ANL, Argonne, USA
RPI, Troy, New York, USA
We examined hybrid parallel infrastructures in order to ensure performance and scalability for beam propagation modeling as we move toward extreme-scale systems. Using an MPI programming interface for parallel algorithms, we expanded the capability of our existing electromagnetic solver to a hybrid (MPI/shared-memory) model that can potentially use the computer resources on future-generation computing architecture more efficiently. As a preliminary step, we discuss a hybrid MPI/OpenMP model and demonstrate performance and analysis on the leadership-class computing systems such as the IBM BG/P, BG/Q, and Cray XK6. Our hybrid MPI/OpenMP model achieves speedup when the computation amounts are large enough to compensate the OMP threading overhead.
