.2893082e+06 | 7.2463768e+06 | 15.22%
18 | 6.3694268e+06 | 7.4626866e+06 | 17.16%
19 | 5.6818182e+06 | 6.6225166e+06 | 16.56%
20 | 5.2083333e+06 | 6.1728395e+06 | 18.52%
21 | 5.0251256e+06 | 5.7471264e+06 | 14.37%
22 | 4.5248869e+06 | 5.1282051e+06 | 13.33%
24 | 4.9261084e+06 | 5.1020408e+06 | 3.57%
25 | 4.6511628e+06 | 4.9504950e+06 | 6.44%
26 | 4.2553191e+06 | 4.6082949e+06 | 8.29%
27 | 3.9682540e+06 | 4.2918455e+06 | 8.15%
28 | 3.8910506e+06 | 4.1322314e+06 | 6.20%
29 | 3.8167939e+06 | 3.7593985e+06 | -1.50%
30 | 3.5842294e+06 | 3.6101083e+06 | 0.72%
31 | 3.1948882e+06 | 3.1645570e+06 | -0.95%
32 | 3.4722222e+06 | 3.7174721e+06 | 7.06%
50 | 1.6474465e+06 | 2.1691974e+06 | 31.67%
100 | 555555.56 | 653594.77 | 17.65%
250 | 109409.19 | 140449.44 | 28.37%
2500 | 1236.09 | 1555.21 | 25.82%
15000 | 34.48 | 43.48 | 26.09%
(36 rows)
> /*
> * Apple M3 Max
> */
NBASE digits | HEAD rate | patch rate | relative difference
--------------+----------------+----------------+---------------------
1 | 4.7169811e+07 | 4.7619048e+07 | 0.95%
2 | 6.0240964e+07 | 5.8479532e+07 | -2.92%
3 | 5.2083333e+07 | 5.3191489e+07 | 2.13%
4 | 4.5871560e+07 | 4.6948357e+07 | 2.35%
5 | 2.2075055e+07 | 2.3529412e+07 | 6.59%
6 | 2.0080321e+07 | 2.1505376e+07 | 7.10%
7 | 1.7301038e+07 | 1.8975332e+07 | 9.68%
8 | 1.6025641e+07 | 1.6556291e+07 | 3.31%
9 | 1.3245033e+07 | 1.3717421e+07 | 3.57%
10 | 1.1709602e+07 | 1.2315271e+07 | 5.17%
11 | 1.0000000e+07 | 1.0989011e+07 | 9.89%
12 | 9.0909091e+06 | 9.7276265e+06 | 7.00%
13 | 8.3333333e+06 | 9.0090090e+06 | 8.11%
14 | 7.6923077e+06 | 8.0645161e+06 | 4.84%
15 | 7.0921986e+06 | 7.5187970e+06 | 6.02%
16 | 6.6666667e+06 | 7.0921986e+06 | 6.38%
17 | 6.2111801e+06 | 6.3694268e+06 | 2.55%
18 | 5.7803468e+06 | 5.9523810e+06 | 2.98%
19 | 5.2910053e+06 | 5.4347826e+06 | 2.72%
20 | 4.7846890e+06 | 5.0505051e+06 | 5.56%
21 | 4.5454545e+06 | 4.6728972e+06 | 2.80%
22 | 4.2372881e+06 | 4.3859649e+06 | 3.51%
24 | 3.7174721e+06 | 3.8759690e+06 | 4.26%
25 | 3.4722222e+06 | 3.6231884e+06 | 4.35%
26 | 3.2894737e+06 | 3.3898305e+06 | 3.05%
27 | 3.0674847e+06 | 3.1847134e+06 | 3.82%
28 | 2.9239766e+06 | 3.0120482e+06 | 3.01%
29 | 2.7548209e+06 | 2.8901734e+06 | 4.91%
30 | 2.6041667e+06 | 2.7322404e+06 | 4.92%
31 | 2.5000000e+06 | 2.5773196e+06 | 3.09%
32 | 4.6082949e+06 | 4.7846890e+06 | 3.83%
50 | 1.7241379e+06 | 2.0703934e+06 | 20.08%
100 | 719424.46 | 869565.22 | 20.87%
250 | 124688.28 | 157977.88 | 26.70%
2500 | 1455.60 | 1811.59 | 24.46%
15000 | 40.00 | 50.00 | 25.00%
(36 rows)
Regards,
Joel
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: Optimize numeric multiplication for one and two base-NBASE digit multiplicands.
