On Mon, 2010-12-20 at 11:46 +0200, Avi Kivity wrote: > On 12/20/2010 11:30 AM, Mike Galbraith wrote: > > > > > > > > Because preempting a perfect stranger is not courteous, all tasks have > > > > to play nice. > > > > > > I don't want to preempt anybody, simply make the task run before me. > > > > I thought you wanted to get the target to the cpu asap? You just can't > > have he runs before me cross cpu. > > You're right, of course. I'm fine with running in parallel. I'm fine > with him running before or instead of me. I'm not fine with running > while the other guy is waiting. Goody, maybe we're headed down a productive path then. > > > Further, this is a kernel internal API, so no need for these types of > > > restrictions. If we expose it to userspace, sure. > > > > Doesn't matter whether it's kernel or not afaikt. If virtualization has > > to coexist peacefully with other loads, it can't just say "my hints are > > the only ones that count", and thus shred other loads throughput. > > What does that have to do with being in the same group or not? I want > to maintain fairness (needed for pure virt workloads, one guest cannot > dominate another), but I don't see how being in the same thread group is > relevant. My thought is that you can shred your own throughput, but not some other concurrent load. I'll have to let that thought stew a bit though. > Again, I don't want more than one entitlement. I want to move part of > my entitlement to another task. Folks can keep trying that, but IMO it's too broken to live. > > > > > > use cfs_rq->next to pass the scheduler a HINT of what you would LIKE to > > > > > > happen. > > > > > > > > > > Hint is fine, so long as the scheduler seriously considers it. > > > > > > > > It will take the hint if the target the target hasn't had too much cpu. > > > > > > Since I'm running and the target isn't, it's clear the scheduler thinks > > > the target had more cpu than I did [73]. That's why I want to donate > > > cpu time. > > > > That's not necessarily true, in fact, it's very often false. Last/next > > buddy will allow a task to run ahead of leftmost so we don't always > > blindly select leftmost and shred cache. > > Ok. > > > > > > > > > What would you suggest? There is no global execution timeline, so if > > > > you want to definitely run after this task, you're stuck with moving to > > > > his timezone or moving him to yours. Well, you could sleep a while, but > > > > we know how productive sleeping is. > > > > > > I don't know. The whole idea of donating runtime was predicated on CFS > > > being completely fair. Now I find that (a) it isn't (b) donating > > > runtimes between tasks on different cpus is problematic. > > > > True and true. However, would you _want_ the scheduler to hold runnable > > tasks hostage, and thus let CPU go to waste in the name of perfect > > fairness? Perfect is the enemy of good applies to that idea imho. > > Sorry, I don't see how it follows. Let's just forget theoretical views, and concentrate on a forward path. > > > Moving tasks between cpus is expensive and sometimes prohibited by > > > pinning. I'd like to avoid it if possible, but it's better than nothing. > > > > Expensive in many ways, so let's try to not do that. > > > > So why do you need this other task to run before you do, even cross cpu? > > If he's a lock holder, getting him to the cpu will give him a chance to > > drop, no? Isn't that what you want to get done? Drop that lock so you > > or someone else can get something other than spinning done? > > Correct. I don't want the other task to run before me, I just don't > want to run before it. OK, so what I gather is that if you can preempt another of your own threads to get the target to cpu, that would be a good thing whether he's on the same cpu as yield_to() caller or not. If the target is sharing a cpu with you, that's even better. Correct? Would a kick/hint option be useful? -Mike -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html