Open Stack

Thanks John. Yes, I also think that this should be a bp as it is going to
make some changes to enable live migration with only one nova compute, will
file a blueprint later.

For your proposal "specify the same host as the instance", this can resolve
the issue of live migration with target host, but what about the case of
live migration without target host? If we still allow "specify the same
host as the instance", the the live migration will goes to dead loop.

So it seems we definitely need to find a way to specify the node for live
migration, hope someone else can show some light here.

Of course, I will file bp and go through the new bp review process for this
feature.

Thanks!


2014-03-31 21:02 GMT+08:00 John Garbutt <john at johngarbutt.com>:

> On 31 March 2014 10:11, Jay Lau <jay.lau.513 at gmail.com> wrote:
> > Hi,
> >
> > Currently with VMWare VCDriver, one nova compute can manage multiple
> > clusters/RPs, this caused cluster admin cannot do live migration between
> > clusters/PRs if those clusters/PRs managed by one nova compute as the
> > current live migration logic request at least two nova computes.
> >
> > A bug [1] was also filed to trace VMWare live migration issue.
> >
> > I'm now trying the following solution to see if it is acceptable for a
> fix,
> > the fix wants enable live migration with one nova compute:
> > 1) When live migration check if host are same, check both host and node
> for
> > the VM instance.
> > 2) When nova scheduler select destination for live migration, the live
> > migration task should put (host, node) to attempted hosts.
> > 3) Nova scheduler needs to be enhanced to support ignored_nodes.
> > 4) nova compute need to be enhanced to check host and node when doing
> live
> > migration.
> >
> > I also uploaded a WIP patch [2] for you to review the idea of the fix and
> > hope can get some comments from you.
> >
> > [1] https://bugs.launchpad.net/nova/+bug/1192192
> > [2] https://review.openstack.org/#/c/84085
>
> Long term, finding a way to unify how cells and the VMware driver
> manages multiple hosts, seems like the best way forward. It would be a
> shame for this API to be different between cells and VMware, although
> right now, that might not work too well :(
>
> A better short term fix, might be to allow you to specify the same
> host as the instance, and the scheduling of the node could be
> delegated to the VMware driver, which might just delegate that to
> vCenter. I assume we still need some way to specify the node, and I
> can't immediately think of a good way forward.
>
> I feel this should really be treated as a blueprint, and go through
> the new blueprint review process. That should help decide the right
> approach to take.
>
> John
>
> _______________________________________________
> OpenStack-dev mailing list
> OpenStack-dev at lists.openstack.org
> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>



-- 
Thanks,

Jay
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openstack.org/pipermail/openstack-dev/attachments/20140331/01b018d5/attachment.html>