Subject: Re: Xen and Raidframe
To: Geert Hendrickx <ghen@NetBSD.org>
From: Thor Lancelot Simon <tls@rek.tjls.com>
List: port-xen
Date: 11/04/2006 13:15:18
On Sat, Nov 04, 2006 at 05:02:12PM +0100, Geert Hendrickx wrote:
> On Sat, Nov 04, 2006 at 08:23:01AM -0500, Steven M. Bellovin wrote:
> > One apparent option, for the production DOMUs, is to have matching
> > partitions on each hard drive, and have the DOMUs do RAID themselves.  A
> > second is to use a large DOM0 partition, with RAID, and assign a DOMu a
> > file in that partition.  This strikes me as slower, but adequate for
> > light-load and experimental servers.
> 
> The third (and I think most-used) option is to run the software RAID in the
> dom0 and provide each domU with a block device on the RAID (e.g. raid0g).

I think this is a poor choice.  You are at the mercy of any of several kinds
of bugs in RAIDframe which might cause it to misbehave when its strategy
routing is called from xbd (I've seen more than one; vnd was specifically
retooled to be safe this way) and the configuration of your domain Us
is forever tied to the disklabels of your domain 0 disks.  What do you do if
you've filled up the partition table on your domain 0's RAID set but you
need to add another domain U?

There is little throughput penalty for using vnd to simply keep your domU
filesystems within a filesystem on the dom0 (which can, and should, be
made redundant with RAIDframe) and there is such a huge latency penalty for
using Xen at all (due to all the MMU manipulation required to get data to
domU) that the added latency of vnd is almost immaterial.

Thor