Re: [PATCH v7 2/2] sched/fair: update scale invariance of PELT

[Vincent Guittot]

On Wed, 28 Nov 2018 at 16:21, Patrick Bellasi <patrick.bellasi@xxxxxxx> wrote:
>
> On 28-Nov 15:55, Vincent Guittot wrote:
> > On Wed, 28 Nov 2018 at 15:40, Patrick Bellasi <patrick.bellasi@xxxxxxx> wrote:
> > >
> > > On 28-Nov 14:33, Vincent Guittot wrote:
> > > > On Wed, 28 Nov 2018 at 12:53, Patrick Bellasi <patrick.bellasi@xxxxxxx> wrote:
> > > > >
> > > > > On 28-Nov 11:02, Peter Zijlstra wrote:
> > > > > > On Wed, Nov 28, 2018 at 10:54:13AM +0100, Vincent Guittot wrote:
> > > > > >
> > > > > > > Is there anything else that I should do for these patches ?
> > > > > >
> > > > > > IIRC, Morten mention they break util_est; Patrick was going to explain.
> > > > >
> > > > > I guess the problem is that, once we cross the current capacity,
> > > > > strictly speaking util_avg does not represent anymore a utilization.
> > > > >
> > > > > With the new signal this could happen and we end up storing estimated
> > > > > utilization samples which will overestimate the task requirements.
> > > > >
> > > > > We will have a spike in estimated utilization at next wakeup, since we
> > > > > use MAX(util_avg@dequeue_time, ewma). Potentially we also inflate the EWMA in
> > > > > case we collect multiple samples above the current capacity.
> > > >
> > > > TBH I don't see how it's different from current implementation with a
> > > > task that was scheduled on big core and now wakes up on little core.
> > > > The util_est is overestimated as well.
> > >
> > > While running below the capacity of a CPU, either big or LITTLE, we
> > > can still measure the actual used bandwidth as long as we have idle
> > > time. If the task is then moved into a lower capacity core, I think
> > > it's still safe to assume that, likely, it would need more capacity.
> > >
> > > Why do you say it's the same ?
> >
> > In the example of a task that runs 39ms in period of 80ms that we used
> > during previous version,
> > the utilization on the big core will reach 709 so will util_est too
> > When the task migrates on little core (512), util_est is higher than
> > current cpu capacity
>
> Right, and what's the problem ?

you worry about an util_est being higher than capacity which is the case there

>
> 1) We know that PELT is calibrated to 32ms period task and in your
>    example, since the runtime is higher then the half-life, it's
>    correct to estimate a utilization higher then 50%.
>
>    PELT utilization is defined _based on the half-life_: thus
>    your task having a 50% duty cycle does not mean we are not correct
>    if report a utilization != 50%.
>    It would be as broken as reporting 10% utilization for a task
>    running 100ms every 1s.
>
> 2) If it was a 70% task on a previous activation, once it's moved into
>    a lower capacity CPU it's still correct to assume that it's likely
>    going to require the same bandwidth and thus will be
>    under-provisioned.
>
> I still don't see where we are wrong in this case :/
>
> To me it looks different then the problem I described.
>
> > > With your new signal instead, once we cross the current capacity,
> > > utilization is just not anymore utilization. Thus, IMHO it make sense
> > > avoid to accumulate a sample for what we call "estimated utilization".

This is not true. With the example above, the util_est will be exactly the same
 on big and little cores with the new signal

> > >
> > > I would also say that, with the current implementation which caps
> > > utilization to the current capacity, we get better estimation in
> > > general. At least we can say with absolute precision:
> > >
> > >    "the task needs _at least_ that amount of capacity".
> > >
> > > Potentially we can also flag the task as being under-provisioned, in
> > > case there was not idle time, and _let a policy_ decide what to do
> > > with it and the granted information we have.
> > >
> > > While, with your new signal, once we are over the current capacity,
> > > the "utilization" is just a sort of "random" number at best useful to
> > > drive some conclusions about how long the task has been delayed.

see my comment above

> > >
> > > IOW, I fear that we are embedding a policy within a signal which is
> > > currently representing something very well defined: how much cpu
> > > bandwidth a task used. While, latency/under-provisioning policies
> > > perhaps should be better placed somewhere else.
> > >
> > > Perhaps I've missed it in some of the previous discussions:
> > > have we have considered/discussed this signal-vs-policy aspect ?
>
> What's your opinion on the above instead ?

It's not a policy but it gives better knowledge about the amount a work done
I have put below discussion on the  subject on previous version

> >
> > With contribution scaling the PELT utilization of a task is a _minimum_
> > utilization. Regardless of where the task is currently/was running (and
> > provided that it doesn't change behaviour) its PELT utilization will
> > approximate its _minimum_ utilization on an idle 1024 capacity CPU.
>
> The main drawback is that the _minimum_ utilization depends on the CPU
> capacity on which the task runs. The two 25% tasks on a 256 capacity
> CPU will have an utilization of 128 as an example
>
> >
> > With time scaling the PELT utilization doesn't really have a meaning on
> > its own. It has to be compared to the capacity of the CPU where it
> > is/was running to know what the its current PELT utilization means. When
>
> I would have said the opposite. The utilization of the task will
> always reflect the same amount of work that has been already done
> whatever the CPU capacity.
> In fact, the new scaling mechanism uses the real amount of work that
> has been already done to compute the utilization signal which is not
> the case currently. This gives more information about the real amount
> of worked that has been computed in the over utilization case.
>
> > the utilization over-shoots the capacity its value is no longer
> > represents utilization, it just means that it has a higher compute
> > demand than is offered on its current CPU and a high value means that it
> > has been suffering longer. It can't be used to predict the actual
> > utilization on an idle 1024 capacity any better than contribution scaled
> > PELT utilization.
>
> I think that it provides earlier detection of over utilization and
> more accurate signal for a longer time duration which can help the
> load balance
> Coming back to 50% task example . I will use a 50ms running time
> during a 100ms period for the example below to make it easier
>
> Starting from 0, the evolution of the utilization is:
>
> With contribution scaling:
>          time  0ms  50ms  100ms  150ms  200ms
> capacity
> 1024           0    666
> 512            0    333   453
> When the CPU start to be over utilized (@100ms), the utilization is
> already too low (453 instead of 666) and scheduler doesn't detect yet
> that we are over utilized
> 256            0    169   226    246    252
> That's even worse with this lower capacity
>
> With time scaling,
>          time  0ms  50ms  100ms  150ms  200ms
> capacity
> 1024           0    666
> 512            0    428   677
> We know that the current capacity is not enough and the utilization
> reflect the correct utilization level compare to 1024 capacity (the
> 666 vs 677 difference comes from the 1024us window so the last window
> is not full in the case of max capacity)
> 256            0    234   468    564    677
> At 100ms, we know that there is not enough capacity. (In fact we know
> that at 56ms). And even at time 200ms, the amount of work is exactly
> what would have been executed on a CPU 4x faster
>
> >
> > This change might not be a showstopper, but it is something to be aware
> > off and take into account wherever PELT utilization is used.
>
> The point above is clearly a big difference between the 2 approaches
> of the no spare cycle case but I think it will help by giving more
> information in the over utilization case
>
> Vincent
> >
> > Morten

>
> --
> #include <best/regards.h>
>
> Patrick Bellasi