On 11 September 2014 18:15, Peter Zijlstra [off-list ref] wrote:

On Tue, Aug 26, 2014 at 01:06:54PM +0200, Vincent Guittot wrote:

quoted

+static inline int group_has_free_capacity(struct sg_lb_stats *sgs,
+                     struct lb_env *env)
 {
+     if ((sgs->group_capacity_orig * 100) >
+                     (sgs->group_utilization * env->sd->imbalance_pct))
+             return 1;
+
+     if (sgs->sum_nr_running < sgs->group_weight)
+             return 1;

+     return 0;
+}

+static inline int group_is_overloaded(struct sg_lb_stats *sgs,
+                     struct lb_env *env)
+{
+     if (sgs->sum_nr_running <= sgs->group_weight)
+             return 0;

+     if ((sgs->group_capacity_orig * 100) <
+                     (sgs->group_utilization * env->sd->imbalance_pct))
+             return 1;

+     return 0;
 }

I'm confused about the utilization vs capacity_orig. I see how we should

1st point is that I should compare utilization vs capacity and not
capacity_orig.
I should have replaced capacity_orig by capacity in the functions
above when i move the utilization statistic from
rq->avg.runnable_avg_sum to cfs.usage_load_avg.
rq->avg.runnable_avg_sum was measuring all activity on the cpu whereas
cfs.usage_load_avg integrates only cfs tasks

With this change, we don't need sgs->group_capacity_orig anymore but
only sgs->group_capacity. So sgs->group_capacity_orig can be removed
as it's no more used in the code as sg_capacity_factor has been
removed

maybe scale things with the capacity when comparing between CPUs/groups,
but not on the same CPU/group.

I would have expected something simple like:

static inline bool group_has_capacity()
{
        /* Is there a spare cycle? */
        if (sgs->group_utilization < sgs->group_weight * SCHED_LOAD_SCALE)
                return true;

        /* Are there less tasks than logical CPUs? */
        if (sgs->sum_nr_running < sgs->group_weight)
                return true;

        return false;
}

Where group_utilization a pure sum of running_avg.

Now this has a hole when there are RT tasks on the system, in that case
the utilization will never hit 1, but we could fix that another way. I
don't think the capacity_orig thing is right.