sched/fair: favor non-idle group in tick preemption

  The non-idle se dominates competition vs the idle se when they
are belong to the same group. We ensure that idle groups would not
preempt non-idle group in wakeup preemption(see check_preempt_wakeup()).
However, this can happen in tick preemption, since check_preempt_tick()
dose not check current/se is idle or not. This patch adds this check.

Signed-off-by: Chuyi Zhou <zhouchuyi@bytedance.com>
---
 kernel/sched/fair.c | 12 +++++++++++-
 1 file changed, 11 insertions(+), 1 deletion(-)
  

Hi Chuyi,

On Thu, Oct 27, 2022 at 1:16 AM Chuyi Zhou <zhouchuyi@bytedance.com> wrote:
>
> The non-idle se dominates competition vs the idle se when they
> are belong to the same group. We ensure that idle groups would not
> preempt non-idle group in wakeup preemption(see check_preempt_wakeup()).
> However, this can happen in tick preemption, since check_preempt_tick()
> dose not check current/se is idle or not. This patch adds this check.
>
> Signed-off-by: Chuyi Zhou <zhouchuyi@bytedance.com>
> ---
>  kernel/sched/fair.c | 12 +++++++++++-
>  1 file changed, 11 insertions(+), 1 deletion(-)
>
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index e4a0b8bd941c..f3324b8753b3 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -4750,6 +4750,7 @@ static void
>  check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
>  {
>         unsigned long ideal_runtime, delta_exec;
> +       int cse_is_idle, pse_is_idle;
>         struct sched_entity *se;
>         s64 delta;
>
> @@ -4779,8 +4780,17 @@ check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
>         if (delta < 0)
>                 return;
>
> -       if (delta > ideal_runtime)
> +       if (delta > ideal_runtime) {
> +               /*
> +                * Favor non-idle group even in tick preemption
> +                */
> +               cse_is_idle = se_is_idle(curr);
> +               pse_is_idle = se_is_idle(se);
> +               if (unlikely(!cse_is_idle && pse_is_idle))
> +                       return;

This would make it so that we never have tick based preemption of a
non-idle entity by an idle entity. That's a recipe for starvation of
the idle entity, if the non-idle entity is cpu bound.

Beyond that though, I'm not quite sure the issue being solved here.
The large difference in weight between the idle and non-idle entity
means that the non-idle entity will not be preempted for quite a while
due to its ideal_runtime being quite high. The idle entity will
quickly be preempted on the next tick it takes due to the smaller
value of sysctl_sched_idle_min_granularity.

The wakeup check is useful for latency sensitive non-idle tasks.
However, in steady state competition between idle and non-idle, we
must allow some amount of round-robin.

> +
>                 resched_curr(rq_of(cfs_rq));
> +       }
>  }
>
>  static void
> --
> 2.20.1
>

Best,
Josh
  

在 2022/10/28 07:34, Josh Don 写道:
> Hi Chuyi,
> 
> On Thu, Oct 27, 2022 at 1:16 AM Chuyi Zhou <zhouchuyi@bytedance.com> wrote:
>>
>> The non-idle se dominates competition vs the idle se when they
>> are belong to the same group. We ensure that idle groups would not
>> preempt non-idle group in wakeup preemption(see check_preempt_wakeup()).
>> However, this can happen in tick preemption, since check_preempt_tick()
>> dose not check current/se is idle or not. This patch adds this check.
>>
>> Signed-off-by: Chuyi Zhou <zhouchuyi@bytedance.com>
>> ---
>>   kernel/sched/fair.c | 12 +++++++++++-
>>   1 file changed, 11 insertions(+), 1 deletion(-)
>>
>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
>> index e4a0b8bd941c..f3324b8753b3 100644
>> --- a/kernel/sched/fair.c
>> +++ b/kernel/sched/fair.c
>> @@ -4750,6 +4750,7 @@ static void
>>   check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
>>   {
>>          unsigned long ideal_runtime, delta_exec;
>> +       int cse_is_idle, pse_is_idle;
>>          struct sched_entity *se;
>>          s64 delta;
>>
>> @@ -4779,8 +4780,17 @@ check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
>>          if (delta < 0)
>>                  return;
>>
>> -       if (delta > ideal_runtime)
>> +       if (delta > ideal_runtime) {
>> +               /*
>> +                * Favor non-idle group even in tick preemption
>> +                */
>> +               cse_is_idle = se_is_idle(curr);
>> +               pse_is_idle = se_is_idle(se);
>> +               if (unlikely(!cse_is_idle && pse_is_idle))
>> +                       return;
> 
Hi Josh, thanks for your reply,
> This would make it so that we never have tick based preemption of a
> non-idle entity by an idle entity. That's a recipe for starvation of
> the idle entity, if the non-idle entity is cpu bound.
> 
> Beyond that though, I'm not quite sure the issue being solved here.
> The large difference in weight between the idle and non-idle entity
> means that the non-idle entity will not be preempted for quite a while
> due to its ideal_runtime being quite high. The idle entity will
> quickly be preempted on the next tick it takes due to the smaller
> value of sysctl_sched_idle_min_granularity.
> 
Actually, I did some tests and traced this issue. the result shows that 
this can happen with small probability. I also do some benchmarks. The 
result seems it has no performance harm, and we can reduce 2%～3% 
context switch when idle group & non-idle group are present at the same 
time. In addition, for throughput concern, I think we better let 
non-idle entity consume it's ideal_runtime. However, as you said, it may 
cause starvation of the idle entity.....

There is another question I would like to take this opportunity to 
consult you. In our production environment, in some cases, we want to 
adjust the weight/shares of the idle-cgroup which means we need to 
change the logic of sched_group_set_shares(), and it can increase the 
probability of the above issue. So, for what reasons did you prohibit 
users from changing weights of idle cgroup.

Thanks again for your review.

