Re: [PATCH 1/1] sched: Consider CPU contention in frequency & load-balance busiest CPU selection

[Chen Yu <yu.c.chen@intel.com>]

On 2023-05-05 at 20:02:36 +0200, Dietmar Eggemann wrote:
> Hi Chenyu,
> 
> On 05/05/2023 09:10, Chen Yu wrote:
> > On 2023-04-06 at 17:50:30 +0200, Dietmar Eggemann wrote:
> >> Use new cpu_boosted_util_cfs() instead of cpu_util_cfs().
> >>
> >> The former returns max(util_avg, runnable_avg) capped by max CPU
> >> capacity. CPU contention is thereby considered through runnable_avg.
> >>
> >> The change in load-balance only affects migration type `migrate_util`.
> >>
> >> Suggested-by: Vincent Guittot <vincent.guittot@linaro.org>
> >> Signed-off-by: Dietmar Eggemann <dietmar.eggemann@arm.com>
> >>
> > Tested on Intel Sapphire Rapids which has 2x56C/112T = 224 CPUs.
> > The test tries to check if this is any impact on find_busiest_queue()
> > so it was tested with cpufreq governor performance.
> > The baseline is the 6.3 sched/core branch on top of
> > Commit 67fff302fc445a ("sched/fair: Introduce SIS_CURRENT to wake up"),
> > and compared to the code with current patch applied.
> > 
> > In summary no obvious difference and some small improvements on tbench
> > were observed so far:
> 
> many thanks for the test results!
> 
> Could you share the parameter lists you use for the individual tests?
> This would make it easier to understand the results and rerun the tests
> on similar machines.
>
Yes. The test parameters are all tested with cpufreq governor performance,
turbo disabled and C-states > C1 are disabled.

netperf:
Loopback test. Launched 1 netserver. Launched 25%, 50%, 75%, 100%, 125%, 150%, 175%, 200%
of the online CPUs number of netperf client. Tested type is TCP_RR and UDP_RR
respectively.

for i in $(seq 1 $nr_netperf); do
    netperf -4 -H 127.0.0.1 -t $type -c -C -l 100 &
done

tbench:
Loopback test. Launched 1 tbench_srv. Launched 25%, 50%, 75%, 100%, 125%, 150%, 175%, 200%
of the online CPUs number of tbench client.

tbench -t 100 $nr_tbench 127.0.0.1


schbench:
Number of worker threads is 25% of the online CPUs. Number of message threads is
1, 2, 4, 8 respectively.

schbench -m $nr_message -t $nr_worker -r 100


hackbench:
Number of hackbench group is 1, 2, 4, 8 respectively. Work types are "process" and "threads"
respectively. IPC mode are "pipe" and "sockets". The number of fd in each group is 12.5% of
the CPU number.

hackbench -g $nr_group --$work_type --$ipc_mode -l 5000000 -s 100 -f $num_fds


thanks,
Chenyu