Horton

Re: [HACKERS] Deadlock in XLogInsert at AIX

Bernd Helmle <mailings@oopsware.de>

View thread View on postgr.es
From: Bernd Helmle <mailings@oopsware.de>
To: "REIX, Tony" <tony.reix@atos.net>, Michael Paquier <michael.paquier@gmail.com>, Noah Misch <noah@leadboat.com>
Cc: Heikki Linnakangas <hlinnaka@iki.fi>, Konstantin Knizhnik <k.knizhnik@postgrespro.ru>, PostgreSQL Hackers <pgsql-hackers@postgresql.org>, "OLIVA, PASCAL" <pascal.oliva@atos.net>, "EMPEREUR-MOT, SYLVIE" <sylvie.empereur-mot@atos.net>
Date: 2018-01-17T11:32:29Z
Lists: pgsql-hackers

Attachments

Am Dienstag, den 16.01.2018, 08:25 +0000 schrieb REIX, Tony:
> I've been able to compare PostgreSQL compiled with XLC vs GCC 7.1
> and, using times outputs provided by PostgreSQL tests, XLC seems to
> provide at least 8% more speed. We also plan to run professional
> performance tests in order to compare PostgreSQL 10.1 on AIX vs
> Linux/Power. I saw some 2017 performance slides, made with older
> versions of PostgreSQL and XLC, that show bad PostgreSQL performance
> on AIX vs Linux/Power, and I cannot believe it. We plan to
> investigate this.

I assume you are referring to the attached graph i've showed on
PGConf.US 2017 ?

The numbers we've got on that E850 machine (pgbench SELECT-only, scale
1000) weren't really good in comparison to Linux on the same machine.

We tried many options to make the performance better, overall the graph
shows the best performance from Linux *and* AIX with gcc. XL C We used
some knobs to get the best out on AIX:

export OBJECT_MODE=64; gcc -m64
ldedit -b forkpolicy:cor -b textpsize:64K -b datapsize:64K -b
stackpsize:64K postgres
export MALLOCOPTIONS=multiheap:16,considersize,pool,no_mallinfo
schedo -p -o vpm_fold_policy=4

There are many other things you can tune on AIX, but they didn't seem
to give the improvement we'd like to see.

Commits

  1. For all ppc compilers, implement compare_exchange and fetch_add with asm.

  2. For PowerPC instruction "addi", use constraint "b".

  3. For all ppc compilers, implement pg_atomic_fetch_add_ with inline asm.

  4. Replace xlc __fetch_and_add() with inline asm.

  5. Test pg_atomic_fetch_add_ with variable addend and 16-bit edge cases.