PostgreSQL and Linux 2.6 kernel. - Mailing list pgsql-performance

From Gary Doades
Subject PostgreSQL and Linux 2.6 kernel.
Date
Msg-id 406C7956.17885.4B2DCD2@localhost
Whole thread Raw
In response to Re: select slow?  (Tom Lane <tgl@sss.pgh.pa.us>)
Responses Re: PostgreSQL and Linux 2.6 kernel.  (Tom Lane <tgl@sss.pgh.pa.us>)
Re: PostgreSQL and Linux 2.6 kernel.  ("Magnus Naeslund(t)" <mag@fbab.net>)
List pgsql-performance
As part of my ongoing evaluation of PostgreSQL I have been doing a little stress testing.
I though I would share an interesting result here..

Machine spec:
500 MHz PIII
256MB RAM
"old-ish" IDE HD (5400RPM)
Linux 2.4.22 kernel (Madrake 9.2)

I have PostgreSQL 7.4.1 installed and have managed to load up a 1.4 GB database
from MS SQLServer. Vaccum analyzed it.

As a test in PosgreSQL I issued a statement to update a single column of a table
containing 2.8 million rows with the values of a column in a table with similar rowcount.
Using the above spec I had to stop the server after 17 hours. The poor thing was
thrashing the hard disk and doing more swapping than useful work.

Having obtained a copy of Mandrake 10.0 with the 2.6 kernal I though I would give it a
go. Same hardware. Same setup. Same database loaded up. Same postgresql.conf file
to make sure all the settings were the same.  Vaccum analyzed it.

same update statement COMPLETED in 2 hours 50 minutes. I'm impressed.

I could see from vmstat that the system was achieving much greater IO thoughput than
the 2.4 kernel. Although the system was still swapping there seems to be a completely
different memory management pattern that suits PostgreSQL very well.

Just to see that this wasn't a coincidence I am repeating the test. It is now into the 14th
hour using the old 2.4 kernel. I'm going to give up.....

Has anyone else done any comparative testing with the 2.6 kernel?

Cheers,
Gary.


pgsql-performance by date:

Previous
From: Christopher Kings-Lynne
Date:
Subject: Re: What index for 'like (%keyword%)' ???
Next
From: Palle Girgensohn
Date:
Subject: single index on more than two coulumns a bad thing?