Re: [PATCH v12 8/8] sched/fair: Add latency list

[Vincent Guittot]

On Tue, 7 Mar 2023 at 11:52, Shrikanth Hegde <sshegde@xxxxxxxxxxxxxxxxxx> wrote:
>
>
> On 3/7/23 3:49 PM, Vincent Guittot wrote:
> > Le mardi 07 mars 2023 à 00:34:49 (+0530), Shrikanth Hegde a écrit :
> >>> Le lundi 06 mars 2023 à 17:03:30 (+0530), Shrikanth Hegde a écrit :
> >>>> On 3/5/23 6:33 PM, Vincent Guittot wrote:
> >>>>> On Sat, 4 Mar 2023 at 16:13, Shrikanth Hegde <sshegde@xxxxxxxxxxxxxxxxxx> wrote:
> >>>>>> On 3/3/23 10:01 PM, Vincent Guittot wrote:
> >>>>>>> Le jeudi 02 mars 2023 à 23:37:52 (+0530), Shrikanth Hegde a écrit :
> >>>>>>>> On 3/2/23 8:30 PM, Shrikanth Hegde wrote:
> >>>>>>>>> On 3/2/23 6:47 PM, Vincent Guittot wrote:
> >>>>>>>>>> On Thu, 2 Mar 2023 at 12:00, Shrikanth Hegde <sshegde@xxxxxxxxxxxxxxxxxx> wrote:
> >>>>>>>>>>> On 3/2/23 1:20 PM, Vincent Guittot wrote:
> >>>>>>>>>>>> On Wed, 1 Mar 2023 at 19:48, shrikanth hegde <sshegde@xxxxxxxxxxxxxxxxxx> wrote:
> >>>>>>>>>>>>> On 2/24/23 3:04 PM, Vincent Guittot wrote:
> >>>>>>> [...]
> >>>>>>>
> >>>>>>>>>>>>> Ran the schbench and hackbench with this patch series. Here comparison is
> >>>>>>>>>>>>> between 6.2 stable tree, 6.2 + Patch and 6.2 + patch + above re-arrange of
> >>>>>>>>>>>>> latency_node. Ran two cgroups, in one cgroup running stress-ng at 50%(group1)
> >>>>>>>>>>>>> and other is running these benchmarks (group2). Set the latency nice
> >>>>>>>>>>>>> of group2 to -20. These are run on Power system with 12 cores with SMT=8.
> >>>>>>>>>>>>> Total of 96 CPU.
> >>>>>>>>>>>>>
> >>>>>>>>>>>>> schbench gets lower latency compared to stabletree. Whereas hackbench seems
> >>>>>>>>>>>>> to regress under this case. Maybe i am doing something wrong. I will re-run
> >>>>>>>>>>>>> and attach the numbers to series.
> >>>>>>>>>>>>> Please suggest if any variation in the test i need to try.
> >>>>>>>>>>>> hackbench takes advanatge of a latency nice 19 as it mainly wants to
> >>>>>>>>>>>> run longer slice to move forward rather than preempting others all the
> >>>>>>>>>>>> time
> >>>>>>>>>>> hackbench still seems to regress in different latency nice values compared to
> >>>>>>>>>>> baseline of 6.2 in this case. up to 50% in some cases.
> >>>>>>>>>>>
> >>>>>>>>>>> 12 core powerpc system  with SMT=8 i.e 96 CPU
> >>>>>>>>>>> running 2 CPU cgroups. No quota assigned.
> >>>>>>>>>>> 1st cgroup is running stress-ng with 48 threads. Consuming 50% of CPU.
> >>>>>>>>>>> latency is not changed for this cgroup.
> >>>>>>>>>>> 2nd cgroup is running hackbench. This cgroup is assigned the different latency
> >>>>>>>>>>> nice values of 0, -20 and 19.
> >>>>>>>>>> According to your other emails, you are using the cgroup interface and
> >>>>>>>>>> not the task's one. Do I get it right ?
