Re: busted CFS group load balancer?

[Peter Zijlstra <a.p.zijlstra@chello.nl>]

On Mon, 2008-11-17 at 23:33 -0800, Ken Chen wrote:
> On Mon, Nov 17, 2008 at 9:19 PM, Peter Zijlstra wrote:
> > Note that with larger cpu count and/or lower group weight we'll quickly
> > run into numerical trouble...
> >
> > I would recommend trying this with the minimum weight in the order of
> > 8-16 times number of cpus on your system.
> >
> > There is only so much one can do with 10 bit fixed precision math :/
> 
> That is probably one of the many problems.  I also found that the
> updates to the per-cpu task_group's sched_entity load weight
> (tg->se[cpu]->load.weight) is very problematic and very erratic.
> 
> The total rq_weight is calculated at one beginning of tg_shares_up(),
> 
>         for_each_cpu_mask(i, sd->span) {
>                 rq_weight += tg->cfs_rq[i]->load.weight;
>                 shares += tg->cfs_rq[i]->shares;
>         }
> 
> However, the scaling of per-cpu se->load.weight in function
> __update_group_shares_cpu() takes another lookup of
> tg->cfs_rq[cpu]->load.weight at a different time.
> cfs_rq[cpu].load.weight aren't always consistent across these two
> times.  Due to these inconsistency of value taken on per cpu cfs_rq,
> I've see tg->se[cpu]->load.weight jumping all over the place.  In our
> environment, the cpu loads are very dynamic.  Process
> queuing/dequeuing at high rate.

Ok, if your load values are very unstable in the order of the
load-balance interval then you're hosed too, the same is true for the
normal smp load-balancer.

The cgroup load-balancer makes that even more problematic.

Again, there's just very little you can do about that, except increase
the coupling between cpus and thereby increase the overhead. Try
decreasing 
sysctl_sched_shares_ratelimit.


> I'm also very troubled with this calculation in __update_group_shares_cpu():
> 
>         shares = (sd_shares * rq_weight) / (sd_rq_weight + 1);
> 
> Won't you have rounding problem here?  value 'shares' will gradually
> decrease for each iteration of __update_group_shares_cpu()?

Yes it will, however at the top of the sched-domain tree its reset. 

        if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
                shares = tg->shares;