Messages in this thread | | | Date | Mon, 22 Oct 2012 13:46:57 +0200 | From | Andrea Righi <> | Subject | Re: [PATCH v2 1/3] sched: introduce distinct per-cpu load average |
| |
On Mon, Oct 22, 2012 at 01:10:40PM +0200, Peter Zijlstra wrote: > On Sat, 2012-10-20 at 21:06 +0200, Andrea Righi wrote: > > @@ -383,13 +383,7 @@ struct rq { > > struct list_head leaf_rt_rq_list; > > #endif > > > > > + unsigned long __percpu *nr_uninterruptible; > > This is O(nr_cpus^2) memory.. >
Correct, this doesn't add too much overhead to the wakeup/sleep path, but it's bad both in terms of memory and performance overhead in the other parts of the code for large SMP systems.
> > > +unsigned long nr_uninterruptible_cpu(int cpu) > > +{ > > + struct rq *this = cpu_rq(cpu); > > + unsigned long val = 0; > > + int i; > > + > > + for_each_online_cpu(i) > > + val += per_cpu(*this->nr_uninterruptible, i); > > + > > + return val; > > +} > > > > > I suspect you've got an accounting leak here on hot-plug.
And I think you're right about the accounting leak with cpu hotplug. I'll do more tests with this part, until I come up with a better idea in general for the nr_uninterruptible accounting.
Thanks! -Andrea
> > > > unsigned long nr_uninterruptible(void) > > { > > unsigned long i, sum = 0; > > > > for_each_possible_cpu(i) > > - sum += cpu_rq(i)->nr_uninterruptible; > > + sum += nr_uninterruptible_cpu(i); > > > > /* > > * Since we read the counters lockless, it might be slightly > > And this makes O(n^2) runtime! >
| |