On Thu, 5 Apr 2018 20:08:59 +1000 Nicholas Piggin <npig...@gmail.com> wrote:
> On Thu, 05 Apr 2018 10:40:20 +0200 > Mike Galbraith <efa...@gmx.de> wrote: > > > On Thu, 2018-04-05 at 10:27 +0200, Peter Zijlstra wrote: > > > On Thu, Apr 05, 2018 at 09:11:38AM +1000, Nicholas Piggin wrote: > > > > Hi, > > > > > > > > I'm seeing some pretty big latencies on a ~idle system when a CPU wakes > > > > out of a nohz idle. Looks like it's due to the taking a lot of remote > > > > locks and cache lines. irqoff trace: > > > > > > On RT I think we default RT_RUNTIME_SHARE to false, maybe we should do > > > the same for mainline. > > > > Probably. My very first enterprise encounter with the thing was it NOT > > saving a box from it's not so clever driver due to that. > > Well I would think a simpler per-cpu limiter might actually stand a > better chance of saving you there. Or even something attached to the > softlockup watchdog. > > I'm still getting a lot of locks coming from sched_rt_period_timer > with RT_RUNTIME_SHARE false, it's just that it's now down to about > NR_CPUS locks rather than 3*NR_CPUS. Oh yeah, putting -1 into sched_rt_runtime_us looks like it fixes it, I didn't look at Mike's patch close enough. If the code stays, it would be nice to be able to default it off at least. Thanks, Nick