[Ksummit-2013-discuss] [ATTEND] Power-aware scheduling

[Catalin Marinas]

On Tue, Jul 16, 2013 at 04:49:04AM +0100, Preeti U Murthy wrote:
> Hi Morten,

(I know I'm not Morten ;) but he's away this week; he will comment more
when he's back)

> It would be good to list out the topics that we need to discuss.
> 
> I would suggest some of the below points.
> 
> 1. What about the relevance of the patches that were posted so far
> (Alex's and Vincent's and your big.LITTLE patches) in the light of
> integration of cpuidle, cpufreq and scheduler. How will their behaviour
> change?

As I see it, there are two main (simplified) aspects to scheduler and
power management (which can be more optimal if unified):

1. Individual CPU C and P state selection
   a) Load tracking
   b) Idle time estimation
2. Optimal task placement
   a) Cache topology
   b) Memory topology (NUMA)
   c) Power topology (individual sockets etc.)

CPUfreq and cpuidle currently take care of (1) but their decision is
affected by (2) and vice-versa, hence the points above need some
unification. The scheduler does 2.a and 2.b but we are missing 2.c.
Morten's patches try to address this but do not exclude other task
packing patches. They currently use cpu_power to give hints to the
scheduler about which CPUs to use (and IIUC a similar cpu_power tweaking
principle is used in Vincent's patches though the latter have additional
optimisations).

That's not related to big.LITTLE but any multi-socket (symmetric) system
(and not necessarily NUMA).

> 2. What about unifying the different patches into one, considering they
> all want power savings. Is it that hard? Can't we extend Alex's patches
> to meet the goals of the other two?

I think they could be extended. For big.LITTLE we have the additional
requirement of favouring little CPUs when packing tasks but I think
Alex's patches could be adapted.  Morten could comment more on Alex's
patches as I haven't looked at them in detail.

-- 
Catalin