> > I see what you mean, but I still don't think it is a problem. If
> > bandwidth matters you will have to use large writes and reads anyways,
> > if bandwidth doesn't matter the number of ctx switches doesn't matter
> > either and latency usually is way more important with small messages.
>
> > Andrea
>
> This is the danger of pre-emption based upon dynamic priorities.
> You can
>get cases where two processes each are permitted to make a very small amount
>of progress in alternation. This can happen just as well with large writes
>as small ones, the amount of data is irrelevent, it's the amount of CPU time
>that's important, or to put it another way, it's how far a process can get
>without suffering a context switch.
>
> I suggest that a process be permitted to use up at least some
> portion of
>its timeslice exempt from any pre-emption based solely on dynamic
>priorities.
Cool.
Thank you for the spiffy idea. I implemented this in my (hack/chop) mm5
tree in about 30 seconds, and it works just fine. Very simple
time_after(jiffies, p->last_run + MIN_TIMESLICE) checks in wake_up_cpu()
plus the obvious dinky change to schedule(), and it worked
instantly. Hmm.. I wonder if this might help some of the other scheduler
(seemingly) bad behavior as well.
Is there any down-side to not preempting quite as often? It seems like
there should be a bandwidth gain.
-Mike
-
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/