How to set good environment variables for the Intel MKL library¶
The Intel MKL library can be linked in a sequential or in a multi-threaded mode (default); the latter is advantageous for multi-CPU architectures.
In order to benefit from the parallelization of MKL, the user should provide appropriate environment variables.
Examples¶
Safe-play default for processor unit with N cores:
export MKL_NUM_THREADS="N"
export MKL_DYNAMIC="FALSE"
export OMP_NUM_THREADS=1
Allow for 4 threads per core in MKL/BLAS routines:
export MKL_NUM_THREADS=4
export MKL_DOMAIN_NUM_THREADS="MKL_BLAS=4"
export OMP_NUM_THREADS=1
export MKL_DYNAMIC="FALSE"
export OMP_DYNAMIC="FALSE"
Note that is important to properly balance MPI and MKL threads. If all cores are already taken by the MPI, the the “unwanted” feature of the threaded-MKL is (sometimes) significant slow down of the code’s execution run.
As an example of proper splitting of CPUs between MPI and MKL, assume a 16 core node where 8 cores are assigned to MPI and the rest to MKL threads:
export MKL_NUM_THREADS="2"
export MKL_DOMAIN_NUM_THREADS="MKL_BLAS=2"
export OMP_NUM_THREADS="1"
export MKL_DYNAMIC="FALSE"
export OMP_DYNAMIC="FALSE"
If you would like to run a sequential job on the same node it would then read:
export MKL_NUM_THREADS="16"
export MKL_DOMAIN_NUM_THREADS="MKL_BLAS=16"
export OMP_NUM_THREADS="1"
export MKL_DYNAMIC="FALSE"
export OMP_DYNAMIC="FALSE"
You can find recommended settings for calling Intel MKL routines from multi-threaded applications on the Intel web page.
