Subject: Re: RaidFrame poor performance
To: Matthias Scheler <email@example.com>
From: Greg Oster <firstname.lastname@example.org>
Date: 01/26/2005 21:18:16
Matthias Scheler writes:
> In article <41F7C157.email@example.com>,
> Jochen Keil <J.Keil@gmx.de> writes:
> > The iozone benchmark are very interesting (at least for me). A 3 disk R5
> > setup (with two usable disks) set up with a correct stripe/fs block size
> > is about 4 times faster than a 4 disk R5 (3 usable disks) setup.
> What's surprising about that? On a 3 disk RAID 5 you need 3 reads and
> 3 writes for 1 write to the RAID. On a 4 disk RAID 5 you need 4 of them.
No. You wouldn't even need that many reads and writes if these sets were
in degraded mode. For the 4-disk RAID 5 set, the worst you'll do in
non-degraded mode for a write less than 1 component worth will be 2 reads
and 2 writes, and in degraded mode it would be at worst 3 reads and 2 writes.
Of course, to get good performance, one wants to be able to skip those
"reads" altogether, and that's much easier to do for a 3 or 5-disk
RAID 5 set (where a "power-of-two-sized" IO is easier to divide by 2 or
4) than it is for a 4-disk RAID 5 set (where a "power-of-two-sized"
IO is tricky to divide evenly by 3...)