Re: [PATCH V2] sched: fair: Use the earliest break even

[Vincent Guittot]

On Fri, 13 Mar 2020 at 13:15, Daniel Lezcano <daniel.lezcano@xxxxxxxxxx> wrote:
>
> On 12/03/2020 13:27, Vincent Guittot wrote:
> > On Thu, 12 Mar 2020 at 11:04, Daniel Lezcano <daniel.lezcano@xxxxxxxxxx> wrote:
> >>
> >> On 12/03/2020 09:36, Vincent Guittot wrote:
> >>> Hi Daniel,
> >>>
> >>> On Wed, 11 Mar 2020 at 21:28, Daniel Lezcano <daniel.lezcano@xxxxxxxxxx> wrote:
> >>>>
> >>>> In the idle CPU selection process occuring in the slow path via the
> >>>> find_idlest_group_cpu() function, we pick up in priority an idle CPU
> >>>> with the shallowest idle state otherwise we fall back to the least
> >>>> loaded CPU.
> >>>
> >>> The idea makes sense but this path is only used by fork and exec so
> >>> I'm not sure about the real impact
> >>
> >> I agree the fork / exec path is called much less often than the wake
> >> path but it makes more sense for the decision.
> >>
> >>>> In order to be more energy efficient but without impacting the
> >>>> performances, let's use another criteria: the break even deadline.
> >>>>
> >>>> At idle time, when we store the idle state the CPU is entering in, we
> >>>> compute the next deadline where the CPU could be woken up without
> >>>> spending more energy to sleep.
> >>>>
> >>>> At the selection process, we use the shallowest CPU but in addition we
> >>>> choose the one with the minimal break even deadline instead of relying
> >>>> on the idle_timestamp. When the CPU is idle, the timestamp has less
> >>>> meaning because the CPU could have wake up and sleep again several times
> >>>> without exiting the idle loop. In this case the break even deadline is
> >>>> more relevant as it increases the probability of choosing a CPU which
> >>>> reached its break even.
> >>>>
> >>>> Tested on:
> >>>>  - a synquacer 24 cores, 6 sched domains
> >>>>  - a hikey960 HMP 8 cores, 2 sched domains, with the EAS and energy probe
> >>>>
> >>>> sched/perf and messaging does not show a performance regression. Ran
> >>>> 50 times schbench, adrestia and forkbench.
> >>>>
> >>>> The tools described at https://lwn.net/Articles/724935/
> >>>>
> >>>>  --------------------------------------------------------------
> >>>> | Synquacer             | With break even | Without break even |
> >>>>  --------------------------------------------------------------
> >>>> | schbench *99.0th      |      14844.8    |         15017.6    |
> >>>> | adrestia / periodic   |        57.95    |              57    |
> >>>> | adrestia / single     |         49.3    |            55.4    |
> >>>>  --------------------------------------------------------------
> >>>
> >>> Have you got some figures or cpuidle statistics for the syncquacer ?
> >>
> >> No, and we just noticed the syncquacer has a bug in the firmware and
> >> does not actually go to the idle states.
> >>
> >>
> >>>> | Hikey960              | With break even | Without break even |
> >>>>  --------------------------------------------------------------
> >>>> | schbench *99.0th      |      56140.8    |           56256    |
> >>>> | schbench energy       |      153.575    |         152.676    |
> >>>> | adrestia / periodic   |         4.98    |             5.2    |
> >>>> | adrestia / single     |         9.02    |            9.12    |
> >>>> | adrestia energy       |         1.18    |           1.233    |
> >>>> | forkbench             |        7.971    |            8.05    |
> >>>> | forkbench energy      |         9.37    |            9.42    |
> >>>>  --------------------------------------------------------------
> >>>>
> >>>> Signed-off-by: Daniel Lezcano <daniel.lezcano@xxxxxxxxxx>
> >>>> ---
> >>>>  kernel/sched/fair.c  | 18 ++++++++++++++++--
> >>>>  kernel/sched/idle.c  |  8 +++++++-
> >>>>  kernel/sched/sched.h | 20 ++++++++++++++++++++
> >>>>  3 files changed, 43 insertions(+), 3 deletions(-)
> >>>>
> >>>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> >>>> index 4b5d5e5e701e..8bd6ea148db7 100644
> >>>> --- a/kernel/sched/fair.c
> >>>> +++ b/kernel/sched/fair.c
> >>>> @@ -5793,6 +5793,7 @@ find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this
> >>>>  {
> >>>>         unsigned long load, min_load = ULONG_MAX;
> >>>>         unsigned int min_exit_latency = UINT_MAX;
> >>>> +       s64 min_break_even = S64_MAX;
> >>>>         u64 latest_idle_timestamp = 0;
> >>>>         int least_loaded_cpu = this_cpu;
> >>>>         int shallowest_idle_cpu = -1;
> >>>> @@ -5810,6 +5811,8 @@ find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this
> >>>>                 if (available_idle_cpu(i)) {
> >>>>                         struct rq *rq = cpu_rq(i);
> >>>>                         struct cpuidle_state *idle = idle_get_state(rq);
> >>>> +                       s64 break_even = idle_get_break_even(rq);
