Re: Scheduling the highest priority task

[Martin Roehricht <ml@felicis.org>]

On 02.08.2007 21:48, Ingo Molnar wrote:
> * Martin Roehricht <ml@felicis.org> wrote:
>
> > On 08/02/2007 05:19 PM, Ingo Molnar wrote:
> > >* Martin Roehricht <ml@felicis.org> wrote:
> > >
> > >>That's fine with me, that within the same priority-queue any task can 
> > >>be chosen. But assume two tasks with highly different priorities, such 
> > >>as 105 and 135 are scheduled on the same processor and one of them is 
> > >>now to be migrated -- shouldn't be the queue with task P=105 
> > >>considered first for migration by this code? Both tasks would use 
> > >>different queues with their own linked lists, right?
> > >
> > >yes. What makes you believe that the lower priority one (prio 135) is 
> > >chosen? [ as i said before, that will only be chosen if all tasks in the 
> > >higher-priority queue (prio 105) are either already running on a CPU or 
> > >have recently run so that the cache-hot logic skips them. ]
> >
> > This believe is primarily based on my observations of multiple 
> > benchmark runs and also on your statement earlier: »in the SMP 
> > migration code, the 'old scheduler' indeed picks the lowest priority 
> > one«.
>
> oh, sorry, that was meant to be the 'highest priority one' :-/
>
> so i think you got it all right, i just typoed that first sentence.

Okay, now I think I understood this part of the code correctly. The 
reason why I observe a continous migration of the _lower_ priority tasks 
is most probably due to the fact that the higher priority one is 
currently running, according to:
	can_migrate_task() in move_tasks(), and therein:

	if (task_running(rq, p))
		return 0;

I tracked down via an extended /proc/schedstats that my tasks fall 
frequently into this pitfall. I basically solved it by making use of the 
more active push-strategy which is called later by load_balance() once 
the move_tasks() function did not succeed. So in case I need the higher 
priority tasks, I return immediately from move_tasks().

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