Performance of FDM Simulation of Seismic Wave Propagation using the ppOpen-APPL/FDM Library on the Intel Xeon Phi Coprocessor

We evaluated the performance of a parallel 3D FDM simulation of seismic wave propagation using the Intel Xeon Phi coprocessor. We confirmed that MPI/OpenMP hybrid parallel computing with hyper-threading is more efficient than pure MPI parallelism. The performance of the thread parallel computing was further improved by fusing the original three DO loops of major kernel routines into two DO loops. The performance of the FDM simulation with two fused DO loops was 1.7-5.1 times faster than the original code with three DO loops however no performance acceleration was achieved for a fused single DO loop calculation. This is probably due to restrictions of the current Fortran compiler optimizations.