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

[Mike Galbraith <efault@gmx.de>]

On Mon, 2010-09-13 at 17:53 +0200, Peter Zijlstra wrote:
> On Mon, 2010-09-13 at 15:56 +0200, Mike Galbraith 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.
> > 
> > That's what I wanted to boil it down to instead of putting the extra
> > preempt check in, but it kills the longish slices of low load.  IIRC,
> > when I tried that, it demolished throughput. 
> 
> Hrm.. yes it would..
> 
> So the reason for all this:
> 
>         /*
>          * Ensure that a task that missed wakeup preemption by a
>          * narrow margin doesn't have to wait for a full slice.
>          * This also mitigates buddy induced latencies under load.
>          */
> 
> Is to avoid tasks getting too far ahead in virtual time due to buddies,
> right?

Yeah, that was the thought anyway.

> Would something like the below work? Don't actually use delta_exec to
> filter, but use wakeup_gran + min_gran on virtual time, (much like Steve
> suggested) and then verify using __sched_gran().
> 
> Or have I now totally confused myself backwards?
> 
>  - delta_exec is walltime, and should thus we compared against a
>    weighted unit like slice,
>  - delta is a vruntime unit, and is thus weight free, hence we can use
>    granularity/unweighted units.

I don't think it really matters.  Distance is weighted when using slice
as the measure.

	-Mike