[pve-devel] Greatest common cluster cpuflags?
markc at unifiedgroup.com
Thu Sep 26 21:45:10 CEST 2013
Thanks for your reply.
I thought that in certain combinations there might be flags that both
physical cpus would support which were not included by the default
qemu64/kvm64 cpu. From the additional research I've done (and from
reading your message) this does not appear to be the case... and if
there are any such combinations they aren't worth the overhead of
regularly evaluating and comparing the flags available to the cpus in
So essentially I under-estimated how close to ideal the default
qemu64/kvm64 already is for most (or perhaps even all) mixed-brand clusters.
On 9/23/2013 10:49 PM, Alexandre DERUMIER wrote:
> If I remember, they are a qemu option at vm start, to test if cpuflags exists on host when starting, and stop the vm if a configured flag is missing.
> If you have homegenous cluster:
> -with exactly same processor : you can choose host
> -with only intel or only amd, but different generation : you can choose the lowest cpu model of the generaton
> -mix of amd and intel : you need to use qemu64/kvm64
> Now what do you want exactly ? only a check before migration ?
> ----- Mail original -----
> De: "Mark Casey" <markc at unifiedgroup.com>
> À: pve-devel at pve.proxmox.com
> Envoyé: Lundi 23 Septembre 2013 21:51:15
> Objet: [pve-devel] Greatest common cluster cpuflags?
> Hello everyone,
> I've read that considerable performance improvements can be achieved
> using a non-default kvm CPU with a VM due to the availability of
> additional flags and extension sets. However I understand that this can
> be troublesome for live-migration when cluster Nodes are not on
> homogenous hardware. My own environment specifies a specific CPU for one
> cluster that is homogenous (though I haven't benchmarked it against
> defaults), but I'm stuck with defaults for another cluster which is
> Intel/AMD-mixed and uses CPUs from two different generations.
> Would it be possible for Proxmox to track the available cpuflags of
> Nodes and store them in the shared cluster filesystem? I think this
> would allow for running VMs with a greater set of cpuflags than the
> default while reducing the chance of panic (or, whatever) during
> live-migration. It may also be possible to compare the flags that a VM
> was started with and warn if the user attempts to live-migrate it to a
> Node that was added since the VM was last started (a Node which may
> support fewer flags than the other Nodes, causing problems).
> I have some interest in playing with this and (I think) a decent idea of
> the scope of work required. Would this idea be considered simply
> impossible, not likely to be adopted/merged under any circumstances, or
> just currently too dangerous without extensive testing? I'm not looking
> for anyone to do my homework for me or write me back a novel...but I
> don't want to waste time if it for sure won't go anywhere. :D
> Thank you,
> pve-devel mailing list
> pve-devel at pve.proxmox.com
More information about the pve-devel