Re: [PATCH v3 06/14] sched/cpufreq: uclamp: add utilization clamping for RT tasks

[Patrick Bellasi]

On 16-Aug 12:34, Dietmar Eggemann wrote:
> On 08/13/2018 05:01 PM, Patrick Bellasi wrote:
> >On 13-Aug 16:06, Vincent Guittot wrote:
> >>On Mon, 13 Aug 2018 at 14:49, Patrick Bellasi <patrick.bellasi@xxxxxxx> wrote:
> >>>On 13-Aug 14:07, Vincent Guittot wrote:
> >>>>On Mon, 13 Aug 2018 at 12:12, Patrick Bellasi <patrick.bellasi@xxxxxxx> wrote:
> >
> >[...]
> >
> >>>Yes I agree that the current behavior is not completely clean... still
> >>>the question is: do you reckon the problem I depicted above, i.e. RT
> >>>workloads eclipsing the min_util required by lower priority classes?
> >>
> >>As said above, I don't think that there is a problem that is specific
> >>to cross class scheduling that can't also happen in the same class.
> >>
> >>Regarding your example:
> >>task TA util=40% with uclamp_min 50%
> >>task TB util=10% with uclamp_min 0%
> >>
> >>If TA and TB are cfs, util=50% and it doesn't seem to be a problem
> >>whereas TB will steal some bandwidth to TA and delay it (and i even
> >>don't speak about the impact of the nice priority of TB)
> >>If TA is cfs and TB is rt, Why util=50% is now a problem for TA ?
> >
> >You right, in the current implementation, where we _do not_
> >distinguish among scheduling classes it's not possible to get a
> >reasonable implementation of a per sched class clamping.
> >
> >>>To a certain extend I see this problem similar to the rt/dl/irq pressure
> >>>in defining cpu_capacity, isn't it?
> >
> >However, I still think that higher priority classes eclipsing the
> >clamping of lower priority classes can still be a problem.
> >
> >In your example above, the main difference between TA and TB being on
> >the same class or different classes is that in the second case TB
> >is granted to always preempt TA. We can end up with a non boosted RT
> >task consuming all the boosted bandwidth required by a CFS task.
> >
> >This does not happen, apart maybe for the corner case of really
> >different nice values, if the tasks are both CFS, since the fair
> >scheduler will grant some progress for both of them.
> >
> >Thus, given the current implementation, I think it makes sense to drop
> >the UCLAMP_SCHED_CLASS policy and stick with a more clean and
> >consistent design.
> 
> I agree with everything said in this thread so far.

Cool!

> So in case you skip UCLAMP_SCHED_CLASS [(B) combine the clamped
> utilizations] in v4, you will only provide [A) clamp the combined
> utilization]?

Right... unless I find time to add support to per scheduling class
tracking of clamps values. It should be relatively simple... but I
guess it's also something we can keep as a really low prio and propose
it once the main bits are not controversial anymore.

> I assume that we don't have to guard the util clamping for rt tasks behind a
> disabled by default sched feature because all runnable rt tasks will have
> util_min = SCHED_CAPACITY_SCALE by default?

So, that's what Quentin also proposed in a previous discussion:
   https://lore.kernel.org/lkml/20180809155551.bp46sixk4u3ilcnh@queper01-lin/

but yes, you're right: it's in my todo list to ensure that by default
RT tasks get a task-specific util_min set SCHED_CAPACITY_SCALE.

> >I'll then see if it makes sense to add a dedicated patch on top of the
> >series to add a proper per-class clamp tracking.
> 
> I assume if you introduce this per-class clamping you will switch to use the
> UCLAMP_SCHED_CLASS approach?

Likely... but at that point we probably don't need the sched feature
anymore and it could be just the default and unique aggregation
policy.

But let see when we will have the patches... and we don't necessarily
need them for v4.

Best,
Patrick

-- 
#include <best/regards.h>

Patrick Bellasi