2019-07-01 20:59:36

On Thu, 2019-07-04 at 11:33 +0200, Vincent Guittot wrote:
> On Fri, 28 Jun 2019 at 22:49, Rik van Riel <[email protected]> wrote:
> > Reducing the overhead of the CPU controller is achieved by not
> > walking
> > all the sched_entities every time a task is enqueued or dequeued.

> > @@ -7687,6 +7700,10 @@ static inline bool cfs_rq_is_decayed(struct
> > cfs_rq *cfs_rq)
> > if (cfs_rq->avg.util_sum)
> > return false;
> >
> > + /* Remove decayed parents once their decayed children are
> > gone. */
> > + if (cfs_rq->children_on_list)
>
> I'm not sure that you really need to count whether childrens are on
> the list.
> Instead you can take advantage of the list ordering and you only have
> to test if the previous cfs_rq in the list is a child. If it's not
> then there is no more child
>
> and you can remove the new field children_on_list and inc/dec it

Good suggestion. I'll do that for v3.

Thank you.

--
All Rights Reversed.

Attachments:

signature.asc (499.00 B)
This is a digitally signed message part

2019-07-05 15:17:31

by Rik van Riel

[permalink] [raw]

Subject: Re: [PATCH 07/10] sched,cfs: fix zero length timeslice calculation

On Fri, 2019-07-05 at 16:51 +0200, Vincent Guittot wrote:
> On Fri, 28 Jun 2019 at 22:49, Rik van Riel <[email protected]> wrote:
> > The way the time slice length is currently calculated, not only do
> > high
> > priority tasks get longer time slices than low priority tasks, but
> > due
> > to fixed point math, low priority tasks could end up with a zero
> > length
> > time slice. This can lead to cache thrashing and other
> > inefficiencies.
> >
> > Simplify the logic a little bit, and cap the minimum time slice
> > length
> > to sysctl_sched_min_granularity.
> >
> > Tasks that end up getting a time slice length too long for their
> > relative
> > priority will simply end up having their vruntime advanced much
> > faster than
> > other tasks, resulting in them receiving time slices less
> > frequently.
> >
> > Signed-off-by: Rik van Riel <[email protected]>
> > ---
> > kernel/sched/fair.c | 25 ++++++++-----------------
> > 1 file changed, 8 insertions(+), 17 deletions(-)
> >
> > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > index d48bff5118fc..8da2823401ca 100644
> > --- a/kernel/sched/fair.c
> > +++ b/kernel/sched/fair.c
> > @@ -671,22 +671,6 @@ static inline u64 calc_delta_fair(u64 delta,
> > struct sched_entity *se)
> > return delta;
> > }
> >
> > -/*
> > - * The idea is to set a period in which each task runs once.
> > - *
> > - * When there are too many tasks (sched_nr_latency) we have to
> > stretch
> > - * this period because otherwise the slices get too small.
> > - *
> > - * p = (nr <= nl) ? l : l*nr/nl
> > - */
> > -static u64 __sched_period(unsigned long nr_running)
> > -{
> > - if (unlikely(nr_running > sched_nr_latency))
> > - return nr_running * sysctl_sched_min_granularity;
> > - else
> > - return sysctl_sched_latency;
> > -}
> > -
> > /*
> > * We calculate the wall-time slice from the period by taking a
> > part
> > * proportional to the weight.
> > @@ -695,7 +679,7 @@ static u64 __sched_period(unsigned long
> > nr_running)
> > */
> > static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity
> > *se)
> > {
> > - u64 slice = __sched_period(cfs_rq->nr_running + !se-
> > >on_rq);
> > + u64 slice = sysctl_sched_latency;
>
> Is the change above and the remove of __sched_period() really needed
> for fixing the null time slice problem ?
> This change impacts how tasks will preempt each other and as a result
> the throughput. It should have it dedicated patch so we can evaluate
> its impact

Good point. I will split this up into two patches for v3.

Thank you.

--
All Rights Reversed.

Attachments:

signature.asc (499.00 B)
This is a digitally signed message part

2019-07-05 15:50:44

by Vincent Guittot

[permalink] [raw]

Subject: Re: [PATCH 07/10] sched,cfs: fix zero length timeslice calculation

On Fri, 28 Jun 2019 at 22:49, Rik van Riel <[email protected]> wrote:
>
> The way the time slice length is currently calculated, not only do high
> priority tasks get longer time slices than low priority tasks, but due
> to fixed point math, low priority tasks could end up with a zero length
> time slice. This can lead to cache thrashing and other inefficiencies.
>
> Simplify the logic a little bit, and cap the minimum time slice length
> to sysctl_sched_min_granularity.
>
> Tasks that end up getting a time slice length too long for their relative
> priority will simply end up having their vruntime advanced much faster than
> other tasks, resulting in them receiving time slices less frequently.
>
> Signed-off-by: Rik van Riel <[email protected]>
> ---
> kernel/sched/fair.c | 25 ++++++++-----------------
> 1 file changed, 8 insertions(+), 17 deletions(-)
>
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index d48bff5118fc..8da2823401ca 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -671,22 +671,6 @@ static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
> return delta;
> }
>
> -/*
> - * The idea is to set a period in which each task runs once.
> - *
> - * When there are too many tasks (sched_nr_latency) we have to stretch
> - * this period because otherwise the slices get too small.
> - *
> - * p = (nr <= nl) ? l : l*nr/nl
> - */
> -static u64 __sched_period(unsigned long nr_running)
> -{
> - if (unlikely(nr_running > sched_nr_latency))
> - return nr_running * sysctl_sched_min_granularity;
> - else
> - return sysctl_sched_latency;
> -}
> -
> /*
> * We calculate the wall-time slice from the period by taking a part
> * proportional to the weight.
> @@ -695,7 +679,7 @@ static u64 __sched_period(unsigned long nr_running)
> */
> static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
> {
> - u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
> + u64 slice = sysctl_sched_latency;

Is the change above and the remove of __sched_period() really needed
for fixing the null time slice problem ?
This change impacts how tasks will preempt each other and as a result
the throughput. It should have it dedicated patch so we can evaluate
its impact

>
> for_each_sched_entity(se) {
> struct load_weight *load;
> @@ -712,6 +696,13 @@ static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
> }
> slice = __calc_delta(slice, se->load.weight, load);
> }
> +
> + /*
> + * To avoid cache thrashing, run at least sysctl_sched_min_granularity.
> + * The vruntime of a low priority task advances faster; those tasks
> + * will simply get time slices less frequently.
> + */
> + slice = max_t(u64, slice, sysctl_sched_min_granularity);
> return slice;
> }
>
> --
> 2.20.1
>