@ 2024-07-09 20:28 Joel Jacobson <[email protected]>
parent: Dean Rasheed <[email protected]>
1 sibling, 0 replies; 103+ messages in thread
From: Joel Jacobson @ 2024-07-09 20:28 UTC (permalink / raw)
To: Dean Rasheed <[email protected]>; +Cc: Dagfinn Ilmari Mannsåker <[email protected]>; pgsql-hackers
On Tue, Jul 9, 2024, at 14:01, Dean Rasheed wrote:
> One thing I noticed while testing the earlier patches on this thread
> was that they were significantly faster if they used unsigned integers
> rather than signed integers. I think the reason is that operations
> like "x / 10000" and "x % 10000" use fewer CPU instructions (on every
> platform, according to godbolt.org) if x is unsigned.
>
> In addition, this reduces the number of times the digit array needs to
> be renormalised, which seems to be the biggest factor.
>
> Another small optimisation that seems to be just about worthwhile is
> to pull the first digit of var1 out of the main loop, so that its
> contributions can be set directly in dig[], rather than being added to
> it. This allows palloc() to be used to allocate dig[], rather than
> palloc0(), and only requires the upper part of dig[] to be initialised
> to zeros, rather than all of it.
>
> Together, these seem to give a decent speed-up:
..
> Attachments:
> * optimise-mul_var.patch
I've reviewed the patch now.
Code is straightforward, and comments easy to understand.
LGTM.
Regards,
Joel
^ permalink raw reply [nested|flat] 103+ messages in thread
* Re: MultiXact\SLRU buffers configuration
@ 2025-07-28 17:23 Andrey Borodin <[email protected]>
0 siblings, 0 replies; 103+ messages in thread
From: Andrey Borodin @ 2025-07-28 17:23 UTC (permalink / raw)
To: Thom Brown <[email protected]>; +Cc: Michael Paquier <[email protected]>; Alvaro Herrera <[email protected]>; vignesh C <[email protected]>; Andrew Borodin <[email protected]>; Yura Sokolov <[email protected]>; Andres Freund <[email protected]>; Thomas Munro <[email protected]>; Gilles Darold <[email protected]>; Alexander Korotkov <[email protected]>; Daniel Gustafsson <[email protected]>; Kyotaro Horiguchi <[email protected]>; pgsql-hackers
Hi Thom!
> On 1 Nov 2024, at 00:25, Thom Brown <[email protected]> wrote:
>
> Unfortunately I didn't gather much information when it was occuring, and prioritised getting rid of the process blocking replay. I just attached gdb to it, got a backtrace and then "print ProcessInterrupts()".
FWIW we diagnosed that problem in the recent thread "IPC/MultixactCreation on the Standby server" [0].
Proposed fix is non-trivial. I'm trying to asses how many reports we had about this problem. Did you ever observe this problem again?
Thanks!
Best regards, Andrey Borodin.
