Re: [RFC] sched: Limit idle_balance() when it is being used too frequently

[Srikar Dronamraju <srikar@linux.vnet.ibm.com>]

* Peter Zijlstra <peterz@infradead.org> [2013-07-18 14:35:31]:

> On Thu, Jul 18, 2013 at 05:45:46PM +0530, Srikar Dronamraju wrote:
> > We take locks if and only if we see imbalance and want to pull the
> > tasks.
> > However if the newly idle balance is not finding an imbalance then this
> > may not be an issue.
> > 
> > Probably /proc/schedstats will give a better picture.
> 
> Right, so we're interested in move_tasks() calls that fail to 'deliver'.
> There's a few conditions in there that can cause us to not move a task,
> most of them not counted.
> 
> The few that are; are from can_mirgrate_task():
> 
>   se.statistics.nr_failed_migrations_affine
>   se.statistics.nr_failed_migrations_running
>   se.statistics.nr_failed_migrations_hot
> 
> If we see significant increments on those we'll be taking locks. 
> 

Agree. 

Even I think number of times no busy group was found, number of times no
busy queue was found also will tell us that locks are not being taken.

In schedstats, I generally see them as overwhelming majority.


> The only one I can see a good way around is the hot one, we could ignore
> hotness in favour of newidle -- although I could see that being
> detrimental, we'll just have to try or so ;-)
> 
> _running shouldn't be much of a problem since we don't bother if
> nr_running <= 1. And _affine is out of our reach anyway.
> 

-- 
Thanks and Regards
Srikar Dronamraju