Re: [PATCH v2 1/1] x86,sched: On AMD EPYC set freq_max = max_boost in schedutil invariant formula

["Rafael J. Wysocki" <rafael@kernel.org>]

On Tue, Feb 2, 2021 at 7:45 PM Rafael J. Wysocki <rafael@kernel.org> wrote:
>
> On Tue, Jan 26, 2021 at 5:19 PM Giovanni Gherdovich <ggherdovich@suse.cz> wrote:
> >
> > On Mon, 2021-01-25 at 11:04 +0100, Peter Zijlstra wrote:
> > > On Fri, Jan 22, 2021 at 09:40:38PM +0100, Giovanni Gherdovich wrote:
> > > > This workload is constant in time, so instead of using the PELT sum we can
> > > > pretend that scale invariance is obtained with
> > > >
> > > >     util_inv = util_raw * freq_curr / freq_max1        [formula-1]
> > > >
> > > > where util_raw is the PELT util from v5.10 (which is to say, not invariant),
> > > > and util_inv is the PELT util from v5.11-rc4. freq_max1 comes from
> > > > commit 976df7e5730e ("x86, sched: Use midpoint of max_boost and max_P for
> > > > frequency invariance on AMD EPYC") and is (P0+max_boost)/2 = (2.25+3.4)/2 =
> > > > 2.825 GHz.  Then we have the schedutil formula
> > > >
> > > >     freq_next = 1.25 * freq_max2 * util_inv            [formula-2]
> > > >
> > > > Here v5.11-rc4 uses freq_max2 = P0 = 2.25 GHz (and this patch changes it to
> > > > 3.4 GHz).
> > > >
> > > > Since all cores are busy, there is no boost available. Let's be generous and say
> > > > the tasks initially get P0, i.e. freq_curr = 2.25 GHz. Combining the formulas
> > > > above and taking util_raw = 825/1024 = 0.8, freq_next is:
> > > >
> > > >     freq_next = 1.25 * 2.25 * 0.8 * 2.25 / 2.825 = 1.79 GHz
> > >
> > > Right, so here's a 'problem' between schedutil and cpufreq, they don't
> > > use the same f_max at all times.
> > >
> > > And this is also an inconsistency between acpi_cpufreq and intel_pstate
> > > (passive). IIRC the intel_pstate cpufreq drivers uses 4C/1C/P0 resp,
> > > while ACPI seems to stick to P0 f_max.
> >
> > That's correct. A different f_max is used depending on the occasion. Let me
> > rephrase with:
>
> OK, I confused the terminology, sorry about that.
>
> > cpufreq core asks the driver what's the f_max. What's the answer?
> >
> > intel_pstate says: 1C
>
> Yes, unless turbo is disabled, in which case it is P0.

BTW, and that actually is quite important, the max_freq reported by
intel_pstate doesn't matter for schedutil after the new ->adjust_perf
callback has been added, because that doesn't even use the frequency.

So, as a long-term remedy, it may just be better to implement
->adjust_perf in acpi_cpufreq().

Again, I'm terribly sorry for missing this thread and the patch.

> > acpi_cpufreq says: P0
>
> This is P0+1, isn't it?
>
> > scheduler asks the freq-invariance machinery what's f_max, because it needs to
> > compute f_curr/f_max. What's the answer?
> >
> > Intel CPUs: 4C in most cases, 1C on Atom, something else on Xeon Phi.
> > AMD CPUs: (P0 + 1C) / 2.
> >
> >
> > Legend:
> > 1C = 1-core boost
> > 4C = 4-cores boost
> > P0 = max non-boost P-States