Distributed-memory parallelization of an explicit time-domain volume integral equation solver on Blue Gene/P

Two distributed-memory schemes for efficiently parallelizing the explicit marching-on in-time based solution of the time domain volume integral equation on the IBM Blue Gene/P platform are presented. In the first scheme, each processor stores the time history of all source fields and only the computationally dominant step of the tested field computations is distributed among processors. This scheme requires all-to-all global communications to update the time history of the source fields from the tested fields. In the second scheme, the source fields as well as all steps of the tested field computations are distributed among processors. This scheme requires sequential global communications to update the time history of the distributed source fields from the tested fields. Numerical results demonstrate that both schemes scale well on the IBM Blue Gene/P platform and the memory efficient second scheme allows for the characterization of transient wave interactions on composite structures discretized using three million spatial elements without an acceleration algorithm.