Re: [PATCH] sched/fair: Update nohz.next_balance for newly NOHZ-idle CPUs

[Valentin Schneider <valentin.schneider@arm.com>]

Hi Vincent,

Thanks for taking a look.

On 15/07/21 09:42, Vincent Guittot wrote:
> On Wed, 14 Jul 2021 at 13:39, Valentin Schneider
> <valentin.schneider@arm.com> wrote:
>>
>> Consider a system with some NOHZ-idle CPUs, such that
>>
>>   nohz.idle_cpus_mask = S
>>   nohz.next_balance = T
>>
>> When a new CPU k goes NOHZ idle (nohz_balance_enter_idle()), we end up
>> with:
>>
>>   nohz.idle_cpus_mask = S \U {k}
>>   nohz.next_balance = T
>>
>> Note that the nohz.next_balance hasn't changed - it won't be updated until
>> a NOHZ balance is triggered. This is problematic if the newly NOHZ idle CPU
>> has an earlier rq.next_balance than the other NOHZ idle CPUs, IOW if:
>>
>>   cpu_rq(k).next_balance < nohz.next_balance
>>
>> In such scenarios, the existing nohz.next_balance will prevent any NOHZ
>> balance from happening, which itself will prevent nohz.next_balance from
>> being updated to this new cpu_rq(k).next_balance. Unnecessary load balance
>> delays of over 12ms caused by this were observed on an arm64 RB5 board.
>
> How many CPUs has the arm64 RB5 ?

That's an 8 CPU DynamIQ system - 4 littles, 3 bigs and one "huge". That
should give us a regular balance_interval of 8ms, but our tests have picked
up CPUs staying idle for >20ms when they really have stuff to pull. In this
case balance_interval increases are involved.

>> @@ -10351,6 +10354,13 @@ static void nohz_balancer_kick(struct rq *rq)
>>  unlock:
>>         rcu_read_unlock();
>>  out:
>> +       /*
>> +        * Some CPUs have recently gone into NOHZ idle; kick a balance to
>> +        * collate the proper next balance interval.
>> +        */
>> +       if (!cpumask_subset(nohz.idle_cpus_mask, nohz.last_balance_mask))
>
> I don't really like having to manipulate a cpumask just to trigger an
> ILB and force the update of nohz.next_balance. Could we use something
> similar to nohz.has_blocked and introduce a nohz.force_update.
> manipulating cpumask will even be more complex if we start to have a
> per node idle_cpus_mask like proposed here:
> https://lore.kernel.org/lkml/20210701055323.2199175-1-npiggin@gmail.com/
>
> Also
>
>
> Something like below is simpler
>
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index 44e44c235f1f..91c314f58982 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -10657,6 +10657,9 @@ static void nohz_newidle_balance(struct rq *this_rq)
>         if (this_rq->avg_idle < sysctl_sched_migration_cost)
>                 return;
>
> +       if (time_before(this_rq->next_balance, READ_ONCE(nohz.next_balance))
> +               WRITE_ONCE(nohz.need_update, 1);
> +

I think we have to do this unconditionally, as we can observe the old
nohz.next_balance while a NOHZ balance is ongoing (which will update
nohz.next_balance without taking into account this newly idle CPU).


>         /* Don't need to update blocked load of idle CPUs*/
>         if (!READ_ONCE(nohz.has_blocked) ||
>             time_before(jiffies, READ_ONCE(nohz.next_blocked)))
>
>
> Then we have to test nohz.need_update in nohz_balancer_kick()
>

But then, when can we safely clear this new nohz.need_update? We can't do
it unconditionally in nohz_idle_balance() as this could race with a new CPU
going NOHZ idle.

Perhaps instead we could have a single nohz.needs_update_mask, the CPU is
set in nohz_newidle_balance(), cleared when iterated over in
_nohz_idle_balance(), and nohz_balancer_kick() can trigger an
e.g. NOHZ_UPDATE_KICK if this new cpumask is non-empty.

>> +               flags |= NOHZ_STATS_KICK;
>
> people complain that an update of blocked load is time consuming so we
> should not kick this update unnecessarily.
> We should introduce a new bit like NOHZ_NEXT_KICK that will only go
> through idle cpus and update nohz.next_balance
>

That sounds reasonable.

>> +
>>         if (flags)
>>                 kick_ilb(flags);
>>  }
>> @@ -10487,6 +10497,7 @@ static bool update_nohz_stats(struct rq *rq)
>>  static void _nohz_idle_balance(struct rq *this_rq, unsigned int flags,
>>                                enum cpu_idle_type idle)
>>  {
>> +       struct cpumask *cpus = this_cpu_cpumask_var_ptr(nohz_balance_mask);
>>         /* Earliest time when we have to do rebalance again */
>>         unsigned long now = jiffies;
>>         unsigned long next_balance = now + 60*HZ;
>> @@ -10518,7 +10529,8 @@ static void _nohz_idle_balance(struct rq *this_rq, unsigned int flags,
>>          * Start with the next CPU after this_cpu so we will end with this_cpu and let a
>>          * chance for other idle cpu to pull load.
>>          */
>> -       for_each_cpu_wrap(balance_cpu,  nohz.idle_cpus_mask, this_cpu+1) {
>> +       cpumask_copy(cpus, nohz.idle_cpus_mask);
>
> we are not sure to go through all idle cpus and ilb can abort
>

Right, this is missing something to re-kick an update, but I think we can
get rid of that entirely...