[Intel-gfx] [RFC 1/8] sched: Add nice value change notifier

[Tvrtko Ursulin]

On 06/10/2021 21:21, Barry Song wrote:
> On Thu, Oct 7, 2021 at 2:44 AM Tvrtko Ursulin
> <tvrtko.ursulin at linux.intel.com> wrote:
>>
>>
>> Hi,
>>
>> On 06/10/2021 08:58, Barry Song wrote:
>>> On Wed, Oct 6, 2021 at 5:15 PM Wanghui (John) <john.wanghui at huawei.com> wrote:
>>>>
>>>> HI Tvrtko
>>>>
>>>> On 2021/10/4 22:36, Tvrtko Ursulin wrote:
>>>>>     void set_user_nice(struct task_struct *p, long nice)
>>>>>     {
>>>>>         bool queued, running;
>>>>> -     int old_prio;
>>>>> +     int old_prio, ret;
>>>>>         struct rq_flags rf;
>>>>>         struct rq *rq;
>>>>>
>>>>> @@ -6915,6 +6947,9 @@ void set_user_nice(struct task_struct *p, long nice)
>>>>>
>>>>>     out_unlock:
>>>>>         task_rq_unlock(rq, p, &rf);
>>>>> +
>>>>> +     ret = atomic_notifier_call_chain(&user_nice_notifier_list, nice, p);
>>>>> +     WARN_ON_ONCE(ret != NOTIFY_DONE);
>>>>>     }
>>>> How about adding a new "io_nice" to task_struct，and move the call chain to
>>>> sched_setattr/getattr, there are two benefits:
>>>
>>> We already have an ionice for block io scheduler. hardly can this new io_nice
>>> be generic to all I/O. it seems the patchset is trying to link
>>> process' nice with
>>> GPU's scheduler, to some extent, it makes more senses than having a
>>> common ionice because we have a lot of IO devices in the systems, we don't
>>> know which I/O the ionice of task_struct should be applied to.
>>>
>>> Maybe we could have an ionice dedicated for GPU just like ionice for CFQ
>>> of bio/request scheduler.
>>
>> Thought crossed my mind but I couldn't see the practicality of a 3rd
>> nice concept. I mean even to start with I struggle a bit with the
>> usefulness of existing ionice vs nice. Like coming up with practical
>> examples of usecases where it makes sense to decouple the two priorities.
>>
>>   From a different angle I did think inheriting CPU nice makes sense for
>> GPU workloads. This is because today, and more so in the future,
>> computations on a same data set do flow from one to the other.
>>
>> Like maybe a simple example of batch image processing where CPU decodes,
>> GPU does a transform and then CPU encodes. Or a different mix, doesn't
>> really matter, since the main point it is one computing pipeline from
>> users point of view.
>>
> 
> I am on it. but I am also seeing two problems here:
> 1. nice is not global in linux. For example, if you have two cgroups, cgroup A
> has more quota then cgroup B. Tasks in B won't win even if it has a lower nice.
> cgroups will run proportional-weight time-based division of CPU.
> 
> 2. Historically, we had dynamic nice which was adjusted based on the average
> sleep/running time; right now, we don't have dynamic nice, but virtual time
> still make tasks which sleep more preempt other tasks with the same nice
> or even lower nice.
> virtual time += physical time/weight by nice
> so, static nice number doesn't always make sense to decide preemption.
> 
> So it seems your patch only works under some simple situation for example
> no cgroups, tasks have similar sleep/running time.

Yes, I broadly agree with your assessment. Although there are plans for 
adding cgroup support to i915 scheduling, I doubt as fine grained 
control and exact semantics as there are on the CPU side will happen.

Mostly because the drive seems to be for more micro-controller managed 
scheduling which adds further challenges in connecting the two sides 
together.

But when you say it is a problem, I would characterize it more a 
weakness in terms of being only a subset of possible control. It is 
still richer (better?) than what currently exists and as demonstrated 
with benchmarks in my cover letter it can deliver improvements in user 
experience. If in the mid term future we can extend it with cgroup 
support then the concept should still apply and get closer to how you 
described nice works in the CPU world.

Main question in my mind is whether the idea of adding the 
sched_attr/priority notifier to the kernel can be justified. Because as 
mentioned before, everything apart from adjusting currently running GPU 
jobs could be done purely in userspace. Stack changes would be quite 
extensive and all, but that is not usually a good enough reason to put 
something in the kernel. That's why it is an RFC an invitation to discuss.

Even ionice inherits from nice (see task_nice_ioprio()) so I think 
argument can be made for drivers as well.

Regards,

Tvrtko

>> In this example perhaps everything could be handled in userspace so
>> that's another argument to be had. Userspace could query the current
>> scheduling attributes before submitting work to the processing pipeline
>> and adjust using respective uapi.
>>
>> Downside would be inability to react to changes after the work is
>> already running which may not be too serious limitation outside the
>> world of multi-minute compute workloads. And latter are probably special
>> case enough that would be configured explicitly.
>>
>>>>
>>>> 1. Decoupled with fair scheduelr. In our use case, high priority tasks often
>>>>       use rt scheduler.
>>>
>>> Is it possible to tell GPU RT as we are telling them CFS nice?
>>
>> Yes of course. We could create a common notification "data packet" which
>> would be sent from both entry points and provide more data than just the
>> nice value. Consumers (of the notifier chain) could then decide for
>> themselves what they want to do with the data.
> 
> RT should have the same problem with CFS once we have cgroups.
> 
>>
>> Regards,
>>
>> Tvrtko
>>
>>>
>>>> 2. The range of value don't need to be bound to -20~19 or 0~139
>>>>
>>>
>>> could build a mapping between the priorities of process and GPU. It seems
>>> not a big deal.
>>>
>>> Thanks
>>> barry
>>>
> 
> Thanks
> barry
>