Hello Marc-Andre.
1) yes, this is nonsense.  I thought that "systimes" was some archaeic 
module that had been discontinued, since I didn't find it in the 
standard modules (where "time" is).  I didn't realize that it was a 
special helper sibling-module to pybench.

3) 'gross', yes, as opposed to 'net'.  The total time of each round.  
For timers which don't update frequently, such as GetProcessTimes(), 
summing up many small contributions compounds the error (in this case, 
+- 10ms per measurement).  Can you be more specific about what it is 
that you don't like?

About the calibration:  I ran into problems where I was tuning the 
memory allocator.  I put HeapAlloc in place of malloc() and found a 
consistent 10% performance degradation, even when using the LFH.  I was 
puzzled by this until I turned off calibration and the effect went away.

The calibration technique is not a bad idea, but it would be helpful, 
when one is comparing runs, to be able to keep the calibration times of 
the previous run, so as not to introduce a random element into the 
comparison.  Alternatively, explain that when doing successive tests, 
the lowest variance in the result is to be expeced with -C 0.

So, how can I improve this?