Messages in this thread | | | Date | Thu, 3 Apr 2008 08:14:11 -0500 | From | Paul Jackson <> | Subject | Re: [PATCH 1/2] Customize sched domain via cpuset |
| |
H.Seto wrote: > I believe there are quite technical reasons why we have no "idle_map."
I would not advocate a single system wide cpumask of idle CPUs. As Peter Zijlstra notes in a follow up post, that's too hot a cache line and clearly doesn't scale.
But I would think it would be ok to have a separate cpumask per node, that marked just the node-local CPUs. We have other per-node data already. If we only support this optional load balancing level across the other CPUs on the same node (or smaller domains, such as the cores in a package), that should work, shouldn't it?
> so I'll update my patch to take levels, getting in your suggestion.
If you see a good solution here that you can provide, good. But if my brain storming ideas have problems, don't hesitate to object to them.
-- I won't rest till it's the best ... Programmer, Linux Scalability Paul Jackson <pj@sgi.com> 1.940.382.4214
| |