Messages in this thread | | | Date | Wed, 8 May 2019 13:41:52 +0100 | From | Quentin Perret <> | Subject | Re: [PATCH V2 1/3] Calculate Thermal Pressure |
| |
Hi Thara,
Sorry for the delayed response.
On Friday 26 Apr 2019 at 10:17:56 (-0400), Thara Gopinath wrote: > On 04/25/2019 08:45 AM, Vincent Guittot wrote: > > Do you mean calling a variant of sched_update_thermal_pressure() in > > update_cpu_capacity() instead of periodic update ? > > Yes , that should be enough > > Hi, > > I do have some concerns in doing this. > 1. Updating thermal pressure does involve some calculations for > accumulating, averaging, decaying etc which in turn could have some > finite and measurable time spent in the function. I am not sure if this > delay will be acceptable for all systems during load balancing (I have > not measured the time involved). We need to decide if this is something > we can live with. > > 2. More importantly, since update can happen from at least two paths ( > thermal fw and periodic timer in case of this patch series)to ensure > mutual exclusion, the update is done under a spin lock. Again calling > from update_cpu_capacity will involve holding the lock in the load > balance path which is possible not for the best. > For me, updating out of load balance minimizes the disruption to > scheduler on the whole. > > But if there is an over whelming support for updating the statistics > from the LB , I can move the code.
If I try to clarify my point a little bit, my observation is really that it's a shame to update the thermal stats often, but to not reflect that in capacity_of().
So in fact there are two alternatives: 1) do the update only during LB (which is what I suggested first) to avoid 'useless' work; or 2) reflect the thermal pressure in the CPU capacity every time the thermal stats are updated.
And thinking more about it, perhaps 2) is actually a better option? With this we could try smaller decay periods than the LB interval (which is most likely useless otherwise) and make sure the capacity considered during wake-up is up-to-date. This should be a good thing for latency sensitive tasks I think. (If you consider a task in the Android display pipeline for example, it needs to run within 16ms or the frame is missed. So, on wake-up, we'd like to know where the task can run fast _now_, not according to the capacities the CPUs had 200ms ago or so).
Thoughts ? Quentin
| |