Messages in this thread | | | From | Al Boldi <> | Subject | Re: [ANNOUNCE][RFC] PlugSched-6.3.1 for 2.6.16-rc5 | Date | Mon, 10 Apr 2006 17:43:23 +0300 |
| |
Peter Williams wrote: > Al Boldi wrote: > > But how does this explain spa_no_frills setting promotion to max not > > having this problem? > > I'm still puzzled by this. The only thing I can think of is that the > promotion mechanism is to simple in that it just moves all promotable > tasks up one slot without regard for how long they've been on the queue. > Doing this was a deliberate decision based on the desire to minimize > overhead and the belief that it wouldn't matter in the grand scheme of > things. I may do some experimenting with slightly more sophisticated > version. > > Properly done, promotion should hardly ever occur but the cost would be > slightly more complex enqueue/dequeue operations. The current version > will do unnecessary promotions but it was felt this was more than > compensated for by the lower enqueue/dequeue costs. We'll see how a > more sophisticated version goes in terms of trade offs.
Would this affect the current, nearly perfect, spa_no_frills rr-behaviour w/ its ability to circumvent the timeslice problem when setting promo to max?
> >> This is one good reason not to use spa_no_frills on > >> production systems. > > > > spa_ebs is great, but rather bursty. Even setting max_ia_bonus=0 > > doesn't fix that. Is there a way to smooth it like spa_no_frills? > > The principal determinant would be the smoothness of the yardstick. > This is supposed to represent the task with the highest (recent) CPU > usage rate per share and is used to determine how fairly CPU is being > distributed among the currently active tasks. Tasks are given a > priority based on how their CPU usage rate per share compares to this > yardstick. This means that as the system load and/or type of task > running changes the priorities of the tasks can change dramatically. > > Is the burstiness that you're seeing just in the observed priorities or > is it associated with behavioural burstiness as well?
It's behavioural, exhibited in a choking style, like a jumpy mouse move during ia boosts.
> >> Perhaps you should consider creating a child > >> scheduler on top of it that meets your needs? > > > > Perhaps. > > Good. I've been hoping that other interested parties might be > encouraged by the small interface to SPA children to try different ideas > for scheduling.
Is there a no-op child skeleton available?
> One thing that could be played with here is to vary the time slice based > on the priority. This would be in the opposite direction to the normal > scheduler with higher priority tasks (i.e. those with lower prio values) > getting smaller time slices. The rationale being: > > 1. stop tasks that have been given large bonuses from shutting out other > tasks for too long, and > 2. reduce the context switch rate for tasks that haven't received bonuses. > > Because tasks that get large bonuses will have short CPU bursts they > should not be adversely effected (if this is done properly) as they will > (except in exceptional circumstances such as a change in behaviour) > surrender the CPU voluntarily before their reduced time slice has > expired. Imaginative use of the available statistics could make this > largely automatic but there would be a need to be aware that the > statistics can be distorted by the shorter time slices. > > On the other hand, giving tasks without bonuses longer time slices > shouldn't adversely effect interactive performance as the interactive > tasks will (courtesy of their bonuses) preempt them.
I couldn't agree more. Tackling the problem on both fronts (prio/tslice) may give us more control, which could result in a more appropriate / fairer / smoother scheduler.
Thanks!
-- Al
- 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/
| |