Messages in this thread | | | Date | Wed, 2 Dec 2020 14:48:18 +0000 | From | Mel Gorman <> | Subject | Re: [sched/fair] 8d86968ac3: netperf.Throughput_tps -29.5% regression |
| |
On Wed, Dec 02, 2020 at 10:29:59PM +0800, Li, Aubrey wrote: > > If the idle mask is not getting cleared then select_idle_cpu() is > > probably returning immediately. select_idle_core() is almost certainly > > failing so that just leaves select_idle_smt() to find a potentially idle > > CPU. That's a limited search space so tasks may be getting stacked and > > missing CPUs that are idling for short periods. > > Vincent suggested we decouple idle cpumask from short idle(stop tick) and > set it every time the CPU enters idle, I'll make this change in V6. >
As a heads-up, I'm trying to prepare a series that alters the time complexity in general of select_idle_sibling(). It would tie into what you are doing with the idle cpumask tracking but would use it as a hint for CPUs to search first. It's still a WIP but I'm hoping to post something tomorrow. It would not replace your patch, just alter it a bit in terms of what happens just before select_idle_cpu().
> > > > On the flip side, I expect cases like hackbench to benefit because it > > can saturate a machine to such a degree that select_idle_cpu() is a waste > > of time. > > Yes, I believe that's also why I saw uperf/netperf improvement at high > load levels. >
Yeah, hackbench is a case where SIS_AVG_CPU shines even though it does not help other cases. It throttles the search. In the series I'm working on right now, I simply kill SIS_AVG_CPU but might incorporate something like it into SIS_PROP as the last patch of the series as an RFC.
> > > > That said, I haven't followed the different versions closely. I know v5 > > got a lot of feedback so will take a closer look at v6. Fundamentally > > though I expect that using the idle mask will be a mixed bag. At low > > utilisation or over-saturation, it'll be a benefit. At the point where > > the machine is almost fully busy, some workloads will benefit (lightly > > communicating workloads that occasionally migrate) and others will not > > (ping-pong workloads looking for CPUs that are idle for very brief > > periods). > > Do you have any interested workload [matrix] I can do the measurement? >
Usually I go with a battery of tests from mmtests instead of one or two specifically to have a mix of wakeup timing, communication patterns and degrees of utilisation. The downside is that they take ages to run.
> > It's tricky enough that it might benefit from a sched_feat() check that > > is default true so it gets tested. For regressions that show up, it'll > > be easy enough to ask for the feature to be disabled to see if it fixes > > it. Over time, that might give an idea of exactly what sort of workloads > > benefit and what suffers. > > Okay, I'll add a sched_feat() for this feature. >
If the series I'm preparing works out ok and your patch can be integrated, the sched_feat() may not be necessary because your patch would further reduce time complexity without worrying about when the information gets reset.
-- Mel Gorman SUSE Labs
| |