`mp.cpu_count()` not reliable on HPC

[chrishalcrow]

PR #4305, added a fast_mode for computing correlograms, which uses mp.cpu_count() to figure out how many threads to use.

On my HPC, this counts ALL the CPUs (hundreds) on the node rather than the ones in my instance (8).

Is there something else we could do @yger @samuelgarcia ? I can just turn fast_mode off for now.

EDIT: I can't turn fast mode off for the correlogram computation in slidingRP_violations.

[yger]

Arg, damned. I'll look into that, other option would be to provide a number here, and let the user decide how many CPU to use fast_mode could be -1, 0 or a number. But then, with your HPC, how does it work if you use global n_jobs params from Spikeinterface ?

[chrishalcrow]

It's fine if I set the n_jobs (if it's not -1). Any functions which try and guess the CPU or memory available tend to fail.

[yger]

Ok, so basically you would be happy with a n_jobs option instead of fast_mode, am I right?

[chrishalcrow]

Yup yup.

[yger]

This has been superseeded by #4442

[alejoe91]

Fixed by #4442