Thread Limit in Numpy

Determining the Underlying Thread Management

The provided output of numpy.show_config() reveals that NumPy is utilizing ATLAS for its linear algebra operations. To limit the number of threads specifically for matrix multiplication, we need to target ATLAS's threading configuration.

Thread Limit Configuration

Before executing your script, set the following environment flags:

• export MKL_NUM_THREADS=1: This flag limits the number of threads used by Intel Math Kernel Library (MKL), which is used for certain operations in NumPy.
• export NUMEXPR_NUM_THREADS=1: This flag constrains the parallelism level in NumExpr, a library employed by NumPy for fast numeric calculations.
• export OMP_NUM_THREADS=1: This flag sets the thread count for OpenMP, a parallel computing platform used by NumPy in some cases.

Troubleshooting Multithreading

If the above flags do not resolve the issue, it is worth investigating if other libraries are introducing multithreading. Explore setting environment flags for those libraries as well, such as:

• export BLAS_NUM_THREADS=1: Specifies the thread count for BLAS, a linear algebra library commonly used by NumPy.
• export LAPACK_NUM_THREADS=1: Sets the thread count for LAPACK, a linear algebra library integrated with BLAS.