Toward large scale Frequency Dependent FDTD for UWB systems

Parallel FD-FDTD method is proposed. Different parallelisation techniques are approached to deal with the large scale computation. Performance measurements indicate that using MPI for shared memory parallelism provides performance benefits for this particular problem. Even if MPI introduces some overhead in data communication, the fact that most of the runtime occurs in calculations and not in communication significantly reduces the ratio of these overheads in total runtime, to the point where OpenMP overheads weigh more heavily. Factors such as problem size, hardware architecture and network interconnect affect the ratio of computation to communication time, so this conclusion cannot be generalised. Since the application has very intensive memory usage requirements, and despite the fact that the algorithm is inherently parallel, scalability bottlenecks occur in architectures where access to memory is shared between more than one PE.