On Fri, Mar 29, 2013 at 10:40:03AM -0700, Tejun Heo wrote:

So the scheduler does know what it's doing after all, which is a nice
news.  Given the result, the best thing to do would be somehow marking
these workqueues as unbound for powersaving?

BTW, recent changes to workqueue means that it shouldn't be too
difficult to create a combined workqueue which distributes work items
to both per-cpu and unbound worker pools depending on cpu_idle(), but,
at least for now, I think we can make do with much coarser approach.
We can refine it afterwards.

Thanks.

-- 
tejun