Messages in this thread | | | Date | Fri, 3 Jul 2015 02:42:34 +0800 | From | Yuyang Du <> | Subject | Re: [PATCH?] Livelock in pick_next_task_fair() / idle_balance() |
| |
On Thu, Jul 02, 2015 at 12:44:55PM +0100, Morten Rasmussen wrote: > On Thu, Jul 02, 2015 at 12:53:59PM +0200, Peter Zijlstra wrote: > > On Thu, Jul 02, 2015 at 07:25:11AM +0800, Yuyang Du wrote: > > > And obviously, the idle balancing livelock SHOULD happen: one CPU pulls > > > tasks from the other, makes the other idle, and this iterates... > > > > > > That being said, it is also obvious to prevent the livelock from happening: > > > idle pulling until the source rq's nr_running is 1, becuase otherwise we > > > just avoid idleness by making another idleness. > > > > Well, ideally the imbalance calculation would be so that it would avoid > > this from happening in the first place. Its a 'balance' operation, not a > > 'steal everything'. > > > > We want to take work -- as we have none -- but we want to ensure that > > afterwards we have equal work, ie we're balanced. > > Agreed, I think this is the true problem. See my other reply.
Yes, this is agreed at all time. Like I said load_balance() (for idle balancing) should compute the right imbalance and move a fair amount, to achieve we are balanced. Whatever is wrong in how much computed and moved "right imbalance" is should be fixed anyway.
But still, I think, even with the above, in idle balancing, pulling until the source rq's nr_running == 1 is not just "a short term fix", but should be there permanently acting like a last guard with no overhead, why not.
> > > > > So clearly that all is hosed. Now Morten was looking into simplifying > > calculate_imbalance() recently. > > Yes. I'm held up doing other stuff at the moment, but I think > calculate_imbalance() needs some attention and I'm planning on looking at > that next.
Thanks, Yuyang
| |