Message-ID: <3D797934.7060001@joeconway.com>
Date: Fri, 06 Sep 2002 20:57:40 -0700
From: Joe Conway <mail@joeconway.com>
User-Agent: Mozilla/5.0 (Windows; U; Windows NT 5.0; en-US;
	rv:1.1) Gecko/20020826
MIME-Version: 1.0
To: Josh Berkus <josh@agliodbs.com>
Cc: pgsql-sql@postgresql.org
Subject: Re: Hardware performance for large updates
References: <web-1635967@davinci.ethosmedia.com>
Content-Type: text/plain; charset=us-ascii; format=flowed
Content-Transfer-Encoding: 7bit

Josh Berkus wrote:
> Particularly, the difficulty is that this application gets many small
> requests during the day (100 simultaneous uses) and shares a server
> with Apache.   So I have to be concerned about how much memory each
> connection soaks up, during the day.   At night, the maintainence tasks
> run a few, really massive procedures.
> 
> So I should probably restart Postgres with different settings at night,
> hey?

Actually, if you can afford the twice daily changes, it sounds like a 
great idea. I think you can get new conf settings to take by sending a 
SIGHUP to the postmaster, so you don't even really need any downtime to 
do it. Yup, here it is:
   http://www.postgresql.org/idocs/index.php?runtime-config.html

>>I do think you probably could increase Shared Buffers, as 256 is
>>pretty small. There's been a lot of debate over the best setting. The
>>usual guidance is start at 25% of physical RAM (16384 == 128MB if you
>>have 512MB RAM), then tweak to optimize performance for your
>>application and hardware. 
> 
> 
> Hmmm... how big is a shared buffer, anyway?   I'm having trouble
> finding actual numbers in the docs.

By default it is 8K. It's mentioned here:
   http://www.postgresql.org/idocs/index.php?kernel-resources.html
So, as I mentioned above, Shared Buffers of 16384 == 128MB if you have a 
default 8K block size.


>>Are you on a Linux server -- if so I found that
>>fdatasync works better than (the default) fsync for wal_sync_method.
> 
> Yes, I am.   Any particular reason why fdatasync works better?

I can't remember the technical reason (although I've seen one on the 
list before), but I have determined it empirically true, at least for my 
setup. Ahh, here we go:
   http://archives.postgresql.org/pgsql-hackers/1998-04/msg00326.php

Joe