[PATCH v3 0/5] Rework system pressure interface to the scheduler

Vincent Guittot vincent.guittot at linaro.org
Fri Jan 19 09:57:41 PST 2024


On Wed, 10 Jan 2024 at 19:10, Dietmar Eggemann <dietmar.eggemann at arm.com> wrote:
>
> On 09/01/2024 14:29, Vincent Guittot wrote:
> > On Tue, 9 Jan 2024 at 12:34, Dietmar Eggemann <dietmar.eggemann at arm.com> wrote:
> >>
> >> On 08/01/2024 14:48, Vincent Guittot wrote:
> >>> Following the consolidation and cleanup of CPU capacity in [1], this serie
> >>> reworks how the scheduler gets the pressures on CPUs. We need to take into
> >>> account all pressures applied by cpufreq on the compute capacity of a CPU
> >>> for dozens of ms or more and not only cpufreq cooling device or HW
> >>> mitigiations. we split the pressure applied on CPU's capacity in 2 parts:
> >>> - one from cpufreq and freq_qos
> >>> - one from HW high freq mitigiation.
> >>>
> >>> The next step will be to add a dedicated interface for long standing
> >>> capping of the CPU capacity (i.e. for seconds or more) like the
> >>> scaling_max_freq of cpufreq sysfs. The latter is already taken into
> >>> account by this serie but as a temporary pressure which is not always the
> >>> best choice when we know that it will happen for seconds or more.
> >>
> >> I guess this is related to the 'user space system pressure' (*) slide of
> >> your OSPM '23 talk.
> >
> > yes
> >
> >>
> >> Where do you draw the line when it comes to time between (*) and the
> >> 'medium pace system pressure' (e.g. thermal and FREQ_QOS).
> >
> > My goal is to consider the /sys/../scaling_max_freq as the 'user space
> > system pressure'
> >
> >>
> >> IIRC, with (*) you want to rebuild the sched domains etc.
> >
> > The easiest way would be to rebuild the sched_domain but the cost is
> > not small so I would prefer to skip the rebuild and add a new signal
> > that keep track on this capped capacity
>
> Are you saying that you don't need to rebuild sched domains since
> cpu_capacity information of the sched domain hierarchy is
> independently updated via:
>
> update_sd_lb_stats() {
>
>   update_group_capacity() {
>
>     if (!child)
>       update_cpu_capacity(sd, cpu) {
>
>         capacity = scale_rt_capacity(cpu) {
>
>           max = get_actual_cpu_capacity(cpu) <- (*)
>         }
>
>         sdg->sgc->capacity = capacity;
>         sdg->sgc->min_capacity = capacity;
>         sdg->sgc->max_capacity = capacity;
>       }
>
>   }
>
> }
>
> (*) influence of temporary and permanent (to be added) frequency
> pressure on cpu_capacity (per-cpu and in sd data)


I'm more concerned by rd->max_cpu_capacity which remains at original
capacity and triggers spurious LB if we take into account the
userspace max freq instead of the original max compute capacity of a
CPU. And also how to manage this in RT and DL

>
>
> example: hackbench on h960 with IPA:
>                                                                                   cap  min  max
> ...
> hackbench-2284 [007] .Ns..  2170.796726: update_group_capacity: sdg !child cpu=7 1017 1017 1017
> hackbench-2456 [007] ..s..  2170.920729: update_group_capacity: sdg !child cpu=7 1018 1018 1018
>     <...>-2314 [007] ..s1.  2171.044724: update_group_capacity: sdg !child cpu=7 1011 1011 1011
> hackbench-2541 [007] ..s..  2171.168734: update_group_capacity: sdg !child cpu=7  918  918  918
> hackbench-2558 [007] .Ns..  2171.228716: update_group_capacity: sdg !child cpu=7  912  912  912
>     <...>-2321 [007] ..s..  2171.352718: update_group_capacity: sdg !child cpu=7  812  812  812
> hackbench-2553 [007] ..s..  2171.476721: update_group_capacity: sdg !child cpu=7  640  640  640
>     <...>-2446 [007] ..s2.  2171.600743: update_group_capacity: sdg !child cpu=7  610  610  610
> hackbench-2347 [007] ..s..  2171.724738: update_group_capacity: sdg !child cpu=7  406  406  406
> hackbench-2331 [007] .Ns1.  2171.848768: update_group_capacity: sdg !child cpu=7  390  390  390
> hackbench-2421 [007] ..s..  2171.972733: update_group_capacity: sdg !child cpu=7  388  388  388
> ...



More information about the linux-arm-kernel mailing list