Re: [PATCH] sched: avoid overpull when pulling RT task

[Mike Galbraith]

On Tue, 2011-05-17 at 22:25 +0800, Hillf Danton wrote:
> On Mon, May 16, 2011 at 10:11 PM, Mike Galbraith <efault@xxxxxx> wrote:
> > On Mon, 2011-05-16 at 21:14 +0800, Hillf Danton wrote:
> >
> >> In the following patch, pulling is played in two rounds. In the first round,
> >> the highest priority task is determined with no pull operation. Pulling is
> >> carried out in the second round, and if the highest priority task is pulled,
> >> pulling could be stopped when overload detected, to relieve overpull.
> >>
> >> Please review again, thanks.
> >
> > Traversing runqueues twice to avoid some potential task bouncing during
> > overload situation seems like a really bad trade.  Not to mention that
> > between pass one and pass two, the world turns under your feet.
> >
> > You could do it in one pass by leaving the victim's runqueue locked
> > unless you find a better victim I suppose.  Dunno, guess it all depends
> > on how much benefit the is to pulling only highest, which I can't answer
> > (my gut says "none, only more pain to be had here").
> >
> Hi Mike
> 
> Efforts are put on the puller's side, but bad result is reached:(

Don't be discouraged by bad results, they're unavoidable.  Keep on
looking, thinking and poking.

> Another patch is prepared, in which pusher is asked to do the hard works,
> say pushees and tasks are selected.
> 
> Unlike puller who only concerns one runqueue that accepts the pulled tasks,
> pusher delivers tasks to more runqueues, so the overpull could get bigger.
> 
> Please review again, thanks.

Sorry, but I don't have any free time atm, my cup runneth over bigtime.
Best suggestion I can offer is measure, and measure again.  No theory
can compete with cold hard numbers.

	-Mike

--
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/