[0] https://www.postgresql.org/message-id/[email protected]
^ permalink raw reply [nested|flat] 103+ messages in thread
end of thread, other threads:[~2025-07-28 17:23 UTC | newest]
Thread overview: 103+ messages (download: mbox mbox.gz follow: Atom feed)
-- links below jump to the message on this page --
2019-03-08 00:31 [PATCH v18 08/18] tableam: finish_bulk_insert(). Andres Freund <[email protected]>
2020-05-08 16:36 MultiXact\SLRU buffers configuration Andrey M. Borodin <[email protected]>
2020-05-11 11:17 ` Re: MultiXact\SLRU buffers configuration Andrey M. Borodin <[email protected]>
2020-05-13 18:08 ` Re: MultiXact\SLRU buffers configuration Andrey M. Borodin <[email protected]>
2020-05-14 01:25 ` Re: MultiXact\SLRU buffers configuration Kyotaro Horiguchi <[email protected]>
2020-05-14 05:19 ` Re: MultiXact\SLRU buffers configuration Andrey M. Borodin <[email protected]>
2020-05-14 06:16 ` Re: MultiXact\SLRU buffers configuration Kyotaro Horiguchi <[email protected]>
2020-05-14 06:44 ` Re: MultiXact\SLRU buffers configuration Andrey M. Borodin <[email protected]>
2020-05-15 00:03 ` Re: MultiXact\SLRU buffers configuration Kyotaro Horiguchi <[email protected]>
2020-05-15 09:01 ` Re: MultiXact\SLRU buffers configuration Andrey M. Borodin <[email protected]>
2020-05-20 04:54 ` Re: MultiXact\SLRU buffers configuration Kyotaro Horiguchi <[email protected]>
2020-07-02 12:02 ` Re: MultiXact\SLRU buffers configuration Daniel Gustafsson <[email protected]>
2020-07-08 07:03 ` Re: MultiXact\SLRU buffers configuration Andrey M. Borodin <[email protected]>
2020-08-02 21:30 ` Re: MultiXact\SLRU buffers configuration Daniel Gustafsson <[email protected]>
2020-08-28 18:08 ` Re: MultiXact\SLRU buffers configuration Anastasia Lubennikova <[email protected]>
2020-09-28 14:41 ` Re: MultiXact\SLRU buffers configuration Andrey M. Borodin <[email protected]>
2020-10-07 14:23 ` Re: MultiXact\SLRU buffers configuration Anastasia Lubennikova <[email protected]>
2020-10-26 01:05 ` Re: MultiXact\SLRU buffers configuration Alexander Korotkov <[email protected]>
2020-10-26 15:45 ` Re: MultiXact\SLRU buffers configuration Andrey Borodin <[email protected]>
2020-10-27 17:02 ` Re: MultiXact\SLRU buffers configuration Alexander Korotkov <[email protected]>
2020-10-27 17:23 ` Re: MultiXact\SLRU buffers configuration Alexander Korotkov <[email protected]>
2020-10-28 01:36 ` Re: MultiXact\SLRU buffers configuration Tomas Vondra <[email protected]>
2020-10-28 07:34 ` Re: MultiXact\SLRU buffers configuration Andrey Borodin <[email protected]>
2020-10-28 23:32 ` Re: MultiXact\SLRU buffers configuration Tomas Vondra <[email protected]>
2020-10-29 07:08 ` Re: MultiXact\SLRU buffers configuration Andrey Borodin <[email protected]>
2020-10-29 13:49 ` Re: MultiXact\SLRU buffers configuration Tomas Vondra <[email protected]>
2020-11-02 12:45 ` Re: MultiXact\SLRU buffers configuration Andrey Borodin <[email protected]>
2020-11-10 00:13 ` Re: MultiXact\SLRU buffers configuration Tomas Vondra <[email protected]>
2020-11-10 06:16 ` Re: MultiXact\SLRU buffers configuration Andrey Borodin <[email protected]>
2020-11-10 18:07 ` Re: MultiXact\SLRU buffers configuration Tomas Vondra <[email protected]>
2020-11-10 18:41 ` Re: MultiXact\SLRU buffers configuration Thomas Munro <[email protected]>
2020-11-13 11:49 ` Re: MultiXact\SLRU buffers configuration Andrey Borodin <[email protected]>
2020-12-08 16:05 ` Re: MultiXact\SLRU buffers configuration Gilles Darold <[email protected]>
2020-12-08 17:52 ` Re: MultiXact\SLRU buffers configuration Andrey Borodin <[email protected]>
2020-12-09 10:51 ` Re: MultiXact\SLRU buffers configuration Gilles Darold <[email protected]>
