public inbox for [email protected]  
help / color / mirror / Atom feed
[PATCH 1/1] resownerbench
5+ messages / 4 participants
[nested] [flat]

* [PATCH 1/1] resownerbench
@ 2020-11-11 22:40 Heikki Linnakangas <[email protected]>
  0 siblings, 0 replies; 5+ messages in thread

From: Heikki Linnakangas @ 2020-11-11 22:40 UTC (permalink / raw)

---
 contrib/resownerbench/Makefile               |  17 ++
 contrib/resownerbench/resownerbench--1.0.sql |  14 ++
 contrib/resownerbench/resownerbench.c        | 154 +++++++++++++++++++
 contrib/resownerbench/resownerbench.control  |   6 +
 contrib/resownerbench/snaptest.sql           |  37 +++++
 5 files changed, 228 insertions(+)
 create mode 100644 contrib/resownerbench/Makefile
 create mode 100644 contrib/resownerbench/resownerbench--1.0.sql
 create mode 100644 contrib/resownerbench/resownerbench.c
 create mode 100644 contrib/resownerbench/resownerbench.control
 create mode 100644 contrib/resownerbench/snaptest.sql

diff --git a/contrib/resownerbench/Makefile b/contrib/resownerbench/Makefile
new file mode 100644
index 00000000000..9b0e1cfee1a
--- /dev/null
+++ b/contrib/resownerbench/Makefile
@@ -0,0 +1,17 @@
+MODULE_big = resownerbench
+OBJS = resownerbench.o
+
+EXTENSION = resownerbench
+DATA = resownerbench--1.0.sql
+PGFILEDESC = "resownerbench - benchmark for ResourceOwners"
+
+ifdef USE_PGXS
+PG_CONFIG = pg_config
+PGXS := $(shell $(PG_CONFIG) --pgxs)
+include $(PGXS)
+else
+subdir = contrib/resownerbench
+top_builddir = ../..
+include $(top_builddir)/src/Makefile.global
+include $(top_srcdir)/contrib/contrib-global.mk
+endif
diff --git a/contrib/resownerbench/resownerbench--1.0.sql b/contrib/resownerbench/resownerbench--1.0.sql
new file mode 100644
index 00000000000..d29182f5982
--- /dev/null
+++ b/contrib/resownerbench/resownerbench--1.0.sql
@@ -0,0 +1,14 @@
+/* contrib/resownerbench/resownerbench--1.0.sql */
+
+-- complain if script is sourced in psql, rather than via CREATE EXTENSION
+\echo Use "CREATE EXTENSION resownerbench" to load this file. \quit
+
+CREATE FUNCTION snapshotbench_lifo(numkeep int, numsnaps int, numiters int)
+RETURNS double precision
+AS 'MODULE_PATHNAME', 'snapshotbench_lifo'
+LANGUAGE C STRICT VOLATILE;
+
+CREATE FUNCTION snapshotbench_fifo(numkeep int, numsnaps int, numiters int)
+RETURNS double precision
+AS 'MODULE_PATHNAME', 'snapshotbench_fifo'
+LANGUAGE C STRICT VOLATILE;
diff --git a/contrib/resownerbench/resownerbench.c b/contrib/resownerbench/resownerbench.c
new file mode 100644
index 00000000000..acfb6c39199
--- /dev/null
+++ b/contrib/resownerbench/resownerbench.c
@@ -0,0 +1,154 @@
+#include "postgres.h"
+
+#include "catalog/pg_type.h"
+#include "catalog/pg_statistic.h"
+#include "executor/spi.h"
+#include "funcapi.h"
+#include "libpq/pqsignal.h"
+#include "utils/catcache.h"
+#include "utils/syscache.h"
+#include "utils/timestamp.h"
+#include "utils/snapmgr.h"
+
+PG_MODULE_MAGIC;
+
+PG_FUNCTION_INFO_V1(snapshotbench_lifo);
+PG_FUNCTION_INFO_V1(snapshotbench_fifo);
+
+/*
+ * ResourceOwner Performance test, using RegisterSnapshot().
+ *
+ * This takes three parameters: numkeep, numsnaps, numiters.
+ *
+ * First, we register 'numkeep' snapshots. They are kept registed
+ * until the end of the test. Then, we repeatedly register and
+ * unregister 'numsnaps - numkeep' additional snapshots, repeating
+ * 'numiters' times. All the register/unregister calls are made in
+ * LIFO order.
+ *
+ * Returns the time spent, in milliseconds.
+ *
+ * The idea is to test the performance of ResourceOwnerRemember()
+ * and ReourceOwnerForget() operations, under different regimes.
+ *
+ * In the old implementation, if 'numsnaps' is small enough, all
+ * the entries fit in the resource owner's small array (it can
+ * hold 64 entries).
+ *
+ * In the new implementation, the array is much smaller, only 8
+ * entries, but it's used together with the hash table so that
+ * we stay in the "array regime" as long as 'numsnaps - numkeep'
+ * is smaller than 8 entries.
+ *
+ * 'numiters' can be adjusted to adjust the overall runtime to be
+ * suitable long.
+ */
+Datum
+snapshotbench_lifo(PG_FUNCTION_ARGS)
+{
+	int			numkeep = PG_GETARG_INT32(0);
+	int			numsnaps = PG_GETARG_INT32(1);
+	int			numiters = PG_GETARG_INT32(2);
