Re: [PATCH RFC 4/4] sched: cpufreq: Apply DVFS headroom to CFS only

[Qais Yousef <qyousef@layalina.io>]

On 08/21/23 18:41, Dietmar Eggemann wrote:
> On 20/08/2023 23:06, Qais Yousef wrote:
> > RT and Deadline have exact performance requirement when running. RT runs
> > at max or a specific OPP defined by uclamp_min. Deadline's OPP is
> > defined by its bandwidth. Both of which are known ahead of time and
> > don't require a headroom to grow into.
> > 
> > IRQs on the other hand have no specific performance requirement and
> > cruises along at whatever the current OPP happens to be when they occur.
> > 
> > Now they all have PELT pressure signals that does impact frequency
> > selection and task placement. The question is do they need DVFS
> > headroom?
> > 
> > I think the answer is no because when CFS is not running at all, these
> > pressure signal has no real impact on performance for RT, DL or IRQ.
> > 
> > If CFS util is not zero, we already add their pressure as an
> > *additional* headroom to account for the lost/stolen time. So I argue
> > that the pressure are headroom themselves and shouldn't need an
> > additional DVFS headroom applied on top.
> > 
> > In summary final outcome should be:
> > 
> > 	CFS + DVFS headroom + (RT, DT, IRQ) pressure headroom
> 
> I assume here you want to align the difference that EAS deals with

This function is used on all systems that use schedutil - EAS being one of them
but not the only one. The definition isn't, and shouldn't, be tied to EAS.
I'm certainly intending this change for all possible users of schedutil.

> `util_cfs` vs `capacity` whereas power deals with `util` vs
> `capacity_orig`? You want that power should only apply the 1.25 to util_cfs?

I don't get what you're saying. But I think it's similar to what I'm saying.

To clarify. What I'm saying is that when we try to calculate the effective
util, CFS is the only entity in practice that interacts with DVFS. DL and RT by
design 'disable' DVFS and when they become runnable set the frequency to
a constant fixed point.  For them DVFS latencies are not acceptable - although
in practice they do take a single hit for the freq change on wake up. IRQ on
the other hand doesn't really care about DVFS. So we end up in practice that
CFS is the only entity that interacts with DVFS, so when we calculate the
DVFS headroom, we should only take its util into account.


Thanks!

--
Qais Yousef