postgresql/doc/TODO.detail/performance
Bruce Momjian b864793e80 update
2001-01-19 19:50:04 +00:00

742 lines
30 KiB
Plaintext

From owner-pgsql-hackers@hub.org Sun Jun 14 18:45:04 1998
Received: from hub.org (hub.org [209.47.148.200])
by candle.pha.pa.us (8.8.5/8.8.5) with ESMTP id SAA03690
for <maillist@candle.pha.pa.us>; Sun, 14 Jun 1998 18:45:00 -0400 (EDT)
Received: from localhost (majordom@localhost) by hub.org (8.8.8/8.7.5) with SMTP id SAA28049; Sun, 14 Jun 1998 18:39:42 -0400 (EDT)
Received: by hub.org (TLB v0.10a (1.23 tibbs 1997/01/09 00:29:32)); Sun, 14 Jun 1998 18:36:06 +0000 (EDT)
Received: (from majordom@localhost) by hub.org (8.8.8/8.7.5) id SAA27943 for pgsql-hackers-outgoing; Sun, 14 Jun 1998 18:36:04 -0400 (EDT)
Received: from angular.illustra.com (ifmxoak.illustra.com [206.175.10.34]) by hub.org (8.8.8/8.7.5) with ESMTP id SAA27925 for <pgsql-hackers@postgresql.org>; Sun, 14 Jun 1998 18:35:47 -0400 (EDT)
Received: from hawk.illustra.com (hawk.illustra.com [158.58.61.70]) by angular.illustra.com (8.7.4/8.7.3) with SMTP id PAA21293 for <pgsql-hackers@postgresql.org>; Sun, 14 Jun 1998 15:35:12 -0700 (PDT)
Received: by hawk.illustra.com (5.x/smail2.5/06-10-94/S)
id AA07922; Sun, 14 Jun 1998 15:35:13 -0700
From: dg@illustra.com (David Gould)
Message-Id: <9806142235.AA07922@hawk.illustra.com>
Subject: [HACKERS] performance tests, initial results
To: pgsql-hackers@postgreSQL.org
Date: Sun, 14 Jun 1998 15:35:13 -0700 (PDT)
Mime-Version: 1.0
Content-Type: text/plain; charset=US-ASCII
Content-Transfer-Encoding: 7bit
Sender: owner-pgsql-hackers@hub.org
Precedence: bulk
Status: RO
I have been playing a little with the performance tests found in
pgsql/src/tests/performance and have a few observations that might be of
minor interest.
The tests themselves are simple enough although the result parsing in the
driver did not work on Linux. I am enclosing a patch below to fix this. I
think it will also work better on the other systems.
A summary of results from my testing are below. Details are at the bottom
of this message.
My test system is 'leslie':
linux 2.0.32, gcc version 2.7.2.3
P133, HX chipset, 512K L2, 32MB mem
NCR810 fast scsi, Quantum Atlas 2GB drive (7200 rpm).
Results Summary (times in seconds)
Single txn 8K txn Create 8K idx 8K random Simple
Case Description 8K insert 8K insert Index Insert Scans Orderby
=================== ========== ========= ====== ====== ========= =======
1 From Distribution
P90 FreeBsd -B256 39.56 1190.98 3.69 46.65 65.49 2.27
IDE
2 Running on leslie
P133 Linux 2.0.32 15.48 326.75 2.99 20.69 35.81 1.68
SCSI 32M
3 leslie, -o -F
no forced writes 15.90 24.98 2.63 20.46 36.43 1.69
4 leslie, -o -F
no ASSERTS 14.92 23.23 1.38 18.67 33.79 1.58
5 leslie, -o -F -B2048
more buffers 21.31 42.28 2.65 25.74 42.26 1.72
6 leslie, -o -F -B2048
more bufs, no ASSERT 20.52 39.79 1.40 24.77 39.51 1.55
Case to Case Difference Factors (+ is faster)
Single txn 8K txn Create 8K idx 8K random Simple
Case Description 8K insert 8K insert Index Insert Scans Orderby
=================== ========== ========= ====== ====== ========= =======
leslie vs BSD P90. 2.56 3.65 1.23 2.25 1.83 1.35
(noflush -F) vs no -F -1.03 13.08 1.14 1.01 -1.02 1.00
No Assert vs Assert 1.05 1.07 1.90 1.06 1.07 1.09
-B256 vs -B2048 1.34 1.69 1.01 1.26 1.16 1.02
Observations:
- leslie (P133 linux) appears to be about 1.8 times faster than the
P90 BSD system used for the test result distributed with the source, not
counting the 8K txn insert case which was completely disk bound.
