Re: [PATCH] sched/fair: Fix nohz.next_balance update

From: Valentin Schneider
Date: Wed May 06 2020 - 16:21:25 EST



On 06/05/20 17:56, Vincent Guittot wrote:
> On Wed, 6 May 2020 at 18:03, Valentin Schneider
> <valentin.schneider@xxxxxxx> wrote:
>>
>>
>> On 06/05/20 14:45, Vincent Guittot wrote:
>> >> But then we may skip an update if we goto abort, no? Imagine we have just
>> >> NOHZ_STATS_KICK, so we don't call any rebalance_domains(), and then as we
>> >> go through the last NOHZ CPU in the loop we hit need_resched(). We would
>> >> end in the abort part without any update to nohz.next_balance, despite
>> >> having accumulated relevant data in the local next_balance variable.
>> >
>> > Yes but on the other end, the last CPU has not been able to run the
>> > rebalance_domain so we must not move nohz.next_balance otherwise it
>> > will have to wait for at least another full period
>> > In fact, I think that we have a problem with current implementation
>> > because if we abort because local cpu because busy we might end up
>> > skipping idle load balance for a lot of idle CPUs
>> >
>> > As an example, imagine that we have 10 idle CPUs with the same
>> > rq->next_balance which equal nohz.next_balance. _nohz_idle_balance
>> > starts on CPU0, it processes idle lb for CPU1 but then has to abort
>> > because of need_resched. If we update nohz.next_balance like
>> > currently, the next idle load balance will happen after a full
>> > balance interval whereas we still have 8 CPUs waiting for running an
>> > idle load balance.
>> >
>> > My proposal also fixes this problem
>> >
>>
>> That's a very good point; so with NOHZ_BALANCE_KICK we can reduce
>> nohz.next_balance via rebalance_domains(), and otherwise we would only
>> increase it if we go through a complete for_each_cpu() loop in
>> _nohz_idle_balance().
>>
>> That said, if for some reason we keep bailing out of the loop, we won't
>> push nohz.next_balance forward and thus may repeatedly nohz-balance only
>> the first few CPUs in the NOHZ mask. I think that can happen if we have
>> say 2 tasks pinned to a single rq, in that case nohz_balancer_kick() will
>> kick a NOHZ balance whenever now >= nohz.next_balance.
>
> If we take my example above and we have CPU0 which is idle at every
> tick and selected as ilb_cpu but unluckily CPU0 has to abort before
> running ilb for CPU1 everytime, I agree that we can end up trying to
> run ilb on CPU0 at every tick without any success. We might consider
> to kick_ilb in _nohz_idle_balance if we have to abort to let another
> CPU handle the ilb

That's an idea; maybe something like the next CPU that was due to be
rebalanced (i.e. the one for which we hit the goto abort).