Re: [RFC] sched: Ensure cpu_power periodic update

From: Vincent Guittot
Date: Thu Dec 15 2011 - 08:36:34 EST


On 15 December 2011 11:08, Peter Zijlstra <a.p.zijlstra@xxxxxxxxx> wrote:
> On Mon, 2011-12-12 at 20:21 +0100, Vincent Guittot wrote:
>> With a lot of small tasks, the softirq sched is nearly never called
>> when no_hz is enable. In this case the load_balance is mainly called with
>> the newly_idle mode which doesn't update the cpu_power.
>> Add a next_update field which ensure a maximum update period when
>> there is short activity
>
>> +     if (local_group) {
>> +             if (idle != CPU_NEWLY_IDLE) {
>> +                     if (balance_cpu != this_cpu) {
>> +                             *balance = 0;
>> +                             return;
>> +                     }
>> +                     update_group_power(sd, this_cpu);
>> +             } else if (time_after_eq(jiffies, group->sgp->next_update))
>> +                     update_group_power(sd, this_cpu);
>>       }
>
> Hmm, I would have expected it to be called from the NOHZ balancing path
> instead of the new_idle path. Your changelog fails to mentions any
> considerations on this..
>

As we are not lucky, the small tasks are mainly running between ticks
and the timer interrupt doesn't fire which implies that both
rebalance_domain of the cpu and nohz_balancer_kick are not called. We
have a lot of call to idle_balance() when cpus become idle and very
few calls to rebalance or nohz_idle_balance. If some tasks are rt
tasks, the cpu_power should be updated regularly to reflect current
use of the cpu by rt scheduler.

I'm using cyclictest to easily reproduce the problem on my dual cortex-A9

> Then again, its probably easier to keep update_group_power on this_cpu
> than to allow a remote update of your cpu_power.
>

This additional path for updating the cpu_power will only be used by
this_cpu because it is called by idle_balance. But we still have a
call to update_group_power by a remote cpu when nohz_idle_balance is
called.

> So I'm not opposed to this patch, I'd just like a little extra
> clarification.

Vincent
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/