- SCSI disks make a big (factor of 3.6) difference. During this test the
disk was hammering and cpu utilization was < 10%.
- Assertion checking seems to cost about 7% except for create index where
it costs 90%
- the -F option to avoid flushing buffers has tremendous effect if there are
many very small transactions. Or, another way, flushing at the end of the
transaction is a major disaster for performance.
- Something is very wrong with our buffer cache implementation. Going from
256 buffers to 2048 buffers costs an average of 25%. In the 8K txn case
it costs about 70%. I see looking at the code and profiling that in the 8K
txn case this is in BufferSync() which examines all the buffers at commit
time. I don't quite understand why it is so costly for the single 8K row
txn (35%) though.
It would be nice to have some more tests. Maybe the Wisconsin stuff will
be useful.
----------------- patch to test harness. apply from pgsql ------------
*** src/test/performance/runtests.pl.orig Sun Jun 14 11:34:04 1998
Differences %
----------------- patch to test harness. apply from pgsql ------------
*** src/test/performance/runtests.pl.orig Sun Jun 14 11:34:04 1998
--- src/test/performance/runtests.pl Sun Jun 14 12:07:30 1998
***************
*** 84,123 ****
open (STDERR, ">$TmpFile") or die;
select (STDERR); $| = 1;
! for ($i = 0; $i <= $#perftests; $i++)
! {
$test = $perftests[$i];
($test, $XACTBLOCK) = split (/ /, $test);
$runtest = $test;
! if ( $test =~ /\.ntm/ )
! {
! #
# No timing for this queries
- #
close (STDERR); # close $TmpFile
open (STDERR, ">/dev/null") or die;
$runtest =~ s/\.ntm//;
}
! else
! {
close (STDOUT);
open(STDOUT, ">&SAVEOUT");
print STDOUT "\nRunning: $perftests[$i+1] ...";
close (STDOUT);
open (STDOUT, ">/dev/null") or die;
select (STDERR); $| = 1;
! printf "$perftests[$i+1]: ";
}
do "sqls/$runtest";
# Restore STDERR to $TmpFile
! if ( $test =~ /\.ntm/ )
! {
close (STDERR);
open (STDERR, ">>$TmpFile") or die;
}
-
select (STDERR); $| = 1;
$i++;
}
--- 84,116 ----
open (STDERR, ">$TmpFile") or die;
select (STDERR); $| = 1;
! for ($i = 0; $i <= $#perftests; $i++) {
$test = $perftests[$i];
($test, $XACTBLOCK) = split (/ /, $test);
$runtest = $test;
! if ( $test =~ /\.ntm/ ) {
# No timing for this queries
close (STDERR); # close $TmpFile
open (STDERR, ">/dev/null") or die;
$runtest =~ s/\.ntm//;
}
! else {
close (STDOUT);
open(STDOUT, ">&SAVEOUT");
print STDOUT "\nRunning: $perftests[$i+1] ...";
close (STDOUT);
open (STDOUT, ">/dev/null") or die;
select (STDERR); $| = 1;
! print "$perftests[$i+1]: ";
}
do "sqls/$runtest";
# Restore STDERR to $TmpFile
! if ( $test =~ /\.ntm/ ) {
close (STDERR);
open (STDERR, ">>$TmpFile") or die;
}
select (STDERR); $| = 1;
$i++;
}
***************
*** 128,138 ****
open (TMPF, "<$TmpFile") or die;
open (RESF, ">$ResFile") or die;
! while (<TMPF>)
! {
! $str = $_;
! ($test, $rtime) = split (/:/, $str);
! ($tmp, $rtime, $rest) = split (/[ ]+/, $rtime);
! print RESF "$test: $rtime\n";
}
--- 121,130 ----
open (TMPF, "<$TmpFile") or die;
open (RESF, ">$ResFile") or die;
! while (<TMPF>) {
! if (m/^(.*: ).* ([0-9:.]+) *elapsed/) {
! ($test, $rtime) = ($1, $2);
! print RESF $test, $rtime, "\n";
! }
}