+	int			i;
+	instr_time	start,
+				duration;
+	Snapshot	lsnap;
+	Snapshot   *rs;
+	int			numregistered = 0;
+
+	rs = palloc(Max(numsnaps, numkeep) * sizeof(Snapshot));
+
+	lsnap = GetLatestSnapshot();
+
+	PG_SETMASK(&BlockSig);
+	INSTR_TIME_SET_CURRENT(start);
+
+	while (numregistered < numkeep)
+	{
+		rs[numregistered] = RegisterSnapshot(lsnap);
+		numregistered++;
+	}
+
+	for (i = 0 ; i < numiters; i++)
+	{
+		while (numregistered < numsnaps)
+		{
+			rs[numregistered] = RegisterSnapshot(lsnap);
+			numregistered++;
+		}
+
+		while (numregistered > numkeep)
+		{
+			numregistered--;
+			UnregisterSnapshot(rs[numregistered]);
+		}
+	}
+
+	while (numregistered > 0)
+	{
+		numregistered--;
+		UnregisterSnapshot(rs[numregistered]);
+	}
+
+	INSTR_TIME_SET_CURRENT(duration);
+	INSTR_TIME_SUBTRACT(duration, start);
+	PG_SETMASK(&UnBlockSig);
+
+	PG_RETURN_FLOAT8(INSTR_TIME_GET_MILLISEC(duration));
+};
+
+
+/*
+ * Same, but do the register/unregister operations in
+ * FIFO order.
+ */
+Datum
+snapshotbench_fifo(PG_FUNCTION_ARGS)
+{
+	int			numkeep = PG_GETARG_INT32(0);
+	int			numsnaps = PG_GETARG_INT32(1);
+	int			numiters = PG_GETARG_INT32(2);
+	int			i,
+				j;
+	instr_time	start,
+				duration;
+	Snapshot	lsnap;
+	Snapshot   *rs;
+	int			numregistered = 0;
+
+	rs = palloc(Max(numsnaps, numkeep) * sizeof(Snapshot));
+
+	lsnap = GetLatestSnapshot();
+
+	PG_SETMASK(&BlockSig);
+	INSTR_TIME_SET_CURRENT(start);
+
+	while (numregistered < numkeep)
+	{
+		rs[numregistered] = RegisterSnapshot(lsnap);
+		numregistered++;
+	}
+
+	for (i = 0 ; i < numiters; i++)
+	{
+		while (numregistered < numsnaps)
+		{
+			rs[numregistered] = RegisterSnapshot(lsnap);
+			numregistered++;
+		}
+
+		for (j = numkeep; j < numregistered; j++)
+			UnregisterSnapshot(rs[j]);
+		numregistered = numkeep;
+	}
+
+	for (j = 0; j < numregistered; j++)
+		UnregisterSnapshot(rs[j]);
+	numregistered = numkeep;
+
+	INSTR_TIME_SET_CURRENT(duration);
+	INSTR_TIME_SUBTRACT(duration, start);
+	PG_SETMASK(&UnBlockSig);
+
+	PG_RETURN_FLOAT8(INSTR_TIME_GET_MILLISEC(duration));
+};
diff --git a/contrib/resownerbench/resownerbench.control b/contrib/resownerbench/resownerbench.control
new file mode 100644
index 00000000000..ada88b8eed8
--- /dev/null
+++ b/contrib/resownerbench/resownerbench.control
@@ -0,0 +1,6 @@
+# resownerbench
+
+comment = 'benchmark for ResourceOwners'
+default_version = '1.0'
+module_pathname = '$libdir/resownerbench'
+relocatable = true
diff --git a/contrib/resownerbench/snaptest.sql b/contrib/resownerbench/snaptest.sql
new file mode 100644
index 00000000000..18c54e13fc9
--- /dev/null
+++ b/contrib/resownerbench/snaptest.sql
@@ -0,0 +1,37 @@
+--
+-- Performance test RegisterSnapshot/UnregisterSnapshot.
+--
+select numkeep, numsnaps,
+       -- numiters,
+       -- round(lifo_time_ms) as lifo_total_time_ms,
+       -- round(fifo_time_ms) as fifo_total_time_ms,
+       round((lifo_time_ms::numeric / (numkeep + (numsnaps - numkeep) * numiters)) * 1000000, 1) as lifo_time_ns,
+       round((fifo_time_ms::numeric / (numkeep + (numsnaps - numkeep) * numiters)) * 1000000, 1) as fifo_time_ns
+from
+(values (0,      1,  10000000),
+        (0,      5,   2000000),
+        (0,     10,   1000000),
+        (0,     60,    100000),
+        (0,     70,    100000),
+        (0,    100,    100000),
+        (0,   1000,     10000),
+        (0,  10000,      1000),
+
+-- These tests keep 9 snapshots registered across the iterations. That
+-- exceeds the size of the little array in the patch, so this exercises
+-- the hash lookups. Without the patch, these still fit in the array
+-- (it's 64 entries without the patch)
+        (9,     10,  10000000),
+        (9,    100,    100000),
+        (9,   1000,     10000),
+        (9,  10000,      1000),
+
+-- These exceed the 64 entry array even without the patch, so these fall
+-- in the hash table regime with and without the patch.
+        (65,    70,   1000000),
+        (65,   100,    100000),
+        (65,  1000,     10000),
+        (65, 10000,      1000)
+) AS params (numkeep, numsnaps, numiters),
+lateral snapshotbench_lifo(numkeep, numsnaps, numiters) as lifo_time_ms,
+lateral snapshotbench_fifo(numkeep, numsnaps, numiters) as fifo_time_ms;
-- 
2.29.2


