Message-ID: <42A639C3.40606@samurai.com>
Date: Wed, 08 Jun 2005 10:20:19 +1000
From: Neil Conway <neilc@samurai.com>
User-Agent: Debian Thunderbird 1.0.2 (X11/20050602)
MIME-Version: 1.0
To: Tom Arthurs <tarthurs@jobflash.com>
Cc: Donald Courtney <Donald.Courtney@Sun.COM>,
 pgsql-performance@postgresql.org
Subject: Re: Postgresql on an AMD64 machine
References: <1118080420.1644.23.camel@penguin1>
	<b918cf3d05060613593a73a18a@mail.gmail.com>
	<42A5F7B6.20502@sun.com> <42A5FB70.8090807@jobflash.com>
	<42A6014C.30108@sun.com> <42A605E8.8040607@jobflash.com>
In-Reply-To: <42A605E8.8040607@jobflash.com>
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
Content-Transfer-Encoding: 7bit

Tom Arthurs wrote:
> Yes, shared buffers in postgres are not used for caching

Shared buffers in Postgres _are_ used for caching, they just form a 
secondary cache on top of the kernel's IO cache. Postgres does IO 
through the filesystem, which is then cached by the kernel. Increasing 
shared_buffers means that less memory is available for the kernel to 
cache IO -- increasing shared_buffers has been shown to be a net 
performance loss beyond a certain point. Still, there is value in 
shared_buffers as it means we can avoid a read() system call for hot 
pages. We can also do better buffer replacement in the PG shared buffer 
than the kernel can do (e.g. treating IO caused by VACUUM specially).

> My biggest challenge with solaris/sparc is trying to reduce context
> switching.

It would be interesting to see if this is improved with current sources, 
as Tom's bufmgr rewrite should have hopefully have reduced this problem.

-Neil