Re: [PATCH 4/7 v3] sched: propagate load during synchronous attach/detach

[Peter Zijlstra <peterz@infradead.org>]

On Thu, Sep 15, 2016 at 02:11:49PM +0100, Dietmar Eggemann wrote:
> On 12/09/16 08:47, Vincent Guittot wrote:

> > +/* Take into account change of load of a child task group */
> > +static inline void
> > +update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se)
> > +{
> > +	struct cfs_rq *gcfs_rq = group_cfs_rq(se);
> > +	long delta, load = gcfs_rq->avg.load_avg;
> > +
> > +	/* If the load of group cfs_rq is null, the load of the
> > +	 * sched_entity will also be null so we can skip the formula
> > +	 */
> > +	if (load) {
> > +		long tg_load;
> > +
> > +		/* Get tg's load and ensure tg_load > 0 */
> > +		tg_load = atomic_long_read(&gcfs_rq->tg->load_avg) + 1;
> > +
> > +		/* Ensure tg_load >= load and updated with current load*/
> > +		tg_load -= gcfs_rq->tg_load_avg_contrib;
> > +		tg_load += load;
> > +
> > +		/* scale gcfs_rq's load into tg's shares*/
> > +		load *= scale_load_down(gcfs_rq->tg->shares);
> > +		load /= tg_load;
> > +
> > +		/*
> > +		 * we need to compute a correction term in the case that the
> > +		 * task group is consuming <1 cpu so that we would contribute
> > +		 * the same load as a task of equal weight.
> 
> Wasn't 'consuming <1' related to 'NICE_0_LOAD' and not
> scale_load_down(gcfs_rq->tg->shares) before the rewrite of PELT (v4.2,
> __update_group_entity_contrib())?


So the approximation was: min(1, runnable_avg) * shares;

And it just so happened that we tracked runnable_avg in 10 bit fixed
point, which then happened to be NICE_0_LOAD.

But here we have load_avg, which already includes a '* shares' factor.
So that then becomes min(shares, load_avg).

We did however loose a lot on why and how min(1, runnable_avg) is a
sensible thing to do...

> > +		*/
> > +		if (tg_load < scale_load_down(gcfs_rq->tg->shares)) {
> > +			load *= tg_load;
> > +			load /= scale_load_down(gcfs_rq->tg->shares);
> > +		}
> > +	}
> 
> [...]