Re: 8.1beta3 performance - Mailing list pgsql-performance

From Tom Lane
Subject Re: 8.1beta3 performance
Date
Msg-id 26529.1130865883@sss.pgh.pa.us
Whole thread Raw
In response to Re: 8.1beta3 performance  (Jon Brisbin <jon.brisbin@npcinternational.com>)
List pgsql-performance
Jon Brisbin <jon.brisbin@npcinternational.com> writes:
> I have a stock FreeBSD 5.4 box that I put 8.1 on last night. I ran
> pgbench against it and my tps dropped from ~300tps in 8.0.3 to 20tps
> in 8.1. That's right. 20. No changes in any system configuration.

You sure about that last?  These numbers are kind of consistent with the
idea that fsync is off in the 8.0 database and on in the 8.1 database.

Using the same test case you mention (pgbench -s 10, -c 25 -t 1000),
I find that 8.1 is a bit faster than 8.0, eg

8.1 fsync off:
tps = 89.831186 (including connections establishing)
tps = 89.865065 (excluding connections establishing)

8.1 fsync on:
tps = 74.865078 (including connections establishing)
tps = 74.889066 (excluding connections establishing)

8.0 fsync off:
tps = 80.271338 (including connections establishing)
tps = 80.302054 (excluding connections establishing)

8.0 fsync on:
tps = 67.405708 (including connections establishing)
tps = 67.426546 (excluding connections establishing)

(All database parameters are defaults except fsync.)

These numbers are with assert-enabled builds, on a cheap PC whose drive
lies about write-complete, so they're not very representative of the
real world I suppose.  But I'm sure not seeing any 10x degradation.

            regards, tom lane

pgsql-performance by date:

Previous
From: Dave Page
Date:
Subject: Re: 8.1beta3 performance
Next
From: "Merlin Moncure"
Date:
Subject: improvise callbacks in plpgsql