2020-12-09 11:06 ` Re: MultiXact\SLRU buffers configuration Gilles Darold <[email protected]>
2020-12-10 14:45 ` Re: MultiXact\SLRU buffers configuration Gilles Darold <[email protected]>
2020-12-11 17:50 ` Re: MultiXact\SLRU buffers configuration Gilles Darold <[email protected]>
2020-12-13 09:17 ` Re: MultiXact\SLRU buffers configuration Gilles Darold <[email protected]>
2020-12-13 17:24 ` Re: MultiXact\SLRU buffers configuration Andrey Borodin <[email protected]>
2020-12-14 06:31 ` Re: MultiXact\SLRU buffers configuration Andrey Borodin <[email protected]>
2020-12-23 16:31 ` Re: MultiXact\SLRU buffers configuration Gilles Darold <[email protected]>
2021-02-15 17:17 ` Re: MultiXact\SLRU buffers configuration Andrey Borodin <[email protected]>
2021-03-11 15:50 ` Re: MultiXact\SLRU buffers configuration Gilles Darold <[email protected]>
2021-03-12 12:44 ` Re: MultiXact\SLRU buffers configuration Andrey Borodin <[email protected]>
2021-03-15 15:41 ` Re: MultiXact\SLRU buffers configuration Gilles Darold <[email protected]>
2021-03-24 21:31 ` Re: MultiXact\SLRU buffers configuration Thomas Munro <[email protected]>
2021-03-25 01:03 ` Re: MultiXact\SLRU buffers configuration Thomas Munro <[email protected]>
2021-03-26 03:46 ` Re: MultiXact\SLRU buffers configuration Thomas Munro <[email protected]>
2021-03-26 06:00 ` Re: MultiXact\SLRU buffers configuration Andrey Borodin <[email protected]>
2021-03-26 15:52 ` Re: MultiXact\SLRU buffers configuration Andrey Borodin <[email protected]>
2021-03-26 20:26 ` Re: MultiXact\SLRU buffers configuration Thomas Munro <[email protected]>
2021-03-27 05:31 ` Re: MultiXact\SLRU buffers configuration Andrey Borodin <[email protected]>
2021-03-28 21:15 ` Re: MultiXact\SLRU buffers configuration Thomas Munro <[email protected]>
2021-03-29 08:26 ` Re: MultiXact\SLRU buffers configuration Andrey Borodin <[email protected]>
2021-03-31 21:09 ` Re: MultiXact\SLRU buffers configuration Andrey Borodin <[email protected]>
2021-04-01 03:40 ` Re: MultiXact\SLRU buffers configuration Thomas Munro <[email protected]>
2021-04-03 19:57 ` Re: MultiXact\SLRU buffers configuration Andrey Borodin <[email protected]>
2021-04-07 05:59 ` Re: MultiXact\SLRU buffers configuration Thomas Munro <[email protected]>
2021-04-07 11:44 ` Re: MultiXact\SLRU buffers configuration Andrey Borodin <[email protected]>
2021-04-07 12:13 ` Re: MultiXact\SLRU buffers configuration Andrey Borodin <[email protected]>
2021-04-08 00:30 ` Re: MultiXact\SLRU buffers configuration Thomas Munro <[email protected]>
2021-04-08 07:24 ` Re: MultiXact\SLRU buffers configuration Andrey Borodin <[email protected]>
2021-04-08 12:22 ` Re: MultiXact\SLRU buffers configuration Thomas Munro <[email protected]>
2021-04-11 18:37 ` Re: MultiXact\SLRU buffers configuration Andrey Borodin <[email protected]>
2021-06-14 17:40 ` Re: MultiXact\SLRU buffers configuration Васильев Дмитрий <[email protected]>
2021-12-26 10:09 ` Re: MultiXact\SLRU buffers configuration Andrey Borodin <[email protected]>
2022-07-21 13:00 ` Re: MultiXact\SLRU buffers configuration Yura Sokolov <[email protected]>
2022-07-23 08:39 ` Re: MultiXact\SLRU buffers configuration Andrey Borodin <[email protected]>
2022-07-23 08:47 ` Re: MultiXact\SLRU buffers configuration Thomas Munro <[email protected]>
2022-12-20 18:39 ` Re: MultiXact\SLRU buffers configuration Andrey Borodin <[email protected]>
2022-08-16 19:36 ` Re: MultiXact\SLRU buffers configuration [email protected]
2022-08-18 03:35 ` Re: MultiXact\SLRU buffers configuration Andrey Borodin <[email protected]>
2022-08-19 15:48 ` Re: MultiXact\SLRU buffers configuration [email protected]
2024-01-20 03:31 ` Re: MultiXact\SLRU buffers configuration vignesh C <[email protected]>
