Subject: Re: Query plan looks OK, but slow I/O - settings advice?
From: "Jeffrey W. Baker" <jwbaker@acm.org>
To: Roger Hand <RHand@kailea.com>
Cc: pgsql-performance@postgresql.org
In-Reply-To: 
 <DB28E9B548192448A4E8C8A3C1B1E475611C66@sj1-exch-01.us.corp.kailea.com>
References: 
 <DB28E9B548192448A4E8C8A3C1B1E475611C66@sj1-exch-01.us.corp.kailea.com>
Content-Type: text/plain
Date: Thu, 18 Aug 2005 23:55:35 -0700
Message-Id: <1124434535.3257.7.camel@noodles>
Mime-Version: 1.0
Content-Transfer-Encoding: 7bit

On Tue, 2005-08-16 at 10:46 -0700, Roger Hand wrote:
> The disks are ext3 with journalling type of ordered, but this was later changed to writeback with no apparent change in speed.
> 
> They're on a Dell poweredge 6650 with LSI raid card, setup as follows:
> 4 disks raid 10 for indexes (145GB) - sdc1
> 6 disks raid 10 for data (220GB) - sdd1
> 2 mirrored disks for logs - sdb1
> 
> stripe size is 32k
> cache policy: cached io (am told the controller has bbu)
> write policy: write-back
> read policy: readahead

I assume you are using Linux 2.6.  Have you considered booting your
machine with elevator=deadline?  You can also change this at runtime
using sysfs.

These read speeds are not too impressive.  Perhaps this is a slow
controller.  Alternately you might need bigger CPUs.

There's a lot of possibilities, obviously :)  I'd start with the
elevator, since that's easily tested.

-jwb