Re: [PATCH v2 0/2] sched: Optionally skip uclamp logic in fast path

[Vincent Guittot]

On Tue, 23 Jun 2020 at 19:40, Dietmar Eggemann <dietmar.eggemann@xxxxxxx> wrote:
>
> On 19/06/2020 19:20, Qais Yousef wrote> This series attempts to address the report that uclamp logic could be expensive
> > sometimes and shows a regression in netperf UDP_STREAM under certain
> > conditions.
> >
> > The first patch is a fix for how struct uclamp_rq is initialized which is
> > required by the 2nd patch which contains the real 'fix'.
> >
> > Worth noting that the root cause of the overhead is believed to be system
> > specific or related to potential certain code/data layout issues, leading to
> > worse I/D $ performance.
> >
> > Different systems exhibited different behaviors and the regression did
> > disappear in certain kernel version while attempting to reporoduce.
> >
> > More info can be found here:
> >
> > https://lore.kernel.org/lkml/20200616110824.dgkkbyapn3io6wik@e107158-lin/
> >
> > Having the static key seemed the best thing to do to ensure the effect of
> > uclamp is minimized for kernels that compile it in but don't have a userspace
> > that uses it, which will allow distros to distribute uclamp capable kernels by
> > default without having to compromise on performance for some systems that could
> > be affected.
>
> My test data indicates that the static key w/o any uclamp users (3)
> brings the performance number for the 'perf bench sched pipe'
> workload back (i.e. from w/ !CONFIG_UCLAMP_TASK) (1).
>
> platform:
>
>     Arm64 Hikey960 (only little CPUs [0-3]), no CPUidle,
>     performance CPUfreq governor
>
> workload:
>
>     perf stat -n -r 20 -- perf bench sched pipe -T -l 100000
>
>
> (A) *** Performance results ***
>
> (1) tip/sched/core
>     # CONFIG_UCLAMP_TASK is not set
>
>     *1.39285* +- 0.00191 seconds time elapsed  ( +-  0.14% )
>
> (2) tip/sched/core
>     CONFIG_UCLAMP_TASK=y
>
>     *1.42877* +- 0.00181 seconds time elapsed  ( +-  0.13% )
>
> (3) tip/sched/core + opt_skip_uclamp_v2
>     CONFIG_UCLAMP_TASK=y
>
>     *1.38833* +- 0.00291 seconds time elapsed  ( +-  0.21% )
>
> (4) tip/sched/core + opt_skip_uclamp_v2
>     CONFIG_UCLAMP_TASK=y
>     echo 512 > /proc/sys/kernel/sched_util_clamp_min (enable uclamp)
>
>     *1.42062* +- 0.00238 seconds time elapsed  ( +-  0.17% )
>
>
> (B) *** Profiling on CPU0 and CPU1  ***
>
>     (further hp'ing out CPU2 and CPU3 to get consistent hit numbers)
>
> (1)
>
> CPU0:  Function             Hit    Time            Avg             s^2
>        --------             ---    ----            ---             ---
>        deactivate_task  1997346    2207642 us      *1.105* us      0.033 us
>        activate_task    1997391    1840057 us      *0.921* us      0.054 us
>
> CPU1:  Function             Hit    Time            Avg             s^2
>        --------             ---    ----            ---             ---
>        deactivate_task  1997455    2225960 us      1.114 us        0.034 us
>        activate_task    1997410    1842603 us      0.922 us        0.052 us
>
> (2)
>
> CPU0:  Function             Hit    Time            Avg             s^2
>        --------             ---    ----            ---             ---
>        deactivate_task  1998538    2419719 us      *1.210* us      0.061 us
>        activate_task    1997119    1960401 us      *0.981* us      0.034 us
>
> CPU1:  Function             Hit    Time            Avg             s^2
>        --------             ---    ----            ---             ---
>        deactivate_task  1996597    2400760 us      1.202 us        0.059 us
>        activate_task    1998016    1985013 us      0.993 us        0.028 us
>
> (3)
>
> CPU0:  Function             Hit    Time            Avg             s^2
>        --------             ---    ----            ---             ---
>        deactivate_task  1997525    2155416 us      *1.079* us      0.020 us
>        activate_task    1997874    1899002 us      *0.950* us      0.044 us
>
> CPU1:  Function             Hit    Time            Avg             s^2
>        --------             ---    ----            ---             ---
>        deactivate_task  1997935    2118648 us      1.060 us        0.017 us
>        activate_task    1997586    1895162 us      0.948 us        0.044 us
>
> (4)
>
> CPU0:  Function             Hit    Time            Avg             s^2
>        --------             ---    ----            ---             ---
>        deactivate_task  1998246    2428121 us      *1.215* us      0.062 us
>        activate_task    1998252    2132141 us      *1.067* us      0.020 us
>
> CPU1:  Function             Hit    Time            Avg             s^2
>        --------             ---    ----            ---             ---
>        deactivate_task  1996154    2414194 us      1.209 us        0.060 us
>        activate_task    1996148    2140667 us      1.072 us        0.021 us

I have rerun the tests that I ran previously on my octo core arm64 (hikey):
20 iteration of perf bench sched pipe -T -l 50000
tip stands for tip/sched/core
uclamp enabled means both uclamp task and uclamp cgroup
the stdev is around 0.25% for all tests

                           root           level 1       level 2       level 3

tip uclamp disabled        50653          47188         44568         41925
tip uclamp enabled         48800(-3.66%)  45600(-3.37%) 42822(-3.92%)
40257(-3.98%)
/w patch uclamp disabled   50615(-0.08%)  47198(+0.02%) 44609(+0.09%)
41735(-0.45%)
/w patch uclamp enabled    49661(-1.96%)  46611(-1.22%) 43803(-1.72%)
41243(-1.63%)

Results are better with your patch