Re: [PATCH v2 2/2] sched/fair: Scan cluster before scanning LLC in wake-up path
From: Barry Song
Date: Wed Feb 16 2022 - 04:13:04 EST
On Tue, Feb 8, 2022 at 6:42 PM Barry Song <21cnbao@xxxxxxxxx> wrote:
>
> On Tue, Feb 8, 2022 at 4:14 AM Gautham R. Shenoy <gautham.shenoy@xxxxxxx> wrote:
> >
> >
> > On Fri, Feb 04, 2022 at 11:28:25PM +1300, Barry Song wrote:
> >
> > > > We already figured out that there are no idle CPUs in this cluster. So dont
> > > > we gain performance by picking a idle CPU/core in the neighbouring cluster.
> > > > If there are no idle CPU/core in the neighbouring cluster, then it does make
> > > > sense to fallback on the current cluster.
> > >
> > > What you suggested is exactly the approach we have tried at the first beginning
> > > during debugging. but we didn't gain performance according to benchmark, we
> > > were actually losing. that is why we added this line to stop ping-pong:
> > > /* Don't ping-pong tasks in and out cluster frequently */
> > > if (cpus_share_resources(target, prev_cpu))
> > > return target;
> > >
> > > If we delete this, we are seeing a big loss of tbench while system
> > > load is medium
> > > and above.
> >
> > Thanks for clarifying this Barry. Indeed, if the workload is sensitive
> > to data ping-ponging across L2 clusters, this heuristic makes sense. I
> > was thinking of workloads that require lower tail latency, in which
> > case exploring the larger LLC would have made more sense, assuming
> > that the larger LLC has an idle core/CPU.
> >
> > In the absence of any hints from the workload, like something that
> > Peter had previous suggested
> > (https://lore.kernel.org/lkml/YVwnsrZWrnWHaoqN@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx/),
> > optimizing for cache-access seems to be the right thing to do.
>
> Thanks, gautham.
>
> Yep. Peter mentioned some hints like SCHED_BATCH and SCHED_IDLE.
> To me, the case we are discussing seems to be more complicated than
> applying some scheduling policy on separate tasks by SCHED_BATCH
> or IDLE.
>
> For example, in case we have a process, and this process has 20 threads.
> thread0-9 might care about cache-coherence latency and want to avoid
> ping-ponging, and thread10-thread19 might want to have tail-latency
> as small as possible. So we need some way to tell kernel, "hey, bro, please
> try to keep thread0-9 still as ping-ponging will hurt them while trying your
> best to find idle cpu in a wider range for thread10-19". But it seems
> SCHED_XXX as a scheduler policy hint can't tell kernel how to organize tasks
> into groups, and is also incapable of telling kernel different groups have
> different needs.
>
> So it seems we want some special cgroups to organize tasks and we can apply
> some special hints on each different group. for example, putting thread0-9
> in a cgroup and thread10-19 in another, then:
> 1. apply "COMMUNCATION-SENSITVE" on the 1st group
> 2. apply "TAIL-LATENCY-SENTIVE" on the 2nd one.
> I am not quite sure how to do this and if this can find its way into
> the mainline.
>
> On the other hand, for this particular patch, the most controversial
> part is those
> two lines to avoid ping-ponging, and I am seeing dropping this can hurt workload
> like tbench only when system load is high, so I wonder if the approach[1] from
> Chen Yu and Tim can somehow resolve the problem alternatively, thus we can
> avoid the controversial part.
> since their patch can also shrink the scanning range while llc load is high.
