Re: [RFC PATCH v2 0/1] cpuidle: teo: Introduce optional util-awareness

[Lukasz Luba]

On 10/20/22 20:52, Daniel Lezcano wrote:
> Hi Kajetan,
>
> On 20/10/2022 18:20, Kajetan Puchalski wrote:
> > Hi Rafael,
> >
> > > The avg_util value tells us nothing about how much the CPU is going to
> > > be idle this time and it also tells us nothing about the
> > > latency-sensitivity of the workload.
> > >
> > > Yes, it tells us how much idle time there was on the given CPU in the
> > > past, on the average, but there is zero information about the
> > > distribution of that idle time in it.
> > >
> > > So in the first place please tell me why it fundamentally makes sense
> > > to use avg_util in CPU idle time management at all.
> >
> > I have an alternative suggestion that could be a reasonable way forward
> > here. Instead of applying util-awareness on top of TEO where it would
> > have to be reconciled with how TEO is currently expected to work, I just
> > wrote a simple completely new governor which operates only on timer
> > events alongside util values.
>
> I second the idea. I took a long time to investigate how to improve the 
> governor and reached the conclusion having a dedicated governor for 
> mobile platform makes sense. Also the behavior is very platform dependent.
>
> Regarding the utilization, one of the issue is the kernel threads 
> preventing a task to wake up on the same CPU and forcing its migration 
> at wake up time. So the prediction is screwed up at that time.
>
> There is a paper talking this issue [1]
>
> I've done a 'mobile' governor, including the next interrupt prediction 
> [2]. It is very simple and almost has the same results as the teo on my 
> platform (rock960).
>
> I'm not planning to upstream it because I don't have spare time to 
> improve the results and take care of the IPIs. part.
>
> Also the paradigm is radically different and you may be interested in 
> the approach.
>
> So if you want to rework, improve, test, upstream it, feel free to reuse 
> the code.
>
>    -- Daniel
>
> [1] Dynamic workload characterization for power efficient scheduling on 
> CMP systems : https://cseweb.ucsd.edu//~tullsen/islped10.pdf
>
> [2] 
> https://git.linaro.org/people/daniel.lezcano/linux.git/commit/?h=cpuidle/mobile-governor-v5.1&id=de1edb05e3c342f0738b414aa84263d6555b7462 

Thanks Daniel! I forgot about your work in this area. As I have
responded in some other email in this thread, we might start
from a new small governor and than others can contribute.

Even this small governor that Kajetan showed me performs really
good on pixel6.

Regards,
Lukasz