Re: [PATCH v5] KVM: halt-polling: poll for the upcoming fire timers

From: Paolo Bonzini
Date: Thu May 26 2016 - 06:30:28 EST

On 26/05/2016 12:26, Paolo Bonzini wrote:
> As discussed on IRC, I would like to understand why the adaptive
> adjustment of halt_poll_ns is failing. It seems like you have so few
> halts that you don't get halt_poll_ns>0. Yet, when the VM halts, it's
> very close to the timer tick---often enough for this patch to have an
> effect.
> Please send a trace of halt_poll_ns_grow and halt_poll_ns_shrink
> tracepoints, so that we can find out more about this.

And 30 seconds after I wrote this email, you told me on IRC that the
guest had HZ=1000 and the module parameter was set to 1 ms in order to
_really_ benefit from the patch. So basically you could obtain the same
effect with idle=poll in the guest.

This explains why your reported results were not so great (as David noted).