[lkml]   [2012]   [Jul]   [20]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [PATCHSET] workqueue: reimplement CPU hotplug to keep idle workers
Hey, again.

On Fri, Jul 20, 2012 at 07:21:17PM +0200, Peter Zijlstra wrote:
> > So, the above was my rationale before this "we need to stop destroying
> > and re-creating kthreads across CPU hotplug events because phones do
> > it gazillion times". Now, I don't think we have any other way.
> OK, so why can't you splice the list of works from the CPU going down
> onto the list of the CPU doing the down and convert any busy worker
> threads to be bound to the cpu doing down?
> That way there's nothing 'left' to get back to on up.

As I wrote above, per-cpu workqueues don't really interact with each
other and there's no mechanism to transfer work items from one to
another, which unfortunately isn't trivial due to backlinks from work
item to cpu workqueue which is necessary for flush / cancel
operations. I'm sure it's doable but that part is already pretty
complex (already was before cmwq and untangling it probably requires
bloating work_struct) and I don't think it's wise to complicate usual
hot paths for hotplug support.

Also, re-binding busy workers is easy. The idle ones are difficult
and we have to do that anyway for PM optimization. What would be the
benefit of not re-binding busy ones at the risk of continually
transferring workers to another CPU given the right workload + CPU
down/up patterns?



 \ /
  Last update: 2012-07-20 20:41    [W:0.096 / U:0.704 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site