When running tests with pytest-xdist on a machine with 12 (physical) CPU machine, the use of OpenMP in HistGradientBoosting seem to lead to significant over-subscription,

pytest sklearn/ensemble/_hist_gradient_boosting/gradient_boosting.py  -v

for me takes 0.85s. This runs 2 docstrings on training GBDT classifier and regressor on iris and boston datasets respectively.

• Running thin on 2 parallel processes with (-n 2) takes 56s (and 50 threads are created).
• Running with 2 processes and OMP_NUM_THREADS=2 takes 0.52s

While I understand the case of catastrophic oversubscription when N_CPU_THREADS**2 threads are created on a machine with many cores, here we create 2*N_CPU_THREADS only as compared to 1*N_CPU_THREADS and get a 10x slowdown.

Can someone reproduce it? Here using scikit-learn master, and a conda env on Linux with latest numpy scipy nomkl python=3.7.

Because pytest-xdist uses its own parallelism system (not sure what it does exactly) I guess this won't be addressed by threadpoolctl #14979?

Edit: Originally reported in https://github.com/tomMoral/loky/issues/224