[openstack-dev] [nova][libvirt] RFC: ensuring live migration ends
vladik.romanovsky at enovance.com
Sat Jan 31 02:55:23 UTC 2015
----- Original Message -----
> From: "Daniel P. Berrange" <berrange at redhat.com>
> To: openstack-dev at lists.openstack.org, openstack-operators at lists.openstack.org
> Sent: Friday, 30 January, 2015 11:47:16 AM
> Subject: [openstack-dev] [nova][libvirt] RFC: ensuring live migration ends
> In working on a recent Nova migration bug
> I had cause to refactor the way the nova libvirt driver monitors live
> migration completion/failure/progress. This refactor has opened the
> door for doing more intelligent active management of the live migration
> As it stands today, we launch live migration, with a possible bandwidth
> limit applied and just pray that it succeeds eventually. It might take
> until the end of the universe and we'll happily wait that long. This is
> pretty dumb really and I think we really ought to do better. The problem
> is that I'm not really sure what "better" should mean, except for ensuring
> it doesn't run forever.
> As a demo, I pushed a quick proof of concept showing how we could easily
> just abort live migration after say 10 minutes
> There are a number of possible things to consider though...
> First how to detect when live migration isn't going to succeeed.
> - Could do a crude timeout, eg allow 10 minutes to succeeed or else.
> - Look at data transfer stats (memory transferred, memory remaining to
> transfer, disk transferred, disk remaining to transfer) to determine
> if it is making forward progress.
I think this is a better option. We could define a timeout for the progress
and cancel if there is no progress. IIRC there were similar debates about it
in Ovirt, we could do something similar:
> - Leave it upto the admin / user to decided if it has gone long enough
> The first is easy, while the second is harder but probably more reliable
> and useful for users.
> Second is a question of what todo when it looks to be failing
> - Cancel the migration - leave it running on source. Not good if the
> admin is trying to evacuate a host.
> - Pause the VM - make it complete as non-live migration. Not good if
> the guest workload doesn't like being paused
> - Increase the bandwidth permitted. There is a built-in rate limit in
> QEMU overridable via nova.conf. Could argue that the admin should just
> set their desired limit in nova.conf and be done with it, but perhaps
> there's a case for increasing it in special circumstances. eg emergency
> evacuate of host it is better to waste bandwidth & complete the job,
> but for non-urgent scenarios better to limit bandwidth & accept failure ?
> - Increase the maximum downtime permitted. This is the small time window
> when the guest switches from source to dest. To small and it'll never
> switch, too large and it'll suffer unacceptable interuption.
In my opinion, it would be great if we could play with bandwidth and downtime
before cancelling the migration or pausing.
However, It makes sense only if there is some kind of a progress in the transfer
stats and not a complete disconnect. In that case we should just cancel it.
> We could do some of these things automatically based on some policy
> or leave them upto the cloud admin/tenant user via new APIs
> Third there's question of other QEMU features we could make use of to
> stop problems in the first place
> - Auto-converge flag - if you set this QEMU throttles back the CPUs
> so the guest cannot dirty ram pages as quickly. This is nicer than
> pausing CPUs altogether, but could still be an issue for guests
> which have strong performance requirements
> - Page compression flag - if you set this QEMU does compression of
> pages to reduce data that has to be sent. This is basically trading
> off network bandwidth vs CPU burn. Probably a win unless you are
> already highly overcomit on CPU on the host
> Fourth there's a question of whether we should give the tenant user or
> cloud admin further APIs for influencing migration
> - Add an explicit API for cancelling migration ?
> - Add APIs for setting tunables like downtime, bandwidth on the fly ?
> - Or drive some of the tunables like downtime, bandwidth, or policies
> like cancel vs paused from flavour or image metadata properties ?
> - Allow operations like evacuate to specify a live migration policy
> eg switch non-live migrate after 5 minutes ?
IMHO, an explicit API for cancelling migration is very much needed.
I remember cases when migrations took about 8 or hours, leaving the
admins helpless :)
Also, I very much like the idea of having tunables and policy to set
in the flavours and image properties.
To allow the administrators to set these as a "template" in the flavour
and also to let the users to update/override or "request" these options
as they should know the best (hopefully) what is running in their guests.
> The current code is so crude and there's a hell of alot of options we
> can take. I'm just not sure which is the best direction for us to go
> What kind of things would be the biggest win from Operators' or tenants'
> POV ?
Thanks for writing this! :)
> |: http://berrange.com -o- http://www.flickr.com/photos/dberrange/ :|
> |: http://libvirt.org -o- http://virt-manager.org :|
> |: http://autobuild.org -o- http://search.cpan.org/~danberr/ :|
> |: http://entangle-photo.org -o- http://live.gnome.org/gtk-vnc :|
> OpenStack Development Mailing List (not for usage questions)
> Unsubscribe: OpenStack-dev-request at lists.openstack.org?subject:unsubscribe
More information about the OpenStack-dev