[RFC] Core Scheduling unnecessary force idle?

[RFC] Core Scheduling unnecessary force idle?

[Benjamin Tang]

When I'm reading the code related to "core scheduling", I have a question.

Say the RQs in a particular core look like this:
Let CFS1 and CFS4 be 2 untagged CFS tasks.
Let CFS2 and CFS3 be 2 untagged CFS tasks.

          rq0                      rq1
     CFS1(no tag)     CFS3(tagged)
     CFS2(tagged)    CFS4(no tag)

Say schedule() runs on rq0. In the core-wide pick logic, if I'm not 
mistaken,
the end result of the selection will be (say prio(CFS1) > prio(CFS3)):

          rq0                 rq1
     CFS1(no tag)    IDLE

Why not consider trying to find untagged tasks for rq1 here?
Is it because it seems less fair, or are there other considerations?

I would be very grateful if someone could give me some suggestions.
Thanks!

Best Wishes

Re: [RFC] Core Scheduling unnecessary force idle?

[Peter Zijlstra]

On Wed, Nov 29, 2023 at 04:53:35PM +0800, Benjamin Tang wrote:
> When I'm reading the code related to "core scheduling", I have a question.
> 
> Say the RQs in a particular core look like this:
> Let CFS1 and CFS4 be 2 untagged CFS tasks.
> Let CFS2 and CFS3 be 2 untagged CFS tasks.
> 
>          rq0                      rq1
>     CFS1(no tag)     CFS3(tagged)
>     CFS2(tagged)    CFS4(no tag)
> 
> Say schedule() runs on rq0. In the core-wide pick logic, if I'm not
> mistaken,
> the end result of the selection will be (say prio(CFS1) > prio(CFS3)):
> 
>          rq0                 rq1
>     CFS1(no tag)    IDLE
> 
> Why not consider trying to find untagged tasks for rq1 here?
> Is it because it seems less fair, or are there other considerations?
> 
> I would be very grateful if someone could give me some suggestions.
> Thanks!

Because it's expensive to unconditionally track the untagged tasks. I
suppose it could be fixed by iterating the task-set when we
enable/disable core-scheduling, but that's going to be somewhat painful.

A work-around would be to always tag everything, eg. have an explicit
'rest' tag.

Re: [RFC] Core Scheduling unnecessary force idle?

[Benjamin Tang]

In general scenarios, the number of tagged tasks should be less.

Is it feasible to maintain the leftmost untagged node?

在 2023/11/29 下午6:17, Peter Zijlstra 写道:
> On Wed, Nov 29, 2023 at 04:53:35PM +0800, Benjamin Tang wrote:
>> When I'm reading the code related to "core scheduling", I have a question.
>>
>> Say the RQs in a particular core look like this:
>> Let CFS1 and CFS4 be 2 untagged CFS tasks.
>> Let CFS2 and CFS3 be 2 untagged CFS tasks.
>>
>>           rq0                      rq1
>>      CFS1(no tag)     CFS3(tagged)
>>      CFS2(tagged)    CFS4(no tag)
>>
>> Say schedule() runs on rq0. In the core-wide pick logic, if I'm not
>> mistaken,
>> the end result of the selection will be (say prio(CFS1) > prio(CFS3)):
>>
>>           rq0                 rq1
>>      CFS1(no tag)    IDLE
>>
>> Why not consider trying to find untagged tasks for rq1 here?
>> Is it because it seems less fair, or are there other considerations?
>>
>> I would be very grateful if someone could give me some suggestions.
>> Thanks!
> Because it's expensive to unconditionally track the untagged tasks. I
> suppose it could be fixed by iterating the task-set when we
> enable/disable core-scheduling, but that's going to be somewhat painful.
>
> A work-around would be to always tag everything, eg. have an explicit
> 'rest' tag.