------------------------------------------------------------------------
------------------------- testcase detail --------------------------
1. from distribution
DBMS: PostgreSQL 6.2b10
OS: FreeBSD 2.1.5-RELEASE
HardWare: i586/90, 24M RAM, IDE
StartUp: postmaster -B 256 '-o -S 2048' -S
Compiler: gcc 2.6.3
Compiled: -O, without CASSERT checking, with
-DTBL_FREE_CMD_MEMORY (to free memory
if BEGIN/END after each query execution)
DB connection startup: 0.20
8192 INSERTs INTO SIMPLE (1 xact): 39.58
8192 INSERTs INTO SIMPLE (8192 xacts): 1190.98
Create INDEX on SIMPLE: 3.69
8192 INSERTs INTO SIMPLE with INDEX (1 xact): 46.65
8192 random INDEX scans on SIMPLE (1 xact): 65.49
ORDER BY SIMPLE: 2.27
2. run on leslie with asserts
DBMS: PostgreSQL 6.3.2 (plus changes to 98/06/01)
OS: Linux 2.0.32 leslie
HardWare: i586/133 HX 512, 32M RAM, fast SCSI, 7200rpm
StartUp: postmaster -B 256 '-o -S 2048' -S
Compiler: gcc 2.7.2.3
Compiled: -O, WITH CASSERT checking, with
-DTBL_FREE_CMD_MEMORY (to free memory
if BEGIN/END after each query execution)
DB connection startup: 0.10
8192 INSERTs INTO SIMPLE (1 xact): 15.48
8192 INSERTs INTO SIMPLE (8192 xacts): 326.75
Create INDEX on SIMPLE: 2.99
8192 INSERTs INTO SIMPLE with INDEX (1 xact): 20.69
8192 random INDEX scans on SIMPLE (1 xact): 35.81
ORDER BY SIMPLE: 1.68
3. with -F to avoid forced i/o
DBMS: PostgreSQL 6.3.2 (plus changes to 98/06/01)
OS: Linux 2.0.32 leslie
HardWare: i586/133 HX 512, 32M RAM, fast SCSI, 7200rpm
StartUp: postmaster -B 256 '-o -S 2048 -F' -S
Compiler: gcc 2.7.2.3
Compiled: -O, WITH CASSERT checking, with
-DTBL_FREE_CMD_MEMORY (to free memory
if BEGIN/END after each query execution)
DB connection startup: 0.10
8192 INSERTs INTO SIMPLE (1 xact): 15.90
8192 INSERTs INTO SIMPLE (8192 xacts): 24.98
Create INDEX on SIMPLE: 2.63
8192 INSERTs INTO SIMPLE with INDEX (1 xact): 20.46
8192 random INDEX scans on SIMPLE (1 xact): 36.43
ORDER BY SIMPLE: 1.69
4. no asserts, -F to avoid forced I/O
DBMS: PostgreSQL 6.3.2 (plus changes to 98/06/01)
OS: Linux 2.0.32 leslie
HardWare: i586/133 HX 512, 32M RAM, fast SCSI, 7200rpm
StartUp: postmaster -B 256 '-o -S 2048' -S
Compiler: gcc 2.7.2.3
Compiled: -O, No CASSERT checking, with
-DTBL_FREE_CMD_MEMORY (to free memory
if BEGIN/END after each query execution)
DB connection startup: 0.10
8192 INSERTs INTO SIMPLE (1 xact): 14.92
8192 INSERTs INTO SIMPLE (8192 xacts): 23.23
Create INDEX on SIMPLE: 1.38
8192 INSERTs INTO SIMPLE with INDEX (1 xact): 18.67
8192 random INDEX scans on SIMPLE (1 xact): 33.79
ORDER BY SIMPLE: 1.58
5. with more buffers (2048 vs 256) and -F to avoid forced i/o
DBMS: PostgreSQL 6.3.2 (plus changes to 98/06/01)
OS: Linux 2.0.32 leslie
HardWare: i586/133 HX 512, 32M RAM, fast SCSI, 7200rpm
StartUp: postmaster -B 2048 '-o -S 2048 -F' -S
Compiler: gcc 2.7.2.3
Compiled: -O, WITH CASSERT checking, with
-DTBL_FREE_CMD_MEMORY (to free memory
if BEGIN/END after each query execution)
DB connection startup: 0.11
8192 INSERTs INTO SIMPLE (1 xact): 21.31
8192 INSERTs INTO SIMPLE (8192 xacts): 42.28
Create INDEX on SIMPLE: 2.65
8192 INSERTs INTO SIMPLE with INDEX (1 xact): 25.74
8192 random INDEX scans on SIMPLE (1 xact): 42.26
ORDER BY SIMPLE: 1.72
6. No Asserts, more buffers (2048 vs 256) and -F to avoid forced i/o
