Re: [PATCH] sched/core: Reduce cost of sched_move_task when config autogroup

[Peter Zijlstra]

On Mon, Mar 20, 2023 at 08:40:51PM +0800, chi wu wrote:
> Peter Zijlstra <peterz@xxxxxxxxxxxxx> 于2023年3月20日周一 16:28写道：
> >
> > On Sun, Mar 19, 2023 at 03:56:43PM +0800, wuchi wrote:
> > > Some sched_move_task calls of autogroup is useless when the
> > > task_struct->sched_task_group isn't changed because of task_group
> > > of cpu_cgroup overlay task_group of autogroup. The overlay key codes
> > > are as follows:
> > >
> > > sched_cgroup_fork->autogroup_task_group->task_wants_autogroup
> > > sched_change_group->autogroup_task_group->autogroup_task_group
> > >
> > > sched_move_task eg:
> > > task A belongs to cpu_cgroup0 and autogroup0, it will always to
> > > cpu_cgroup0 when doing exit. So there is no need to do {de|en}queue.
> > > The call graph is as follow.
> > >
> > > do_exit
> > >   sched_autogroup_exit_task
> > >     sched_move_task
> > >       dequeue_task
> > >         sched_change_group
> > >         A.sched_task_group = sched_get_task_group
> > >       enqueue_task
> > >
> > > So do some check before dequeue task in sched_move_task.
> >
> > No immediate objection; but the $subject seems to suggest you did this
> > for performance reasons, yet there are no performance results in this
> > Changelog. How much does this save under what load?
> 
> cpu:bogomips=4600.00
> kernel:6.3.0-rc3
> cpu cgroup: 6:cpu,cpuacct:/user.slice
> 
> run cmds:
> 
> #!/bin/bash
> for i in {0..10000}; do
> sleep 0 &
> done
> wait
> 
> bpftrace -e 'k:sched_move_task { @ts[tid] = nsecs; }
> kr:sched_move_task /@ts[tid]/ { @ns += nsecs - @ts[tid];
> delete(@ts[tid]); }'
> 
> cost time(ns):
>  without patch: 43528033
>  with      patch: 18934496
>                   diff: -24593537  -56.5%
> 
> About the change, move sched_task_group_changed before
> update_rq_clock as following and add the performance results in
> the Changelog?

Yes please, that gives good motivation for the change.

Thanks!