On Wed, Apr 13, 2016 at 05:18:51AM +0200, Mike Galbraith wrote: > On Tue, 2016-04-12 at 16:07 -0400, Chris Mason wrote: > > > I think that if we're worried about the cost of the idle scan for this > > workload, find_idlest_group() is either going to hurt much more, or not > > search enough CPUs to find the idle one. > > find_idlest_group()? No no no, that's not what I mean at all. > > wake_wide() identifies loads that really want to spread out, thus turns > off affine wakeups. We still call select_idle_sibling(), only > difference being that target is the original cpu, not the waking cpu.
Ah ok, I see what you mean now. > Given making that wide connection bidirectional helped FB's load, it > seems reasonable that passing wide information to select_idle_sibling() > would have a good chance of hitting the candidate that stands to gain > from a full socket scan, while also keeping that cache scrambling scan > far away from the rest. > > > But I'm happy to try patches or other ideas, I have a fixed version of > > the bitmap one going through production benchmarks now. [ benchmarks say it needs more fixing, ick ] > > Making that wide/full search cheap is still good, because wake_wide() > also identifies interrupt sources that are waking many, so cheap wide > search should increase utilization there as well. The thought was to > just make the wide thing have a tad wider effect on what it already > does affect.. and hope that doesn't demolish anything. So you're interested in numbers where we pass the wake_wide decision into select_idle_sibling(), and then use that instead of (or in addition to?) my should_scan_idle() function? I agree we may need to tweak wake_wide, since most of our wakeups now are failed affine wakeups. But, the differences are in p99, so I'll probably need to get some better metrics. -chris
