Re: [PATCH v2 2/7] sched/fair: Decay task PELT values during migration

[Vincent Guittot]

On Tue, 18 Jan 2022 at 11:56, Vincent Donnefort
<vincent.donnefort@xxxxxxx> wrote:
>
> On Mon, Jan 17, 2022 at 06:31:25PM +0100, Vincent Guittot wrote:
> > On Wed, 12 Jan 2022 at 17:14, Vincent Donnefort
> > <vincent.donnefort@xxxxxxx> wrote:
> > >
> > > Before being migrated to a new CPU, a task sees its PELT values
> > > synchronized with rq last_update_time. Once done, that same task will also
> > > have its sched_avg last_update_time reset. This means the time between
> > > the migration and the last clock update (B) will not be accounted for in
> > > util_avg and a discontinuity will appear. This issue is amplified by the
> > > PELT clock scaling. If the clock hasn't been updated while the CPU is
> > > idle, clock_pelt will not be aligned with clock_task and that time (A)
> > > will be also lost.
> > >
> > >    ---------|----- A -----|-----------|------- B -----|>
> > >         clock_pelt   clock_task     clock            now
> > >
> > > This is especially problematic for asymmetric CPU capacity systems which
> > > need stable util_avg signals for task placement and energy estimation.
> > >
> > > Ideally, this problem would be solved by updating the runqueue clocks
> > > before the migration. But that would require taking the runqueue lock
> > > which is quite expensive [1]. Instead estimate the missing time and update
> > > the task util_avg with that value:
> > >
> > >   A + B = clock_task - clock_pelt + sched_clock_cpu() - clock
> > >
> > > Neither clock_task, clock_pelt nor clock can be accessed without the
> > > runqueue lock. The new runqueue clock_pelt_lag is therefore created and
> > > encode those three values.
> > >
> > >   clock_pelt_lag = clock - clock_task + clock_pelt
> > >
> > > And we can then write the missing time as follow:
> > >
> > >   A + B = sched_clock_cpu() - clock_pelt_lag
> > >
> > > The B. part of the missing time is however an estimation that doesn't take
> > > into account IRQ and Paravirt time.
> > >
> > > Now we have an estimation for A + B, we can create an estimator for the
> > > PELT value at the time of the migration. We need for this purpose to
> > > inject last_update_time which is a combination of both clock_pelt and
> > > lost_idle_time. The latter is a time value which is completely lost form a
> > > PELT point of view and must be ignored. And finally, we can write:
> > >
> > >   rq_clock_pelt_estimator() = last_update_time + A + B
> > >                             = last_update_time +
> > >                                    sched_clock_cpu() - clock_pelt_lag
> > >
> > > [1] https://lore.kernel.org/all/20190709115759.10451-1-chris.redpath@xxxxxxx/
> > >
> > > Signed-off-by: Vincent Donnefort <vincent.donnefort@xxxxxxx>
> > >
> > > diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> > > index 06cf7620839a..11c6aeef4583 100644
> > > --- a/kernel/sched/core.c
> > > +++ b/kernel/sched/core.c
> > > @@ -618,6 +618,12 @@ struct rq *task_rq_lock(struct task_struct *p, struct rq_flags *rf)
> > >         }
> > >  }
> > >
> > > +static void update_rq_clock_pelt_lag(struct rq *rq)
> > > +{
> > > +       u64_u32_store(rq->clock_pelt_lag,
> > > +                     rq->clock - rq->clock_task + rq->clock_pelt);
> >
> > This has several shortfalls:
> > - have a look at cfs_rq_clock_pelt() and rq_clock_pelt(). What you
> > name clock_pelt in your commit message and is used to update PELT and
> > saved in se->avg.last_update_time is : rq->clock_pelt -
> > rq->lost_idle_time - cfs_rq->throttled_clock_task_time
>
> That's why, the PELT "lag" is added onto se->avg.last_update_time. (see the last
> paragraph of the commit message) The estimator is just a time delta, that is
> added on top of the entity's last_update_time. I don't see any problem with the
> lost_idle_time here.

lost_idle_time is updated before entering idle and after your
clock_pelt_lag has been updated. This means that the delta that you
are computing can be wrong

I haven't look in details but similar problem probably happens for
throttled_clock_task_time

>
> > - you are doing this whatever the state of the cpu : idle or not. But
> > the clock cycles are not accounted for in the same way in both cases.
>
> If the CPU is idle and clock_pelt == clock_task, the component A of the
> estimator would be 0 and we only would account for how outdated is the rq's
> clock, i.e. component B.