DBMS: PostgreSQL 6.3.2 (plus changes to 98/06/01)
OS: Linux 2.0.32 leslie
HardWare: i586/133 HX 512, 32M RAM, fast SCSI, 7200rpm
StartUp: postmaster -B 2048 '-o -S 2048 -F' -S
Compiler: gcc 2.7.2.3
Compiled: -O, No CASSERT checking, with
-DTBL_FREE_CMD_MEMORY (to free memory
if BEGIN/END after each query execution)
DB connection startup: 0.11
8192 INSERTs INTO SIMPLE (1 xact): 20.52
8192 INSERTs INTO SIMPLE (8192 xacts): 39.79
Create INDEX on SIMPLE: 1.40
8192 INSERTs INTO SIMPLE with INDEX (1 xact): 24.77
8192 random INDEX scans on SIMPLE (1 xact): 39.51
ORDER BY SIMPLE: 1.55
---------------------------------------------------------------------
-dg
David Gould dg@illustra.com 510.628.3783 or 510.305.9468
Informix Software (No, really) 300 Lakeside Drive Oakland, CA 94612
"Don't worry about people stealing your ideas. If your ideas are any
good, you'll have to ram them down people's throats." -- Howard Aiken
From owner-pgsql-hackers@hub.org Tue Oct 19 10:31:10 1999
Received: from renoir.op.net (root@renoir.op.net [209.152.193.4])
by candle.pha.pa.us (8.9.0/8.9.0) with ESMTP id KAA29087
for <maillist@candle.pha.pa.us>; Tue, 19 Oct 1999 10:31:08 -0400 (EDT)
Received: from hub.org (hub.org [216.126.84.1]) by renoir.op.net (o1/$Revision: 1.5 $) with ESMTP id KAA27535 for <maillist@candle.pha.pa.us>; Tue, 19 Oct 1999 10:19:47 -0400 (EDT)
Received: from localhost (majordom@localhost)
by hub.org (8.9.3/8.9.3) with SMTP id KAA30328;
Tue, 19 Oct 1999 10:12:10 -0400 (EDT)
(envelope-from owner-pgsql-hackers)
Received: by hub.org (bulk_mailer v1.5); Tue, 19 Oct 1999 10:11:55 -0400
Received: (from majordom@localhost)
by hub.org (8.9.3/8.9.3) id KAA30030
for pgsql-hackers-outgoing; Tue, 19 Oct 1999 10:11:00 -0400 (EDT)
(envelope-from owner-pgsql-hackers@postgreSQL.org)
Received: from sss.sss.pgh.pa.us (sss.pgh.pa.us [209.114.166.2])
by hub.org (8.9.3/8.9.3) with ESMTP id KAA29914
for <pgsql-hackers@postgreSQL.org>; Tue, 19 Oct 1999 10:10:33 -0400 (EDT)
(envelope-from tgl@sss.pgh.pa.us)
Received: from sss.sss.pgh.pa.us (localhost [127.0.0.1])
by sss.sss.pgh.pa.us (8.9.1/8.9.1) with ESMTP id KAA09038;
Tue, 19 Oct 1999 10:09:15 -0400 (EDT)
To: "Hiroshi Inoue" <Inoue@tpf.co.jp>
cc: "Vadim Mikheev" <vadim@krs.ru>, pgsql-hackers@postgreSQL.org
Subject: Re: [HACKERS] mdnblocks is an amazing time sink in huge relations
In-reply-to: Your message of Tue, 19 Oct 1999 19:03:22 +0900
<000801bf1a19$2d88ae20$2801007e@cadzone.tpf.co.jp>
Date: Tue, 19 Oct 1999 10:09:15 -0400
Message-ID: <9036.940342155@sss.pgh.pa.us>
From: Tom Lane <tgl@sss.pgh.pa.us>
Sender: owner-pgsql-hackers@postgreSQL.org
Status: OR
"Hiroshi Inoue" <Inoue@tpf.co.jp> writes:
> 1. shared cache holds committed system tuples.
> 2. private cache holds uncommitted system tuples.
> 3. relpages of shared cache are updated immediately by
> phisical change and corresponding buffer pages are
> marked dirty.
> 4. on commit, the contents of uncommitted tuples except
> relpages,reltuples,... are copied to correponding tuples
> in shared cache and the combined contents are
> committed.
> If so,catalog cache invalidation would be no longer needed.
> But synchronization of the step 4. may be difficult.