2024-01-27 04:58 ` Re: MultiXact\SLRU buffers configuration Andrey Borodin <[email protected]>
2024-01-28 12:49 ` Re: MultiXact\SLRU buffers configuration Alvaro Herrera <[email protected]>
2024-01-28 18:17 ` Re: MultiXact\SLRU buffers configuration Andrey M. Borodin <[email protected]>
2020-10-28 19:36 ` Re: MultiXact\SLRU buffers configuration Alexander Korotkov <[email protected]>
2020-10-28 23:21 ` Re: MultiXact\SLRU buffers configuration Tomas Vondra <[email protected]>
2024-03-28 10:30 [PATCH v15 2/8] Row pattern recognition patch (parse/analysis). Tatsuo Ishii <[email protected]>
2024-07-03 11:17 Re: Optimize numeric multiplication for one and two base-NBASE digit multiplicands. Dean Rasheed <[email protected]>
2024-07-03 11:43 ` Re: Optimize numeric multiplication for one and two base-NBASE digit multiplicands. Ranier Vilela <[email protected]>
2024-07-03 13:48 ` Re: Optimize numeric multiplication for one and two base-NBASE digit multiplicands. Joel Jacobson <[email protected]>
2024-07-03 18:57 ` Re: Optimize numeric multiplication for one and two base-NBASE digit multiplicands. Dean Rasheed <[email protected]>
2024-07-03 20:27 ` Re: Optimize numeric multiplication for one and two base-NBASE digit multiplicands. Joel Jacobson <[email protected]>
2024-07-03 20:45 ` Re: Optimize numeric multiplication for one and two base-NBASE digit multiplicands. Joel Jacobson <[email protected]>
2024-07-04 18:43 ` Re: Optimize numeric multiplication for one and two base-NBASE digit multiplicands. Dean Rasheed <[email protected]>
2024-07-05 11:56 ` Re: Optimize numeric multiplication for one and two base-NBASE digit multiplicands. Joel Jacobson <[email protected]>
2024-07-05 15:41 ` Re: Optimize numeric multiplication for one and two base-NBASE digit multiplicands. Dean Rasheed <[email protected]>
2024-07-05 16:42 ` Re: Optimize numeric multiplication for one and two base-NBASE digit multiplicands. Joel Jacobson <[email protected]>
2024-07-05 17:37 ` Re: Optimize numeric multiplication for one and two base-NBASE digit multiplicands. Joel Jacobson <[email protected]>
2024-07-06 09:34 ` Re: Optimize numeric multiplication for one and two base-NBASE digit multiplicands. Dean Rasheed <[email protected]>
2024-07-06 11:16 ` Re: Optimize numeric multiplication for one and two base-NBASE digit multiplicands. Joel Jacobson <[email protected]>
2024-07-09 09:11 ` Re: Optimize numeric multiplication for one and two base-NBASE digit multiplicands. Dean Rasheed <[email protected]>
2024-07-09 12:01 ` Re: Optimize numeric multiplication for one and two base-NBASE digit multiplicands. Dean Rasheed <[email protected]>
2024-07-09 14:11 ` Re: Optimize numeric multiplication for one and two base-NBASE digit multiplicands. Joel Jacobson <[email protected]>
2024-07-09 19:48 ` Re: Optimize numeric multiplication for one and two base-NBASE digit multiplicands. Joel Jacobson <[email protected]>
2024-07-09 20:28 ` Re: Optimize numeric multiplication for one and two base-NBASE digit multiplicands. Joel Jacobson <[email protected]>
2024-07-03 19:05 ` Re: Optimize numeric multiplication for one and two base-NBASE digit multiplicands. Joel Jacobson <[email protected]>
2024-07-04 07:38 ` Re: Optimize numeric multiplication for one and two base-NBASE digit multiplicands. Joel Jacobson <[email protected]>
2024-07-04 10:38 ` Re: Optimize numeric multiplication for one and two base-NBASE digit multiplicands. Joel Jacobson <[email protected]>
2025-07-28 17:23 Re: MultiXact\SLRU buffers configuration Andrey Borodin <[email protected]>
This inbox is served by agora; see mirroring instructions
for how to clone and mirror all data and code used for this inbox