And if cpu is not idle, you can't apply the diff between clk_pelt and clock_task

>
> > - (B) doesn't seem to be accurate as you skip irq and steal time
> > accounting and you don't apply any scale invariance if the cpu is not
> > idle
>
> The missing irq and paravirt time is the reason why it is called "estimator".
> But maybe there's a chance of improving this part with a lockless version of
> rq->prev_irq_time and rq->prev_steal_time_rq?
>
> > - IIUC your explanation in the commit message above, the (A) period
> > seems to be a problem only when idle but you apply it unconditionally.
>
> If the CPU is idle (and clock_pelt == clock_task), only the B part would be
> worth something:
>
>   A + B = [clock_task - clock_pelt] + [sched_clock_cpu() - clock]
>                       A                            B
>
> > If cpu is idle you can assume that clock_pelt should be equal to
> > clock_task but you can't if cpu is not idle otherwise your sync will
> > be inaccurate and defeat the primary goal of this patch. If your
> > problem with clock_pelt is that the pending idle time is not accounted
> > for when entering idle but only at the next update (update blocked
> > load or wakeup of a thread). This patch below should fix this and
> > remove your A.
>
> That would help slightly the current situation, but this part is already
> covered by the estimator.

But the estimator, as you name it, is wrong beaus ethe A part can't be
applied unconditionally

