numpy - ATLAS Linear algebra and openmpi on multicore with python -
i use mpi4py , openmpi on multi-cpu/core machine linear algebra. numpy built using atlas. suppose have 4 core machine , run 4 node python script linear algebra on each node using numpy.
how can ensure atlas not use more 1 core when doing linear algebra in each node? when build atlas, there seems no option have configured running on 1 core @ time. intel mkl, think can set omp_num_threads=1, , behavior guaranteed. there way build atlas purpose? there doesn't seem environment variable equivalent.
i guessing running several blas operations simultaneously on each core of multicore cpu not strategy. can comment on or give rules of thumb or bad idea?
unfortunately maximum number of threads atlas uses cannot controlled environment variable such omp_num_threads. you'll have relink numpy against serial atlas library. see
Comments
Post a Comment