> > I did some experiements with forcing different frequencies.  I haven't
> > finished processing the results, but I notice that as the frequency goes
> > up, the utilization (specifically the value of
> > map_util_perf(sg_cpu->util) at the point of the call to
> > cpufreq_driver_adjust_perf in sugov_update_single_perf) goes up as well.
> > Is this expected?
>
> It isn't, as long as the scale-invariance mechanism mentioned in my
> previous message works properly.

The results are attached.  I have now tried four machines:

2 x Intel Xeon Gold 6130, 16 cores/CPU
2 x Intel Xeon Gold 5218R, 20 cores/CPU
Intel Xeon Gold 5220, 18 cores/CPU
2 x Intel Xeon Gold 6130, 16 cores/CPU

In the graphs, for a particular execution, I take all of the reported
utilizations at the call to cpufreq_driver_adjust_perf associated with a
given application, and then take the average.  Only the five applications
with the most data points are included.  Typically, most of the execution
time is associated with the first one (blue).  For each application, the
legend contains the least to the greatest utilization when fixing the
pstates as indicated in the x axis.

julia