>
> >
> >
> > diff --git a/kernel/sched/pelt.h b/kernel/sched/pelt.h
> > index e06071bf3472..855877be4dd8 100644
> > --- a/kernel/sched/pelt.h
> > +++ b/kernel/sched/pelt.h
> > @@ -114,6 +114,7 @@ static inline void update_idle_rq_clock_pelt(struct rq *rq)
> >  {
> >         u32 divider = ((LOAD_AVG_MAX - 1024) << SCHED_CAPACITY_SHIFT)
> > - LOAD_AVG_MAX;
> >         u32 util_sum = rq->cfs.avg.util_sum;
> > +       u64 now = rq_clock_task(rq);
> >         util_sum += rq->avg_rt.util_sum;
> >         util_sum += rq->avg_dl.util_sum;
> >
> > @@ -127,7 +128,10 @@ static inline void update_idle_rq_clock_pelt(struct rq *rq)
> >          * rq's clock_task.
> >          */
> >         if (util_sum >= divider)
> > -               rq->lost_idle_time += rq_clock_task(rq) - rq->clock_pelt;
> > +               rq->lost_idle_time += now - rq->clock_pelt;
> > +
> > +       /* The rq is idle, we can sync to clock_task */
> > +       rq->clock_pelt  = now;
> >  }
> >
> >  static inline u64 rq_clock_pelt(struct rq *rq)
> >
> > ---
> >
> >
> > > +}
> > > +
> > >  /*
> > >   * RQ-clock updating methods:
> > >   */
> > > @@ -674,6 +680,7 @@ static void update_rq_clock_task(struct rq *rq, s64 delta)
> > >                 update_irq_load_avg(rq, irq_delta + steal);
> > >  #endif
> > >         update_rq_clock_pelt(rq, delta);
> > > +       update_rq_clock_pelt_lag(rq);
> > >  }
> > >
> > >  void update_rq_clock(struct rq *rq)
> > > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > > index 99ea9540ece4..046d5397eb8a 100644
> > > --- a/kernel/sched/fair.c
> > > +++ b/kernel/sched/fair.c
> > > @@ -6852,6 +6852,14 @@ select_task_rq_fair(struct task_struct *p, int prev_cpu, int wake_flags)
> > >
> > >  static void detach_entity_cfs_rq(struct sched_entity *se);
> > >
> > > +static u64 rq_clock_pelt_estimator(struct rq *rq, u64 last_update_time)
> > > +{
> > > +       u64 pelt_lag = sched_clock_cpu(cpu_of(rq)) -
> > > +                      u64_u32_load(rq->clock_pelt_lag);
> >
> > Have you evaluated the impact of calling sched_clock_cpu(cpu_of(rq))
> > for a remote cpu ? especially with a huge number of migration and
> > concurrent access from several cpus
>
> I have not, but I will have a look.
>
> >
> > > +
> > > +       return last_update_time + pelt_lag;
> > > +}
> > > +
> > >  /*
> > >   * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
> > >   * cfs_rq_of(p) references at time of call are still valid and identify the
> > > @@ -6859,6 +6867,9 @@ static void detach_entity_cfs_rq(struct sched_entity *se);
> > >   */
> > >  static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
> > >  {
> > > +       struct sched_entity *se = &p->se;
> > > +       struct rq *rq = task_rq(p);
> > > +
> > >         /*
> > >          * As blocked tasks retain absolute vruntime the migration needs to
> > >          * deal with this by subtracting the old and adding the new
> > > @@ -6866,7 +6877,6 @@ static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
> > >          * the task on the new runqueue.
> > >          */
> > >         if (READ_ONCE(p->__state) == TASK_WAKING) {
> > > -               struct sched_entity *se = &p->se;
> > >                 struct cfs_rq *cfs_rq = cfs_rq_of(se);
> > >
> > >                 se->vruntime -= u64_u32_load(cfs_rq->min_vruntime);
> > > @@ -6877,26 +6887,32 @@ static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
> > >                  * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old'
> > >                  * rq->lock and can modify state directly.
> > >                  */
> > > -               lockdep_assert_rq_held(task_rq(p));
> > > -               detach_entity_cfs_rq(&p->se);
> > > +               lockdep_assert_rq_held(rq);
> > > +               detach_entity_cfs_rq(se);
> > >
> > >         } else {
> > > +               u64 now;
> > > +
> > > +               remove_entity_load_avg(se);
> > > +
> > >                 /*
> > > -                * We are supposed to update the task to "current" time, then
> > > -                * its up to date and ready to go to new CPU/cfs_rq. But we
> > > -                * have difficulty in getting what current time is, so simply
> > > -                * throw away the out-of-date time. This will result in the
> > > -                * wakee task is less decayed, but giving the wakee more load
> > > -                * sounds not bad.
> > > +                * Here, the task's PELT values have been updated according to
> > > +                * the current rq's clock. But if that clock hasn't been
> > > +                * updated in a while, a substantial idle time will be missed,
> > > +                * leading to an inflation after wake-up on the new rq.
> > > +                *
> > > +                * Estimate the PELT clock lag, and update sched_avg to ensure
> > > +                * PELT continuity after migration.
> > >                  */
> > > -               remove_entity_load_avg(&p->se);
> > > +               now = rq_clock_pelt_estimator(rq, se->avg.last_update_time);
> > > +               __update_load_avg_blocked_se(now, se);
> > >         }
> > >
> > >         /* Tell new CPU we are migrated */
> > > -       p->se.avg.last_update_time = 0;
> > > +       se->avg.last_update_time = 0;
> > >
> > >         /* We have migrated, no longer consider this task hot */
> > > -       p->se.exec_start = 0;
> > > +       se->exec_start = 0;
> > >
> > >         update_scan_period(p, new_cpu);
> > >  }
> > > diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h
> > > index f1a445efdc63..fdf2a9e54c0e 100644
> > > --- a/kernel/sched/sched.h
> > > +++ b/kernel/sched/sched.h
> > > @@ -1027,8 +1027,13 @@ struct rq {
> > >         /* Ensure that all clocks are in the same cache line */
> > >         u64                     clock_task ____cacheline_aligned;
> > >         u64                     clock_pelt;
> > > +       u64                     clock_pelt_lag;
> > >         unsigned long           lost_idle_time;
> > >
> > > +#ifndef CONFIG_64BIT
> > > +       u64                     clock_pelt_lag_copy;
> > > +#endif
> > > +
> > >         atomic_t                nr_iowait;
> > >
> > >  #ifdef CONFIG_SCHED_DEBUG
> > > --
> > > 2.25.1
> > >