Re: [RESEND][PATCH v2 3/3] schedutil: trace: Add tracing to capture filter out requests

From: Qais Yousef
Date: Fri Jun 30 2023 - 08:01:35 EST


On 06/20/23 18:52, Lukasz Luba wrote:
> Hi Qais,
>
> I have somehow missed your feedback on this series.
>
> On 5/31/23 19:31, Qais Yousef wrote:
> > On 05/22/23 15:57, Lukasz Luba wrote:
> > > Some of the frequency update requests coming form the task scheduler
> > > might be filter out. It can happen when the previous request was served
> > > not that long ago (in a period smaller than provided by the cpufreq driver
> > > as minimum for frequency update). In such case, we want to know if some of
> > > the frequency updates cannot make through.
> > > Export the new tracepoint as well. That would allow to handle it by a
> > > toolkit for trace analyzes.
> > >
> > > Reported-by: kernel test robot <lkp@xxxxxxxxx> # solved tricky build
> > > Signed-off-by: Lukasz Luba <lukasz.luba@xxxxxxx>
> > > ---
> > > include/trace/events/sched.h | 4 ++++
> > > kernel/sched/cpufreq_schedutil.c | 10 ++++++++--
> > > 2 files changed, 12 insertions(+), 2 deletions(-)
> > >
> > > diff --git a/include/trace/events/sched.h b/include/trace/events/sched.h
> > > index dbfb30809f15..e34b7cd5de73 100644
> > > --- a/include/trace/events/sched.h
> > > +++ b/include/trace/events/sched.h
> > > @@ -739,6 +739,10 @@ DECLARE_TRACE(uclamp_update_tsk_tp,
> > > TP_PROTO(struct task_struct *tsk, int uclamp_id, unsigned int value),
> > > TP_ARGS(tsk, uclamp_id, value));
> > > +DECLARE_TRACE(schedutil_update_filtered_tp,
> > > + TP_PROTO(int cpu),
> > > + TP_ARGS(cpu));
> > > +
> > > #endif /* _TRACE_SCHED_H */
> > > /* This part must be outside protection */
> > > diff --git a/kernel/sched/cpufreq_schedutil.c b/kernel/sched/cpufreq_schedutil.c
> > > index f462496e5c07..4f9daf258a65 100644
> > > --- a/kernel/sched/cpufreq_schedutil.c
> > > +++ b/kernel/sched/cpufreq_schedutil.c
> > > @@ -6,6 +6,8 @@
> > > * Author: Rafael J. Wysocki <rafael.j.wysocki@xxxxxxxxx>
> > > */
> > > +EXPORT_TRACEPOINT_SYMBOL_GPL(schedutil_update_filtered_tp);
> > > +
> > > #define IOWAIT_BOOST_MIN (SCHED_CAPACITY_SCALE / 8)
> > > struct sugov_tunables {
> > > @@ -318,8 +320,10 @@ static inline bool sugov_update_single_common(struct sugov_cpu *sg_cpu,
> > > ignore_dl_rate_limit(sg_cpu);
> > > - if (!sugov_should_update_freq(sg_cpu->sg_policy, time))
> > > + if (!sugov_should_update_freq(sg_cpu->sg_policy, time)) {
> > > + trace_schedutil_update_filtered_tp(sg_cpu->cpu);
> > > return false;
> > > + }
> >
> > Can't we have something more generic here too? Are you interested to count
> > these events? How do you plan to use it?
>
> The plan is to record those events, count them and maybe adjust the FW
> if the frequency switching capabilities are too low, e.g. 4ms...

You mean as part of tuning step for the system or at runtime? The latter seems
to indicate for a proper interface instead.

IMHO I think the current filtering mechanism needs a bit of a massage.

One thing I think we must do is to ignore the filter if there's a big sudden
change in requested frequency. Like for instance if a big task migrates. Then
prev_cpu should go to lower freq sooner, and new_cpu should change to higher
frequency sooner too. The filtering makes sense only in steady state situation
where we are ramping up or down gradually.

If no one beats me to it, I'll propose something in that regard.

>
> We need those numbers to point out that there is a need for faster
> FW micro-controller to serve those incoming requests.

I think there's a big assumption here that the filter is always set correctly
;-)

>
> >
> > I think this will be a very noisy event by the way.
>
> Could be, but on the other hand for those statistical analysis
> 'the more the better'. It will also depend on number of
> CPUs in the cluster, e.g. 4 CPUs vs 1 CPU.
>
> I don't know when we will switch to this per-cpu cpufreq mode
> when all CPUs behave like independent DVFS. Juno mainline kernel and FW
> supports that mode. We would have to compare those two modes and
> measure how much we gain/loose when using one and not the other.
>
> Furthermore, we already suspect some of our integration testing for
> EAS-mainline (on Juno) failing due to filtered out requests. How much
> that would impact other boards - it would be nice to see in traces.

Another problem I think we have is that the DVFS headroom value should be
a function of this filter. At the moment it is hardcoded to a random value
which causes power issue.

So to summarize I think there are two improvements required (and if anyone has
the time to try them out go ahead otherwise I'll get to it):

1. The filter should only be applied if the history hasn't changed. ie: we are
gradually increasing or decreasing PELT. Otherwise we should honour sudden
changes ASAP.
2. DVFS headroom should be a function of the filter. 25% is too high for
500us. It could be too low for 10ms (I don't know).


Thanks!

--
Qais Yousef

>
> Thanks for your feedback!
> Lukasz