2022-04-29 Meeting meeting notes

 Date

Apr 26, 2022

 Participants

@Jonathan Cameron

@Lorenzo Pieralisi (Deactivated)

@Vincent Guittot

@yangyicong@hisilicon.com

@hesham.almatary@huawei.com

@21cnbao@gmail.com

@dietmar.eggemann@arm.com

@ilkka@os.amperecomputing.com

@Joyce Qi

 Goals

  •  

 Discussion topics

Time

Item

Presenter

Notes

Time

Item

Presenter

Notes

 

RMR

 




Scheduler


Yichong Yang



 Action items


RMR:
- (Shameer) Request to look at Patch 4 as some changes from earlier versions.  Warning fix for next version.
- Lorenzo will check with Robin.
- Need to ask Joerg to pull the series.
vCPU HP.
Commits · vcpu-poc-1 · Jonathan / Qemu · GitLab
- Kernel Patches - hopefully posted soon.  Not Lorenzo as not available for a few months.
- _MAT needed or not.
SPDM
- Random discussion of the difficulty of debugging the rather complex exchanges. Nothing worth
  noting.

Scheduler:
This is yangyicong's topic and some summarize:

https://lore.kernel.org/lkml/ef3b3e55-8be9-595f-6d54-886d13a7e2fd@hisilicon.com/

Yangyicong's slides provide a good summary of the question.
- Observed that any topology based estimate is going to be challenging!
- Vincent - set migration cost per sched domain level?
 - Dietmar - workloads weren't clear in original thread.
   Question of taskhot or new Idle Balance as relevant to workload.
   Tighter description of what is going on in the benchmarks needed.
 - Vincent - Uarch + cache etc relevant.
 - Hesham - potentially user perf counters to get some more info.

Takeways
 1. Need to figure out the underlying reason of the performance variation of certain benchmark:
    If it's because of task hotness or newidle_balance(), etc.
 2. Effect may related to the micro-arch, cache, and task's states. Also can be tuned according
    to the cpu numbers and scheduler domain levels.
 3. Can be get from some firmware reports, to avoid the long time measurement in booting.
 4. May be possible to calibrate during the boot time, but narrow the scope of measured CPUs
    and test time. This won't take long, but is susceptible to noise from other sources.
 5. Hardware counters or profilings help understand what is going on, but unlikely to be
     consistently available for use in the loop.
 6. Maybe make migration cost per sched domain on implementation.

 

Recordings:

Video Conferencing, Web Conferencing, Webinars, Screen Sharing
密码:f*i0EwCG

 

 Decisions