Re: [PATCH v2 net-next] virtio: Fix affinity for #VCPUs != #queue pairs

["Michael S. Tsirkin" <mst@redhat.com>]

On Wed, Feb 15, 2017 at 08:50:34AM -0800, Willem de Bruijn wrote:
> On Tue, Feb 14, 2017 at 1:05 PM, Michael S. Tsirkin <mst@redhat.com> wrote:
> > On Tue, Feb 14, 2017 at 11:17:41AM -0800, Benjamin Serebrin wrote:
> >> On Wed, Feb 8, 2017 at 11:37 AM, Michael S. Tsirkin <mst@redhat.com> wrote:
> >>
> >> > IIRC irqbalance will bail out and avoid touching affinity
> >> > if you set affinity from driver.  Breaking that's not nice.
> >> > Pls correct me if I'm wrong.
> >>
> >>
> >> I believe you're right that irqbalance will leave the affinity alone.
> >>
> >> Irqbalance has had changes that may or may not be in the versions bundled with
> >> various guests, and I don't have a definitive cross-correlation of irqbalance
> >> version to guest version.  But in the existing code, the driver does
> >> set affinity for #VCPUs==#queues, so that's been happening anyway.
> >
> > Right - the idea being we load all CPUs equally so we don't
> > need help from irqbalance - hopefully packets will be spread
> > across queues in a balanced way.
> >
> > When we have less queues the load isn't balanced so we
> > definitely need something fancier to take into account
> > the overall system load.
> 
> For pure network load, assigning each txqueue IRQ exclusively
> to one of the cores that generates traffic on that queue is the
> optimal layout in terms of load spreading. Irqbalance does
> not have the XPS information to make this optimal decision.

Try to add hints for it?

> Overall system load affects this calculation both in the case of 1:1
> mapping uneven queue distribution. In both cases, irqbalance
> is hopefully smart enough to migrate other non-pinned IRQs to
> cpus with lower overall load.

Not if everyone starts inserting hacks like this one in code.

> > But why the first N cpus? That's more or less the same as assigning them
> > at random.
> 
> CPU selection is an interesting point. Spreading equally across numa
> nodes would be preferable over first N. Aside from that, the first N
> should work best to minimize the chance of hitting multiple
> hyperthreads on the same core -- if all architectures lay out
> hyperthreads in the same way as x86_64.

That's another problem with this patch. If you care about hyperthreads
you want an API to probe for that.

-- 
MST