public inbox for [email protected]  
help / color / mirror / Atom feed
From: Maxim Orlov <[email protected]>
To: Postgres hackers <[email protected]>
Subject: old_snapshot_threshold bottleneck on replica
Date: Mon, 23 Jan 2023 17:40:15 +0300
Message-ID: <CACG=ezYV+EvO135fLRdVn-ZusfVsTY6cH1OZqWtezuEYH6ciQA@mail.gmail.com> (raw)

Hi!

One of our customers stumble onto a significant performance degradation
while running multiple OLAP-like queries on a replica.
After some investigation, it became clear that the problem is in accessing
old_snapshot_threshold parameter.

Accessing old_snapshot_threshold parameter is guarded by mutex_threshold.
This is not a problem on primary
server, since we rarely call GetOldSnapshotThresholdTimestamp:

5028 void
5029 TestForOldSnapshot_impl(Snapshot snapshot, Relation relation)
5030 {
5031 ····if (RelationAllowsEarlyPruning(relation)
5032 ········&& (snapshot)->whenTaken < GetOldSnapshotThresholdTimestamp())
5033 ········ereport(ERROR,
5034 ················(errcode(ERRCODE_SNAPSHOT_TOO_OLD),
5035 ················ errmsg("snapshot too old")));

But in case of a replica, we have to call GetOldSnapshotThresholdTimestamp
much often. So, this become a
bottleneck. The customer solve this issue by setting old_snapshot_threshold
to 0. But, I think, we can
do something about it.

Some more investigation:

-- On primary --
$ ./bin/psql postgres -c "create database benchmark"
CREATE DATABASE
$ ./bin/pgbench -i -Uorlov -s300 benchmark
dropping old tables...
NOTICE:  table "pgbench_accounts" does not exist, skipping
...
creating tables...
generating data (client-side)...
30000000 of 30000000 tuples (100%) done (elapsed 142.37 s, remaining 0.00 s)
vacuuming...
creating primary keys...
done in 177.67 s (drop tables 0.00 s, create tables 0.01 s, client-side
generate 144.45 s, vacuum 0.59 s, primary keys 32.61 s).

-- On secondary --
$ touch 1.sql
$ vim 1.sql
$ cat 1.sql
\set bid random(1, 300)
BEGIN;
SELECT sum(aid) FROM pgbench_accounts where bid = :bid GROUP BY bid;
END;
$ ./bin/pgbench -f 1.sql -p5433 -Uorlov -j10 -c100 -T720 -P1 -n benchmark
pgbench (16devel)
progress: 1.0 s, 0.0 tps, lat 0.000 ms stddev 0.000, 0 failed
...
progress: 20.0 s, 0.0 tps, lat 0.000 ms stddev 0.000, 0 failed

$ perf record -F 99 -a -g --call-graph=dwarf sleep 5
$ perf script --header --fields comm,pid,tid,time,event,ip,sym,dso > file
$ grep s_lock file | wc -l

3486


My proposal is to use atomic for threshold_timestamp and threshold_xid. PFA
0001 patch.
With patch 0001 we got:

$ grep s_lock file2 | wc -l
8


Maybe, we shall go farther and remove mutex_threshold here? This will lead
to inconsistency of
threshold_timestamp and threshold_xid, but is this really a problem?

Thoughts?

-- 
Best regards,
Maxim Orlov.


Attachments:

  [application/octet-stream] 0001-PGPRO-7624-use-atomic-old_snapshot_threshold.patch (4.4K, ../CACG=ezYV+EvO135fLRdVn-ZusfVsTY6cH1OZqWtezuEYH6ciQA@mail.gmail.com/3-0001-PGPRO-7624-use-atomic-old_snapshot_threshold.patch)
  download | inline diff:
From 98fa70754a8c68cdfb20f27bd7f6e6e3e5f8ed92 Mon Sep 17 00:00:00 2001
From: Maxim Orlov <[email protected]>
Date: Wed, 11 Jan 2023 18:07:10 +0300
Subject: [PATCH] [PGPRO-7624] use atomic old_snapshot_threshold

Using spinlock to access old_snapshot_threshold lead to the bottleneck on
replica, since GetOldSnapshotThresholdTimestamp is called too often. So, switch
to an atomic values.

tags: commitfest_hotfix
---
 src/backend/utils/time/snapmgr.c | 31 +++++++++++++++++--------------
 src/include/utils/old_snapshot.h |  5 +++--
 2 files changed, 20 insertions(+), 16 deletions(-)

diff --git a/src/backend/utils/time/snapmgr.c b/src/backend/utils/time/snapmgr.c
index 7d11ae34781..0902e29f224 100644
--- a/src/backend/utils/time/snapmgr.c
+++ b/src/backend/utils/time/snapmgr.c
@@ -230,8 +230,8 @@ SnapMgrInit(void)
 		oldSnapshotControl->latest_xmin = InvalidTransactionId;
 		oldSnapshotControl->next_map_update = 0;
 		SpinLockInit(&oldSnapshotControl->mutex_threshold);
-		oldSnapshotControl->threshold_timestamp = 0;
-		oldSnapshotControl->threshold_xid = InvalidTransactionId;
+		pg_atomic_init_u64(&oldSnapshotControl->threshold_timestamp, 0);
+		pg_atomic_init_u32(&oldSnapshotControl->threshold_xid, InvalidTransactionId);
 		oldSnapshotControl->head_offset = 0;
 		oldSnapshotControl->head_timestamp = 0;
 		oldSnapshotControl->count_used = 0;
@@ -1706,9 +1706,7 @@ GetOldSnapshotThresholdTimestamp(void)
 {
 	TimestampTz threshold_timestamp;
 