--------------1104342304EC6341F5D7C60C--





^ permalink  raw  reply  [nested|flat] 5+ messages in thread

* Re: request for database identifier in the startup packet
@ 2024-05-09 18:20 Dave Cramer <[email protected]>
  2024-05-09 19:14 ` Re: request for database identifier in the startup packet Andres Freund <[email protected]>
  0 siblings, 1 reply; 5+ messages in thread

From: Dave Cramer @ 2024-05-09 18:20 UTC (permalink / raw)
  To: Robert Haas <[email protected]>; +Cc: PostgreSQL Hackers <[email protected]>

Dave Cramer


On Thu, 9 May 2024 at 12:22, Robert Haas <[email protected]> wrote:

> On Thu, May 9, 2024 at 8:06 AM Dave Cramer <[email protected]> wrote:
> > The JDBC driver is currently keeping a per connection cache of types in
> the driver. We are seeing cases where the number of columns is quite high.
> In one case Prevent fetchFieldMetaData() from being run when unnecessary. ·
> Issue #3241 · pgjdbc/pgjdbc (github.com) 2.6 Million columns.
> >
> > If we knew that we were connecting to the same database we could use a
> single cache across connections.
> >
> > I think we would require a server/database identifier in the startup
> message.
>
> I understand the desire to share the cache, but not why that would
> require any kind of change to the wire protocol.
>
> The server identity is actually useful for many things such as knowing
which instance of a cluster you are connected to.
For the cache however we can't use the IP address to determine which server
we are connected to as we could be connected to a pooler.
Knowing exactly which server/database makes it relatively easy to have a
common cache across connections. Getting that in the startup message seems
like a good place

Dave


^ permalink  raw  reply  [nested|flat] 5+ messages in thread

* Re: request for database identifier in the startup packet
  2024-05-09 18:20 Re: request for database identifier in the startup packet Dave Cramer <[email protected]>
@ 2024-05-09 19:14 ` Andres Freund <[email protected]>
  2024-05-09 19:19   ` Re: request for database identifier in the startup packet Robert Haas <[email protected]>
  0 siblings, 1 reply; 5+ messages in thread

From: Andres Freund @ 2024-05-09 19:14 UTC (permalink / raw)
  To: Dave Cramer <[email protected]>; +Cc: Robert Haas <[email protected]>; PostgreSQL Hackers <[email protected]>

