Date:   Thu, 9 Aug 2018 16:23:13 +0100
From:   Patrick Bellasi <patrick.bellasi@....com>
To:     Juri Lelli <juri.lelli@...hat.com>
Cc:     linux-kernel@...r.kernel.org, linux-pm@...r.kernel.org,
        Ingo Molnar <mingo@...hat.com>,
        Peter Zijlstra <peterz@...radead.org>,
        Tejun Heo <tj@...nel.org>,
        "Rafael J . Wysocki" <rafael.j.wysocki@...el.com>,
        Viresh Kumar <viresh.kumar@...aro.org>,
        Vincent Guittot <vincent.guittot@...aro.org>,
        Paul Turner <pjt@...gle.com>,
        Dietmar Eggemann <dietmar.eggemann@....com>,
        Morten Rasmussen <morten.rasmussen@....com>,
        Todd Kjos <tkjos@...gle.com>,
        Joel Fernandes <joelaf@...gle.com>,
        Steve Muckle <smuckle@...gle.com>,
        Suren Baghdasaryan <surenb@...gle.com>
Subject: Re: [PATCH v3 01/14] sched/core: uclamp: extend sched_setattr to
 support utilization clamping

On 09-Aug 11:50, Juri Lelli wrote:
> On 09/08/18 10:14, Patrick Bellasi wrote:
> > On 07-Aug 14:35, Juri Lelli wrote:
> > > On 06/08/18 17:39, Patrick Bellasi wrote:

[...]

> > 1) make CAP_SYS_NICE protected the clamp groups, with an optional boot
> >    time parameter to relax this check
> 
> It seems to me that this might work well with that the intended usage of
> the interface that you depict above. SMS only (or any privileged user)
> will be in control of how groups are configured, so no problem for
> normal users.

Yes, well... apart normal users still getting a -ENOSPC is they are
requesting one of the not pre-configured clamp values. Which is why
the following bits can be helpful.

> > 2) add discretization support to clamp groups allocation
> 
> And this might also work well if we feel that we don't want to restrict
> usage of the interface to admin only, however...
> 
> > This second feature specifically, will ensure that clamp values are
> > always mapped into one of the available clamp groups. While the exact
> > clamp value can always be used for tasks placement biasing, when it
> > comes to frequency selection biasing, depending on concurrently
> > running tasks, you can end up with an effective clamp value which is a
> > rounded up.
> 
> what I'm not so sure about is that we might lose in flexibility if the
> number of available discrete clamp groups is too small compared to the
> number of available OPP on the platform.

Regarding this concern, I would say that we should consider that, for
frequency biasing, we are in general not interested in nailing down
the single 1% difference and/or exact OPP capacities

A certain coarse grained resolution is usually acceptable for many
different reasons:
a) schedutil already uses a 20% margin which can potentially eclipse
   few OPP when we scale up/down
b) tasks/CPUs utilization are good enough but never exact and precise
   values
c) reducing the number of OPP switches could have some benefits on
   stability/latencies
d) clamping is actually defining minimum/maximum preferred values, is
   not to be considered a tool for "precise control"

All that considered, I would say that maybe a 5% resolution could
still be considered an acceptable _worst case_ rounding since we don't
have always to round up to the next 5%.

For example, if we have:
 - TaskA: util_min=41%
 - TaskB: util_nin=44%
they will be both accounted in the 40-45% clamp group but the clamp
group value can be modulated at run-time depending on RUNNABLE
tasks. When TaskA is running alone, we can still set util_min to
41%, while we will use 44% (not 45%) when TaskB is (also) running.

It's worth to notice that we pre-allocated at compile time 20 clamp
groups, but not necessarily all of them will be used at run-time.
Indeed, we will still use a policy where only the actual required
values are allocated at the beginning of the clamps map, thus
optimizing max updates.

-- 
#include <best/regards.h>

Patrick Bellasi

Powered by blists - more mailing lists