Re: [PATCH v3 3/5] sched/fair: Switch to task based throttle model

[Aaron Lu <ziqianlu@bytedance.com>]

On Thu, Aug 14, 2025 at 05:54:34PM +0200, Valentin Schneider wrote:
> On 12/08/25 16:48, Aaron Lu wrote:
> > On Fri, Aug 08, 2025 at 01:45:11PM +0200, Valentin Schneider wrote:
> >> On 08/08/25 18:13, Aaron Lu wrote:
> >> > Let me run some scheduler benchmark to see how it impacts performance.
> >> >
> >> > I'm thinking maybe running something like hackbench on server platforms,
> >> > first with quota not set and see if performance changes; then also test
> >> > with quota set and see how performance changes.
> >> >
> >> > Does this sound good to you? Or do you have any specific benchmark and
> >> > test methodology in mind?
> >> >
> >>
> >> Yeah hackbench is pretty good for stressing the EQ/DQ paths.
> >>
> >
> > Tested hackbench/pipe and netperf/UDP_RR on Intel EMR(2 sockets/240
> > cpus) and AMD Genoa(2 sockets/384 cpus), the tldr is: there is no clear
> > performance change between base and this patchset(head). Below is
> > detailed test data:
> > (turbo/boost disabled, cpuidle disabled, cpufreq set to performance)
> >
> > hackbench/pipe/loops=150000
> > (seconds, smaller is better)
> >
> > On Intel EMR:
> >
> > nr_group          base             head          change
> >  1              3.62±2.99%      3.61±10.42%      +0.28%
> >  8              8.06±1.58%      7.88±5.82%       +2.23%
> > 16             11.40±2.57%     11.25±3.72%       +1.32%
> >
> > For nr_group=16 case, configure a cgroup and set quota to half cpu and
> > then let hackbench run in this cgroup:
> >
> >                  base             head           change
> > quota=50%      18.35±2.40%     18.78±1.97%       -2.34%
> >
> > On AMD Genoa:
> >
> > nr_group          base             head          change
> >  1             17.05±1.92%     16.99±2.81%       +0.35%
> >  8             16.54±0.71%     16.73±1.18%       -1.15%
> > 16             27.04±0.39%     26.72±2.37%       +1.18%
> >
> > For nr_group=16 case, configure a cgroup and set quota to half cpu and
> > then let hackbench run in this cgroup:
> >
> >                   base             head          change
> > quota=50%      43.79±1.10%     44.65±0.37%       -1.96%
> >
> > Netperf/UDP_RR/testlen=30s
> > (throughput, higher is better)
> >
> > 25% means nr_clients set to 1/4 nr_cpu, 50% means nr_clients is 1/2
> > nr_cpu, etc.
> >
> > On Intel EMR:
> >
> > nr_clients     base                 head             change
> >   25%       83,567±0.06%         84,298±0.23%        +0.87%
> >   50%       61,336±1.49%         60,816±0.63%        -0.85%
> >   75%       40,592±0.97%         40,461±0.14%        -0.32%
> >  100%       31,277±2.11%         30,948±1.84%        -1.05%
> >
> > For nr_clients=100% case, configure a cgroup and set quota to half cpu
> > and then let netperf run in this cgroup:
> >
> > nr_clients     base                 head             change
> >  100%       25,532±0.56%         26,772±3.05%        +4.86%
> >
> > On AMD Genoa:
> >
> > nr_clients     base                 head             change
> >  25%        12,443±0.40%         12,525±0.06%        +0.66%
> >  50%        11,403±0.35%         11,472±0.50%        +0.61%
> >  75%        10,070±0.19%         10,071±0.95%         0.00%
> > 100%         9,947±0.80%          9,881±0.58%        -0.66%
> >
> > For nr_clients=100% case, configure a cgroup and set quota to half cpu
> > and then let netperf run in this cgroup:
> >
> > nr_clients     base                 head             change
> > 100%         4,954±0.24%          4,952±0.14%         0.00%
> 
> Thank you for running these, looks like mostly slightly bigger variance on
> a few of these but that's about it.
> 
> I would also suggest running similar benchmarks but with deeper
> hierarchies, to get an idea of how much worse unthrottle_cfs_rq() can get
> when tg_unthrottle_up() goes up a bigger tree.

No problem.

I suppose I can reuse the previous shared test script:
https://lore.kernel.org/lkml/CANCG0GdOwS7WO0k5Fb+hMd8R-4J_exPTt2aS3-0fAMUC5pVD8g@mail.gmail.com/

There I used:
nr_level1=2
nr_level2=100
nr_level3=10

But I can tweak these numbers for this performance evaluation. I can make
the leaf level to be 5 level deep and place tasks in leaf level cgroups
and configure quota on 1st level cgroups.

I'll get back to you once I finished collecting data, feel free to let
me know if you have other idea testing this :)