kvm migration fails with large-memory VMs

I just tried with the Hardy PPA KVM:84. It doesn't have this migration issue, and migration claims to go to completion, however:

1) The first time I tried migrating, after migration completed, the guest VM, also running Hardy, had done a fresh boot.

2) I tried migrating again, this time with a vncviewer attached to the destination. The vncviewer went away on migration completion, but when I started the vncviewer again, the guest VM, now on the destination, had panicked. See screenshot, attached, for what I was able to see of the kernel panic.

PS A minor nuisance with KVM:84, as packaged: if you don't specify script= in your -net options, kvm looks for /etc/kvm-ifup, which does not exist, rather than /etc/kvm/kvm-ifup.

I tried again to migrate with Hardy PPA KVM:84 and got a different oops. However, if I first stop the guest VM, and then migrate, it does work fine, and it's actually pretty quick, such that an ssh would see just a few seconds delay. So, offline migration does work, and isn't bad. I would love have live migration working properly, but the offline migration delay was brief enough that I'll probably cancel plans to experiment with Xen and just stick with KVM (the newer version in PPA).

One thing I didn't like with the migration is that the migrated guest takes the full amount of RAM the guest was allocated, rather than just the RAM the guest was using prior to the migration. Good thing I already bought more RAM, although maybe there's something I can do with the new memory ballooning feature...

The live migration crashes with KVM:84 are specific to SMP guests. This is already logged in the KVM bug database. I've seen mention that KVM:85 is soon to be release, but I do not know whether or not this bug is fixed.

Hi Brent-

Upstream indicates that this is solved in qemu 0.11, which is what we have in Karmic right now.

Any chance you can test this on a Karmic system running qemu-kvm 0.11?

With regard to your ifup/ifdown question, I'm uploading a fix right now for Bug #376387, which is that bug.

:-Dustin

I'm marking "won't fix" against kvm, since we won't be backporting this feature.

And I'm marking "fix released" against qemu-kvm, since upstream says that this fixed there. If you can reproduce this on an up-to-date karmic system, please move the qemu-kvm task back to "confirmed".

Thanks,
:-Dustin

On Tue, 8 Sep 2009, Dustin Kirkland wrote:

> Hi Brent-
>
> Upstream indicates that this is solved in qemu 0.11, which is what we
> have in Karmic right now.
>
> Any chance you can test this on a Karmic system running qemu-kvm 0.11?
>
> With regard to your ifup/ifdown question, I'm uploading a fix right now
> for Bug #376387, which is that bug.
>
> :-Dustin
>

The KVM 84 PPA you posted for Hardy fixed the memory issue (although live
migration of SMP systems is still broken; this is fixed somewhat after KVM
84, apparently). I don't have any Koala to test with (trying to hold out
for the next LTS release), but it appears to be the same version or
better, so I would expect that it resolves the issue with migration of
large-memory VMs.

Thanks,

Brent