This is an old revision of the document!
Comsol is licensed software, users will need to supply their own license or have permission from the license holder. The performance of Comsol is highly sensitive to how you allocate the available cores between MPI processes (comsol -nn runs MPI like mpirun -np) and OpenMP threads, and Comsol is generally much faster on Intel than on AMD. We have tested on our nodes as follows:
Each 2-socket Intel server (standard queues): Use -nn 1 and OMP_NUM_THREADS=(cores per node)=(12 or 16) Each 4-socket Intel server (768 GB): Use -nn 2 and OMP_NUM_THREADS=(cores per node)/nn=16 Each 4-socket AMD server (512 GB): Use -nn 2 and OMP_NUM_THREADS=(cores per node)/nn/2=8 Each 4-socket AMD server (64 GB Trestles): Not installed
Example with two-socket 16-core server:
export OMP_NUM_THREADS=16 /share/apps/comsol/comsol52/multiphysics/bin/glnxa64/comsol -nn 1 batch -inputfile dipole_antenna.mph >logfile
Example with multiple nodes
export OMP_NUM_THREADS=8 module load impi/5.1.2 /share/apps/comsol/comsol52/multiphysics/bin/glnxa64/comsol -nn 4 batch -f $PBS_NODEFILE \ -inputfile inkjet_nozzle_ls.mph >logfile