I think the main problem is that relpages and reltuples shouldn't
be kept in pg_class columns at all, because they need to have
very different update behavior from the other pg_class columns.
The rest of pg_class is update-on-commit, and we can lock down any one
row in the normal MVCC way (if transaction A has modified a row and
transaction B also wants to modify it, B waits for A to commit or abort,
so it can know which version of the row to start from). Furthermore,
there can legitimately be several different values of a row in use in
different places: the latest committed, an uncommitted modification, and
one or more old values that are still being used by active transactions
because they were current when those transactions started. (BTW, the
present relcache is pretty bad about maintaining pure MVCC transaction
semantics like this, but it seems clear to me that that's the direction
we want to go in.)
relpages cannot operate this way. To be useful for avoiding lseeks,
relpages *must* change exactly when the physical file changes. It
matters not at all whether the particular transaction that extended the
file ultimately commits or not. Moreover there can be only one correct
value (per relation) across the whole system, because there is only one
length of the relation file.
If we want to take reltuples seriously and try to maintain it
on-the-fly, then I think it needs still a third behavior. Clearly
it cannot be updated using MVCC rules, or we lose all writer
concurrency (if A has added tuples to a rel, B would have to wait
for A to commit before it could update reltuples...). Furthermore
"updating" isn't a simple matter of storing what you think the new
value is; otherwise two transactions adding tuples in parallel would
leave the wrong answer after B commits and overwrites A's value.
I think it would work for each transaction to keep track of a net delta
in reltuples for each table it's changed (total tuples added less total
tuples deleted), and then atomically add that value to the table's
shared reltuples counter during commit. But that still leaves the
problem of how you use the counter during a transaction to get an
accurate answer to the question "If I scan this table now, how many tuples
will I see?" At the time the question is asked, the current shared
counter value might include the effects of transactions that have
committed since your transaction started, and therefore are not visible
under MVCC rules. I think getting the correct answer would involve
making an instantaneous copy of the current counter at the start of
your xact, and then adding your own private net-uncommitted-delta to
the saved shared counter value when asked the question. This doesn't
look real practical --- you'd have to save the reltuples counts of
*all* tables in the database at the start of each xact, on the off
chance that you might need them. Ugh. Perhaps someone has a better
idea. In any case, reltuples clearly needs different mechanisms than
the ordinary fields in pg_class do, because updating it will be a
performance bottleneck otherwise.
If we allow reltuples to be updated only by vacuum-like events, as
it is now, then I think keeping it in pg_class is still OK.
In short, it seems clear to me that relpages should be removed from
pg_class and kept somewhere else if we want to make it more reliable
than it is now, and the same for reltuples (but reltuples doesn't
behave the same as relpages, and probably ought to be handled
differently).
regards, tom lane
************
From owner-pgsql-hackers@hub.org Tue Oct 19 21:25:30 1999
Received: from renoir.op.net (root@renoir.op.net [209.152.193.4])
by candle.pha.pa.us (8.9.0/8.9.0) with ESMTP id VAA28130
for <maillist@candle.pha.pa.us>; Tue, 19 Oct 1999 21:25:26 -0400 (EDT)
Received: from hub.org (hub.org [216.126.84.1]) by renoir.op.net (o1/$Revision: 1.5 $) with ESMTP id VAA10512 for <maillist@candle.pha.pa.us>; Tue, 19 Oct 1999 21:15:28 -0400 (EDT)
Received: from localhost (majordom@localhost)
by hub.org (8.9.3/8.9.3) with SMTP id VAA50745;
Tue, 19 Oct 1999 21:07:23 -0400 (EDT)
(envelope-from owner-pgsql-hackers)
Received: by hub.org (bulk_mailer v1.5); Tue, 19 Oct 1999 21:07:01 -0400
Received: (from majordom@localhost)
by hub.org (8.9.3/8.9.3) id VAA50644
