To: Gavin Hamill <gdh@laterooms.com>
Cc: pgsql-performance@postgresql.org
Subject: Re: pg 8.1.3, AIX, huge box, painfully slow.
References: <20060407185823.abe93017.gdh@laterooms.com>
	<15233.1144435299@sss.pgh.pa.us>
	<20060407204526.197afc19.gdh@laterooms.com>
In-Reply-To: <20060407204526.197afc19.gdh@laterooms.com>
From: Greg Stark <gsstark@mit.edu>
Organization: The Emacs Conspiracy; member since 1992
Date: 07 Apr 2006 16:30:15 -0400
Message-ID: <871ww9m8mg.fsf@stark.xeocode.com>
Lines: 20
User-Agent: Gnus/5.09 (Gnus v5.9.0) Emacs/21.4
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii


Gavin Hamill <gdh@laterooms.com> writes:

> This is one hell of a moving target and I can't help but think I'm just
> missing something that's right in front of my nose, too close to see.

I'm assuming you compiled postgres yourself? Do you have the output from the
configure script? I'm wondering if it failed to find a good spinlock match for
the architecture. Not sure if that's really likely but it's a possibility.

Also, I'm pretty sure IBM has tools that would let you disable some of the
processors to see if maybe it's a shared memory bus issue. If I understand you
right the machine isn't in production yet? In which case I would get timing
information for a single processor, two processors, four processors, and eight
processors. If you see it max out and start dropping then that would point
towards a hardware/low level postgres issue like spinlocks or shared memory
rather than a high level database issue like stats.

-- 
greg