Best,
Chuyi
> The wakeup check is useful for latency sensitive non-idle tasks.
> However, in steady state competition between idle and non-idle, we
> must allow some amount of round-robin.
> 
>> +
>>                  resched_curr(rq_of(cfs_rq));
>> +       }
>>   }
>>
>>   static void
>> --
>> 2.20.1
>>
> 
> Best,
> Josh
  

On Thu, Oct 27, 2022 at 8:57 PM Chuyi Zhou <zhouchuyi@bytedance.com> wrote:
>
>
>
> 在 2022/10/28 07:34, Josh Don 写道:
> > Hi Chuyi,
> >
> > On Thu, Oct 27, 2022 at 1:16 AM Chuyi Zhou <zhouchuyi@bytedance.com> wrote:
> >>
> >> The non-idle se dominates competition vs the idle se when they
> >> are belong to the same group. We ensure that idle groups would not
> >> preempt non-idle group in wakeup preemption(see check_preempt_wakeup()).
> >> However, this can happen in tick preemption, since check_preempt_tick()
> >> dose not check current/se is idle or not. This patch adds this check.
> >>
> >> Signed-off-by: Chuyi Zhou <zhouchuyi@bytedance.com>
> >> ---
> >>   kernel/sched/fair.c | 12 +++++++++++-
> >>   1 file changed, 11 insertions(+), 1 deletion(-)
> >>
> >> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> >> index e4a0b8bd941c..f3324b8753b3 100644
> >> --- a/kernel/sched/fair.c
> >> +++ b/kernel/sched/fair.c
> >> @@ -4750,6 +4750,7 @@ static void
> >>   check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
> >>   {
> >>          unsigned long ideal_runtime, delta_exec;
> >> +       int cse_is_idle, pse_is_idle;
> >>          struct sched_entity *se;
> >>          s64 delta;
> >>
> >> @@ -4779,8 +4780,17 @@ check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
> >>          if (delta < 0)
> >>                  return;
> >>
> >> -       if (delta > ideal_runtime)
> >> +       if (delta > ideal_runtime) {
> >> +               /*
> >> +                * Favor non-idle group even in tick preemption
> >> +                */
> >> +               cse_is_idle = se_is_idle(curr);
> >> +               pse_is_idle = se_is_idle(se);
> >> +               if (unlikely(!cse_is_idle && pse_is_idle))
> >> +                       return;
> >
> Hi Josh, thanks for your reply,
> > This would make it so that we never have tick based preemption of a
> > non-idle entity by an idle entity. That's a recipe for starvation of
> > the idle entity, if the non-idle entity is cpu bound.
> >
> > Beyond that though, I'm not quite sure the issue being solved here.
> > The large difference in weight between the idle and non-idle entity
> > means that the non-idle entity will not be preempted for quite a while
> > due to its ideal_runtime being quite high. The idle entity will
> > quickly be preempted on the next tick it takes due to the smaller
> > value of sysctl_sched_idle_min_granularity.
> >
> Actually, I did some tests and traced this issue. the result shows that
> this can happen with small probability. I also do some benchmarks. The
> result seems it has no performance harm, and we can reduce 2%～3%
> context switch when idle group & non-idle group are present at the same
> time. In addition, for throughput concern, I think we better let
> non-idle entity consume it's ideal_runtime. However, as you said, it may
> cause starvation of the idle entity.....

I don't doubt it improves the performance of the non-idle entity, but
it is never advisable to indefinitely starve threads. That's a recipe
for hard lockups, priority inversion, etc.

Does your non-idle entity have a reasonable number of cpu.shares? You
can push out the round-robin interval by tuning CFS parameters without
completely starving the idle entity.

> There is another question I would like to take this opportunity to
> consult you. In our production environment, in some cases, we want to
> adjust the weight/shares of the idle-cgroup which means we need to
> change the logic of sched_group_set_shares(), and it can increase the
> probability of the above issue. So, for what reasons did you prohibit
> users from changing weights of idle cgroup.

The reason for limiting the control of weight for idle cgroups is to
match the semantics of the per-task SCHED_IDLE api, which gives
SCHED_IDLE threads minimum weight. The idea behind SCHED_IDLE is that
these entities are intended to soak "unused" cpu cycles, and should
give minimal interference to any non-idle thread. However, we don't
have strict priority between idle and non-idle, due to the potential
for starvation issues.

Perhaps you could clarify your use case a bit further. Why do you want
to change the weight? Is it to adjust the competition between two idle
groups, or something else?

>
> Thanks again for your review.
>
> Best,
> Chuyi
> > The wakeup check is useful for latency sensitive non-idle tasks.
> > However, in steady state competition between idle and non-idle, we
> > must allow some amount of round-robin.
> >
> >> +
> >>                  resched_curr(rq_of(cfs_rq));
> >> +       }
> >>   }
> >>
> >>   static void
> >> --
> >> 2.20.1
> >>
> >
> > Best,
> > Josh
  

