Possible regression in nested parallelism #12389
Description
I think that we have a possible regression linked to the nested parallelism. Let's take the following minimum example:
import numpy as np
from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import cross_val_score
X = np.random.randn(500000, 80)
y = np.random.randint(0, 2, 500000)
clf = RandomForestClassifier(n_jobs=-1)
cross_val_score(clf, X, y, n_jobs=1) In 0.19.2, htop shows activity for all processors. In 0.20, a single core is active. Is it link to the heuristic to avoid to over-subscription?
Can anyone try to reproduce it? ping @tomMoral @ogrisel @pierreglaser