Re: [RFC PATCH v2] sched/fair: select idle cpu from idle cpumask in sched domain

["Li, Aubrey" <aubrey.li@linux.intel.com>]

On 2020/9/26 0:45, Vincent Guittot wrote:
> Le vendredi 25 sept. 2020 à 17:21:46 (+0800), Li, Aubrey a écrit :
>> Hi Vicent,
>>
>> On 2020/9/24 21:09, Vincent Guittot wrote:
>>>>>>
>>>>>> Would you mind share uperf(netperf load) result on your side? That's the
>>>>>> workload I have seen the most benefit this patch contributed under heavy
>>>>>> load level.
>>>>>
>>>>> with uperf, i've got the same kind of result as sched pipe
>>>>> tip/sched/core: Throughput 24.83Mb/s (+/- 0.09%)
>>>>> with this patch:  Throughput 19.02Mb/s (+/- 0.71%) which is a 23%
>>>>> regression as for sched pipe
>>>>>
>>>> In case this is caused by the logic error in this patch(sorry again), did
>>>> you see any improvement in patch V2? Though it does not helps for nohz=off
>>>> case, just want to know if it helps or does not help at all on arm platform.
>>>
>>> With the v2 which rate limit the update of the cpumask (but doesn't
>>> support sched_idle stask),  I don't see any performance impact:
>>
>> I agree we should go the way with cpumask update rate limited.
>>
>> And I think no performance impact for sched-pipe is expected, as this workload
>> has only 2 threads and the platform has 8 cores, so mostly previous cpu is
>> returned, and even if select_idle_sibling is called, select_idle_core is hit
>> and rarely call select_idle_cpu.
> 
> my platform is not smt so select_idle_core is nop. Nevertheless select_idle_cpu
> is almost never called because prev is idle and selected before calling it in
> our case
> 
>>
>> But I'm more curious why there is 23% performance penalty? So for this patch, if
>> you revert this change but keep cpumask updated, is 23% penalty still there?
>>
>> -       cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
>> +       cpumask_and(cpus, sds_idle_cpus(sd->shared), p->cpus_ptr);
> 
> I was about to say that reverting this line should not change anything because
> we never reach this point but it does in fact. And after looking at a trace,
> I can see that the 2 threads of perf bench sched pipe are on the same CPU and
> that the sds_idle_cpus(sd->shared) is always empty. In fact, the rq->curr is
> not yet idle and still point to the cfs task when you call update_idle_cpumask().
> This means that once cleared, the bit will never be set
> You can remove the test in update_idle_cpumask() which is called either when
> entering idle or when there is only sched_idle tasks that are runnable.
> 
> @@ -6044,8 +6044,7 @@ void update_idle_cpumask(struct rq *rq)
>         sd = rcu_dereference(per_cpu(sd_llc, cpu));
>         if (!sd || !sd->shared)
>                 goto unlock;
> -       if (!available_idle_cpu(cpu) || !sched_idle_cpu(cpu))
> -               goto unlock;
> +
>         cpumask_set_cpu(cpu, sds_idle_cpus(sd->shared));
>  unlock:
>         rcu_read_unlock();
> 
> With this fix, the performance decrease is only 2%
> 
>>
>> I just wonder if it's caused by the atomic ops as you have two cache domains with
>> sd_llc(?). Do you have a x86 machine to make a comparison? It's hard for me to find
>> an ARM machine but I'll try.
>>
>> Also, for uperf(task thread num = cpu num) workload, how is it on patch v2? no any
>> performance impact?
> 
> with v2 :  Throughput 24.97Mb/s (+/- 0.07%) so there is no perf regression
> 

Thanks Vincent, let me try to refine this patch.

-Aubrey