for pgsql-hackers-outgoing; Tue, 19 Oct 1999 21:06:06 -0400 (EDT)
(envelope-from owner-pgsql-hackers@postgreSQL.org)
Received: from sd.tpf.co.jp (sd.tpf.co.jp [210.161.239.34])
by hub.org (8.9.3/8.9.3) with ESMTP id VAA50584
for <pgsql-hackers@postgreSQL.org>; Tue, 19 Oct 1999 21:05:26 -0400 (EDT)
(envelope-from Inoue@tpf.co.jp)
Received: from cadzone ([126.0.1.40] (may be forged))
by sd.tpf.co.jp (2.5 Build 2640 (Berkeley 8.8.6)/8.8.4) with SMTP
id KAA01715; Wed, 20 Oct 1999 10:05:14 +0900
From: "Hiroshi Inoue" <Inoue@tpf.co.jp>
To: "Tom Lane" <tgl@sss.pgh.pa.us>
Cc: <pgsql-hackers@postgreSQL.org>
Subject: RE: [HACKERS] mdnblocks is an amazing time sink in huge relations
Date: Wed, 20 Oct 1999 10:09:13 +0900
Message-ID: <000501bf1a97$b925a860$2801007e@cadzone.tpf.co.jp>
MIME-Version: 1.0
Content-Type: text/plain;
charset="iso-8859-1"
Content-Transfer-Encoding: 7bit
X-Priority: 3 (Normal)
X-MSMail-Priority: Normal
X-Mailer: Microsoft Outlook 8.5, Build 4.71.2173.0
X-Mimeole: Produced By Microsoft MimeOLE V4.72.2106.4
Importance: Normal
Sender: owner-pgsql-hackers@postgreSQL.org
Status: ORr
> -----Original Message-----
> From: Hiroshi Inoue [mailto:Inoue@tpf.co.jp]
> Sent: Tuesday, October 19, 1999 6:45 PM
> To: Tom Lane
> Cc: pgsql-hackers@postgreSQL.org
> Subject: RE: [HACKERS] mdnblocks is an amazing time sink in huge
> relations
>
>
> >
> > "Hiroshi Inoue" <Inoue@tpf.co.jp> writes:
>
> [snip]
>
> >
> > > Deletion is necessary only not to consume disk space.
> > >
> > > For example vacuum could remove not deleted files.
> >
> > Hmm ... interesting idea ... but I can hear the complaints
> > from users already...
> >
>
> My idea is only an analogy of PostgreSQL's simple recovery
> mechanism of tuples.
>
> And my main point is
> "delete fails after commit" doesn't harm the database
> except that not deleted files consume disk space.
>
> Of cource,it's preferable to delete relation files immediately
> after(or just when) commit.
> Useless files are visible though useless tuples are invisible.
>
Anyway I don't need "DROP TABLE inside transactions" now
and my idea is originally for that issue.
After a thought,I propose the following solution.
1. mdcreate() couldn't create existent relation files.
If the existent file is of length zero,we would overwrite
the file.(seems the comment in md.c says so but the
code doesn't do so).
If the file is an Index relation file,we would overwrite
the file.
2. mdunlink() couldn't unlink non-existent relation files.
mdunlink() doesn't call elog(ERROR) even if the file
doesn't exist,though I couldn't find where to change
now.
mdopen() doesn't call elog(ERROR) even if the file
doesn't exist and leaves the relation as CLOSED.
Comments ?
Regards.
Hiroshi Inoue
Inoue@tpf.co.jp
************
From pgsql-hackers-owner+M6267@hub.org Sun Aug 27 21:46:37 2000
Received: from hub.org (root@hub.org [216.126.84.1])
by candle.pha.pa.us (8.9.0/8.9.0) with ESMTP id UAA07972
for <pgman@candle.pha.pa.us>; Sun, 27 Aug 2000 20:46:36 -0400 (EDT)
Received: from hub.org (majordom@localhost [127.0.0.1])
by hub.org (8.10.1/8.10.1) with SMTP id e7S0kaL27996;
Sun, 27 Aug 2000 20:46:36 -0400 (EDT)
Received: from sss2.sss.pgh.pa.us (sss.pgh.pa.us [209.114.166.2])
by hub.org (8.10.1/8.10.1) with ESMTP id e7S05aL24107
for <pgsql-hackers@postgreSQL.org>; Sun, 27 Aug 2000 20:05:36 -0400 (EDT)
Received: from sss2.sss.pgh.pa.us (tgl@localhost [127.0.0.1])
by sss2.sss.pgh.pa.us (8.9.3/8.9.3) with ESMTP id UAA01604
for <pgsql-hackers@postgreSQL.org>; Sun, 27 Aug 2000 20:05:29 -0400 (EDT)
To: pgsql-hackers@postgreSQL.org
Subject: [HACKERS] Possible performance improvement: buffer replacement policy
Date: Sun, 27 Aug 2000 20:05:29 -0400
Message-ID: <1601.967421129@sss.pgh.pa.us>
From: Tom Lane <tgl@sss.pgh.pa.us>
X-Mailing-List: pgsql-hackers@postgresql.org
Precedence: bulk
Sender: pgsql-hackers-owner@hub.org
Status: ORr
Those of you with long memories may recall a benchmark that Edmund Mergl
drew our attention to back in May '99. That test showed extremely slow
performance for updating a table with many indexes (about 20). At the
time, it seemed the problem was due to bad performance of btree with
many equal keys, so I thought I'd go back and retry the benchmark after
this latest round of btree hackery.
