Re: [RFC PATCH 0/7] Introduce thermal pressure

[Vincent Guittot]

On Wed, 10 Oct 2018 at 12:36, Quentin Perret <quentin.perret@xxxxxxx> wrote:
>
> On Wednesday 10 Oct 2018 at 12:14:32 (+0200), Vincent Guittot wrote:
> > On Wed, 10 Oct 2018 at 11:55, Quentin Perret <quentin.perret@xxxxxxx> wrote:
> > >
> > > Hi Vincent,
> > >
> > > On Wednesday 10 Oct 2018 at 10:50:05 (+0200), Vincent Guittot wrote:
> > > > The problem with reflecting directly the capping is that it happens
> > > > far more often than the pace at which cpu_capacity_orig is updated in
> > > > the scheduler.
> > >
> > > Hmm, how can you be so sure ? That most likely depends on the workload,
> > > the platform and the thermal governor. Some platforms heat up slowly,
> > > some quickly. The pace at which the thermal governor will change things
> > > should depend on that I assume.
> > >
> > > > This means that at the moment when scheduler uses the
> > > > value, it might not be correct anymore.
> > >
> > > And OTOH, when you remove a cap for example, it will take time before
> > > the scheduler can see the newly available capacity if you need to wait
> > > for the signal to decay. So you are using a wrong information too in
> > > that scenario.
> >
> > But we stay consistant with all other metrics. The same happen when a
> > task decide to stay idle for a long time after some activity... You
> > will have to wait for the signal to decay
>
> Because you see that as a task going idle. You could also say that
> removing the cap from a CPU is equivalent to migrating a task off that
> CPU. In this case you should see the newly available cap pretty much
> immediately.
>
> Also, I feel like the whole thermal capping story could interest the
> Intel folks who, IIRC, were discussing how to represent their 'boosted'
> OPPs in the scheduler some time ago. You can see the Intel boost thing
> as a cap I think -- some OPPs can be inaccessible in some case. I'd be
> interested to see what's their take on this.
>
> > > > Then, this value are also used
> > > > when building the sched_domain and setting max_cpu_capacity which
> > > > would also implies the rebuilt the sched_domain topology ...
> > >
> > > Wait what ? I thought the thermal cap was reflected in capacity_of, not
> > > capacity_orig_of ... You need to rebuild the sched_domain in case of
> > > thermal pressure ?
>
> I can't locate where you need to do this in the series. Do you actually
> need to rebuild the sd hierarchy ?

This patchset doesn't touch cpu_capacity_orig and doesn't need to  as
it assume that the max capacity is unchanged but some capacity is
momentary stolen by thermal.
 If you want to reflect immediately all thermal capping change, you
have to update this field and all related fields and struct around

>
> > > Hmm, let me have a closer look at the patches, I must have missed
> > > something ...
> > >
> > > > The pace of changing the capping is to fast to reflect that in the
> > > > whole scheduler topology
> > >
> > > That's probably true in some cases, but it'd be cool to have numbers to
> > > back up that statement, I think.
> > >
> > > Now, if you do need to rebuild the sched domain topology every time you
> > > update the thermal pressure, I think the PELT HL is _way_ too short for
> > > that ... You can't rebuild the whole thing every 32ms or so. Or am I
> > > misunderstanding something ?
> > >
> > > > > Thara, have you tried to experiment with a simpler implementation as
> > > > > suggested by Ingo ?
> > > > >
> > > > > Also, assuming that we do want to average things, do we actually want to
> > > > > tie the thermal ramp-up time to the PELT half life ? That provides
> > > > > nice maths properties wrt the other signals, but it's not obvious to me
> > > > > that this thermal 'constant' should be the same on all platforms. Or
> > > > > maybe it should ?
> > > >
> > > > The main interest of using PELT signal is that thermal pressure will
> > > > evolve at the same pace as other signals used in the scheduler.
> > >
> > > Right, I think this is a nice property too (assuming that we actually
> > > want to average things out).
> > >
> > > > With
> > > > thermal pressure, we have the exact same problem as with RT tasks. The
> > > > thermal will cap the max frequency which will cap the utilization of
> > > > the tasks running on the CPU
> > >
> > > Well, the nature of the signal is slightly different IMO. Yes it's
> > > capacity, but you're no actually measuring time spent on the CPU. All
> > > other PELT signals are based on time, this thermal thing isn't, so it is
> > > kinda different in a way. And I'm still wondering if it could be helpful
> >
> > hmmm ... it is based on time too.
>
> You're not actually measuring the time spent on the CPU by the 'thermal
> task'. There is no such thing as a 'thermal task'. You're just trying to
> model things like that, but the thermal stuff isn't actually
> interrupting running tasks to eat CPU cycles. It just makes thing run
> slower, which isn't exactly the same thing IMO.
>
> But maybe that's a detail.
>
> > Both signals (current ones and thermal one) are really close. The main
> > difference with other utilization signal is that instead of providing
> > a running/not running boolean that is then weighted by the current
> > capacity, the signal uses direclty the capped max capacity to reflect
> > the amount of cycle that is stolen by thermal mitigation.
> >
> > > to be able to have a different HL for that thermal signal. That would
> > > 'break' the nice maths properties we have, yes, but is it a problem or is
> > > it actually helpful to cope with the thermal characteristics of
> > > different platforms ?
> >
> > If you don't use the sign kind of signal with the same responsiveness,
> > you will start to see some OPP toggles as an example when the thermal
> > state change because one metrics will change faster than the other one
> > and you can't have a correct view of the system. Same problem was
> > happening with rt task.
>
> Well, that wasn't the problem with rt tasks. The problem with RT tasks
> was that the time they spend on the CPU wasn't accounted _at all_ when
> selecting frequency for CFS, not that the accounting was at a different
> pace ...

The problem was the same with RT, the cfs utilization was lower than
reality because RT steals soem cycle to CFS
So schedutil was selecting a lower frequency when cfs was running
whereas the CPU was fully used.
The same can happen with thermal:
cap the max freq because of thermal
the utilization with decrease.
remove the cap
the utilization is still low and you will select a low OPP because you
don't take into account cycle stolen by thermal like with RT

Regards,
Vincent
>
> Thanks,
> Quentin