This is an archived copy of the Xen.org mailing list, which we have preserved to ensure that existing links to archives are not broken. The live archive, which contains the latest emails, can be found at http://lists.xen.org/
Home Products Support Community News


Re: [Xen-users] Xen and iSCSI


Am Samstag, 28. Januar 2006 21:46 schrieb Javier Guerra:
> On Saturday 28 January 2006 1:18 pm, Markus Hochholdinger wrote:
> > so the cpu has to work a lot for iscsi. But should performance be better
> > in dom0 than in domU? I'm planning to use iscsi in domU, if possible.
> if you do heavy network in domU, either it has to pass through the dom0
> kernel, or you have to dedicate a NIC to each domU. so, i think it could be
> slower (or more CPU demanding) to do the initiator in domU than dom0.

yes, that's also what my tests show.

> > > what i plan to do (once i get a few test boxes) is to use coraid's AoE
> > > protocol (the driver is in the main Linux tree).  it's designed to be
> > > software driven, without any special card; and it's CPU utilization
> > > seems to be around ATA levels.
> > I also read about this. My thoughts were that iscsi exists longer and is
> > a standard. AoE is no standard and can only be used by linux by now (i
> > think). You also can make iscsi targets and initiators without any
> > special hardware.
> all very true; but software-only iSCSI is a _real_ cpu hog.  i've read
> about 40-50% CPU usage for 1Gb on 3GHz P4 (a couple of years ago, might
> have improved with better FSBs or PCI busses)

Hm, i've not found anything about cpu load with software iscsi. Maybe i'll 
test this.

> most 1Gb NICs do some offloading, but only iSCSI ones manage the SCSI part.
> and with the paravirtualized network drivers, i think you can't use either
> TCP or IP processing in the card; at best the card will do only the
> Ethernet frame checksum (i think even 100BaseT NICs do that).  i wouldn't
> be surprised if iSCSI on domU uses twice as much CPU than on dom0 (where it
> would use the full offloading capabilities of the NIC)

Yes, ack.

> > > with any kind of SAN, i think the best strategy would be to use CLVM in
> > > dom0 and export the logical volumes to the domUs; this way you only
> > > have to setup the SAN drivers in dom0, the domUs would treat those as
> > > local disks.
> > Is this safe for migration? How can you get the block device, which is
> > exported to the domU, to other Xen hosts when migrating an domU? If all
> with CLVM you can get several boxes looking at the same SAN drives and
> seeing the same LVs.  of course you can't mount a filesystem on two boxes
> (unless it's a cluster filesystem, like GFS); but its perfectly safe to use
> a LV for a VM on box A, then migrate to box B, where it can find the same
> LV.  at no point in time the LV is used from two different machines (real
> or virtual) at the same time.

Yes. This was a misunderstanding for my part. I thought this is not possible, 
but as i see the responds here it is possible to make it so.
Just the block devices exportet in dom0 to domU have to be equal on each Xen 
host (each dom0).
So i will think over my ideas.



Attachment: pgplSfsTxxqJy.pgp
Description: PGP signature

Xen-users mailing list
<Prev in Thread] Current Thread [Next in Thread>