Message-ID: <45749A0E.3010707@imageworks.com>
Date: Mon, 04 Dec 2006 13:58:38 -0800
From: Matt Chambers <chambers@imageworks.com>
Reply-To: chambers@imageworks.com
Organization: Sony Pictures Imageworks
User-Agent: Thunderbird 1.5.0.5 (X11/20060719)
MIME-Version: 1.0
To: pgsql-performance@postgresql.org
Subject: pgsql upgrade
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
Content-Transfer-Encoding: 7bit

Hi all, I've run into an issue with a Pg 7.4.6 to 8.1.5 upgrade along 
with hardware upgrade. 
I moved the database from a 2x 3.0ghz Xeon (512kb w/HT) to a 2x Opteron 
250.  The database is in
memory on a tmpfs partition. (the application can rebuild the db during 
total system failure)

When I first switched it over, the results were exactly what I expected. 
I was sustaining about
a 2.2 on the Xeon and with the Opteron, about a 1.5 with the same 
traffic. The box is highload,
it ouputs about 15mbps.

After a couple hours working perfectly the ' system' (vs user) load 
jumped from
a 3% of total CPU usage, to 30%, a 10x increase, and postgres started to 
write out
to data/base at a fairly fast rate.  The CPU context switch rate doubled 
at the same time.
Iowait, which was historically 0 on the 7.4 box, went to 0.08.

Strangely enough, a vacuum (not full or analyze) stopped postgres from 
writing to data/base
but the strange load pattern remains. (system is ~30% of the overall 
load, vs 3% before)

So, my question is, what happened, and how can I get it back to the same 
load pattern
7.4.6 had, and the same pattern I had for 4 hours before it went crazy?


Matt