>
> [1] https://lore.kernel.org/lkml/20220207034013.599214-1-yu.c.chen@xxxxxxxxx/
Yicong's testing shows the patch from Chen Yu and Tim can somehow resolve the
problem and make sure there is no performance regression for tbench
while load is
high after we remove the code to avoid ping-pong:
5.17-rc1: vanilla
rc1 + chenyu: vanilla + chenyu's LLC overload patch
rc1+chenyu+cls: vanilla + chenyu's patch + my this patchset
rc1+chenyu+cls-pingpong: vanilla + chenyu's patch + my this patchset -
the code avoiding ping-pong
rc1+cls: vanilla + my this patchset
tbench running on numa 0 &1:
5.17-rc1 rc1 + chenyu
rc1+chenyu+cls rc1+chenyu+cls-pingpong rc1+cls
Hmean 1 320.01 ( 0.00%) 318.03 * -0.62%*
357.15 * 11.61%* 375.43 * 17.32%* 378.44 * 18.26%*
Hmean 2 643.85 ( 0.00%) 637.74 * -0.95%*
714.36 * 10.95%* 745.82 * 15.84%* 752.52 * 16.88%*
Hmean 4 1287.36 ( 0.00%) 1285.20 * -0.17%*
1431.35 * 11.18%* 1481.71 * 15.10%* 1505.62 * 16.95%*
Hmean 8 2564.60 ( 0.00%) 2551.02 * -0.53%*
2812.74 * 9.68%* 2921.51 * 13.92%* 2955.29 * 15.23%*
Hmean 16 5195.69 ( 0.00%) 5163.39 * -0.62%*
5583.28 * 7.46%* 5726.08 * 10.21%* 5814.74 * 11.91%*
Hmean 32 9769.16 ( 0.00%) 9815.63 * 0.48%*
10518.35 * 7.67%* 10852.89 * 11.09%* 10872.63 * 11.30%*
Hmean 64 15952.50 ( 0.00%) 15780.41 * -1.08%*
10608.36 * -33.50%* 17503.42 * 9.72%* 17281.98 * 8.33%*
Hmean 128 13113.77 ( 0.00%) 12000.12 * -8.49%*
13095.50 * -0.14%* 13991.90 * 6.70%* 13895.20 * 5.96%*
Hmean 256 10997.59 ( 0.00%) 12229.20 * 11.20%*
11902.60 * 8.23%* 12214.29 * 11.06%* 11244.69 * 2.25%*
Hmean 512 14623.60 ( 0.00%) 15863.25 * 8.48%*
14103.38 * -3.56%* 16422.56 * 12.30%* 15526.25 * 6.17%*
tbench running on numa 0 only:
5.17-rc1 rc1 + chenyu
rc1+chenyu+cls rc1+chenyu+cls-pingpong rc1+cls
Hmean 1 324.73 ( 0.00%) 330.96 * 1.92%*
358.97 * 10.54%* 376.05 * 15.80%* 378.01 * 16.41%*
Hmean 2 645.36 ( 0.00%) 643.13 * -0.35%*
710.78 * 10.14%* 744.34 * 15.34%* 754.63 * 16.93%*
Hmean 4 1302.09 ( 0.00%) 1297.11 * -0.38%*
1425.22 * 9.46%* 1484.92 * 14.04%* 1507.54 * 15.78%*
Hmean 8 2612.03 ( 0.00%) 2623.60 * 0.44%*
2843.15 * 8.85%* 2937.81 * 12.47%* 2982.57 * 14.19%*
Hmean 16 5307.12 ( 0.00%) 5304.14 * -0.06%*
5610.46 * 5.72%* 5763.24 * 8.59%* 5886.66 * 10.92%*
Hmean 32 9354.22 ( 0.00%) 9738.21 * 4.11%*
9360.21 * 0.06%* 9699.05 * 3.69%* 9908.13 * 5.92%*
Hmean 64 7240.35 ( 0.00%) 7210.75 * -0.41%*
6992.70 * -3.42%* 7321.52 * 1.12%* 7278.78 * 0.53%*
Hmean 128 6186.40 ( 0.00%) 6314.89 * 2.08%*
6166.44 * -0.32%* 6279.85 * 1.51%* 6187.85 ( 0.02%)
Hmean 256 9231.40 ( 0.00%) 9469.26 * 2.58%*
9134.42 * -1.05%* 9322.88 * 0.99%* 9448.61 * 2.35%*
Hmean 512 8907.13 ( 0.00%) 9130.46 * 2.51%*
9023.87 * 1.31%* 9276.19 * 4.14%* 9397.22 * 5.50%*
as you can see rc1+chenyu+cls-pingpong still shows similar improvement
like rc1+cls, in some
cases(256, 512 threads on numa0&1), it is even much better.
Thanks
Barry