On 29 May 2014 16:04, Peter Zijlstra [off-list ref] wrote:

On Fri, May 23, 2014 at 05:53:04PM +0200, Vincent Guittot wrote:

quoted

@@ -7282,6 +7289,12 @@ static inline int nohz_kick_needed(struct rq *rq)

              if (nr_busy > 1)
                      goto need_kick_unlock;
+
+             if ((rq->cfs.h_nr_running >= 1)
+              && ((rq->cpu_power * sd->imbalance_pct) <
+                                     (rq->cpu_power_orig * 100)))
+                     goto need_kick_unlock;
+
      }

      sd = rcu_dereference(per_cpu(sd_asym, cpu));

So what happens when a cpu is consistently low on power (say due to a
pinned RT task) the balancer would quickly adjust the load level, but
this would endlessly kick things into action, even though we're balanced
just fine.

If there is more than 1 running task or more than 1 busy CPU, we will
kick the ilb because of the former conditions. Then, if there is only
1 task and no other busy cpu, we should trig the ILB. Nevertheless, I
can add a test to check that there is an idle cpu in the sched_domain