bss_eval_sources gets slower with more CPUs

What ?

bss_eval_sources gets slower as the number of available CPUs grow. We got up to 4x slower computation with 32 cores compared to only one core (thx @sunits).

Expected behaviour

If not speeding up with number of available CPU cores, at least not slowing down.

How to reproduce :

• Create a file sdr_speed_test.py containing

import numpy as np
from mir_eval.separation import bss_eval_sources

def time_sdr():
    ref = np.random.randn(2, 64000)
    est = np.random.randn(2, 64000)
    sdrs = bss_eval_sources(ref, est, compute_permutation=True)

• Time the execution with all available cores :

python -m timeit -s "from sdr_speed_test import time_sdr" "time_sdr()" 

• Time the execution with a single core:

taskset --cpu-list 0 python -m timeit -s "from sdr_speed_test import time_sdr" "time_sdr()"

The difference is even greater if you have a large number of cores on your machine but I could see the difference on my 8 core laptop. This was replicated on three different machines.

Additional info

Out np.__config__.show() looks like this :

blas_mkl_info:
  NOT AVAILABLE
blis_info:
  NOT AVAILABLE
openblas_info:
    libraries = ['openblas', 'openblas']
    library_dirs = ['/usr/local/lib']
    language = c
    define_macros = [('HAVE_CBLAS', None)]
blas_opt_info:
    libraries = ['openblas', 'openblas']
    library_dirs = ['/usr/local/lib']
    language = c
    define_macros = [('HAVE_CBLAS', None)]
lapack_mkl_info:
  NOT AVAILABLE
openblas_lapack_info:
    libraries = ['openblas', 'openblas']
    library_dirs = ['/usr/local/lib']
    language = c
    define_macros = [('HAVE_CBLAS', None)]
lapack_opt_info:
    libraries = ['openblas', 'openblas']
    library_dirs = ['/usr/local/lib']
    language = c
    define_macros = [('HAVE_CBLAS', None)]

[faroit]

did you try to disable BLAS for testing?

[mpariente]

Thanks for trying. NumPy sends me to OpenBLAS, I’ll close this.