On Thu, Apr 1, 2021 at 10:09 AM Andrey Borodin <x4...@yandex-team.ru> wrote:
> > 29 марта 2021 г., в 11:26, Andrey Borodin <x4...@yandex-team.ru> 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.
From 05ce938fef85a5f23407d6977165319650b05d63 Mon Sep 17 00:00:00 2001
From: Thomas Munro <thomas.mu...@gmail.com>
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 <x4...@yandex-team.ru>
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
From edd07254e3f018ee3afe3f4251ab2cc0406d8b7a Mon Sep 17 00:00:00 2001
From: Andrey Borodin <amboro...@acm.org>
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 <x4...@yandex-team.ru>
Reviewed-by: Anastasia Lubennikova <a.lubennik...@postgrespro.ru>
Reviewed-by: Tomas Vondra <tomas.von...@2ndquadrant.com>
Reviewed-by: Alexander Korotkov <aekorot...@gmail.com>
Reviewed-by: Gilles Darold <gil...@darold.net>
Reviewed-by: Thomas Munro <thomas.mu...@gmail.com>
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
