Re: [RFC PATCH v2 0/6] Introduce Active Stats framework with CPU performance statistics

From: Lukasz Luba
Date: Tue Jul 06 2021 - 12:51:37 EST




On 7/6/21 5:34 PM, Rafael J. Wysocki wrote:
On Tue, Jul 6, 2021 at 5:56 PM Lukasz Luba <lukasz.luba@xxxxxxx> wrote:

On 7/6/21 4:28 PM, Rafael J. Wysocki wrote:
On Tue, Jul 6, 2021 at 3:18 PM Lukasz Luba <lukasz.luba@xxxxxxx> wrote:

Hi all,

This patch set introduces a new mechanism: Active Stats framework (ASF), which
gathers and maintains statistics of CPU performance - time residency at each
performance state.

The ASF tracks all the frequency transitions as well as CPU
idle entry/exit events for all CPUs. Based on that it accounts the active
(non-idle) residency time for each CPU at each frequency. This information can
be used by some other subsystems (like thermal governor) to enhance their
estimations about CPU usage at a given period.

This seems to mean that what is needed is something like the cpufreq
stats but only collected during the time when CPUs are not in idle
states.

Yes

So this is a clear problem statement: the cpufreq statistics cover the
time when CPUs are in idle states, so they are not suitable for
certain purposes, like thermal control.

Agree, it's better described problem statement.


The most straightforward approach to address it seems to be to modify
the collection of cpufreq statistics so that they don't include the
time spent by CPUs in idle states, or to make it possible to
distinguish the time spent in idle states from the "active" time.

Does it fix something in mainline?
Yes, there is thermal governor Intelligent Power Allocation (IPA), which
estimates the CPUs power used in the past. IPA is sampling the CPU utilization
and frequency and relies on the info available at the time of sampling
and this imposes the estimation errors.
The use of ASF solve the issue and enables IPA to make better estimates.

Obviously the IPA is not used on all platforms where cpufreq and
cpuidle are used. What platforms are going to benefit from this
change?

Arm platforms which still use kernel thermal to control temperature,
such as Chromebooks or mid-, low-end phones.

Which means that this feature is not going to be universally useful.

However, if the time spent by CPUs in idle states were accounted for
in the cpufreq statistics, that would be universally useful.

True



Why it couldn't be done using existing frameworks?
The CPUFreq and CPUIdle statistics are not combined, so it is not possible
to derive the information on how long exactly the CPU was running with a given
frequency.

But it doesn't mean that the statistics could not be combined.

For instance, the frequency of the CPU cannot change via cpufreq when
active_stats_cpu_idle_enter() is running, so instead of using an
entirely new framework for collecting statistics it might update the
existing cpufreq stats to register that event.

True, but keep in mind that the cpufreq most likely works for a few
CPUs (policy::related_cpus), while cpuidle in a per-cpu fashion.
I would say that cpuidle should check during enter/exit what is
the currently set frequency for cluster and account its active
period.

Yes, that's not particularly difficult to achieve in principle: on
idle entry and exit, update the cpufreq statistics of the policy
including the current CPU.

Sounds good.



And analogously for the wakeup.

This new framework combines that information and provides
it in a handy way.

I'm not convinced about the last piece.

The handy structure is called Active Stats Monitor. It samples
the stats gathered after processing idle. That private
structure maintains statistics which are for a given period
(current snapshot - previous snapshot).

So collecting the statistics should be fast and simple and processing
them need not be.

Ideally, they should be processed only when somebody asks the data.

Correct.


I'm not sure if that is the case in the current patchset.

Which is the case in current implementation, where the Active Stats
Monitor (ASM) is run in the context of thermal workqueue. It is
scheduled every 100ms to run IPA throttling, which does the
statistics gathering and calculation in the ASM. Time accounted for this
major calculation is moved to the 'client' (like IPA) context.



IMHO it has to be implemented as a new framework, next to
CPUFreq and CPUIdle, due to a clean design and not just hooks from thermal
governor into the frequency change and idle code paths.

As far as the design is concerned, I'm not sure if I agree with it.

From my perspective it's all a 1000-line patch that I have to read and
understand to figure out what the design is.

I can help you with understanding it with some design docs if you want.

That may help, but let's avoid doing extra work just yet.

Understood



Tha patch 4/6 introduces a new API for cooling devices, which allows to
stop tracking the freq and idle statistics.

The patch set contains also a patches 5/6 6/6 which adds the new power model
based on ASF into the cpufreq cooling (used by thermal governor IPA).
It is added as ifdef option, since Active Stats might be not compiled in.
The ASF is a compile time option, but that might be changed and IPA could
select it, which would allow to remove some redundant code from
cpufreq_cooling.c.

Comments and suggestions are very welcome.

I'm totally not convinced that it is necessary to put the extra 1000
lines of code into the kernel to address the problem at hand.


I understand your concerns. If you have another idea than this framework
I'm happy to hear it. Maybe better stats in cpuidle, which would be
are of the cpufreq?

One idea that I have is outlined above and I'm not seeing a reason to
put cpufreq statistics into cpuidle.


I'm happy to prepare such RFC if you like. I would just need a bit more
information. It sounds your proposed solution might be smaller in code
size, since the client's statistics accounting might be moved to the
cpufreq_cooling.c (which now live in the ASM). Also, there won't be a
new framework to maintain (which is a big plus).