> >>>>>>>>> right. I create cgroup, attach bash command with echo $$,
> >>>>>>>>> assign the latency nice to cgroup, and run hackbench from that bash prompt.
> >>>>>>>>>
> >>>>>>>>>> I haven't run test such tests in a cgroup but at least the test with
> >>>>>>>>>> latency_nice == 0 should not make any noticeable difference. Does this
> >>>>>>>>>> include the re-arrange patch that you have proposed previously ?
> >> Ran the test on a different system altogether. I don't see similar regression there.
> >> In fact latency nice is helping in reducing the latency as expected.
> >> It is much bigger system with 60 cores. i.e total of 480 cpu.
> >>
> >> Tested in the same way. created two cgroups. one is running the micro benchmarks
> >> and other is running stress-ng at different utilization point.
> >> This data is at 50% utilization point. Similar observations w.r.t latency
> >> is seen at 0%, 25%, 75% and 100% utilization as well.
> >>
> > Thanks for testing on a different system which seems to get results aligned with what
> > Prateek and I have seen on our system.
> >
> >
> >> ==========
> >> schbench
> >> ==========
> >>             6.2            6.2 + V12 + LN=0
> >> Groups: 1
> >> 50.0th:                 14.0             12.5
> >> 75.0th:                 16.5             14.0
> >> 90.0th:                 18.5             15.5
> >> 95.0th:                 20.5             17.0
> >> 99.0th:                 27.5             21.0
> >> 99.5th:                 36.0             23.5
> >> Groups: 2
> >> 50.0th:                 14.0             16.0
> >> 75.0th:                 17.0             18.0
> >> 90.0th:                 20.0             21.0
> >> 95.0th:                 23.0             23.0
> >> 99.0th:                 71.0             34.0
> >> 99.5th:               1170.0             96.0
> >> 99.9th:               5088.0           3212.0
> >> Groups: 4
> >> 50.0th:                 20.5             19.5
> >> 75.0th:                 24.5             22.5
> >> 90.0th:                 31.0             26.0
> >> 95.0th:                260.5             28.0
> >> 99.0th:               3644.0             35.0
> >> 99.5th:               5152.0             44.5
> >> 99.9th:               8076.0            168.5
> >> Groups: 8
> >> 50.0th:                 26.0             25.5
> >> 75.0th:                 32.5             31.5
> >> 90.0th:                 41.5             36.5
> >> 95.0th:                794.0             39.5
> >> 99.0th:               5992.0             66.0
> >> 99.5th:               7208.0            159.0
> >> 99.9th:               9392.0           1604.0
> >> Groups: 16
> >> 50.0th:                 37.5             34.0
> >> 75.0th:                 49.5             44.5
> >> 90.0th:                 70.0             53.5
> >> 95.0th:               1284.0             58.5
> >> 99.0th:               5600.0            102.5
> >> 99.5th:               7216.0            368.5
> >> 99.9th:               9328.0           5192.0
> >> Groups: 32
> >> 50.0th:                 59.0             54.5
> >> 75.0th:                 83.0             74.5
> >> 90.0th:                118.5             91.0
> >> 95.0th:               1921.0            100.5
> >> 99.0th:               6672.0            317.0
> >> 99.5th:               8252.0           2264.0
> >> 99.9th:              10448.0           8388.0
> >>
> >>
> >> ===========
> >> hackbench
> >> ==========
> >>
> >> type                 Groups      6.2      | 6.2 + V12 + LN=0
> >> Process               10         0.19     |  0.19
> >> Process               20         0.34     |  0.34
> >> Process               30         0.45     |  0.44
> >> Process               40         0.58     |  0.57
> >> Process               50         0.70     |  0.69
> >> Process               60         0.82     |  0.80
> >> thread                10         0.20     |  0.20
> >> thread                20         0.36     |  0.36
> >> Process(Pipe)         10         0.24     |  0.21
> >> Process(Pipe)         20         0.46     |  0.40
> >> Process(Pipe)         30         0.65     |  0.58
> >> Process(Pipe)         40         0.90     |  0.68
> >> Process(Pipe)         50         1.04     |  0.83
> >> Process(Pipe)         60         1.16     |  0.86
> >> thread(Pipe)          10         0.19     |  0.18
> >> thread(Pipe)          20         0.46     |  0.37
> >>
> >>
> > [...]
