Re: [PATCH v2] sched/core: Preempt current task in favour of bound kthread

[Vincent Guittot]

On Tue, 10 Dec 2019 at 11:11, Srikar Dronamraju
<srikar@xxxxxxxxxxxxxxxxxx> wrote:
>
> * Vincent Guittot <vincent.guittot@xxxxxxxxxx> [2019-12-10 10:43:46]:
>
> > On Tue, 10 Dec 2019 at 06:43, Srikar Dronamraju
> > <srikar@xxxxxxxxxxxxxxxxxx> wrote:
> > >
> > > This is more prone to happen if the current running task is CPU
> > > intensive and the sched_wake_up_granularity is set to larger value.
> > > When the sched_wake_up_granularity was relatively small, it was observed
> > > that the bound thread would complete before the load balancer would have
> > > chosen to move the cache hot task to a different CPU.
> > >
> > > To deal with this situation, the current running task would yield to a
> > > per CPU bound kthread, provided kthread is not CPU intensive.
> > >
> > > /pboffline/hwcct_prg_old/lib/fsperf -t overwrite --noclean -f 5g -b 4k /pboffline
> > >
> > > (With sched_wake_up_granularity set to 15ms)
> >
> > So you increase sched_wake_up_granularity to a high level to ensure
> > that current is no preempted by waking thread but then you add a way
> > to finally preempt it which is somewhat weird IMO
> >
>
> Yes, setting to a smaller value will help mitigate/solve the problem.
> There may be folks out who have traditionally set a high wake_up_granularity
> (and have seen better performance with it), who may miss out that when using
> blk-mq, such settings will cause more harm. And they may continue to see
> some performance regressions when they move to a lower wake_up_granularity.
>
> > Have you tried to increase the priority of workqueue thread  (decrease
> > nice priority) ? This is the right way to reduce the impact of the
> > sched_wake_up_granularity on the wakeup of your specific kthread.
> > Because what you want at the end is keeping a low wakeup granularity
> > for these io workqueues
> >
>
> Yes, people can tune the priority of workqueue threads and infact it may be
> easier to set wake_up_granularity to a lower value. However the point is how
> do we make everyone aware that they are running into a performance issue
> with a higher wakeup_granularity?

I did the test on my local setup to change the nice priority of io
workqueue and the active migrations are removed even with high
wakeup_granularity because IO workqueue can still preempt normal task
but let other workqueue behave normally.

>
> --
> Thanks and Regards
> Srikar Dronamraju
>