Parallelization

chmpy is pure Python and gets its parallelism from threads. numpy and scipy release the GIL while they work, so the threads run side by side.

The workers argument

Functions that can use several cores take a workers argument, which means the same thing it does in scipy: the number of threads to use, or -1 for one per CPU. -1 is the default.

Function	Parallel part
chmpy.calc.nearest_neighbors	the KD-tree search
chmpy.descriptors.adaptive_cna, interval_cna, fixed_cna	the neighbour search, then the analysis in chunks of atoms
chmpy.descriptors.cna.neighbor_bonds, cna_signatures, classify	chunks of atoms

The answer does not depend on workers. Pass workers=1 when you run several analyses side by side yourself, for example one per trajectory frame across a process pool, so that each one stays on a single core:

from chmpy.descriptors import adaptive_cna

types = adaptive_cna(positions, cell, workers=1)

BLAS threads

numpy hands matrix products to a BLAS library, which keeps its own pool of threads. Most of chmpy's products are too small for that to matter, but if it competes with your own parallelism, limit it with the environment variable for your BLAS, set before Python starts:

export OMP_NUM_THREADS=1         # OpenBLAS, as shipped in numpy's wheels
export VECLIB_MAXIMUM_THREADS=1  # Apple Accelerate