Repository navigation
Conversation
fd65ca9 to
0250b51
Compare
|
It might be simpler to e.g. changing each The multi-process benchmarking appears to be useful also for allowing a better sampling of the measurement noise. |
|
The first thing that I tried was a I'll see how things will look with the changes from the statistics pull and maybe just drop support for combining |
|
Just wanted to comment that having this would be great! |
|
@pv - It seems this would be super useful for a number of projects, and I'm happy to rebase to the rid of the conflicts if that would be helpful to get it merged. |
Imports require a fairly specific treatment:
goal_timemay be approximated by adjustingrepeat, butnumberin most cases should be 1.There are a couple of approaches:
subprocessargument;Py_NewInterpreter).The
subprocessflag would be slightly more general, but for what else would it be useful? On the other hand, a specialized benchmark could be extended to also evaluate memory usage of imports (with #179).Properly resetting the interpreter (if at all possible, cleaning
sys.modulesis not sufficient) or benchmarking in sub-interpreters may be hard to keep compatible across multiple versions of Python.Some discussion can be found in numpy/numpy#8159 and astropy/astropy-benchmarks#15.