在 2022/10/29 06:40, Josh Don 写道:
> On Thu, Oct 27, 2022 at 8:57 PM Chuyi Zhou <zhouchuyi@bytedance.com> wrote:
>>
>>
>>
>> 在 2022/10/28 07:34, Josh Don 写道:
>>> Hi Chuyi,
>>>
>>> On Thu, Oct 27, 2022 at 1:16 AM Chuyi Zhou <zhouchuyi@bytedance.com> wrote:
>>>>
>>>> The non-idle se dominates competition vs the idle se when they
>>>> are belong to the same group. We ensure that idle groups would not
>>>> preempt non-idle group in wakeup preemption(see check_preempt_wakeup()).
>>>> However, this can happen in tick preemption, since check_preempt_tick()
>>>> dose not check current/se is idle or not. This patch adds this check.
>>>>
>>>> Signed-off-by: Chuyi Zhou <zhouchuyi@bytedance.com>
>>>> ---
>>>>    kernel/sched/fair.c | 12 +++++++++++-
>>>>    1 file changed, 11 insertions(+), 1 deletion(-)
>>>>
>>>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
>>>> index e4a0b8bd941c..f3324b8753b3 100644
>>>> --- a/kernel/sched/fair.c
>>>> +++ b/kernel/sched/fair.c
>>>> @@ -4750,6 +4750,7 @@ static void
>>>>    check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
>>>>    {
>>>>           unsigned long ideal_runtime, delta_exec;
>>>> +       int cse_is_idle, pse_is_idle;
>>>>           struct sched_entity *se;
>>>>           s64 delta;
>>>>
>>>> @@ -4779,8 +4780,17 @@ check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
>>>>           if (delta < 0)
>>>>                   return;
>>>>
>>>> -       if (delta > ideal_runtime)
>>>> +       if (delta > ideal_runtime) {
>>>> +               /*
>>>> +                * Favor non-idle group even in tick preemption
>>>> +                */
>>>> +               cse_is_idle = se_is_idle(curr);
>>>> +               pse_is_idle = se_is_idle(se);
>>>> +               if (unlikely(!cse_is_idle && pse_is_idle))
>>>> +                       return;
>>>
>> Hi Josh, thanks for your reply,
>>> This would make it so that we never have tick based preemption of a
>>> non-idle entity by an idle entity. That's a recipe for starvation of
>>> the idle entity, if the non-idle entity is cpu bound.
>>>
>>> Beyond that though, I'm not quite sure the issue being solved here.
>>> The large difference in weight between the idle and non-idle entity
>>> means that the non-idle entity will not be preempted for quite a while
>>> due to its ideal_runtime being quite high. The idle entity will
>>> quickly be preempted on the next tick it takes due to the smaller
>>> value of sysctl_sched_idle_min_granularity.
>>>
>> Actually, I did some tests and traced this issue. the result shows that
>> this can happen with small probability. I also do some benchmarks. The
>> result seems it has no performance harm, and we can reduce 2%～3%
>> context switch when idle group & non-idle group are present at the same
>> time. In addition, for throughput concern, I think we better let
>> non-idle entity consume it's ideal_runtime. However, as you said, it may
>> cause starvation of the idle entity.....
> 
> I don't doubt it improves the performance of the non-idle entity, but
> it is never advisable to indefinitely starve threads. That's a recipe
> for hard lockups, priority inversion, etc.
> 
> Does your non-idle entity have a reasonable number of cpu.shares? You
> can push out the round-robin interval by tuning CFS parameters without
> completely starving the idle entity.
> 
>> There is another question I would like to take this opportunity to
>> consult you. In our production environment, in some cases, we want to
>> adjust the weight/shares of the idle-cgroup which means we need to
>> change the logic of sched_group_set_shares(), and it can increase the
>> probability of the above issue. So, for what reasons did you prohibit
>> users from changing weights of idle cgroup.
> 
> The reason for limiting the control of weight for idle cgroups is to
> match the semantics of the per-task SCHED_IDLE api, which gives
> SCHED_IDLE threads minimum weight. The idea behind SCHED_IDLE is that
> these entities are intended to soak "unused" cpu cycles, and should
> give minimal interference to any non-idle thread. However, we don't
> have strict priority between idle and non-idle, due to the potential
> for starvation issues.
> 
> Perhaps you could clarify your use case a bit further. Why do you want
> to change the weight? Is it to adjust the competition between two idle
> groups, or something else?
> 
Suppose we have two cgroups（idle & non-idle）in /sys/fs/cgroup/cpu.
Idle cgroup contains some offline service, such as beg data processing; 
non-idle cgroup contains some online service which have
higher priority to users and are sensitive to latency. We set
quota/period for idle cgroup which indicates it's *cpu limit*.
In general, we consider that the idle cgroup's cpu usage
closer to the limit, the better. However, when the system is busy,
the idle cgroups can only get little cpu resources with minimum weight. 
To cope with the above situation, we changed the default weight.

One more question is, why you think this patch can strave idle entity?
	
	/*
	 * Ensure that a task that missed wakeup preemption by a
	 * narrow margin doesn't have to wait for a full slice.
	 * This also mitigates buddy induced latencies under load.
	 */
	se = __pick_first_entity(cfs_rq);
	delta = curr->vruntime - se->vruntime;

	if (delta < 0)
		return;

	if (delta > ideal_runtime)
		resched_curr(rq_of(cfs_rq));

se can preempt curr only when
	curr->vruntime > se->vruntime &&
		curr->vruntime - se->vruntime > ideal_runtime
is true. I think the original purpose is that se doesn't have to wait 
for a full slice, reduce response time if se is latency sensitive.
This patch just let curr exhaust it's idleal_runtime when se is idle and 
curr is non-idle. Normally se will be choosed by pick_next_entity().

Maybe I missed something ?
Thanks


	

>>
>> Thanks again for your review.
>>
>> Best,
>> Chuyi
>>> The wakeup check is useful for latency sensitive non-idle tasks.
>>> However, in steady state competition between idle and non-idle, we
>>> must allow some amount of round-robin.
>>>
>>>> +
>>>>                   resched_curr(rq_of(cfs_rq));
>>>> +       }
>>>>    }
>>>>
>>>>    static void
>>>> --
>>>> 2.20.1
>>>>
>>>
>>> Best,
>>> Josh
  