Hi,

On 2024-05-09 14:20:49 -0400, Dave Cramer wrote:
> On Thu, 9 May 2024 at 12:22, Robert Haas <[email protected]> wrote:
> > On Thu, May 9, 2024 at 8:06 AM Dave Cramer <[email protected]> wrote:
> > > The JDBC driver is currently keeping a per connection cache of types in
> > the driver. We are seeing cases where the number of columns is quite high.
> > In one case Prevent fetchFieldMetaData() from being run when unnecessary. ·
> > Issue #3241 · pgjdbc/pgjdbc (github.com) 2.6 Million columns.
> > >
> > > If we knew that we were connecting to the same database we could use a
> > single cache across connections.
> > >
> > > I think we would require a server/database identifier in the startup
> > message.
> >
> > I understand the desire to share the cache, but not why that would
> > require any kind of change to the wire protocol.
> >
> > The server identity is actually useful for many things such as knowing
> which instance of a cluster you are connected to.
> For the cache however we can't use the IP address to determine which server
> we are connected to as we could be connected to a pooler.
> Knowing exactly which server/database makes it relatively easy to have a
> common cache across connections. Getting that in the startup message seems
> like a good place

ISTM that you could just as well query the information you'd like after
connecting. And that's going to be a lot more flexible than having to have
precisely the right information in the startup message, and most clients not
needing it.

Greetings,

Andres Freund






^ permalink  raw  reply  [nested|flat] 5+ messages in thread

* Re: request for database identifier in the startup packet
  2024-05-09 18:20 Re: request for database identifier in the startup packet Dave Cramer <[email protected]>
  2024-05-09 19:14 ` Re: request for database identifier in the startup packet Andres Freund <[email protected]>
@ 2024-05-09 19:19   ` Robert Haas <[email protected]>
  2024-05-09 19:33     ` Re: request for database identifier in the startup packet Dave Cramer <[email protected]>
  0 siblings, 1 reply; 5+ messages in thread

From: Robert Haas @ 2024-05-09 19:19 UTC (permalink / raw)
  To: Andres Freund <[email protected]>; +Cc: Dave Cramer <[email protected]>; PostgreSQL Hackers <[email protected]>

On Thu, May 9, 2024 at 3:14 PM Andres Freund <[email protected]> wrote:
> ISTM that you could just as well query the information you'd like after
> connecting. And that's going to be a lot more flexible than having to have
> precisely the right information in the startup message, and most clients not
> needing it.

I agree with this.

-- 
Robert Haas
EDB: http://www.enterprisedb.com






^ permalink  raw  reply  [nested|flat] 5+ messages in thread

* Re: request for database identifier in the startup packet
  2024-05-09 18:20 Re: request for database identifier in the startup packet Dave Cramer <[email protected]>
  2024-05-09 19:14 ` Re: request for database identifier in the startup packet Andres Freund <[email protected]>
  2024-05-09 19:19   ` Re: request for database identifier in the startup packet Robert Haas <[email protected]>
@ 2024-05-09 19:33     ` Dave Cramer <[email protected]>
  0 siblings, 0 replies; 5+ messages in thread

From: Dave Cramer @ 2024-05-09 19:33 UTC (permalink / raw)
  To: Robert Haas <[email protected]>; +Cc: Andres Freund <[email protected]>; PostgreSQL Hackers <[email protected]>

On Thu, 9 May 2024 at 15:19, Robert Haas <[email protected]> wrote:

> On Thu, May 9, 2024 at 3:14 PM Andres Freund <[email protected]> wrote:
> > ISTM that you could just as well query the information you'd like after
> > connecting. And that's going to be a lot more flexible than having to
> have
> > precisely the right information in the startup message, and most clients
> not
> > needing it.
>
> I agree with this.
>
> Well other than the extra round trip.

Thanks,
Dave


^ permalink  raw  reply  [nested|flat] 5+ messages in thread


end of thread, other threads:[~2024-05-09 19:33 UTC | newest]

Thread overview: 5+ messages (download: mbox mbox.gz follow: Atom feed)
-- links below jump to the message on this page --
2020-11-11 22:40 [PATCH 1/1] resownerbench Heikki Linnakangas <[email protected]>
2024-05-09 18:20 Re: request for database identifier in the startup packet Dave Cramer <[email protected]>
2024-05-09 19:14 ` Re: request for database identifier in the startup packet Andres Freund <[email protected]>
2024-05-09 19:19   ` Re: request for database identifier in the startup packet Robert Haas <[email protected]>
2024-05-09 19:33     ` Re: request for database identifier in the startup packet Dave Cramer <[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