Messages in this thread | | | Date | Fri, 9 Oct 2015 13:49:34 +0100 | From | Morten Rasmussen <> | Subject | Re: [RFCv5 PATCH 25/46] sched: Add over-utilization/tipping point indicator |
| |
On Tue, Sep 29, 2015 at 01:08:46PM -0700, Steve Muckle wrote: > On 08/14/2015 06:02 AM, Morten Rasmussen wrote: > > To be sure not to break smp_nice, we have defined over-utilization as > > when: > > > > cpu_rq(any)::cfs::avg::util_avg + margin > cpu_rq(any)::capacity > > > > is true for any cpu in the system. IOW, as soon as one cpu is (nearly) > > 100% utilized, we switch to load_avg to factor in priority. > > > > Now with this definition, we can skip periodic load-balance as no cpu > > has an always-running task when the system is not over-utilized. All > > tasks will be periodic and we can balance them at wake-up. This > > conservative condition does however mean that some scenarios that could > > benefit from energy-aware decisions even if one cpu is fully utilized > > would not get those benefits. > > > > For system where some cpus might have reduced capacity on some cpus > > (RT-pressure and/or big.LITTLE), we want periodic load-balance checks as > > soon a just a single cpu is fully utilized as it might one of those with > > reduced capacity and in that case we want to migrate it. > > > > I haven't found any reasonably easy-to-track conditions that would work > > better. Suggestions are very welcome. > > Workloads with a single heavy task and many small tasks are pretty > common. I'm worried about the single heavy task tripping the > over-utilization condition on a b.L system, EAS getting turned off, and > small tasks running on big CPUs, leading to an increase in power > consumption. > > Perhaps an extension to the over-utilization logic such as the following > could cause big CPUs being saturated by a single task to be ignored? > > util(cpu X) + margin > capacity(cpu X) && > (capacity(cpu X) != max_capacity ? 1 : nr_running(cpu X) > 1)
I have had the same thought as well. I think it could work. nr_running() doesn't take into account blocked tasks, so we could in theory see fewer tasks than there is, but those scenarios are currently ignored by load-balancing anyway and if the cpu is seriously over-utilized we are quite likely to have nr_running() > 1.
I'm in favor of giving it a try and see what explodes :-)
| |