xfs
[Top] [All Lists]

Re: poor io performance with xfs+raid5

To: Mike Eldridge <diz@xxxxxxxxx>
Subject: Re: poor io performance with xfs+raid5
From: Thor Lancelot Simon <tls@xxxxxxxxxxxx>
Date: Fri, 26 Apr 2002 07:30:46 -0400
Cc: linux-xfs@xxxxxxxxxxx
In-reply-to: <20020425154135.B14120@ornery.cafes.net>; from diz@cafes.net on Thu, Apr 25, 2002 at 03:41:35PM -0500
References: <Pine.BSO.4.44.0204251216360.25324-100000@paperboy.websocietyinc.com> <3CC85999.501431E5@ch.sauter-bc.com> <20020425144025.N16048@ornery.cafes.net> <1019765945.12905.102.camel@jen.americas.sgi.com> <20020425154135.B14120@ornery.cafes.net>
Sender: owner-linux-xfs@xxxxxxxxxxx
User-agent: Mutt/1.2.5i
On Thu, Apr 25, 2002 at 03:41:35PM -0500, Mike Eldridge wrote:
> On Thu, Apr 25, 2002 at 03:19:05PM -0500, Steve Lord wrote:
> > Well, Duh! I should have seen that first time around, I get into the
> > habit of reading my email too fast!
> > 
> > We may be able to fix some things, if we can remake the filesystem.
> > First you need to know the stripe unit of your raid - we can feed
> > this into XFS to make it do stripe aligned allocations. This has
> > to be done by hand on linux. Take a look at the mkfs.xfs man page
> > and the section on sunit and swidth options. Probably bump your log size
> > up from the default somewhat, not sure how it ended up as 1839
> > that is scary.
> 
> RAID5 on this card offers only a 64K stripe size.  however, i will be

And is also notorious for mediocre performance.  It appears to have been
implemented largely as a checkbox feature.

Doing unaligned writes to a hardware RAID device can still be a problem,
if the hardware RAID device can't at least cache enough data to coalesce
writes up to its stripe size; one does, from time to time, see cheap RAID
hardware in PCs that has exactly this problem, though IIRC there is another
cause for the Escalade's crummy RAID5 performance.

Thor


<Prev in Thread] Current Thread [Next in Thread>