Re: [RFC patch 1/2] sched: dynamically adapt granularity withnr_running

[Peter Zijlstra]

On Mon, 2010-09-13 at 09:56 -0400, Mathieu Desnoyers wrote:
> 
> > One option is to simply get rid of that stuff in check_preempt_tick()
> > and instead do a wakeup-preempt check on the leftmost task instead.
> > 
> > The code as it stands today does that delta_exec < min_gran check to
> > ensure current gets some runtime before doing that second preemption
> > check, which compares vruntime with a wall-time measure.
> > 
> > Making that gran more complex doesn't really buy us much because for a
> > system with different weights in the gran and slice lengths don't match
> > up anyway.
> 
> So I bet this last sentence is about the example of a system with many nice 19
> processes I told you about on IRC. Yes, this one is a bummer, as we would not
> like to count them as running threads at all.

Of course we would. But the same is true for -5 and 5 threads together.

> >  static void
> >  check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
> >  {
> > -     unsigned long ideal_runtime, delta_exec;
> > +     unsigned long slice = sched_slice(cfs_rq, curr);
> 
> So you still compute the sched_slice(), based on sched_period(), based on
> sysctl_sched_min_granularity *= nr_running when there are more than nr_latency
> running threads.

What's wrong with that? I keep asking you, you keep not giving an
answer. Stop focussing on nr_latency, its an by produce not a
fundamental entity.

 period := max(latency, min_gran * nr_running)

See, no nr_latency -- the one and only purpose of nr_latency is avoiding
that multiplication when possible.


> > -     if (delta_exec < sysctl_sched_min_granularity)
> > -             return;
> 
> Well, the reason why this test is here seems to be that we don't want to trigger
> "resched_task" more often than needed, and here it's defined by the granularity.

Right, but its wrong for the weighted case. Letting a light task run
that long will make its latency suck.

> I don't quite see with what you are replacing this, other than "let's set the
> resched flag all the time to save a 32-bit division". I figure out it's more
> expensive the call the scheduler than to do a 32-bit div.

The more divs we put it, the more expensive it all becomes.


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/