Date: Mon, 14 Apr 2008 11:42:50 -0400 (EDT)
From: Greg Smith <gsmith@gregsmith.com>
To: Gaetano Mendola <mendola@gmail.com>
cc: "pgsql-performance@postgresql.org" <pgsql-performance@postgresql.org>
Subject: Re: shared_buffers performance
In-Reply-To: <48032021.90704@gmail.com>
Message-ID: <Pine.GSO.4.64.0804141132580.3587@westnet.com>
References: <48032021.90704@gmail.com>
MIME-Version: 1.0
Content-Type: TEXT/PLAIN; charset=US-ASCII; format=flowed

On Mon, 14 Apr 2008, Gaetano Mendola wrote:

> I'm using postgres 8.2.3 on Red Hat compiled with GCC 3.4.6.

8.2.3 has a performance bug that impacts how accurate pgbench results are; 
you really should be using a later version.

> http://img84.imageshack.us/my.php?image=totalid7.png
> as you can see using 64MB as value for shared_buffers I'm obtaining 
> better results.

I'm assuming you've read my scaling article at 
http://www.westnet.com/~gsmith/content/postgresql/pgbench-scaling.htm 
since you're using the graph template I suggest there.

If you look carefully at your results, you are getting better results for 
higher shared_buffers values in the cases where performance is memory 
bound (the lower scale numbers).  Things reverse so that more buffers 
gives worse performance only when your scale >100.  I wouldn't conclude 
too much from that.  The pgbench select test is doing a low-level 
operation that doesn't benefit as much from having more memory available 
to PostgreSQL instead of the OS as a real-world workload will.

--
* Greg Smith gsmith@gregsmith.com http://www.gregsmith.com Baltimore, MD