The good news is that btree itself seems to be pretty well fixed; the
bad news is that the benchmark is still slow for large numbers of rows.
The problem is I/O: the CPU mostly sits idle waiting for the disk.
As best I can tell, the difficulty is that the working set of pages
needed to update this many indexes is too large compared to the number
of disk buffers Postgres is using. (I was running with -B 1000 and
looking at behavior for a 100000-row test table. This gave me a table
size of 3876 pages, plus 11526 pages in 20 indexes.)
Of course, there's only so much we can do when the number of buffers
is too small, but I still started to wonder if we are using the buffers
as effectively as we can. Some tracing showed that most of the pages
of the indexes were being read and written multiple times within a
single UPDATE query, while most of the pages of the table proper were
fetched and written only once. That says we're not using the buffers
as well as we could; the index pages are not being kept in memory when
they should be. In a query like this, we should displace main-table
pages sooner to allow keeping more index pages in cache --- but with
the simple LRU replacement method we use, once a page has been loaded
it will stay in cache for at least the next NBuffers (-B) page
references, no matter what. With a large NBuffers that's a long time.
I've come across an interesting article:
The LRU-K Page Replacement Algorithm For Database Disk Buffering
Elizabeth J. O'Neil, Patrick E. O'Neil, Gerhard Weikum
Proceedings of the 1993 ACM SIGMOD international conference
on Management of Data, May 1993
(If you subscribe to the ACM digital library, you can get a PDF of this
from there.) This article argues that standard LRU buffer management is
inherently not great for database caches, and that it's much better to
replace pages on the basis of time since the K'th most recent reference,
not just time since the most recent one. K=2 is enough to get most of
the benefit. The big win is that you are measuring an actual page
interreference time (between the last two references) and not just
dealing with a lower-bound guess on the interreference time. Frequently
used pages are thus much more likely to stay in cache.
It looks like it wouldn't take too much work to replace shared buffers
on the basis of LRU-2 instead of LRU, so I'm thinking about trying it.
Has anyone looked into this area? Is there a better method to try?
regards, tom lane
From prlw1@newn.cam.ac.uk Fri Jan 19 12:54:45 2001
Received: from henry.newn.cam.ac.uk (henry.newn.cam.ac.uk [131.111.204.130])
by candle.pha.pa.us (8.9.0/8.9.0) with ESMTP id MAA29822
for <pgman@candle.pha.pa.us>; Fri, 19 Jan 2001 12:54:44 -0500 (EST)
Received: from [131.111.204.180] (helo=quartz.newn.cam.ac.uk)
by henry.newn.cam.ac.uk with esmtp (Exim 3.13 #1)
id 14JfkU-0001WA-00; Fri, 19 Jan 2001 17:54:54 +0000
Received: from prlw1 by quartz.newn.cam.ac.uk with local (Exim 3.13 #1)
id 14Jfj6-0001cq-00; Fri, 19 Jan 2001 17:53:28 +0000
Date: Fri, 19 Jan 2001 17:53:28 +0000
From: Patrick Welche <prlw1@newn.cam.ac.uk>
To: Bruce Momjian <pgman@candle.pha.pa.us>
Cc: Tom Lane <tgl@sss.pgh.pa.us>, pgsql-hackers@postgreSQL.org
Subject: Re: [HACKERS] Possible performance improvement: buffer replacement policy
Message-ID: <20010119175328.A6223@quartz.newn.cam.ac.uk>
Reply-To: prlw1@cam.ac.uk
References: <1601.967421129@sss.pgh.pa.us> <200101191703.MAA25873@candle.pha.pa.us>
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
User-Agent: Mutt/1.2i
In-Reply-To: <200101191703.MAA25873@candle.pha.pa.us>; from pgman@candle.pha.pa.us on Fri, Jan 19, 2001 at 12:03:58PM -0500
Status: OR
On Fri, Jan 19, 2001 at 12:03:58PM -0500, Bruce Momjian wrote:
>
> Tom, did we ever test this? I think we did and found that it was the
> same or worse, right?
