Date: Thu, 11 Aug 2005 10:07:04 -0400
From: Michael Stone <mstone+postgres@mathom.us>
Subject: Re: PG8 Tuning
In-reply-to: <1536.217.45.209.171.1123763001.squirrel@www.gradwell.com>
To: pgsql-performance@postgresql.org
Mail-Followup-To: pgsql-performance@postgresql.org
Message-id: <20050811140704.GL19080@mathom.us>
MIME-version: 1.0
Content-type: text/plain; charset=us-ascii; format=flowed
Content-disposition: inline
References: <1536.217.45.209.171.1123763001.squirrel@www.gradwell.com>
User-Agent: Mutt/1.5.9i

On Thu, Aug 11, 2005 at 01:23:21PM +0100, Paul Johnson wrote:
>I'm guessing that this is because pg_xlog has gone from a 9 spindle LUN to
>a single spindle disk?
>
>In cases such as this, where an external storage array with a hardware
>RAID controller is used, the normal advice to separate the data from the
>pg_xlog  seems to come unstuck

Yes. That's the downside to dogma. If you're writing pg_xlog to a
battery-backed ram buffer you'll see faster commits than you will with a
write to a disk, even if you've got a dedicated spindle, unless you've
got constant write activity. (Because once the buffer fills you're
limited to disk speed as you wait for buffer flushes.) If you've got a
lot of system RAM, a battery-backed disk buffer, an OS/filesystem than
effectively delays writes, and bursty transactional writes it's quite
possible you'll get better performance putting everything on one array
rather than breaking it up to follow the "rules". You might get a
performance boost by putting the transaction log on a seperate partition
or lun on the external array, depending on how the fs implements syncs
or whether you can optimize the filsystem choice for each partition. The
correct approach is to run comparative benchmarks of each configuration.
:-)

Mike Stone