Profiling Macro Data Flow Graphs for Parallel Implementation of FDTD Computations

In this paper, we present methodology, which enables designing and profiling macro data flow graphs that represent computation and communication patterns for the FDTD (finite difference time domain) problem in irregular computational areas. Optimized macro dataflow graphs (MDFG) for FDTD computations are generated in three main phases: generation of initial MDFG based on wave propagation area partitioning, MDFG nodes merging with load balancing to obtain given number of macro nodes and communication optimization to minimize and balance internode data transmissions. The computation efficiency for several communication systems (MPI, RDMA RB, SHMEM) is discussed. Relations between communication optimization algorithms and overall FDTD computation efficiency are shown. Experimental results obtained by simulation are presented