Re: [RFC PATCH 1/9] sched,cgroup: Add interface for latency-nice

[Parth Shah <parth@linux.ibm.com>]

Hi Joel,

On 4/18/20 9:31 PM, Joel Fernandes wrote:
> Hi Dietmar,
> 
> On Thu, Apr 16, 2020 at 1:23 PM Dietmar Eggemann
> <dietmar.eggemann@arm.com> wrote:
>>
>> Hi Joel,
>>
>> On 16.04.20 02:02, Joel Fernandes wrote:
>>> On Thu, Sep 05, 2019 at 12:47:26PM +0100, Qais Yousef wrote:
>>>> On 09/05/19 13:30, Peter Zijlstra wrote:
>>>>> On Thu, Sep 05, 2019 at 12:13:47PM +0100, Qais Yousef wrote:
>>>>>> On 09/05/19 12:46, Peter Zijlstra wrote:
>>>>>
>>>>>>> This is important because we want to be able to bias towards less
>>>>>>> importance to (tail) latency as well as more importantance to (tail)
>>>>>>> latency.
>>>>>>>
>>>>>>> Specifically, Oracle wants to sacrifice (some) latency for throughput.
>>>>>>> Facebook OTOH seems to want to sacrifice (some) throughput for latency.
>>>>>>
>>>>>> Another use case I'm considering is using latency-nice to prefer an idle CPU if
>>>>>> latency-nice is set otherwise go for the most energy efficient CPU.
>>>>>>
>>>>>> Ie: sacrifice (some) energy for latency.
>>>>>>
>>>>>> The way I see interpreting latency-nice here as a binary switch. But
>>>>>> maybe we can use the range to select what (some) energy to sacrifice
>>>>>> mean here. Hmmm.
>>>>>
>>>>> It cannot be binary, per definition is must be ternary, that is, <0, ==0
>>>>> and >0 (or middle value if you're of that persuasion).
>>>>
>>>> I meant I want to use it as a binary.
>>>>
>>>>>
>>>>> In your case, I'm thinking you mean >0, we want to lower the latency.
>>>>
>>>> Yes. As long as there's an easy way to say: does this task care about latency
>>>> or not I'm good.
>>>
>>> Qais, Peter, all,
>>>
>>> For ChromeOS (my team), we are planning to use the upstream uclamp mechanism
>>> instead of the out-of-tree schedtune mechanism to provide EAS with the
>>> latency-sensitivity (binary/ternary) hint. ChromeOS is thankfully quite a bit
>>> upstream focussed :)
>>>
>>> However, uclamp is missing an attribute to provide this biasing to EAS as we
>>> know.
>>>
>>> What was the consensus on adding a per-task attribute to uclamp for providing
>>> this? Happy to collaborate on this front.
>>
>> We're planning to have a session about this topic (latency-nice
>> attribute per task group) during the virtual Pisa OSPM summit
>> retis.sssup.it/ospm-summit in May this year.
> 
> Cool, I registered as well.
> 
>>
>> There are two presentations/discussions planned:
>>
>> "Introducing Latency Nice for Scheduler Hints and Optimizing Scheduler
>> Task Wakeup" and "The latency nice use case for Energy-Aware-Scheduling
>> (EAS) in Android Common Kernel (ACK)"
>>
>> We'll probably merge those two into one presentation/discussion.
>>
>> So far we have Parth's per-task implementation
>>
>> https://lore.kernel.org/lkml/20200228090755.22829-1-parth@linux.ibm.com
> 
> Cool, I see it has some Reviewed-by tags so that's a good sign. Will
> look more into that.
> 
>> What's missing is the per-taskgroup implementation, at least from the
>> standpoint of ACK.
>>
>> The (mainline) EAS use-case for latency nice is already in ACK
>> (android-5.4):
>>
>> https://android.googlesource.com/kernel/common/+/760b82c9b88d2c8125abfc5f732cc3cd460b2a54
> 
> Yes, I was aware of this. But if we use task groups, then the
> transition from schedtune -> uclamp means now the tasks that use
> uclamp would also be subjected to cpu.shares. That's why we were
> looking into the per-task interface and glad there's some work on this
> already done.
> 

Yes, that series of latency_nice seems to be in good shape to be used for
any usecases. Hopefully, OSPM will lead to its upstreaming sooner :-)
But at the end, we aim to have both the per-task and cgroup based interface
to mark the latency_nice value of a task.
Till, then I'm finding some generic use-cases to show benefits of such task
attribute to increase community interest.


Thanks,
Parth