On Mon, Oct 31, 2022 at 1:39 AM Chuyi Zhou <zhouchuyi@bytedance.com> wrote:
>
> >> 在 2022/10/28 07:34, Josh Don 写道:
> > The reason for limiting the control of weight for idle cgroups is to
> > match the semantics of the per-task SCHED_IDLE api, which gives
> > SCHED_IDLE threads minimum weight. The idea behind SCHED_IDLE is that
> > these entities are intended to soak "unused" cpu cycles, and should
> > give minimal interference to any non-idle thread. However, we don't
> > have strict priority between idle and non-idle, due to the potential
> > for starvation issues.
> >
> > Perhaps you could clarify your use case a bit further. Why do you want
> > to change the weight? Is it to adjust the competition between two idle
> > groups, or something else?
> >
> Suppose we have two cgroups（idle & non-idle）in /sys/fs/cgroup/cpu.
> Idle cgroup contains some offline service, such as beg data processing;
> non-idle cgroup contains some online service which have
> higher priority to users and are sensitive to latency. We set
> quota/period for idle cgroup which indicates it's *cpu limit*.
> In general, we consider that the idle cgroup's cpu usage
> closer to the limit, the better. However, when the system is busy,
> the idle cgroups can only get little cpu resources with minimum weight.
> To cope with the above situation, we changed the default weight.

I see. So you want the part of SCHED_IDLE that makes the entity highly
preemptible (and avoids preemption of non idle entities), but want to
adjust weight to reach a target cpu split? That seems a bit
counterintuitive to me, since by giving the idle entities higher
weight, you'll end up pushing out the round-robin latency for the
non-idle entities.

Worth noting that SCHED_IDLE is a bit of a CFS hack, but the intended
semantics of it are that these threads soak only "remaining cycles".
This comes with many implications beyond just weight. For example, a
cpu running only SCHED_IDLE entities is considered as "idle" from the
perspective of non-idle entities. If we give these idle entities
meaningful weight, we start to break assumptions there, for example
see sched_idle_cpu() and load balancing.

I wonder if maybe dusting off SCHED_BATCH is a better answer here, for
this type of use case (some amount of throughput "guarantee", but with
preemption properties similar to SCHED_IDLE). Peter, thoughts?

> One more question is, why you think this patch can strave idle entity?
>
>         /*
>          * Ensure that a task that missed wakeup preemption by a
>          * narrow margin doesn't have to wait for a full slice.
>          * This also mitigates buddy induced latencies under load.
>          */
>         se = __pick_first_entity(cfs_rq);
>         delta = curr->vruntime - se->vruntime;
>
>         if (delta < 0)
>                 return;
>
>         if (delta > ideal_runtime)
>                 resched_curr(rq_of(cfs_rq));
>
> se can preempt curr only when
>         curr->vruntime > se->vruntime &&
>                 curr->vruntime - se->vruntime > ideal_runtime
> is true. I think the original purpose is that se doesn't have to wait
> for a full slice, reduce response time if se is latency sensitive.
> This patch just let curr exhaust it's idleal_runtime when se is idle and
> curr is non-idle. Normally se will be choosed by pick_next_entity().
>
> Maybe I missed something ?
> Thanks

No that was my mistake, I accidentally thought this delta was being
applied to the 'if (delta_exec > ideal_runtime) {' above in
check_preempt_tick().

Some weirdness about this change though, is that if there is a
non-idle current entity, and the two next entities on the cfs_rq are
idle and non-idle respectively, we'll now take longer to preempt the
on-cpu non-idle entity, because the non-idle entity on the cfs_rq is
'hidden' by the idle 'first' entity. Wakeup preemption is different
because we're always directly comparing the current entity with the
newly woken entity.

Best,
Josh
  

在 2022/11/1 06:44, Josh Don 写道:
> On Mon, Oct 31, 2022 at 1:39 AM Chuyi Zhou <zhouchuyi@bytedance.com> wrote:
>>
>>>> 在 2022/10/28 07:34, Josh Don 写道:
>>> The reason for limiting the control of weight for idle cgroups is to
>>> match the semantics of the per-task SCHED_IDLE api, which gives
>>> SCHED_IDLE threads minimum weight. The idea behind SCHED_IDLE is that
>>> these entities are intended to soak "unused" cpu cycles, and should
>>> give minimal interference to any non-idle thread. However, we don't
>>> have strict priority between idle and non-idle, due to the potential
>>> for starvation issues.
>>>
>>> Perhaps you could clarify your use case a bit further. Why do you want
>>> to change the weight? Is it to adjust the competition between two idle
>>> groups, or something else?
>>>
>> Suppose we have two cgroups（idle & non-idle）in /sys/fs/cgroup/cpu.
>> Idle cgroup contains some offline service, such as beg data processing;
>> non-idle cgroup contains some online service which have
>> higher priority to users and are sensitive to latency. We set
>> quota/period for idle cgroup which indicates it's *cpu limit*.
>> In general, we consider that the idle cgroup's cpu usage
>> closer to the limit, the better. However, when the system is busy,
>> the idle cgroups can only get little cpu resources with minimum weight.
>> To cope with the above situation, we changed the default weight.
> 
> I see. So you want the part of SCHED_IDLE that makes the entity highly
> preemptible (and avoids preemption of non idle entities), but want to
> adjust weight to reach a target cpu split? That seems a bit
> counterintuitive to me, since by giving the idle entities higher
> weight, you'll end up pushing out the round-robin latency for the
> non-idle entities.
> 
> Worth noting that SCHED_IDLE is a bit of a CFS hack, but the intended
> semantics of it are that these threads soak only "remaining cycles".
> This comes with many implications beyond just weight. For example, a
> cpu running only SCHED_IDLE entities is considered as "idle" from the
> perspective of non-idle entities. If we give these idle entities
> meaningful weight, we start to break assumptions there, for example
> see sched_idle_cpu() and load balancing.
> 
> I wonder if maybe dusting off SCHED_BATCH is a better answer here, for
> this type of use case (some amount of throughput "guarantee", but with
> preemption properties similar to SCHED_IDLE). Peter, thoughts?
> 
>> One more question is, why you think this patch can strave idle entity?
>>
>>          /*
>>           * Ensure that a task that missed wakeup preemption by a
>>           * narrow margin doesn't have to wait for a full slice.
>>           * This also mitigates buddy induced latencies under load.
>>           */
>>          se = __pick_first_entity(cfs_rq);
>>          delta = curr->vruntime - se->vruntime;
>>
>>          if (delta < 0)
>>                  return;
>>
>>          if (delta > ideal_runtime)
>>                  resched_curr(rq_of(cfs_rq));
>>
>> se can preempt curr only when
>>          curr->vruntime > se->vruntime &&
>>                  curr->vruntime - se->vruntime > ideal_runtime
>> is true. I think the original purpose is that se doesn't have to wait
>> for a full slice, reduce response time if se is latency sensitive.
>> This patch just let curr exhaust it's idleal_runtime when se is idle and
>> curr is non-idle. Normally se will be choosed by pick_next_entity().
>>
>> Maybe I missed something ?
>> Thanks
> 
> No that was my mistake, I accidentally thought this delta was being
> applied to the 'if (delta_exec > ideal_runtime) {' above in
> check_preempt_tick().
> 
> Some weirdness about this change though, is that if there is a
> non-idle current entity, and the two next entities on the cfs_rq are
> idle and non-idle respectively, we'll now take longer to preempt the
> on-cpu non-idle entity, because the non-idle entity on the cfs_rq is
> 'hidden' by the idle 'first' entity. Wakeup preemption is different
> because we're always directly comparing the current entity with the
> newly woken entity.
> 
You are right, this can happen with high probability.
This patch just compared the curr with the first entity in
the tick, and it seems hard to consider all the other entity
in cfs_rq.

