Re: [PATCH 1/3] sched/fair: Prepare variables for increased precision of EAS estimated energy

[Vincent Guittot]

On Wed, 7 Jul 2021 at 11:48, Lukasz Luba <lukasz.luba@xxxxxxx> wrote:
>
>
>
> On 7/7/21 10:37 AM, Vincent Guittot wrote:
> > On Wed, 7 Jul 2021 at 10:23, Lukasz Luba <lukasz.luba@xxxxxxx> wrote:
> >>
> >>
> >>
> >> On 7/7/21 9:00 AM, Vincent Guittot wrote:
> >>> On Wed, 7 Jul 2021 at 09:49, Lukasz Luba <lukasz.luba@xxxxxxx> wrote:
> >>>>
> >>>>
> >>>>
> >>>> On 7/7/21 8:07 AM, Vincent Guittot wrote:
> >>>>> On Fri, 25 Jun 2021 at 17:26, Lukasz Luba <lukasz.luba@xxxxxxx> wrote:
> >>>>>>
> >>>>>> The Energy Aware Scheduler (EAS) tries to find best CPU for a waking up
> >>>>>> task. It probes many possibilities and compares the estimated energy values
> >>>>>> for different scenarios. For calculating those energy values it relies on
> >>>>>> Energy Model (EM) data and em_cpu_energy(). The precision which is used in
> >>>>>> EM data is in milli-Watts (or abstract scale), which sometimes is not
> >>>>>> sufficient. In some cases it might happen that two CPUs from different
> >>>>>> Performance Domains (PDs) get the same calculated value for a given task
> >>>>>> placement, but in more precised scale, they might differ. This rounding
> >>>>>> error has to be addressed. This patch prepares EAS code for better
> >>>>>> precision in the coming EM improvements.
> >>>>>
> >>>>> Could you explain why 32bits results are not enough and you need to
> >>>>> move to 64bits ?
> >>>>>
> >>>>> Right now the result is in the range [0..2^32[ mW. If you need more
> >>>>> precision and you want to return uW instead, you will have a result in
> >>>>> the range  [0..4kW[ which seems to be still enough
> >>>>>
> >>>>
> >>>> Currently we have the max value limit for 'power' in EM which is
> >>>> EM_MAX_POWER 0xffff (64k - 1). We allow to register such big power
> >>>> values ~64k mW (~64Watts) for an OPP. Then based on 'power' we
> >>>> pre-calculate 'cost' fields:
> >>>> cost[i] = power[i] * freq_max / freq[i]
> >>>> So, for max freq the cost == power. Let's use that in the example.
> >>>>
> >>>> Then the em_cpu_energy() calculates as follow:
> >>>> cost * sum_util / scale_cpu
> >>>> We are interested in the first part - the value of multiplication.
> >>>
> >>> But all these are internal computations of the energy model. At the
> >>> end, the computed energy that is returned by compute_energy() and
> >>> em_cpu_energy(), fits in a long
> >>
> >> Let's take a look at existing *10000 precision for x CPUs:
> >> cost * sum_util / scale_cpu =
> >> (64k *10000) * (x * 800) / 1024
> >> which is:
> >> x * ~500mln
> >>
> >> So to be close to overflowing u32 the 'x' has to be > (?=) 8
> >> (depends on sum_util).
> >
> > Sorry but I don't get your point.
> > This patch is about the return type of compute_energy() and
> > em_cpu_energy(). And even if we decide to return uW instead of mW,
> > there is still a lot of margin.
> >
> > It's not because you need u64 for computing intermediate value that
> > you must returns u64
>
> The example above shows the need of u64 return value for platforms
> which are:
> - 32bit
> - have e.g. 16 CPUs
> - has big power value e.g. ~64k mW
> Then let's to the calc:
> (64k * 10000) * (16 * 800) / 1024 = ~8000mln = ~8bln

so you return a power consumption of 8kW !!!

>
> The returned value after applying the whole patch set
> won't fit in u32 for such cluster.
>
> We might make *assumption* that the 32bit platforms will not
> have bigger number of CPUs in the cluster or won't report
> big power values. But I didn't wanted to make such assumption.
>
>