> >
> >>>>>>
> >>>>>> Do you want me to try any other experiment on this further?
> >>>>> Yes, would be good to know which of the 3 changes in the patch create
> >>>>> the regression
> >>>>>
> >>>>> I suspect the 1st change to be the root cause of your problem but It
> >>>>> would be good if you can confirm my assumption with some tests
> >>>>>
> >>>>> Thanks
> >>>> Applied each change individually. 3rd change seems to cause the regression.
> >>>> Kept only the 3rd change and numbers are same as stable 6.2 for latency nice
> >>>> value of 0.
> >>> Ok, it's the patch 1 that aims to prevent some unfairness with low weight
> >>> waking task. And your platform probably falls in the last part of the commit:
> >>>
> >>> " Strictly speaking, we should use cfs->min_vruntime instead of
> >>> curr->vruntime but it doesn't worth the additional overhead and complexity
> >>> as the vruntime of current should be close to min_vruntime if not equal."
> >>>
> >>> Could you try the patch below on top of v12 ?
> >>>
> >>> ---
> >>>  kernel/sched/fair.c | 21 +++++++++++----------
> >>>  1 file changed, 11 insertions(+), 10 deletions(-)
> >>>
> >>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> >>> index e2aeb4511686..77b03a280912 100644
> >>> --- a/kernel/sched/fair.c
> >>> +++ b/kernel/sched/fair.c
> >>> @@ -5049,7 +5049,7 @@ set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
> >>>  }
> >>>
> >>>  static int
> >>> -wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
> >>> +wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se, struct cfs_rq *cfs_rq);
> >>>
> >>>  /*
> >>>   * Pick the next process, keeping these things in mind, in this order:
> >>> @@ -5088,16 +5088,16 @@ pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
> >>>                             second = curr;
> >>>             }
> >>>
> >>> -           if (second && wakeup_preempt_entity(second, left) < 1)
> >>> +           if (second && wakeup_preempt_entity(second, left, cfs_rq) < 1)
> >>>                     se = second;
> >>>     }
> >>>
> >>> -   if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) {
> >>> +   if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left, cfs_rq) < 1) {
> >>>             /*
> >>>              * Someone really wants this to run. If it's not unfair, run it.
> >>>              */
> >>>             se = cfs_rq->next;
> >>> -   } else if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) {
> >>> +   } else if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left, cfs_rq) < 1) {
> >>>             /*
> >>>              * Prefer last buddy, try to return the CPU to a preempted task.
> >>>              */
> >>> @@ -5107,7 +5107,7 @@ pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
> >>>     /* Check for latency sensitive entity waiting for running */
> >>>     latency = __pick_first_latency(cfs_rq);
> >>>     if (latency && (latency != se) &&
> >>> -       wakeup_preempt_entity(latency, se) < 1)
> >>> +       wakeup_preempt_entity(latency, se, cfs_rq) < 1)
> >>>             se = latency;
> >>>
> >>>     return se;
> >>> @@ -7808,7 +7808,7 @@ static unsigned long wakeup_gran(struct sched_entity *se)
> >>>   *
> >>>   */
> >>>  static int
> >>> -wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
> >>> +wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se, struct cfs_rq *cfs_rq)
> >>>  {
> >>>     s64 gran, vdiff = curr->vruntime - se->vruntime;
> >>>     s64 offset = wakeup_latency_gran(curr, se);
> >>> @@ -7818,6 +7818,8 @@ wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
> >>>
> >>>     gran = offset + wakeup_gran(se);
> >>>
> >>> +   if (vdiff > gran)
> >>> +           return 1;
> >>>     /*
> >>>      * At wake up, the vruntime of a task is capped to not be older than
> >>>      * a sched_latency period compared to min_vruntime. This prevents long
> >>> @@ -7827,9 +7829,8 @@ wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
> >>>      * for low priority task. Make sure that long sleeping task will get a
> >>>      * chance to preempt current.
