Setting the Number of
OpenMP* Threads
Setting
the Number of
OpenMP* Threads The
OpenMP* run-time library responds to the environment variable
also has other mechanisms to set the number of OpenMP threads, such as the
OMP_NUM_THREADS
. Intel® oneAPI Math Kernel Library
MKL_NUM_THREADS
or
MKL_DOMAIN_NUM_THREADS
environment variables (see
Using Additional Threading Control).
Make sure that the relevant environment variables have the same and correct values on all the nodes. does not set the default number of OpenMP threads to one, but depends on the OpenMP libraries used with the compiler to set the default number. For the threading layer based on the Intel compiler (
Intel® oneAPI Math Kernel Library
lib
mkl_intel_thread.
a
), this value is
the number of CPUs according to the OS.
Avoid over-prescribing the number of OpenMP
threads, which may occur, for instance, when the number of MPI ranks per node
and the number of OpenMP threads per node are both greater than one. The number
of MPI ranks per node multiplied by the number of
OpenMP threads per node should not exceed the number of hardware
threads per node.
If
you are using your login environment to set an environment variable, such as
OMP_NUM_THREADS
,
remember that changing the value on the head node and then doing your run, as
you do on a shared-memory (SMP) system, does not change the variable on all the
nodes because
mpirun
starts a fresh
default shell on all the nodes. To change the number of OpenMP threads on all
the nodes, in
.bashrc
, add a line
at the top, as follows:OMP_NUM_THREADS=1; export
OMP_NUM_THREADS
You can run multiple CPUs per
node using MPICH. To do this, build MPICH to enable multiple CPUs per node. Be
aware that certain MPICH applications may fail to work perfectly in a threaded
environment (see the Known Limitations section in the
Release Notes
. If you
encounter problems with MPICH and setting of the number of OpenMP threads is
greater than one, first try setting the number of threads to one and see
whether the problem persists.
For Cluster Sparse Solver, set the number of OpenMP
threads to a number greater than one because the implementation of the solver
only supports a multithreaded algorithm.