Redundant Design of Parallel MLFMA Distributed Levels on a Cluster with Distributed Memory

Abstract

in this paper, we study and analyze the distributed levels of MLFMA and propose an efficient scheme. This scheme uses the top-down method to construct the distributed levels from the coarsest distributed level which is equally divided. The method can reduce the communications of the ghost boxes compared to the bottom-up method during the aggregation phase. And the non local far-field boxes are redundantly stored and computed using shifting originally during the disaggregation phase, so the downward pass operation does not require communications to obtain the outgoing patterns of these boxes. The numerical results demonstrate that, this strategy can greatly reduce the iterative time and improve parallel efficiency of the whole algorithm effectively.