Assessing gains from parallel computation on a supercomputer

We assess gains from parallel computation on Backlight supercomputer. The information transfers are expensive. We find that to make parallel computation efficient, a task per core must be sufficiently large, ranging from few seconds to one minute depending on the number of cores employed. For small problems, the shared memory programming (OpenMP) and a hybrid of shared and distributive memory programming (OpenMP&MPI) leads to a higher efficiency of parallelization than the distributive memory programming (MPI) alone.