Re: [PATCH] PM / EM: Inefficient OPPs detection

[Vincent Donnefort <vincent.donnefort@arm.com>]

On Thu, Apr 15, 2021 at 03:04:34PM +0000, Quentin Perret wrote:
> On Thursday 15 Apr 2021 at 15:12:08 (+0100), Vincent Donnefort wrote:
> > On Thu, Apr 15, 2021 at 01:12:05PM +0000, Quentin Perret wrote:
> > > Hi Vincent,
> > > 
> > > On Thursday 08 Apr 2021 at 18:10:29 (+0100), Vincent Donnefort wrote:
> > > > Some SoCs, such as the sd855 have OPPs within the same performance domain,
> > > > whose cost is higher than others with a higher frequency. Even though
> > > > those OPPs are interesting from a cooling perspective, it makes no sense
> > > > to use them when the device can run at full capacity. Those OPPs handicap
> > > > the performance domain, when choosing the most energy-efficient CPU and
> > > > are wasting energy. They are inefficient.
> > > > 
> > > > Hence, add support for such OPPs to the Energy Model, which creates for
> > > > each OPP a performance state. The Energy Model can now be read using the
> > > > regular table, which contains all performance states available, or using
> > > > an efficient table, where inefficient performance states (and by
> > > > extension, inefficient OPPs) have been removed.
> > > > 
> > > > Currently, the efficient table is used in two paths. Schedutil, and
> > > > find_energy_efficient_cpu(). We have to modify both paths in the same
> > > > patch so they stay synchronized. The thermal framework still relies on
> > > > the original table and hence, DevFreq devices won't create the efficient
> > > > table.
> > > > 
> > > > As used in the hot-path, the efficient table is a lookup table, generated
> > > > dynamically when the perf domain is created. The complexity of searching
> > > > a performance state is hence changed from O(n) to O(1). This also
> > > > speeds-up em_cpu_energy() even if no inefficient OPPs have been found.
> > > 
> > > Interesting. Do you have measurements showing the benefits on wake-up
> > > duration? I remember doing so by hacking the wake-up path to force tasks
> > > into feec()/compute_energy() even when overutilized, and then running
> > > hackbench. Maybe something like that would work for you?
> > 
> > I'll give a try and see if I get improved numbers.
> > 
> > > 
> > > Just want to make sure we actually need all that complexity -- while
> > > it's good to reduce the asymptotic complexity, we're looking at a rather
> > > small problem (max 30 OPPs or so I expect?), so other effects may be
> > > dominating. Simply skipping inefficient OPPs could be implemented in a
> > > much simpler way I think.
> > 
> > I could indeed just skip the perf state if marked as ineffective. But the idea
> > was to avoid bringing another for loop in this hot-path.
> 
> Right, though it would just extend a little bit the existing loop, so
> the overhead is unlikely to be noticeable.

In the case where we let cpufreq_table resolution, it's a whole new loop that we
would bring. In the case where we rely only on the EM resolution and bypass the
cpufreq_table though it would be even. But with the look-up table, we're winning
everywhere :-) Anyway I'll see if I can measure any improvement here.

-- 
Vincent

> 
> > Also, not covered by this patch but probably we could get rid of the EM
> > complexity limit as the table resolution is way faster with this change.
> 
> Probably yeah. I was considering removing it since eb92692b2544
> ("sched/fair: Speed-up energy-aware wake-ups") but ended up keeping it
> as it's entirely untested on large systems. But maybe we can reconsider.
> 
> Thanks,
> Quentin