On 30/04/2020 13:00, Pavan Kondeti wrote: > On Wed, Apr 29, 2020 at 07:39:50PM +0200, Dietmar Eggemann wrote: >> On 27/04/2020 16:17, luca abeni wrote:
[...] >>> On Mon, 27 Apr 2020 15:34:38 +0200 >>> Juri Lelli <juri.le...@redhat.com> wrote: [...] >>>> On 27/04/20 10:37, Dietmar Eggemann wrote: >>>>> From: Luca Abeni <luca.ab...@santannapisa.it> [...] >>>>> - if (!cpumask_empty(later_mask)) >>>>> - return 1; >>>>> + if (cpumask_empty(later_mask)) >>>>> + cpumask_set_cpu(max_cpu, later_mask); >>>> >>>> Think we touched upon this during v1 review, but I'm (still?) >>>> wondering if we can do a little better, still considering only free >>>> cpus. >>>> >>>> Can't we get into a situation that some of the (once free) big cpus >>>> have been occupied by small tasks and now a big task enters the >>>> system and it only finds small cpus available, were it could have fit >>>> into bigs if small tasks were put onto small cpus? >>>> >>>> I.e., shouldn't we always try to best fit among free cpus? >>> >>> Yes; there was an additional patch that tried schedule each task on the >>> slowest core where it can fit, to address this issue. >>> But I think it will go in a second round of patches. >> >> Yes, we can run into this situation in DL, but also in CFS or RT. >> > In CFS case, the misfit task handling in load balancer should help pulling > the BIG task running on the little CPUs. I get your point that we can run > into the same scenario with other scheduling class tasks. Yes, the CPU stopper (i.e. CFS's active load balance) can help here. IMHO, using the CPU stopper in RT/DL for moving the running task (next to using best fit rather than just fit CPU) is considered future work. AFAICS, push/pull is not designed for migration of running tasks. [...] >> You did spot the rt-app 'delay' for the small tasks in the test case ;-) > > Thanks for the hint. It was not clear to me why 1 msec delay is given for > the small tasks in the rt-app json description in the cover letter. > I get it now :-) So far Capacity awareness in RT/DL means that as long as there are CPUs available which fit the task, use one of them. This is already beneficial for a lot of use cases on CPU asymmetric systems since it offers more predictable behavior. I'll add a note to the cover letter in the next version about the reason of the rt-app 'delay'.