So, what specific negative effects this situation would cause?
For example, the "hidden" non-idle entity's latency will be worse
than before?

Thanks for your patient reply and guidance!

BenchMark
=======================================================
All the benchmark are done in /sys/fs/cgroup/cpu/online,
which is a a normal cpu cgroup. In /sys/fs/cgroup/cpu/
offline, 'perf bench sched messaging -g 1 -l 1000000000'
is ran continuously. Besides, we set offline cgroup as idle.


schbench
                      baseline patched
Lat 50.0th-qrtle-1	6	6
Lat 75.0th-qrtle-1	6	7
Lat 90.0th-qrtle-1	7	7
Lat 95.0th-qrtle-1	7	7
Lat 99.0th-qrtle-1	10	10
Lat 99.5th-qrtle-1	11	11
Lat 99.9th-qrtle-1	12	12
Lat 50.0th-qrtle-2	6	7
Lat 75.0th-qrtle-2	7	8
Lat 90.0th-qrtle-2	8	9
Lat 95.0th-qrtle-2	9	10
Lat 99.0th-qrtle-2	11	12
Lat 99.5th-qrtle-2	12	13
Lat 99.9th-qrtle-2	19	18
Lat 50.0th-qrtle-4	8	9
Lat 75.0th-qrtle-4	10	11
Lat 90.0th-qrtle-4	11	13
Lat 95.0th-qrtle-4	12	14
Lat 99.0th-qrtle-4	15	16
Lat 99.5th-qrtle-4	17	18
Lat 99.9th-qrtle-4	25	23
Lat 50.0th-qrtle-8	11	13
Lat 75.0th-qrtle-8	15	17
Lat 90.0th-qrtle-8	18	20
Lat 95.0th-qrtle-8	19	22
Lat 99.0th-qrtle-8	23	26
Lat 99.5th-qrtle-8	25	28
Lat 99.9th-qrtle-8	41	48
Lat 50.0th-qrtle-16	20	21
Lat 75.0th-qrtle-16	27	28
Lat 90.0th-qrtle-16	32	33
Lat 95.0th-qrtle-16	35	36
Lat 99.0th-qrtle-16	44	43
Lat 99.5th-qrtle-16	53	47
Lat 99.9th-qrtle-16	1310	247
Lat 50.0th-qrtle-23	28	28
Lat 75.0th-qrtle-23	39	39
Lat 90.0th-qrtle-23	46	46
Lat 95.0th-qrtle-23	50	50
Lat 99.0th-qrtle-23	62	58
Lat 99.5th-qrtle-23	449	67
Lat 99.9th-qrtle-23	5516	906

hackbench-process-pipes

                   baseline                patched
Amean     1        0.6540 (   0.00%)      0.6480 (   0.92%)
Amean     4        0.8023 (   0.00%)      0.7860 (   2.04%)
Amean     7        1.3543 (   0.00%)      1.3780 (  -1.75%)
Amean     12       2.2653 (   0.00%)      2.2853 (  -0.88%)
Amean     21       4.7187 (   0.00%)      5.7217 * -21.26%*
Amean     30       7.3217 (   0.00%)      7.5797 (  -3.52%)
Amean     48       7.8410 (   0.00%)      7.7687 (   0.92%)
Amean     79      10.6037 (   0.00%)     10.9147 (  -2.93%)
Amean     110     11.7623 (   0.00%)     12.2150 *  -3.85%*
Amean     141     11.9980 (   0.00%)     13.0153 (  -8.48%)
Amean     172     13.6023 (   0.00%)     13.3613 (   1.77%)
Amean     203     15.2607 (   0.00%)     15.5010 (  -1.57%)
Amean     234     17.1007 (   0.00%)     17.2960 (  -1.14%)
Amean     265     18.1547 (   0.00%)     18.2500 (  -0.53%)

