Re: [PATCH V2 1/2] sched/fair: Prevent cpu_busy_time from exceeding actual_cpu_capacity

[Qais Yousef <qyousef@layalina.io>]

On 07/04/24 19:01, Pierre Gondois wrote:

> I thought the EAS was comparing instantaneous power and not energy,
> i.e. how the energy computation is done:
> 
>   *             ps->power * cpu_max_freq
>   *   cpu_nrg = ------------------------ * cpu_util           (3)
>   *               ps->freq * scale_cpu
> 
> cpu_nrg should have the same dimension as ps->power (i.e. energy/s).
>  From this PoV, the energy computation should not take into account how
> much time a task is expected to run. But it might be a side discussion,

I had this discussion with Quentin recently as I indicated in another reply on
this thread I think we do have inaccuracies here in terms of how we try to
represent the running time (or busy time) of the cpu/pd.

AFAIU this is supposed to be computing energy, but we don't explicitly multiply
with any time value and there's an assumed time multiplication with
unspecified period. Maybe it's PELT HF, maybe it's something else. But I think
we do have sources of inaccuracies here, but I need to analyse and dig more.

The cpu_util/cpu_cap, or sum_util/pd->cap is assumed to represent the
percentage of time we are busy during this unspecified period

	cpu_nrg = ps->power x (cpu_util/cpu_cap) * T

where T is a constant value of some period, hence it is omitted from all
calculations. I'm of course thinking this is not best; but sure keeps it
a 'simple energy model' :)

I'm not looking in this area at the moment, but if someone is, it'd be great to
consider the impact of this properly.