Parallel Implementation of ADI-FDTD on Shared and Distributed Memory Computers

We present results of a parallel ADI-FDTD implementation on shared and distributed memory computers using the Message Passing Interface (MPI) library. The ADI-FDTD method involves solving in every time step a tridiagonal matrix system of equations over the xy, xz, yz planes of the domain to update the electric field components. In our parallel implementation, the three-dimensional computational domain was spatially decomposed in two directions. Each tridiagonal matrix system is solved by a single processor, with the parallel computer architecture being exploited to solve multiple systems at the same time. This avoids inefficiency associated with using parallelised tridiagonal solver on a small matrices. The amount of data exchanged between processes was proportional to the volume of the local sub- domain. The code was developed in C programming language using the Message Passing Interface (MPI) library and tested on two high performance computers (IBM p690 symmetric multi-processor and a general purpose distributed memory cluster). Benchmarking simulations ran faster as the number of processors was increased, with a maximum speedup factor of 11 for the shared memory system and 4 for the distributed memory system. For the particular simulations that we ran, the speedup factor did not significantly improve with the use of more than 10 processors for the shared memory system and 7 processors for the distributed memory system.