Messages in this thread | | | Subject | Re: 4.2.2: NR_CPUS effectively being 1 bug | From | Mike Galbraith <> | Date | Sat, 03 Oct 2015 14:49:01 +0200 |
| |
On Sat, 2015-10-03 at 14:31 +0200, Frederic Weisbecker wrote: > On Sat, Oct 03, 2015 at 02:04:44PM +0200, Mike Galbraith wrote: > > On Sat, 2015-10-03 at 10:52 +0200, Ingo Molnar wrote: > > > Ok, looks like a patch with good intentions but bad effects. Mind sending a revert > > > patch, changelogged, signed off? > > > > No. They know. > > The reason for this patch is that NO_HZ_FULL is only useful on a CPU if no task > other than the desired one can be scheduled on it. Hence the cpu_isolated_map.
Yes, but makes it needlessly static.
> Only those who enable NO_HZ_FULL_ALL by accident do complain, not those who > really use it so far. At least it makes people realize their mistake. > > That said I never liked that cpu_isolated_map. And some regular non-isolation > work may be needed to be done even on NO_HZ_FULL_ALL machines and it that > case we get screwed.
ATM, using nohz_full CPUs for generic work has a high price, but those CPUs work just fine. Andy is allegedly gonna make that overhead go away, at which time dynamic sets become a much more attractive, but you can do that now.
> So I should revert that and defer that isolation work to explicit affinity > setting or cpusets.
Yay.
-Mike
| |