Re: [RFC PATCH 2/8] cpuidle: Prefer teo over menu governor

["Rafael J. Wysocki" <rafael@kernel.org>]

On Mon, Sep 30, 2024 at 6:12 PM Christian Loehle
<christian.loehle@arm.com> wrote:
>
> On 9/30/24 16:06, Rafael J. Wysocki wrote:
> > On Thu, Sep 5, 2024 at 11:27 AM Christian Loehle
> > <christian.loehle@arm.com> wrote:
> >>
> >> Since menu no longer has the interactivity boost teo works better
> >> overall, so make it the default.
> >>
> >> Signed-off-by: Christian Loehle <christian.loehle@arm.com>
>
> First of all thank you for taking a look.
>
> >
> > I know that this isn't strictly related to the use of iowait in menu,
> > but I'd rather wait with this one until the previous change in menu
> > settles down.
>
> Sure, I will look at any regressions that are reported, although "teo
> is performing better/worse/eqyal" would already be a pretty helpful hint
> and for me personally, if they do both perform badly I find debugging
> teo way easier.
>
> >
> > Also it would be good to provide some numbers to support the "teo
> > works better overall" claim above.
>
> Definitely, there are some in the overall cover-letter if you just
> compare equivalent menu/teo columns, but with the very fragmented
> cpuidle world this isn't anywhere near enough to back up that claim.
> We have found it to provide better results in both mobile and infra/
> server workloads on common arm64 platforms.

So why don't you add some numbers to the patch changelog?

If you can at least demonstrate that they are on par with each other
in some relevant benchmarks, then you can use the argument of teo
being more straightforward and so easier to reason about.

> That being said, I don't mind menu being around or even the default
> per-se, but would encourage anyone to give teo a try.

Fair enough.