Virtualization and HotPlugging/swapping

[Daniel P. Berrange]

On Fri, Jan 21, 2011 at 12:34:29PM -0600, Renich Bon Ciric wrote:
> On Fri, Jan 21, 2011 at 5:04 AM, Daniel P. Berrange <berrange at redhat.com> wrote:
> > That is correct. The live memory adjustment for KVM/QEMU via
> > libvirt is ballooning, not memory hotplug. Likewise for Xen
> > and VMWare. Real memory hotplug is a future RFE.
> >
> >> I've read some stuff here:
> >> http://www.linux-kvm.org/page/CPUHotPlug
> >
> > Yep, CPU hotplug is available for KVM/libvirt, but support among
> > common guest OS is flakey. It doesn't work at all for Windows,
> > and for Linux only hotplug works, but not unplug. Xen however
> > did allow hot-unplug because it was using a different mechanism
> > IIUC.
> >
> > For KVM, HDD hotplug already works for virtio disks, or SCSI
> > disks, or USB disks. It doesn't work for IDE disks. For Xen
> > hotplug works for xen paravirt disks.
> >
> > For KVM, NIC hotplug also works, and in general any kind of
> > emulated PCI device can be hotplugged. For Xen, again their
> > xen paravirt NICs could be hotplugged.
> 
> So, in practical terms,
> 
> RAM "extension" is done by ballooning

NB, you can't ever extend RAM beyond initial boot amount with
ballooning. It merely lets you adjust it to any value below
the initial boot amount, and back up.

> CPU hotplugging is possible but not unplugging.
> And HDD hotplugging is possible...
> 
> I read a thread (2008) that offered a patch on this ACPI technique for
> RAM HotPlugging. Check it out:
> http://www.spinics.net/lists/kvm/msg27395.html

Yep, not merged though. It will likely come eventually
but for now the focus has been in other areas of the
RAM subsystem like transparent huge pages and KSM
page merging which have a far greater benefit.

> 
> CPU UnPlugging is not considered as a needed feature is it?
> http://www.linux-kvm.org/page/CPUHotPlug
> "Hot-unplug does not work, and actually, does not even make that much sense."

I wouldn't neccessarily agree with the statement on
that page. I think it is perfectly valid to be able
to want to unplug CPUs. If you have a 4 CPU guest on
a 4 CPU host, and need to migrate the guest to a 2
CPU host, then it could well be a better idea to
unplug 2 of the guest CPUs than leave it overcommitted.
Thus I expect someone will make hot-unplug work eventually

Daniel