Re: [memcg] 45208c9105: aim7.jobs-per-min -14.0% regression

[Shakeel Butt]

On Mon, Sep 13, 2021 at 12:42 PM Tejun Heo <tj@xxxxxxxxxx> wrote:
>
> Hello,
>
> On Mon, Sep 13, 2021 at 12:40:06PM -0700, Shakeel Butt wrote:
> > I did one more experiment with same workload but with system_wq
> > instead system_unbound_wq and there is clear difference in profile:
> >
> > With system_unbound_wq:
> > -    4.63%     0.33%  mmap  [kernel.kallsyms]  [k] queue_work_on
> >      4.29% queue_work_on
> >       - __queue_work
> >          - 3.45% wake_up_process
> >             - try_to_wake_up
> >                - 2.46% ttwu_queue
> >                   - 1.66% ttwu_do_activate
> >                      - 1.14% activate_task
> >                         - 0.97% enqueue_task_fair
> >                              enqueue_entity
> >
> > With system_wq:
> > -    1.36%     0.06%  mmap  [kernel.kallsyms]  [k] queue_work_on
> >      1.30% queue_work_on
> >       - __queue_work
> >          - 1.03% wake_up_process
> >             - try_to_wake_up
> >                - 0.97% ttwu_queue
> >                     0.66% ttwu_do_activate
> >
> > Tejun, is this expected? i.e. queuing work on system_wq has a
> > different performance impact than on system_unbound_wq?
>
> Yes, system_unbound_wq is putting the work item on the global shared
> workqueue while the system_wq is per-cpu, so on a loaded system, overhead
> difference showing up isn't too surprising.
>

Thanks a lot for the explanation. Are there any concerns to call
cgroup_rstat_flush_irqsafe(root_mem_cgroup->css.cgroup) in system_wq?
This will be called every 2 seconds, so, we can assume the updated
tree would be small most of the time.

Thanks,
Shakeel