On Mon, Nov 24, 2014 at 02:41:28PM +0000, Vincent Guittot wrote:

On 21 November 2014 at 13:37, Morten Rasmussen [off-list ref] wrote:

quoted

On Mon, Nov 03, 2014 at 04:54:45PM +0000, Vincent Guittot wrote:

[snip]

quoted

quoted

                 */
                if (prefer_sibling && sds->local &&
-                   sds->local_stat.group_has_free_capacity)
-                       sgs->group_capacity_factor = min(sgs->group_capacity_factor, 1U);
+                   group_has_capacity(env, &sds->local_stat) &&
+                   (sgs->sum_nr_running > 1)) {
+                       sgs->group_no_capacity = 1;
+                       sgs->group_type = group_overloaded;
+               }

I'm still a bit confused about SD_PREFER_SIBLING. What is the flag
supposed to do and why?

The goal is to spread tasks across the group even if the the latter is
not overloaded. for SMT level, the goal is to have 1 task per core
before 1 task per HW thread

That makes more sense and is in line with how I understand SMT
scheduling. So we try to have at least one task per group. Where each
group is a domain with SD_PREFER_SIBLING.

Anyway, you don't change the prefer-sibling behaviour in this patch set.
I was just wondering how it would work for SMT balancing.

Thanks,
Morten