Hi Ingo,
On Wed, Oct 4, 2023 at 12:26 AM Ingo Molnar <mingo@xxxxxxxxxx> wrote:
>
>
> * Namhyung Kim <namhyung@xxxxxxxxxx> wrote:
>
> > AFAIK we don't have a tool to measure the context switch overhead
> > directly. (I think I should add one to perf ftrace latency). But I can
> > see it with a simple perf bench command like this.
> >
> > $ perf bench sched pipe -l 100000
> > # Running 'sched/pipe' benchmark:
> > # Executed 100000 pipe operations between two processes
> >
> > Total time: 0.650 [sec]
> >
> > 6.505740 usecs/op
> > 153710 ops/sec
> >
> > It runs two tasks communicate each other using a pipe so it should
> > stress the context switch code. This is the normal numbers on my
> > system. But after I run these two perf stat commands in background,
> > the numbers vary a lot.
> >
> > $ sudo perf stat -a -e cycles -G user.slice -- sleep 100000 &
> > $ sudo perf stat -a -e uncore_imc/cas_count_read/ -- sleep 10000 &
> >
> > I will show the last two lines of perf bench sched pipe output for
> > three runs.
> >
> > 58.597060 usecs/op # run 1
> > 17065 ops/sec
> >
> > 11.329240 usecs/op # run 2
> > 88267 ops/sec
> >
> > 88.481920 usecs/op # run 3
> > 11301 ops/sec
> >
> > I think the deviation comes from the fact that uncore events are managed
> > a certain number of cpus only. If the target process runs on a cpu that
> > manages uncore pmu, it'd take longer. Otherwise it won't affect the
> > performance much.
>
> The numbers of pipe-message context switching will vary a lot depending on
> CPU migration patterns as well.
>
> The best way to measure context-switch overhead is to pin that task
> to a single CPU with something like:
>
> $ taskset 1 perf stat --null --repeat 10 perf bench sched pipe -l 10000 >/dev/null
>
> Performance counter stats for 'perf bench sched pipe -l 10000' (10 runs):
>
> 0.049798 +- 0.000102 seconds time elapsed ( +- 0.21% )
>
> as you can see the 0.21% stddev is pretty low.
>
> If we allow 2 CPUs, both runtime and stddev is much higher:
>
> $ taskset 3 perf stat --null --repeat 10 perf bench sched pipe -l 10000 >/dev/null
>
> Performance counter stats for 'perf bench sched pipe -l 10000' (10 runs):
>
> 1.4835 +- 0.0383 seconds time elapsed ( +- 2.58% )
Thanks for taking your time. I should have said I also tried this.
But the problem is that it doesn't need the pure context switch.
It needs to switch to a different cgroup to trigger the overhead.
For example, I counted the number of context switches.
$ perf stat -e context-switches,cgroup-switches \
> perf bench sched pipe -l 10000 > /dev/null
Performance counter stats for 'perf bench sched pipe -l 10000':
20,001 context-switches
20,001 cgroup-switches
But if I use the taskset,
$ perf stat -e context-switches,cgroup-switches \
> taskset -c 0 perf bench sched pipe -l 10000 > /dev/null
Performance counter stats for 'taskset -c 0 perf bench sched pipe -l 10000':
20,003 context-switches
2 cgroup-switches
So the regression didn't happen when I used taskset because
the two tasks run on the cpu without changing cgroups.
Maybe I can add an option to perf bench sched to place
senders and receivers in different cgroups.
Thanks,
Namhyung
