Messages in this thread | | | Subject | Re: [PATCH] sched: balance_cpu to consider other cpus in its group as target of (pinned) task migration | From | Peter Zijlstra <> | Date | Mon, 04 Jun 2012 17:33:34 +0200 |
| |
On Mon, 2012-06-04 at 20:55 +0530, Srivatsa Vaddagiri wrote: > * Peter Zijlstra <peterz@infradead.org> [2012-06-04 17:21:11]: > > > That is not to say you couldn't contrive a scenario where it would be > > needed.. > > Right ..what if B0. B1 are pinned? !! I think most of the current > weakness lies in handling tasks that have affinity set.
Yeah, affinity is a pain.. the whole group_imb crap is due to that as well.
Now I'm not as opposed to this as Mike is, the load cpu_power thing can also happen due to excessive IRQ time, and hopefully we'll get to have an unpriv. SCHED_DEADLINE at some point as well.
But we should try to keep the stuff reasonably sane and very much consider the worst case compute time of the load-balancer. All that redo logic can be triggered on purpose.
Thing is, you can create an arbitrary hard problem by creating lots of tasks with tricky masks, we shouldn't bend over backwards trying to solve it just because.
[ Also, I suspect I wrecked the ALL_PINNED muck, shouldn't we reset env.loop_break? ]
| |