hackbench-process-sockets

                   baseline                patched
Amean     1        0.8340 (   0.00%)      0.8290 (   0.60%)
Amean     4        1.7260 (   0.00%)      1.7340 (  -0.46%)
Amean     7        2.8583 (   0.00%)      2.8520 (   0.22%)
Amean     12       4.6833 (   0.00%)      4.7387 *  -1.18%*
Amean     21       7.9787 (   0.00%)      7.9723 (   0.08%)
Amean     30      11.1853 (   0.00%)     11.2573 *  -0.64%*
Amean     48      17.7430 (   0.00%)     17.7410 (   0.01%)
Amean     79      29.4067 (   0.00%)     29.4900 (  -0.28%)
Amean     110     41.4427 (   0.00%)     41.6630 *  -0.53%*
Amean     141     53.0730 (   0.00%)     53.4000 *  -0.62%*
Amean     172     64.6120 (   0.00%)     65.0123 *  -0.62%*
Amean     203     76.3320 (   0.00%)     76.6423 (  -0.41%)
Amean     234     87.9563 (   0.00%)     88.4263 *  -0.53%*
Amean     265     99.5607 (   0.00%)    100.3303 *  -0.77%*
Amean     296    111.0987 (   0.00%)    112.1843 *  -0.98%*

hackbench-thread-pipes

                   baseline                patched
Amean     1        0.7007 (   0.00%)      0.7087 (  -1.14%)
Amean     4        0.8553 (   0.00%)      0.8197 (   4.17%)
Amean     7        1.4327 (   0.00%)      1.4503 (  -1.23%)
Amean     12       2.3497 (   0.00%)      2.3653 (  -0.67%)
Amean     21       5.7677 (   0.00%)      5.7340 (   0.58%)
Amean     30       7.4670 (   0.00%)      7.4693 (  -0.03%)
Amean     48       8.2720 (   0.00%)      8.2417 (   0.37%)
Amean     79      11.3247 (   0.00%)     11.1930 (   1.16%)
Amean     110     13.2153 (   0.00%)     13.1750 (   0.31%)
Amean     141     14.3250 (   0.00%)     14.9840 (  -4.60%)
Amean     172     16.7150 (   0.00%)     15.8843 (   4.97%)
Amean     203     17.7397 (   0.00%)     17.4787 (   1.47%)
Amean     234     19.4990 (   0.00%)     21.0840 *  -8.13%*
Amean     265     21.6867 (   0.00%)     22.1387 (  -2.08%)
Amean     296     24.5460 (   0.00%)     25.0093 (  -1.89%)

hackbench-thread-sockets

                   baseline                patched
Amean     1        0.8930 (   0.00%)      0.8937 (  -0.07%)
Amean     4        1.7803 (   0.00%)      1.7853 (  -0.28%)
Amean     7        2.9317 (   0.00%)      2.9563 (  -0.84%)
Amean     12       4.8313 (   0.00%)      4.8227 (   0.18%)
Amean     21       8.1783 (   0.00%)      8.1900 (  -0.14%)
Amean     30      11.5653 (   0.00%)     11.5180 (   0.41%)
Amean     48      18.3100 (   0.00%)     18.2730 (   0.20%)
Amean     79      30.2047 (   0.00%)     30.2960 (  -0.30%)
Amean     110     42.3330 (   0.00%)     42.3503 (  -0.04%)
Amean     141     54.1853 (   0.00%)     54.1483 (   0.07%)
Amean     172     66.0127 (   0.00%)     65.8043 *   0.32%*
Amean     203     77.9667 (   0.00%)     77.6113 (   0.46%)
Amean     234     89.6077 (   0.00%)     89.3893 *   0.24%*
Amean     265    101.2477 (   0.00%)    101.2773 (  -0.03%)
Amean     296    113.1667 (   0.00%)    113.2967 (  -0.11%)

perfpipe
                      baseline                patched
Min       Time        8.86 (   0.00%)        8.83 (   0.39%)
1st-qrtle Time        8.92 (   0.00%)        9.00 (  -0.94%)
2nd-qrtle Time        8.99 (   0.00%)        9.06 (  -0.76%)
3rd-qrtle Time        9.06 (   0.00%)        9.08 (  -0.17%)
Max-1     Time        8.86 (   0.00%)        8.83 (   0.39%)
Max-5     Time        8.86 (   0.00%)        8.83 (   0.39%)
Max-10    Time        8.88 (   0.00%)        8.94 (  -0.70%)
Max-90    Time        9.10 (   0.00%)        9.13 (  -0.40%)
Max-95    Time        9.13 (   0.00%)        9.20 (  -0.75%)
Max-99    Time        9.15 (   0.00%)        9.22 (  -0.84%)
Max       Time        9.16 (   0.00%)        9.28 (  -1.23%)
Amean     Time        9.00 (   0.00%)        9.06 *  -0.65%*
Stddev    Time        0.09 (   0.00%)        0.09 (  -3.71%)
CoeffVar  Time        0.95 (   0.00%)        0.98 (  -3.04%)
BAmean-99 Time        8.99 (   0.00%)        9.05 (  -0.63%)
BAmean-95 Time        8.99 (   0.00%)        9.04 (  -0.62%)
BAmean-90 Time        8.98 (   0.00%)        9.04 (  -0.62%)
BAmean-75 Time        8.96 (   0.00%)        9.02 (  -0.67%)
BAmean-50 Time        8.93 (   0.00%)        9.00 (  -0.75%)
BAmean-25 Time        8.89 (   0.00%)        8.95 (  -0.68%)

tbench4 Latency
                              baseline                patched
Amean     latency-1          0.73 (   0.00%)        0.24 *  66.90%*
Amean     latency-2          0.93 (   0.00%)        0.42 *  54.94%*
Amean     latency-4          0.99 (   0.00%)        0.41 *  58.49%*
Amean     latency-8          1.03 (   0.00%)        0.82 *  20.47%*
Amean     latency-16         1.36 (   0.00%)        1.07 *  21.12%*
Amean     latency-32         2.93 (   0.00%)        1.38 *  52.74%*
Amean     latency-64         4.21 (   0.00%)        5.53 * -31.25%*
Amean     latency-128        9.76 (   0.00%)       10.67 *  -9.28%*
Amean     latency-256       23.68 (   0.00%)       12.17 *  48.59%*
Amean     latency-384       61.68 (   0.00%)       57.65 (   6.54%)

