Performance regression in MNE connectivity

[toni-neurosc]

I was trying to get some performance metrics to present my contributions at the lab meeting next week, and I noticed that while most features are faster to calculate now than they were before, the opposite happened with MNE connecivity:

• v0.02: 3.52s for 5ch x 100k artificial data points
• Current: 20.3s for the same data

This is a tremendous performance regression so some kind of bug must have been introduced at some point, need to investigate.

[toni-neurosc]

Ok, I did the investigation and concluded that there was no regression, it was an artifact in the profiling results caused by the logging to the console that MNE was doing in prior versions (before I set it to verbose=False). It probably has to do with the specific profiler I'm using (yappi) or how i configured it (yappi.set_clock_type("cpu")).

I need to find a more reliable way to profile. Perhaps try other profilers:

• py-spy by Ben Frederickson from NVIDIA: https://github.com/benfred/py-spy
• py-instrument by Joe Rick from PyPA https://github.com/joerick/pyinstrument
• Scalene by the PLASMA team at UMass https://github.com/plasma-umass/scalene
• Good ol' cProfile (python built-in)