On Fri, Mar 4, 2016 at 12:26 PM, Juri Lelli [off-list ref] wrote:

Hi Rafael,

Hi,

On 04/03/16 04:35, Rafael J. Wysocki wrote:

quoted

From: Rafael J. Wysocki <redacted>

Add a new cpufreq scaling governor, called "schedutil", that uses
scheduler-provided CPU utilization information as input for making
its decisions.

Doing that is possible after commit fe7034338ba0 (cpufreq: Add
mechanism for registering utilization update callbacks) that
introduced cpufreq_update_util() called by the scheduler on
utilization changes (from CFS) and RT/DL task status updates.
In particular, CPU frequency scaling decisions may be based on
the the utilization data passed to cpufreq_update_util() by CFS.

The new governor is relatively simple.

The frequency selection formula used by it is

      next_freq = util * max_freq / max

where util and max are the utilization and CPU capacity coming from CFS.

The formula looks better to me now. However, problem is that, if you
have freq. invariance, util will slowly saturate to the current
capacity. So, we won't trigger OPP changes for a task that for example
starts light and then becomes big.

This is the same problem we faced with schedfreq. The current solution
there is to use a margin for calculating a threshold (80% of current
capacity ATM). Once util goes above that threshold we trigger an OPP
change.  Current policy is pretty aggressive, we go to max_f and then
adapt to the "real" util during successive enqueues. This was also
tought to cope with the fact that PELT seems slow to react to abrupt
changes in tasks behaviour.

I'm not saying this is the definitive solution, but I fear something
along this line is needed when you add freq invariance in the mix.

I really would like to avoid adding factors that need to be determined
experimentally, because the result of that tends to depend on the
system where the experiment is carried out and tunables simply don't
work (99% or maybe even more users don't change the defaults anyway).

So I would really like to use a formula that's based on some science
and doesn't depend on additional input.

Now, since the equation generally is f = a * x + b (f - frequency, x =
util/max) and there are good arguments for b = 0, it all boils down to
what number to take as a.  a = max_freq is a good candidate (that's
what I'm using right now), but it may turn out to be too small.
Another reasonable candidate is a = min_freq + max_freq, because then
x = 0.5 selects the frequency in the middle of the available range,
but that may turn out to be way too big if min_freq is high (like
higher that 50% of max_freq).

I need to think more about that and admittedly my understanding of the
frequency invariance consequences is limited ATM.

Thanks,
Rafael