-	SpinLockAcquire(&oldSnapshotControl->mutex_threshold);
-	threshold_timestamp = oldSnapshotControl->threshold_timestamp;
-	SpinLockRelease(&oldSnapshotControl->mutex_threshold);
+	threshold_timestamp = pg_atomic_read_u64(&oldSnapshotControl->threshold_timestamp);
 
 	return threshold_timestamp;
 }
@@ -1716,11 +1714,18 @@ GetOldSnapshotThresholdTimestamp(void)
 void
 SetOldSnapshotThresholdTimestamp(TimestampTz ts, TransactionId xlimit)
 {
+	TimestampTz threshold_timestamp;
+	TransactionId threshold_xid;
+
 	SpinLockAcquire(&oldSnapshotControl->mutex_threshold);
-	Assert(oldSnapshotControl->threshold_timestamp <= ts);
-	Assert(TransactionIdPrecedesOrEquals(oldSnapshotControl->threshold_xid, xlimit));
-	oldSnapshotControl->threshold_timestamp = ts;
-	oldSnapshotControl->threshold_xid = xlimit;
+	threshold_timestamp = pg_atomic_read_u64(&oldSnapshotControl->threshold_timestamp);
+	threshold_xid = pg_atomic_read_u32(&oldSnapshotControl->threshold_xid);
+
+	Assert(threshold_timestamp <= ts);
+	Assert(TransactionIdPrecedesOrEquals(threshold_xid, xlimit));
+
+	pg_atomic_write_u64(&oldSnapshotControl->threshold_timestamp, ts);
+	pg_atomic_write_u32(&oldSnapshotControl->threshold_xid, xlimit);
 	SpinLockRelease(&oldSnapshotControl->mutex_threshold);
 }
 
@@ -1739,9 +1744,7 @@ SnapshotTooOldMagicForTest(void)
 
 	ts -= 5 * USECS_PER_SEC;
 
-	SpinLockAcquire(&oldSnapshotControl->mutex_threshold);
-	oldSnapshotControl->threshold_timestamp = ts;
-	SpinLockRelease(&oldSnapshotControl->mutex_threshold);
+	pg_atomic_write_u64(&oldSnapshotControl->threshold_timestamp, ts);
 }
 
 /*
@@ -1846,8 +1849,8 @@ TransactionIdLimitedForOldSnapshots(TransactionId recentXmin,
 
 		/* Check for fast exit without LW locking. */
 		SpinLockAcquire(&oldSnapshotControl->mutex_threshold);
-		threshold_timestamp = oldSnapshotControl->threshold_timestamp;
-		threshold_xid = oldSnapshotControl->threshold_xid;
+		threshold_timestamp = pg_atomic_read_u64(&oldSnapshotControl->threshold_timestamp);
+		threshold_xid = pg_atomic_read_u32(&oldSnapshotControl->threshold_xid);
 		SpinLockRelease(&oldSnapshotControl->mutex_threshold);
 
 		if (ts == threshold_timestamp)
diff --git a/src/include/utils/old_snapshot.h b/src/include/utils/old_snapshot.h
index f1978a28e1c..3dd31d721c7 100644
--- a/src/include/utils/old_snapshot.h
+++ b/src/include/utils/old_snapshot.h
@@ -16,6 +16,7 @@
 #define OLD_SNAPSHOT_H
 
 #include "datatype/timestamp.h"
+#include "port/atomics.h"
 #include "storage/s_lock.h"
 
 /*
@@ -33,8 +34,8 @@ typedef struct OldSnapshotControlData
 	TransactionId latest_xmin;	/* latest snapshot xmin */
 	TimestampTz next_map_update;	/* latest snapshot valid up to */
 	slock_t		mutex_threshold;	/* protect threshold fields */
-	TimestampTz threshold_timestamp;	/* earlier snapshot is old */
-	TransactionId threshold_xid;	/* earlier xid may be gone */
+	pg_atomic_uint64 threshold_timestamp;	/* earlier snapshot is old */
+	pg_atomic_uint32 threshold_xid;	/* earlier xid may be gone */
 
 	/*
 	 * Keep one xid per minute for old snapshot error handling.
-- 
2.34.1



view thread (2+ messages)

reply

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Reply to all the recipients using the --to and --cc options:
  reply via email

  To: [email protected]
  Cc: [email protected], [email protected]
  Subject: Re: old_snapshot_threshold bottleneck on replica
  In-Reply-To: <CACG=ezYV+EvO135fLRdVn-ZusfVsTY6cH1OZqWtezuEYH6ciQA@mail.gmail.com>

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

This inbox is served by agora; see mirroring instructions
for how to clone and mirror all data and code used for this inbox