On Wed, Jun 01, 2016 at 09:57:23PM +0200, Peter Zijlstra wrote: > On Mon, May 23, 2016 at 01:00:10PM +0100, Morten Rasmussen wrote: > > On Mon, May 23, 2016 at 01:12:07PM +0200, Mike Galbraith wrote: > > > On Mon, 2016-05-23 at 11:58 +0100, Morten Rasmussen wrote: > > > > wake_wide() is based on task wakee_flips of the waker and the wakee to > > > > decide whether an affine wakeup is desirable. On lightly loaded systems > > > > the waker is frequently the idle task (pid=0) which can accumulate a lot > > > > of wakee_flips in that scenario. It makes little sense to prevent affine > > > > wakeups on an idle cpu due to the idle task wakee_flips, so it makes > > > > more sense to ignore them in wake_wide(). > > > > > > You sure? What's the difference between a task flipping enough to > > > warrant spreading the load, and an interrupt source doing the same? > > > I've both witnessed firsthand, and received user confirmation of this > > > very thing improving utilization. > > > > Right, I didn't consider the interrupt source scenario, my fault. > > > > 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 :-) > > Seeing how we always so select_idle_siblings() after affine_sd, the only > wake_affine movement that matters is cross-llc. > > So intra-llc wakeups can avoid the movement, no?
Won't help I think; the interrupt that got us in this situation will already have wrecked your idle time/state to begin with. You really want to help interrupt routing.
