BENCH scaling of a GridSearch with n_jobs

[tomMoral]

Add a benchmark to compare joblib scaling for a GridSearch on synthetic data.
This benchmark allows to compare 2 versions of joblib by running it multiple times.

[codecov]

Codecov Report

Patch coverage has no change and project coverage change: -0.06 ⚠️

Comparison is base (fc6c4ce) 94.66% compared to head (cba3957) 94.61%.

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #1427      +/-   ##
==========================================
- Coverage   94.66%   94.61%   -0.06%     
==========================================
  Files          45       44       -1     
  Lines        6918     7205     +287     
==========================================
+ Hits         6549     6817     +268     
- Misses        369      388      +19     

see 6 files with indirect coverage changes

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

☔ View full report in Codecov by Sentry.
📢 Do you have feedback about the report comment? Let us know in this issue.

[GaelVaroquaux]

Very nice!

The bump between 6 and 8 jobs is quite curious.

[ogrisel]

Thanks, here is what I get on my laptop with 4 performance cores and 4 efficiency cores.

I cannot reproduce the plateauing effect in thomas' plot.

[ogrisel]

Here is the same result with a ylim to start at 0.

I find it less misleading.

[ogrisel]

I messed up with git, but I finally pushed a simpler ylim.

[ogrisel]

I think this kind of plots would be more informative as speed-ups factor vs a chosen baseline (e.g. the slowest or the single thread result from the current branch) but this is already useful as it is so I merged.