Bug 789345

Description of problem:
I'm a member of the procps devel group and we're facing the following issue. It isn't easy to make it clear since we had to internally exchange many emails to avoid confusions.

With introduction of dynamic frequency scaling (cpu throttling) our procps statistics became distorted by the automatic CPU frequency changes (which usually happen several times per sampling interval and thus can't be reliably measured). The main problem lies in the /proc/stat counters. They represent time units (hundreths of seconds) "consumed" by the processes (or time units spent in the idle state) whilst we need to know how many work units (or cpu ticks) were consumed by the task to make the stats frequency independent. Time units can be reliably used in case of CPUs with fixed frequencies, but in case of dynamic frequency scaling these values become useless, because they can't reliably tell us how many reserves we actually have and thus they can't be used for monitoring platforms.
We'd like to know if it's easily possible to measure and export also work units. Maybe the time units are internally evaluated using the work units?

In order to explain the impact I've created a testing application producing a fixed amount of work units each second and then forced CPU to stay at the following fixed frequencies...

The procps results were:
f[GHz]  load[%]
 2.66       34
 1.73       52

whilst in fact the real load is 34% in both cases, but this is distorted by the scaling and the load then looks higher.

The following table shows differences in /proc/stat counters between two subsequent samples (interval 1.0s)

Results (all 4 cores - first /proc/stat line):
f[GHz]  user[]  idle[]  total[]
 1.73     218     180      398
 2.66     142     257      399

Results (first core - second /proc/stat line):
f[GHz]  user[]  idle[]  total[]
 1.73      55      43       98
 2.66      37      63      100

We don't want to break a backward compatibility and thus we don't want to change behaviour or the current counters. We want an introduction of new counters which could be used for frequency independent statistics. Additionally, if You could export a theoretical maximum number of work units which can be processed by a CPU core running at it's maximum frequency, then it would be really great. Such value is constant and could be used for evaluation of the load in percentage. We encourage any reasonable unit prefixes to make the increments not too high and not too low. They could be eventually prescaled to always have any reasonable range.

Please, do not hesitate to ask if anything isn't clear.

Regards,
Jaromir.