Power Scheduler Design

Power Scheduler Design

[Abel Vesa]

For a while now, I've started studying the power aware scheduling problem.
And like many other rookies out there I took all the lkml mails related
and read them all (well, almost all) and I saw that there are some
debating on the implementation.I even look over the implementation
proposed of Preeti U Murthy. I also worked (just for fun) for a while on
some ideas of my own (nothing worth sharing, yet) but I have problem
understanding the design requirements. Here is one. 

  Some of you (even Ingo) said  that the scheduler should be the one to
manage the cpu P/C states. In this case the governors of the cpuidle and
cpufreq would not make any sense anymore.  Does that mean they will not
be a part of this scheduling solution anymore?

Re: Power Scheduler Design

[Alex Shi]

于 9/7/14, 4:47, Abel Vesa 写道:
> For a while now, I've started studying the power aware scheduling problem.
> And like many other rookies out there I took all the lkml mails related
> and read them all (well, almost all) and I saw that there are some
> debating on the implementation.I even look over the implementation
> proposed of Preeti U Murthy. I also worked (just for fun) for a while on
> some ideas of my own (nothing worth sharing, yet) but I have problem
> understanding the design requirements. Here is one.
>
>    Some of you (even Ingo) said  that the scheduler should be the one to
> manage the cpu P/C states. In this case the governors of the cpuidle and
> cpufreq would not make any sense anymore.  Does that mean they will not
> be a part of this scheduling solution anymore?
CPUIDLE and CPUFREQ are used for cpu power saving when related CPU is 
not busy.
Scheduling is coordinate the system load and cpu load. Currently. 
Scheduling has no much
idea of CPUIDLE/CPUFREQ status, then may give task to a cpu which in 
poor latency or poor
powersaving status. That leads to poor latency and high cost of power.

The power aware scheduling target is to know and coordinate the 
cpuidle/cpufreq in scheduling.
Then tasks will be assigned to a cpu unit with better 
latency/powersaving consideration.

>
> _______________________________________________
> linaro-kernel mailing list
> linaro-kernel@lists.linaro.org
> http://lists.linaro.org/mailman/listinfo/linaro-kernel

Re: Power Scheduler Design

[Mike Turquette]

Quoting Abel Vesa (2014-09-07 04:47:13)
> For a while now, I've started studying the power aware scheduling problem.
> And like many other rookies out there I took all the lkml mails related
> and read them all (well, almost all) and I saw that there are some
> debating on the implementation.I even look over the implementation
> proposed of Preeti U Murthy. I also worked (just for fun) for a while on
> some ideas of my own (nothing worth sharing, yet) but I have problem
> understanding the design requirements. Here is one. 
> 
>   Some of you (even Ingo) said  that the scheduler should be the one to
> manage the cpu P/C states. In this case the governors of the cpuidle and
> cpufreq would not make any sense anymore.  Does that mean they will not
> be a part of this scheduling solution anymore?

Correct. The current thinking from the energy-aware scheduling (eas)
workshop in September is that the CPUfreq governors will go away, in
time. This won't happen soon.

Of course making smart choices on how and when to change cpu frequency
involves some platform-specific knowledge, and this will likely be
handled by the in-kernel energy model. The energy model will be
per-platform or per-machine. See the recent RFCs from Morten Rasmussen
to get more info on this.

The latest efforts are focused on task placement, but C- and P-states
will come along in the future.

Regards,
Mike