Re: [PATCH 2/2] sched/fair: Relax task_hot() for misfit tasks

[Dietmar Eggemann <dietmar.eggemann@arm.com>]

On 21/04/2021 12:52, Valentin Schneider wrote:
> On 20/04/21 16:33, Vincent Guittot wrote:
>> On Mon, 19 Apr 2021 at 19:13, Valentin Schneider
>> <valentin.schneider@arm.com> wrote:
>>>
>>> On 16/04/21 15:51, Vincent Guittot wrote:
>>>> Le jeudi 15 avril 2021 � 18:58:46 (+0100), Valentin Schneider a �crit :
>>>>> +
>>>>> +/*
>>>>> + * What does migrating this task do to our capacity-aware scheduling criterion?
>>>>> + *
>>>>> + * Returns 1, if the task needs more capacity than the dst CPU can provide.
>>>>> + * Returns 0, if the task needs the extra capacity provided by the dst CPU
>>>>> + * Returns -1, if the task isn't impacted by the migration wrt capacity.
>>>>> + */
>>>>> +static int migrate_degrades_capacity(struct task_struct *p, struct lb_env *env)
>>>>> +{
>>>>> +    if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
>>>>> +            return -1;
>>>>> +
>>>>> +    if (!task_fits_capacity(p, capacity_of(env->src_cpu))) {
>>>>> +            if (cpu_capacity_greater(env->dst_cpu, env->src_cpu))
>>>>> +                    return 0;
>>>>> +            else if (cpu_capacity_greater(env->src_cpu, env->dst_cpu))
>>>>> +                    return 1;
>>>>> +            else
>>>>> +                    return -1;
>>>>> +    }
>>>>
>>>> Being there means that task fits src_cpu capacity so why testing p against dst_cpu ?
>>>>
>>>
>>> Because if p fits on src_cpu, we don't want to move it to a dst_cpu on
>>> which it *doesn't* fit.
>>
>> OK. I was confused because I thought that this was only to force
>> migration in case of group_misfit_task but you tried to extend to
>> other cases... I'm not convinced that you succeeded to cover all cases
>>
>> Also I found this function which returns 3 values a bit disturbing.
>> IIUC you tried to align to migrate_degrades_capacity but you should
>> have better aligned to task_hot and return only 0 or 1. -1 is not used
>>
> 
> Ack, will do.
> 
>>>>> @@ -7672,6 +7698,15 @@ int can_migrate_task(struct task_struct *p, struct lb_env *env)
>>>>>      if (tsk_cache_hot == -1)
>>>>>              tsk_cache_hot = task_hot(p, env);
>>>>>
>>>>> +    /*
>>>>> +     * On a (sane) asymmetric CPU capacity system, the increase in compute
>>>>> +     * capacity should offset any potential performance hit caused by a
>>>>> +     * migration.
>>>>> +     */
>>>>> +    if ((env->dst_grp_type == group_has_spare) &&
>>>>
>>>> Shouldn't it be env->src_grp_type == group_misfit_task to only care of misfit task case as
>>>> stated in $subject
>>>>
>>>
>>> Previously this was env->idle != CPU_NOT_IDLE, but I figured dst_grp_type
>>> could give us a better picture. Staring at this some more, this isn't so
>>> true when the group size goes up - there's no guarantees the dst_cpu is the
>>> one that has spare cycles, and the other CPUs might not be able to grant
>>> the capacity uplift dst_cpu can.
>>
>> yeah you have to keep checking for env->idle != CPU_NOT_IDLE
>>
>>>
>>> As for not using src_grp_type == group_misfit_task, this is pretty much the
>>> same as [1]. CPU-bound (misfit) task + some other task on the same rq
>>> implies group_overloaded classification when balancing at MC level (no SMT,
>>> so one group per CPU).
>>
>> Is it something that happens often or just a sporadic/transient state
>> ? I mean does it really worth the extra complexity and do you see
>> performance improvement ?
>>
> 
> "Unfortunately" yes, this is a relatively common scenario when running "1
> big task per CPU" types of workloads. The expected behaviour for big.LITTLE
> systems is to upmigrate tasks stuck on the LITTLE CPUs as soon as a big CPU
> becomes free, usually via newidle balance (which, since they process work
> faster than the LITTLEs, is bound to happen), and an extra task being
> enqueued at "the wrong time" can prevent this from happening.
> 
> This usually means a misfit task can take a few dozen extra ms than it
> should to be migrated - in the tests I run (which are pretty much this 1
> hog per CPU workload) this happens about ~20% of the time.
> 
>> You should better focus on fixing the simple case of group_misfit_task
>> task. This other cases looks far more complex with lot of corner cases
>>
>>>
>>> [1]: http://lore.kernel.org/r/jhjblcuv2mo.mognet@arm.com


Just to make sure I can follow the conversation ...

In case you:

(1) return 1 instead of -1

(2) keep the `env->idle != CPU_NOT_IDLE` check

(3) and remove the `dst_grp_type == group_has_spare` check

you are pretty much back to what you had in [PATCH v3 7/7] directly in
task_hot() except:

(4) the 'if (p fits on src_cpu && p !fits dst_cpu) => tsk_cache_hot) check?