Parallelization of a 3D FD-TD Code for the Maxwell Equations Using MPI

We have parallelized an existing code that solves the Maxwell equations in the time domain. The code uses finite differences in 3D for discretization. The method is based on a leap frog scheme introduced by Yee in 1966. The first order Mur scheme is used as outer (absorbing) boundary condition. Wave excitation is done with dipoles, i.e. using point sources. All parts of this scheme are local in space. It is therefore suitable to parallelize the code using domain decomposition. This is done with MPI. It is possible to achieve negligible communication time on an IBM SP-2 when each node houses a large enough problem. This is demonstrated with a problem size of 100×p·100×100 where p is the number of processors. We also give speed-up results for the IBM SP-2 and a Cray J932.