(Funnily enough, I just read that message:)
To: Bruce Momjian <pgman@candle.pha.pa.us>
cc: pgsql-hackers@postgreSQL.org
Subject: Re: [HACKERS] Possible performance improvement: buffer replacement policy
In-reply-to: <200010161541.LAA06653@candle.pha.pa.us>
References: <200010161541.LAA06653@candle.pha.pa.us>
Comments: In-reply-to Bruce Momjian <pgman@candle.pha.pa.us>
message dated "Mon, 16 Oct 2000 11:41:41 -0400"
Date: Mon, 16 Oct 2000 11:49:52 -0400
Message-ID: <26100.971711392@sss.pgh.pa.us>
From: Tom Lane <tgl@sss.pgh.pa.us>
X-Mailing-List: pgsql-hackers@postgresql.org
Precedence: bulk
Sender: pgsql-hackers-owner@hub.org
Status: RO
Content-Length: 947
Lines: 19
Bruce Momjian <pgman@candle.pha.pa.us> writes:
>> It looks like it wouldn't take too much work to replace shared buffers
>> on the basis of LRU-2 instead of LRU, so I'm thinking about trying it.
>>
>> Has anyone looked into this area? Is there a better method to try?
> Sounds like a perfect idea. Good luck. :-)
Actually, the idea went down in flames :-(, but I neglected to report
back to pghackers about it. I did do some code to manage buffers as
LRU-2. I didn't have any good performance test cases to try it with,
but Richard Brosnahan was kind enough to re-run the TPC tests previously
published by Great Bridge with that code in place. Wasn't any faster,
in fact possibly a little slower, likely due to the extra CPU time spent
on buffer freelist management. It's possible that other scenarios might
show a better result, but right now I feel pretty discouraged about the
LRU-2 idea and am not pursuing it.
regards, tom lane
From pgsql-hackers-owner+M3455@postgresql.org Fri Jan 19 13:18:12 2001
Received: from mail.postgresql.org (webmail.postgresql.org [216.126.85.28])
by candle.pha.pa.us (8.9.0/8.9.0) with ESMTP id NAA02092
for <pgman@candle.pha.pa.us>; Fri, 19 Jan 2001 13:18:12 -0500 (EST)
Received: from mail.postgresql.org (webmail.postgresql.org [216.126.85.28])
by mail.postgresql.org (8.11.1/8.11.1) with SMTP id f0JIFJ037872;
Fri, 19 Jan 2001 13:15:19 -0500 (EST)
(envelope-from pgsql-hackers-owner+M3455@postgresql.org)
Received: from sectorbase2.sectorbase.com ([208.48.122.131])
by mail.postgresql.org (8.11.1/8.11.1) with ESMTP id f0JI7V036780
for <pgsql-hackers@postgreSQL.org>; Fri, 19 Jan 2001 13:07:31 -0500 (EST)
(envelope-from vmikheev@SECTORBASE.COM)
Received: by sectorbase2.sectorbase.com with Internet Mail Service (5.5.2653.19)
id <DG1W4LRZ>; Fri, 19 Jan 2001 09:46:14 -0800
Message-ID: <8F4C99C66D04D4118F580090272A7A234D329F@sectorbase1.sectorbase.com>
From: "Mikheev, Vadim" <vmikheev@SECTORBASE.COM>
To: "'Tom Lane'" <tgl@sss.pgh.pa.us>, Bruce Momjian <pgman@candle.pha.pa.us>
Cc: pgsql-hackers@postgresql.org
Subject: RE: [HACKERS] Possible performance improvement: buffer replacemen
t policy
Date: Fri, 19 Jan 2001 10:07:27 -0800
MIME-Version: 1.0
X-Mailer: Internet Mail Service (5.5.2653.19)
Content-Type: text/plain;
charset="iso-8859-1"
Precedence: bulk
Sender: pgsql-hackers-owner@postgresql.org
Status: OR
> > Tom, did we ever test this? I think we did and found that
> > it was the same or worse, right?
>
> I tried it and didn't see any noticeable improvement on the particular
> test case I was using, so I got discouraged and didn't pursue the idea
> further. I'd like to come back to it someday, though.
I don't know how much useful could be LRU-2 but with WAL we should try
to reuse undirty free buffers first, not dirty ones, just to postpone
writes as long as we can. (BTW, this is what Oracle does.)
So, we probably should put new unfree dirty buffer just before first
dirty one in LRU.
Vadim