Re: [PATCH 03/16] sched/fair: Disregard idle task wakee_flips in wake_wide

[Yuyang Du]

On Mon, May 23, 2016 at 05:42:20PM +0200, Mike Galbraith wrote:
> On Mon, 2016-05-23 at 15:10 +0100, Morten Rasmussen wrote:
> > On Mon, May 23, 2016 at 03:00:46PM +0200, Mike Galbraith wrote:
> > > On Mon, 2016-05-23 at 13:00 +0100, Morten Rasmussen wrote:
> > > 
> > > > The problem then seems to be distinguishing truly idle and busy doing
> > > > interrupts. The issue that I observe is that wake_wide() likes pushing
> > > > tasks around in lightly scenarios which isn't desirable for power
> > > > management. Selecting the same cpu again may potentially let others
> > > > reach deeper C-state.
> > > > 
> > > > With that in mind I will if I can do better. Suggestions are welcome :-)
> > > 
> > > None here.  For big boxen that are highly idle, you'd likely want to
> > > shut down nodes and consolidate load, but otoh, all that slows response
> > > to burst, which I hate.  I prefer race to idle, let power gating do its
> > > job.  If I had a server farm with enough capacity vs load variability
> > > to worry about, I suspect I'd become highly interested in routing.
> > 
> > I don't disagree for systems of that scale, but at the other end of the
> > spectrum it is a single SoC we are trying squeeze the best possible
> > mileage out of. That implies optimizing for power gating to reach deeper
> > C-states when possible by consolidating idle-time and grouping
> > idle cpus. Migrating task unnecessarily isn't helping us in achieving
> > that, unfortunately :-(
> 
> Yup, the goals are pretty much mutually exclusive.  For your goal, you
> want more of an allocator like behavior, where stacking of tasks is bad
> only once there's too much overlap (ie latency, defining is hard), and
> allocation always has the same order (expand rightward or such for the
> general case, adding little/big complexity for arm).  For mine, current
> behavior is good, avoid stacking like the plague.

I'd be happy to have a switch to either one goal.