openstack team mailing list archive

["Daniel P. Berrange" <berrange@xxxxxxxxxx>]

On Wed, Aug 08, 2012 at 09:50:20AM +0800, Huang Zhiteng wrote:
> > But to the contrary. I tested live-migrate (without block migrate)
> > last night using a guest with 8GB RAM (almost fully committed) and
> > lost any access/contact with the guest for over 4 minutes - it was
> > paused for the duration. Not something I'd want to do to a user's
> > web-server on a regular basis...
> 
> 4 minutes of pause (down time)?  That's way too long.  Even there was
> crazy memory intensive workload inside the VM being migrated, the
> worst case is KVM has to pause VM and transmit all 8 GB memory (all
> memory are dirty, which is very rare).  If you have 1GbE link between
> two host, that worst case pause period (down time) is less than 2
> minutes.  My previous experience is: the down time for migrating one
> idle (almost no memory access) 8GB VM via 1GbE is less than 1 second;
> the down time for migrating a 8 GB VM that page got dirty really
> quickly is <60 seconds.  FYI.

KVM has a tunable setting for the maximum allowable live migration
downtime, which IIRC defaults to something very small like 250ms.

If the migration can't be completed within this downtime limit,
KVM will simply never complete migration. Given that Nova does
not tune this downtime setting, I don't see how you'd see 4 mins
downtime unless it was not truely live migration, or there was
something else broken (eg the network bridge device had a delay
inserted by the STP protocol which made the VM /appear/ to be
unreponsive on the network even though it was running fine).

Regards,
Daniel
-- 
|: 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 :|