Re: [Linux-ia64] Re: web page on O(1) scheduler

[Mike Galbraith]

At 10:05 AM 6/2/2003 +0200, Ingo Molnar wrote:

>On Thu, 29 May 2003, Mike Galbraith wrote:
>
> > [...] What makes more sense to me than the current implementation is to
> > rotate the entire peer queue when a thread expires... ie pull in the
> > head of the expired queue into the tail of the active queue at the same
> > time so you always have a player if one exists.  (you'd have to select
> > queues based on used cpu time to make that work right though)
>
>we have tried all sorts of more complex yield() schemes before - they
>sucked for one or another workload. So in 2.5 i took the following path:
>make yield() _simple_ and effective, ie. expire the yielding task (push it
>down the runqueue roughly halfway, statistically) and dont try to be too
>smart doing it. All the real yield() users (mostly in the kernel) want it
>to be an efficient way to avoid livelocks. The old 2.4 yield
>implementation had the problem of enabling a ping-pong between two
>higher-prio yielding processes, until they use up their full timeslice.

(yeah, i looked at that in ktracer logs.  cpu hot-potato sucks;)

>(we could do one more thing that still keeps the thing simple: we could
>re-set the yielding task's timeslice instead of the current 'keep the
>previous timeslice' logic.)

(more consistent)