Subject: Re: Low AAC performance but only when tested through the file system
To: Olaf Seibert <firstname.lastname@example.org>
From: Greg 'groggy' Lehey <grog@NetBSD.org>
Date: 12/05/2003 12:46:30
Content-Type: text/plain; charset=us-ascii
On Thursday, 4 December 2003 at 14:02:53 +0100, Olaf Seibert wrote:
> On Mon 01 Dec 2003 at 22:13:11 -0800, Jason Thorpe wrote:
>> You could probably get some improvement by tweaking your stripe depth
>> on the card. You really want to be writing an entire stripe at a time,
>> so setting your stripe depth to (64k / ndisks-1) for RAID-5 should do
>> the trick (64k because that is the maximum I/O size the NetBSD kernel
>> will currently issue).
>> NetBSD really needs to do some work tuning the file system to play nice
>> with RAID. In particular, tweaking the file system layout for the RAID
>> layout, and making sure to issue I/O that is stripe-aligned and
>> multiple-of-stripe-size would really improve the performance of the
>> file system on RAID-4 and RAID-5.
> So, I did some further experiments yesterday. I changed two things, each
> of which should improve the performance: I changed the stripe size to
> 16K, and I changed the geometry in the disklabel to 64 heads and 64
> sectors/track. This way the partitions, which are "cylinder"-aligned,
> ought to be better aligned than in the previous 255x63 arrangement.
> Strangely enough, performance was 25% *lower* than before!
That doesn't surprise me at all. My experiments have shown that
optimum stripe sizes are in the 256 kB to 512 kB range, though this
depends on the implementation, of course.
See complete headers for address and phone numbers.
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.2.0 (FreeBSD)
-----END PGP SIGNATURE-----