IRQ balancing within the hypervisor

IRQ balancing within the hypervisor

[Jan Beulich]

While pirq_guest_bind() makes a (lame) attempt at setting a new IRQ's
affinity reasonably, subsequent adjustments to the (supposedly)
handling vCPU's processor assignment as well as any domain specified
vCPU affinity changes don't have any effect on the IRQ's affinity, not
the least because there's neither a hypercall interface to allow the
handling domain to inform Xen of its (virtual) affinity, nor a scheduler
interface to adjust the affinity when the vCPU placement changes.

On large and/or heavily NUMA systems it may be quite desirable to
get this changed - are there any plans, or was this considered/
discussed in the past and rejected for some reason?

Thanks, Jan

Re: IRQ balancing within the hypervisor

[Keir Fraser]

On 31/05/2010 10:46, "Jan Beulich" <JBeulich@novell.com> wrote:

> While pirq_guest_bind() makes a (lame) attempt at setting a new IRQ's
> affinity reasonably, subsequent adjustments to the (supposedly)
> handling vCPU's processor assignment as well as any domain specified
> vCPU affinity changes don't have any effect on the IRQ's affinity, not
> the least because there's neither a hypercall interface to allow the
> handling domain to inform Xen of its (virtual) affinity, nor a scheduler
> interface to adjust the affinity when the vCPU placement changes.
> 
> On large and/or heavily NUMA systems it may be quite desirable to
> get this changed - are there any plans, or was this considered/
> discussed in the past and rejected for some reason?

Noone's got round to it. It would be good to have.

 -- Keir

> Thanks, Jan
> 
> 
> _______________________________________________
> Xen-devel mailing list
> Xen-devel@lists.xensource.com
> http://lists.xensource.com/xen-devel