Re: [PATCH] cpuidle: use high confidence factors only when considering polling

[Rik van Riel <riel@redhat.com>]

[-- Attachment #1: Type: text/plain, Size: 1488 bytes --]

On Sat, 2016-03-19 at 03:06 +0100, Rafael J. Wysocki wrote:
> On Sat, Mar 19, 2016 at 2:53 AM, Rik van Riel <riel@redhat.com>
> wrote:
> > 
> > The kernel test robot also showed over 6% improvement in (IIRC)
> > the wikibench test, with polling being done a little more often.
> > 
> > Other parts of the kernel do pretty much anything for 2% extra
> > performance, so if we can find a way to preserve that 6%, without
> > hurting power usage significantly, I suspect it may be worthwhile.
> Well. OK.  That is a worthy goal, but things in this thread look a
> bit
> too ad-hoc to me to be honest.
> 
> I'd like to take one step at a time and in such a way that
> improvements are actually confirmed in every step.
> 
> Does this sound reasonable?
> 
> If so, I first would like to restore the energy consumption in the
> Doug's case at least to the level from before commit a9ceb78bc75c
> (cpuidle,menu: use interactivity_req to disable polling).  That's
> what
> I'd like to do for 4.6 and possibly before the merge window is over.
> 
> After 4.6-rc1, in the 4.7 cycle, we can think about improving
> performance here and maybe sacrificing some energy for that, but in a
> controlled way as I said above.
> 
> What do you think?

That sounds perfectly reasonable to me.

Count me in for any efforts to improve performance with
the cpuidle code, and hopefully allow more people to enable
c-states on their servers.

-- 
All Rights Reversed.


[-- Attachment #2: This is a digitally signed message part --]
[-- Type: application/pgp-signature, Size: 473 bytes --]