Re: [PATCH 2/2] sched: add statistic for rq->max_idle_balance_cost

From: Morten Rasmussen
Date: Thu Jan 23 2014 - 07:32:57 EST


On Thu, Jan 23, 2014 at 10:49:35AM +0000, Alex Shi wrote:
> On 01/23/2014 05:54 PM, Peter Zijlstra wrote:
> > On Thu, Jan 23, 2014 at 02:49:25PM +0800, Alex Shi wrote:
> >> On 01/23/2014 02:10 AM, Jason Low wrote:
> >>>>>>> P64(avg_idle);
> >>>>>>> + P64(max_idle_balance_cost);
> >>>>>>> #endif
> >>>>>>> P(ttwu_count);
> >>>>>
> >>>>> Not also the per-sd value in sd_alloc_ctl_domain_table() ?
> >>> Yeah, tracking the sd->max_newidle_lb_cost can also be useful.
> >>>
> >>>
> >>
> >> Thanks for suggestion!
> >>
> >> I thought the sd->max_newidle_lb_cost keep changing. But yes, it's
> >> still meaningful.
> >> BTW, in the pandaboard ES, sd balance cost is about 1~2ms.
> >
> > That's an insane amount of time for only 2 cpus.
>
> maybe, :(
>
> but it is the data.
> $ cat /proc/sys/kernel/sched_domain/cpu0/domain0/max_newidle_lb_cost
> 1873393

What sort of workload are you running? It seems rather extreme.

I did a quick test on TC2 and got max_newidle_lb_cost in the range
~7000-80000 when having a constantly running task on each cpu for a few
seconds.

I also traced curr_cost just to get an idea of the cost of idle_balance.
I get an average cost across all five cpus of around 7000.

Morten
--
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/