Re: [PATCH 4/8] cpufreq/schedutil: sysfs capacity margin tunable
From: Rafael J. Wysocki
Date: Thu Mar 17 2016 - 18:34:56 EST
On Thu, Mar 17, 2016 at 7:56 PM, Michael Turquette
> Quoting Juri Lelli (2016-03-17 10:54:07)
>> On 17/03/16 15:53, Patrick Bellasi wrote:
>> > On 17-Mar 06:55, Steve Muckle wrote:
>> > > On 03/17/2016 02:40 AM, Juri Lelli wrote:
>> > > >> Could the default schedtune value not serve as the out of the box margin?
>> > > >>
>> > > > I'm not sure I understand you here. For me schedtune should be disabled
>> > > > by default, so I'd say that it doesn't introduce any additional margin
>> > > > by default. But we still need a margin to make the governor work without
>> > > > schedtune in the mix.
>> > >
>> > > Why not have schedtune be enabled always, and use it to add the margin?
>> > > It seems like it'd simplify things.
>> > Actually one of the effects we noticed when SchedTune and SchedFreq
>> > are both in use is that we have a sort of "double boosting" effect.
>> > SchedTune boosts the CPU utilization signal, thus already providing a
>> > sort of margin for the selection of the OPP. This margin overlaps with
>> > the SchedFreq margin, which in turns could results in the selection of
>> > an OPP even more higher than required (with boost already accouned).
>> > > I haven't looked at the schedtune code at all so I don't know whether
>> > > this makes sense given its current implementation.
>> > The current implementation requires review, of course ;-)
>> > Last (and only) posting is based on top of SchedFreq code, as it was
>> > at that time.
>> > > But conceptually I don't know why we'd need or want one margin in
>> > > schedutil which will be tunable, and then another mechanism for
>> > > tuning as well.
>> > I agree with Steve on the conceptual standpoint. The main goal of
>> > SchedTune is actually to provide a "single tunable" to bias many
>> > different subsystem in a "consistent" way. Thus, from a conceptual
>> > standpoint, IMO it makes sens to investigate better how the boost value
>> > can be linked with SchedFreq.
>> > A possible option can be to:
>> > 1. use an hardcoded margin (M) defined by SchedFreq
>> > this margin is used to trigger OPP jumps
>> > when SchedTune _is not_ in use
>> > 2. "compose" the M margin with a boost value defined margin (B)
>> > when SchedTune _is_ in use
>> > This means, e.g.
>> > schedfreq_margin = max(M, B)
>> > Thus:
>> > a) non boosted tasks (and in general when SchedTune is not in use)
>> > gets OPPs jumps based on the hardcoded M margin
>> > b) boosted tasks can get more aggressive OPPs jumps based on the B
>> > margin
>> > While the M margin is hardcoded, the B one is defined via CGroups
>> > depending on the how much tasks needs to be boosted.
>> Makes sense to me. And I think M margin is the one we don't want to make
>> part of the ABI and only play with it under DEBUG.
> Regarding "composing" the margin, schedtune could even overwrite the
> margin entirely via cpufreq_set_cfs_capacity_margin (see patch #2 in
> this series). This avoids complications around a "double boosting"
> Either way, it sounds like the schedtune angle is something that we can
> figure out in due time and change the code as needed later on. For
> schedutil to make sense for frequency-invariant platforms we do need a
> margin today, and there is desire to tune it easily, so I will move this
> sysfs knob to a debug knob in v2.
Also, if you look at the latest iteration of the schedutil patch
(https://patchwork.kernel.org/patch/8612561/), it maps the choice of
the margin to the choice of the frequency tipping point. That is, the
value of (util / max) for which the frequency will stay the same as it
was before. [For (util / max) below the tipping point the new
frequency will be less than the old one (unless it already is minimum)
and for (util / max) above it, the new frequency will be greater than
the old one.]
The tipping point seems to be a good candidate for a tunable to me,
because its meaning is well defined and the range of values that make
sense is quite easy to figure out too.