> >>>> +
> >>>>                         if (idle && idle->exit_latency < min_exit_latency) {
> >>>>                                 /*
> >>>>                                  * We give priority to a CPU whose idle state
> >>>> @@ -5817,10 +5820,21 @@ find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this
> >>>>                                  * of any idle timestamp.
> >>>>                                  */
> >>>>                                 min_exit_latency = idle->exit_latency;
> >>>> +                               min_break_even = break_even;
> >>>>                                 latest_idle_timestamp = rq->idle_stamp;
> >>>>                                 shallowest_idle_cpu = i;
> >>>> -                       } else if ((!idle || idle->exit_latency == min_exit_latency) &&
> >>>> -                                  rq->idle_stamp > latest_idle_timestamp) {
> >>>> +                       } else if ((idle && idle->exit_latency == min_exit_latency) &&
> >>>> +                                  break_even < min_break_even) {
> >>>> +                               /*
> >>>> +                                * We give priority to the shallowest
> >>>> +                                * idle states with the minimal break
> >>>> +                                * even deadline to decrease the
> >>>> +                                * probability to choose a CPU which
> >>>> +                                * did not reach its break even yet
> >>>> +                                */
> >>>> +                               min_break_even = break_even;
> >>>> +                               shallowest_idle_cpu = i;
> >>>> +                       } else if (!idle && rq->idle_stamp > latest_idle_timestamp) {
> >>>>                                 /*
> >>>>                                  * If equal or no active idle state, then
> >>>>                                  * the most recently idled CPU might have
> >>>> diff --git a/kernel/sched/idle.c b/kernel/sched/idle.c
> >>>> index b743bf38f08f..3342e7bae072 100644
> >>>> --- a/kernel/sched/idle.c
> >>>> +++ b/kernel/sched/idle.c
> >>>> @@ -19,7 +19,13 @@ extern char __cpuidle_text_start[], __cpuidle_text_end[];
> >>>>   */
> >>>>  void sched_idle_set_state(struct cpuidle_state *idle_state)
> >>>>  {
> >>>> -       idle_set_state(this_rq(), idle_state);
> >>>> +       struct rq *rq = this_rq();
> >>>> +
> >>>> +       idle_set_state(rq, idle_state);
> >>>
> >>> Shouldn't the state be set after setting break even otherwise you will
> >>> have a time window with an idle_state != null but the break_even still
> >>> set to the previous value
> >>
> >> IIUC we are protected in this section. Otherwise the routine above would
> >> be also wrong [if (idle && idle->exit_latency)], no?
> >
> > no there are not the same because it uses the idle pointer to read
> > exit_latency so we are sure to use exit_latency related to the idle
> > pointer.
> >
> > In your case it checks idle is not null but then it uses rq to read
> > break_even but it might not have been already updated
>
> Ok I will invert the lines.
>
> >>>> +
> >>>> +       if (idle_state)
> >>>> +               idle_set_break_even(rq, ktime_get_ns() +
> >>>
> >>> What worries me a bit is that it adds one ktime_get call each time a
> >>> cpu enters idle
> >>
> >> Right, we can improve this in the future by folding the local_clock() in
> >> cpuidle when entering idle with this ktime_get.
> >
> > Using local_clock() would be more latency friendly
>
> Unfortunately we are comparing the deadline across CPUs, so the
> local_clock() can not be used here.
>
> But if we have one ktime_get() instead of a local_clock() + ktime_get(),
> that should be fine, no?

Can't this computation of break_even be done in cpuidle framework
while computing other statistics for selecting the idle state instead
? cpuidle already uses ktime_get for next hrtimer as an example.
So cpuidle compute break_even and make it available to scheduler like
exit_latency. And I can imagine that system wide time value will also
be needed when looking at next wakeup event of cluster/group of CPUs

>
> >>>> +                                   idle_state->exit_latency_ns);
> >>>>  }
> >>
> >> [ ... ]
> >>
> >>
> >> --
> >>  <http://www.linaro.org/> Linaro.org â Open source software for ARM SoCs
> >>
> >> Follow Linaro:  <http://www.facebook.com/pages/Linaro> Facebook |
> >> <http://twitter.com/#!/linaroorg> Twitter |
> >> <http://www.linaro.org/linaro-blog/> Blog
> >>
>
>
> --
>  <http://www.linaro.org/> Linaro.org â Open source software for ARM SoCs
>
> Follow Linaro:  <http://www.facebook.com/pages/Linaro> Facebook |
> <http://twitter.com/#!/linaroorg> Twitter |
> <http://www.linaro.org/linaro-blog/> Blog
>