Re: [PATCH 8/9] sched/psi: add kernel cmdline parameter psi_inner_cgroup

[Chengming Zhou <zhouchengming@bytedance.com>]

On 2022/8/4 03:22, Johannes Weiner wrote:
> On Wed, Aug 03, 2022 at 07:58:27AM -1000, Tejun Heo wrote:
>> Hello,
>>
>> On Wed, Aug 03, 2022 at 08:17:22PM +0800, Chengming Zhou wrote:
>>>> Assuming the above isn't wrong, if we can figure out how we can re-enable
>>>> it, which is more difficult as the counters need to be resynchronized with
>>>> the current state, that'd be ideal. Then, we can just allow each cgroup to
>>>> enable / disable PSI reporting dynamically as they see fit.
>>>
>>> This method is more fine-grained but more difficult like you said above.
>>> I think it may meet most needs to disable PSI stats in intermediate cgroups?
>>
>> So, I'm not necessarily against implementing something easier but we at
>> least wanna get the interface right, so that if we decide to do the full
>> thing later we can easily expand on the existing interface. ie. let's please
>> not be too hacky. I don't think it'd be that difficult to implement
>> per-cgroup disable-only operation that we can later expand to allow
>> re-enabling, right?
> 
> It should be relatively straight-forward to disable and re-enable
> state aggregation, time tracking, averaging on a per-cgroup level, if
> we can live with losing history from while it was disabled. I.e. the
> avgs will restart from 0, total= will have gaps - should be okay, IMO.
> 
> Where it gets trickier is also stopping the tracking of task counts in
> a cgroup. For re-enabling afterwards, we'd have to freeze scheduler
> and cgroup state and find all tasks of interest across all CPUs for
> the given cgroup to recreate the counts. I'm not quite sure whether
> that's feasible, and if so, whether it's worth the savings.
> 
> It might be good to benchmark the two disabling steps independently.
> Maybe stopping aggregation while keeping task counts is good enough,
> and we can commit to a disable/re-enable interface from the start.
> 
> Or maybe it's all in the cachelines and iteration, and stopping the
> aggregation while still writing task counts isn't saving much. In that
> case we'd have to look closer at reconstructing task counts, to see if
> later re-enabling is actually a practical option or whether a one-off
> kill switch is more realistic.
> 
> Chengming, can you experiment with disabling: record_times(), the
> test_state() loop and state_mask construction, and the averaging
> worker - while keeping the groupc->tasks updates?

Hello,

I did this experiment today with disabling record_times(), test_state()
loop and averaging worker, while only keeping groupc->tasks[] updates,
the results look promising.

mmtests/config-scheduler-perfpipe on Intel Xeon Platinum with 3 levels of cgroup:

perfpipe
                                  tip                    tip                patched
                              psi=off                 psi=on      only groupc->tasks[]
Min       Time        7.99 (   0.00%)        8.86 ( -10.95%)        8.31 (  -4.08%)
1st-qrtle Time        8.11 (   0.00%)        8.94 ( -10.22%)        8.39 (  -3.46%)
2nd-qrtle Time        8.17 (   0.00%)        9.02 ( -10.42%)        8.44 (  -3.37%)
3rd-qrtle Time        8.20 (   0.00%)        9.08 ( -10.72%)        8.48 (  -3.43%)
Max-1     Time        7.99 (   0.00%)        8.86 ( -10.95%)        8.31 (  -4.08%)
Max-5     Time        7.99 (   0.00%)        8.86 ( -10.95%)        8.31 (  -4.08%)
Max-10    Time        8.09 (   0.00%)        8.89 (  -9.96%)        8.35 (  -3.22%)
Max-90    Time        8.31 (   0.00%)        9.13 (  -9.90%)        8.55 (  -2.95%)
Max-95    Time        8.32 (   0.00%)        9.14 (  -9.88%)        8.55 (  -2.81%)
Max-99    Time        8.39 (   0.00%)        9.26 ( -10.30%)        8.57 (  -2.09%)
Max       Time        8.56 (   0.00%)        9.26 (  -8.23%)        8.72 (  -1.90%)
Amean     Time        8.19 (   0.00%)        9.03 * -10.26%*        8.45 *  -3.27%*


Tejun suggested using a bitmap in task to remember whether the task is accounted
at a given level or not, which I think also is a very good idea, but I haven't
clearly figure out how to do it.

The above performance test result looks good to me, so I think we can implement this
per-cgroup "cgroup.psi" interface to disable/re-enable PSI stats from the start,
and we can change to a better implementation if needed later?

Thanks!