tbench4 Throughput (misleading but traditional)
                    baseline                patched
Hmean     1        287.60 (   0.00%)      287.83 (   0.08%)
Hmean     2        567.87 (   0.00%)      582.14 *   2.51%*
Hmean     4       1126.12 (   0.00%)     1143.32 *   1.53%*
Hmean     8       2138.48 (   0.00%)     2212.82 *   3.48%*
Hmean     16      3955.46 (   0.00%)     4180.57 *   5.69%*
Hmean     32      6838.94 (   0.00%)     6879.27 *   0.59%*
Hmean     64      9462.36 (   0.00%)     9267.99 *  -2.05%*
Hmean     128    17156.42 (   0.00%)    17901.01 *   4.34%*
Hmean     256    15823.09 (   0.00%)    14647.72 *  -7.43%*
Hmean     384    15366.71 (   0.00%)    16592.95 *   7.98%*

> Best,
> Josh
  

Hi Josh,

On 11/1/22 6:44 AM, Josh Don wrote:
> ...nit...
> Some weirdness about this change though, is that if there is a
> non-idle current entity, and the two next entities on the cfs_rq are
> idle and non-idle respectively, we'll now take longer to preempt the
> on-cpu non-idle entity, because the non-idle entity on the cfs_rq is
> 'hidden' by the idle 'first' entity. Wakeup preemption is different
> because we're always directly comparing the current entity with the
> newly woken entity.

Indeed. The hidden non-idle entity might run longer at the cost of
delayed preemption. This behavior is not compliant to the SCHED_NORMAL
semantics. But we also can't tell from here that the non-idle entity
contains SCHED_NORMAL tasks or not. As long as the entities do not
aware of the schedule policies, this ambiguity exists.

Best,
Abel
  

> > Some weirdness about this change though, is that if there is a
> > non-idle current entity, and the two next entities on the cfs_rq are
> > idle and non-idle respectively, we'll now take longer to preempt the
> > on-cpu non-idle entity, because the non-idle entity on the cfs_rq is
> > 'hidden' by the idle 'first' entity. Wakeup preemption is different
> > because we're always directly comparing the current entity with the
> > newly woken entity.
> >
> You are right, this can happen with high probability.
> This patch just compared the curr with the first entity in
> the tick, and it seems hard to consider all the other entity
> in cfs_rq.
>
> So, what specific negative effects this situation would cause?
> For example, the "hidden" non-idle entity's latency will be worse
> than before?

As Abel points out in his email, it can push out the time it'll take
to switch to the other non-idle entity. The change might boost some
benchmarks numbers, but I don't think it is conclusive enough to say
it is a generically beneficial improvement that should be integrated.

By the way, I'm curious if you modified any of the sched_idle_cpu()
and related load balancing around idle entities given that you've made
it so that idle entities can have arbitrary weight (since, as I
described in my prior email, this can otherwise cause issues there).
  

在 2022/11/2 07:39, Josh Don 写道:
>>> Some weirdness about this change though, is that if there is a
>>> non-idle current entity, and the two next entities on the cfs_rq are
>>> idle and non-idle respectively, we'll now take longer to preempt the
>>> on-cpu non-idle entity, because the non-idle entity on the cfs_rq is
>>> 'hidden' by the idle 'first' entity. Wakeup preemption is different
>>> because we're always directly comparing the current entity with the
>>> newly woken entity.
>>>
>> You are right, this can happen with high probability.
>> This patch just compared the curr with the first entity in
>> the tick, and it seems hard to consider all the other entity
>> in cfs_rq.
>>
>> So, what specific negative effects this situation would cause?
>> For example, the "hidden" non-idle entity's latency will be worse
>> than before?
> 
> As Abel points out in his email, it can push out the time it'll take
> to switch to the other non-idle entity. The change might boost some
> benchmarks numbers, but I don't think it is conclusive enough to say
> it is a generically beneficial improvement that should be integrated.
> 
> By the way, I'm curious if you modified any of the sched_idle_cpu()
> and related load balancing around idle entities given that you've made
> it so that idle entities can have arbitrary weight (since, as I
> described in my prior email, this can otherwise cause issues there).
Hi Josh,

Indeed, there are some subtle influences here. For example,
find_idlest_group_cpu() may return i if sched_idle_cpu(i) is true.
In nromal cases, it would work well, however it would be unpredictable
if we change the default minumum weight. I have not found a suitable 
solution yet....

Thanks for your guidance.
  

On 2022/11/2 Josh Don wrote:
>>> Some weirdness about this change though, is that if there is a
>>> non-idle current entity, and the two next entities on the cfs_rq are
>>> idle and non-idle respectively, we'll now take longer to preempt the
>>> on-cpu non-idle entity, because the non-idle entity on the cfs_rq is
>>> 'hidden' by the idle 'first' entity. Wakeup preemption is different
>>> because we're always directly comparing the current entity with the
>>> newly woken entity.
>>>
>> You are right, this can happen with high probability.
>> This patch just compared the curr with the first entity in
>> the tick, and it seems hard to consider all the other entity
>> in cfs_rq.
>>
>> So, what specific negative effects this situation would cause?
>> For example, the "hidden" non-idle entity's latency will be worse
>> than before?
> 
> As Abel points out in his email, it can push out the time it'll take
> to switch to the other non-idle entity. The change might boost some
> benchmarks numbers, but I don't think it is conclusive enough to say
> it is a generically beneficial improvement that should be integrated.
> 
> By the way, I'm curious if you modified any of the sched_idle_cpu()
> and related load balancing around idle entities given that you've made
> it so that idle entities can have arbitrary weight (since, as I
> described in my prior email, this can otherwise cause issues there).