> >>>      */
> >>> -   gran = min_t(s64, gran, get_latency_max());
> >>> -
> >>> -   if (vdiff > gran)
> >>> +   vdiff = cfs_rq->min_vruntime - se->vruntime;
> >>> +   if (vdiff > get_latency_max())
> >>>             return 1;
> >>>
> >>>     return 0;
> >>> @@ -7933,7 +7934,7 @@ static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_
> >>>             return;
> >>>
> >>>     update_curr(cfs_rq_of(se));
> >>> -   if (wakeup_preempt_entity(se, pse) == 1) {
> >>> +   if (wakeup_preempt_entity(se, pse, cfs_rq_of(se)) == 1) {
> >>>             /*
> >>>              * Bias pick_next to pick the sched entity that is
> >>>              * triggering this preemption.
> >>> --
> >>> 2.34.1
> >> Tried above patch on top of V12. Numbers are worse than V12. We maybe running into
> >> a corner case on this system.
> > Yes it can be a corner case.
> >
> > Nevertheless, the patch above has a problem and does an unsigned comparison instead of a signed
> > one. I have forced the signed comparison in the patch below to be applied on top  of
> > previous one:
> >
> > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > index 77b03a280912..22a497f92dbb 100644
> > --- a/kernel/sched/fair.c
> > +++ b/kernel/sched/fair.c
> > @@ -7830,7 +7830,7 @@ wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se, struct
> >          * chance to preempt current.
> >          */
> >         vdiff = cfs_rq->min_vruntime - se->vruntime;
> > -       if (vdiff > get_latency_max())
> > +       if (vdiff > (s64)get_latency_max())
> >                 return 1;
> >
> >         return 0;
>
> Tested the above patch on top of previous patch + V12.
> Numbers are still worse than V12. Same as V12+previous patch.

So It really looks like a corner case for this system and I'm not sure
we can do anything as others don't face he problem
>
>
> >
> >
> >> Type                Groups       6.2     | 6.2+V12
> >>
> >>  Process              10        0.33     |  0.44
> >>  Process              20        0.61     |  0.90
> >>  Process              30        0.87     |  1.29
> >>  Process              40        1.10     |  1.69
> >>  Process              50        1.34     |  2.08
> >>  Process              60        1.58     |  2.39
> >>  thread               10        0.36     |  0.53
> >>  thread               20        0.64     |  0.94
> >>  Process(Pipe)        10        0.18     |  0.46
> >>  Process(Pipe)        20        0.32     |  0.75
> >>  Process(Pipe)        30        0.42     |  1.01
> >>  Process(Pipe)        40        0.56     |  1.15
> >>  Process(Pipe)        50        0.68     |  1.38
> >>  Process(Pipe)        60        0.80     |  1.56
> >>
> >>
> >>>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> >>>> index cdcd991bbcf1..c89c201dd164 100644
> >>>> --- a/kernel/sched/fair.c
> >>>> +++ b/kernel/sched/fair.c
> >>>> @@ -7827,7 +7827,6 @@ wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
> >>>>          * for low priority task. Make sure that long sleeping task will get a
> >>>>          * chance to preempt current.
> >>>>          */
> >>>> -       gran = min_t(s64, gran, get_latency_max());
> >>>>
> >>>>         if (vdiff > gran)
> >>>>                 return 1;
> >>>>
> >>>>
> >>>>
> > [...]
> >
> >>
>