Re: [Patch] Idle balancer: cache align nohz structure to improve idle load balancing scalability

[Suresh Siddha <suresh.b.siddha@intel.com>]

On Mon, 2011-11-14 at 01:32 -0800, Peter Zijlstra wrote:
> On Tue, 2011-11-01 at 16:52 -0700, Suresh Siddha wrote:
> > @@ -3317,6 +3317,7 @@ static void update_cpu_power(struct sched_domain *sd, int cpu)
> >  
> >         cpu_rq(cpu)->cpu_power = power;
> >         sdg->sgp->power = power;
> > +       atomic_set(&sdg->sgp->nr_busy_cpus, sdg->group_weight);
> >  }
> >  
> >  static void update_group_power(struct sched_domain *sd, int cpu)
> > @@ -3339,6 +3340,7 @@ static void update_group_power(struct sched_domain *sd, int cpu)
> >         } while (group != child->groups);
> >  
> >         sdg->sgp->power = power;
> > +       atomic_set(&sdg->sgp->nr_busy_cpus, sdg->group_weight);
> >  } 
> 
> So we run this rather frequently, and it will trample all over:
> 
> > +                */
> > +               for_each_domain(cpu, sd)
> > +                       atomic_dec(&sd->groups->sgp->nr_busy_cpus);
> 
> because I cannot see any serialization between those sites.

That was an overlook from myside. I moved the initialization of this to
init_sched_groups_power() now and there is no need to re-do this
everytime we call the update_group_power().

> Also, isn't it rather weird to just assume all cpus are busy in
> update_group_power()? If you would actually set the right value in
> update_cpu_power() you could use a straight sum in update_group_power()
> and get a more or less accurate number out.

I will post an updated patch soon (once we complete the performance
analysis of this patch with respect to other workloads) but the below
hunk gives an idea of what I am planning to do now.

@@ -7369,6 +7384,7 @@ static void init_sched_groups_power(int cpu, struct sched_
                return;
 
        update_group_power(sd, cpu);
+       atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
 }
 
 /*

thanks,
suresh