If we want to make it easier for non-idle tasks to preempt idle tasks in 
tick, maybe we can consider lowering sysctl_sched_idle_min_granularity. 
Of course this may not ensure that non-idle tasks successfully preempt 
idle tasks every time, but it seems to be more beneficial for non-idle 
tasks.

IMHO, even if it is allowed to increase the weight of non-idle, it seems 
that we can make it easier for non-idle tasks to preempt idle tasks by 
lowering sysctl_sched_idle_min_granularity.

Thanks,
Hao
  

On Thu, Nov 3, 2022 at 8:49 PM Hao Jia <jiahao.os@bytedance.com> wrote:
>
>
>
> On 2022/11/2 Josh Don wrote:
> >>> Some weirdness about this change though, is that if there is a
> >>> non-idle current entity, and the two next entities on the cfs_rq are
> >>> idle and non-idle respectively, we'll now take longer to preempt the
> >>> on-cpu non-idle entity, because the non-idle entity on the cfs_rq is
> >>> 'hidden' by the idle 'first' entity. Wakeup preemption is different
> >>> because we're always directly comparing the current entity with the
> >>> newly woken entity.
> >>>
> >> You are right, this can happen with high probability.
> >> This patch just compared the curr with the first entity in
> >> the tick, and it seems hard to consider all the other entity
> >> in cfs_rq.
> >>
> >> So, what specific negative effects this situation would cause?
> >> For example, the "hidden" non-idle entity's latency will be worse
> >> than before?
> >
> > As Abel points out in his email, it can push out the time it'll take
> > to switch to the other non-idle entity. The change might boost some
> > benchmarks numbers, but I don't think it is conclusive enough to say
> > it is a generically beneficial improvement that should be integrated.
> >
> > By the way, I'm curious if you modified any of the sched_idle_cpu()
> > and related load balancing around idle entities given that you've made
> > it so that idle entities can have arbitrary weight (since, as I
> > described in my prior email, this can otherwise cause issues there).
>
> If we want to make it easier for non-idle tasks to preempt idle tasks in
> tick, maybe we can consider lowering sysctl_sched_idle_min_granularity.
> Of course this may not ensure that non-idle tasks successfully preempt
> idle tasks every time, but it seems to be more beneficial for non-idle
> tasks.
>
> IMHO, even if it is allowed to increase the weight of non-idle, it seems
> that we can make it easier for non-idle tasks to preempt idle tasks by
> lowering sysctl_sched_idle_min_granularity.

Yep, although the effectiveness is partially limited by whatever the
HZ is set to for the scheduling tick.

>
> Thanks,
> Hao
  

On 11/2/22 7:39 AM, Josh Don wrote:
>>> Some weirdness about this change though, is that if there is a
>>> non-idle current entity, and the two next entities on the cfs_rq are
>>> idle and non-idle respectively, we'll now take longer to preempt the
>>> on-cpu non-idle entity, because the non-idle entity on the cfs_rq is
>>> 'hidden' by the idle 'first' entity. Wakeup preemption is different
>>> because we're always directly comparing the current entity with the
>>> newly woken entity.
>>>
>> You are right, this can happen with high probability.
>> This patch just compared the curr with the first entity in
>> the tick, and it seems hard to consider all the other entity
>> in cfs_rq.
>>
>> So, what specific negative effects this situation would cause?
>> For example, the "hidden" non-idle entity's latency will be worse
>> than before?
> 
> As Abel points out in his email, it can push out the time it'll take
> to switch to the other non-idle entity. The change might boost some
> benchmarks numbers, but I don't think it is conclusive enough to say
> it is a generically beneficial improvement that should be integrated.

Agree.

> 
> By the way, I'm curious if you modified any of the sched_idle_cpu()
> and related load balancing around idle entities given that you've made
> it so that idle entities can have arbitrary weight (since, as I
> described in my prior email, this can otherwise cause issues there).

Being able to change idle entities' weight can bring nothing but
convenience, because it can also be achieved by modifying all their
siblings' weight. Which seems not a strong reason to get merged.

And I'm also thinking that, although rare, a non-idle group can also
have a weight close or even equal to 3. I guess some users who made
this kind of setting might only want to benefit from the preemption
at wakeup? Nevertheless this setting is supported now :)

Best,
	Abel
  

On Thu, Nov 10, 2022 at 7:50 PM Abel Wu <wuyun.abel@bytedance.com> wrote:
>
> > By the way, I'm curious if you modified any of the sched_idle_cpu()
> > and related load balancing around idle entities given that you've made
> > it so that idle entities can have arbitrary weight (since, as I
> > described in my prior email, this can otherwise cause issues there).
>
> Being able to change idle entities' weight can bring nothing but
> convenience, because it can also be achieved by modifying all their
> siblings' weight. Which seems not a strong reason to get merged.
>
> And I'm also thinking that, although rare, a non-idle group can also
> have a weight close or even equal to 3. I guess some users who made
> this kind of setting might only want to benefit from the preemption
> at wakeup? Nevertheless this setting is supported now :)

Strongly disagree with this; part of the semantics for idle relies on
the minimum weight value. It is true that this behavior gets a little
weirder if siblings also have close to min weight, but this is an
artifact of the fact that SCHED_IDLE is built into CFS rather than
being a separate scheduling class. The minimum weight in general is
assumed by load balance, etc. for the purpose of placing non-idle
entities.

For consistency with the per-task SCHED_IDLE behavior, which
effectively makes idle tasks have the max nice value, we also need to
match the cgroup idle behavior.

I think the use case you're describing would honestly be better served
by extending SCHED_BATCH with the preemption properties and using
that, rather than try to overload SCHED_IDLE here. There's a
difference between non-interactive entities that are ok getting
aggressively preempted, vs "idle" entities that should really only be
soaking the remaining cycles on the machine.

Best,
Josh