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

Peter Zijlstra peterz at infradead.org
Tue Jul 16 15:37:45 UTC 2013


On Tue, Jul 16, 2013 at 02:05:43PM +0200, Rafael J. Wysocki wrote:
> Anyway, it all boils down to the "How are we going to measure success?"
> question I mentioned in another message.

Specifically for the power stuff, I'd say be able to do more 'work' on the same
battery charge.

Lasting longer on the same battery charge isn't a good measure as a scheduler
that simply refuses to run anything will last longest I'd think. So we need to
measure power used against work done.

Now this could be a 'virtual' battery through use of things like the ACPI power
meter or an actual battery, whatever people like.

Typically I expect workloads where this can be true to be 'light' loads. I mean
there's really nothing you can do about a kernel build -j64 on your laptop,
that's going to keep all cores/threads burning until its done.

Now we don't actually have benchmarks like that -- we typically like to burn
cpu and measure completion time etc. I remember asking the folks working on
this to create such benchmarks (more than a year ago in SFO some time or so).

At that point there was talk that there might be browser simulator things that
could be (ab)used to do this. Not sure anything ever came of it.

At the time PJT also suggested measuring 'typical' mobile workloads using perf
sched and use that to create artificial loads using linsched tools that could be
ran both in linsched as well as on the actual hardware.


I'm not too bothered about proving 100% coverage and no regressions as that's
not ever going to happen anyway. We'll work on it when there's problems is I
think the best we can do.

What I do care about is tackling the problem in a coherent fashion considering
everything that goes into power management.

But I think people care enough about their battery operated devices that
they're going to try and work on this. All I must ensure is that all interested
parties work together and don't go off and do their own little thing.


More information about the Ksummit-2013-discuss mailing list