Spatial Partitioning Strategy for Parallelization of MLFMA with Reduced Communication

The bottleneck of the spatial partitioning for parallelizing the multilevel fast multipole algorithm (MLFMA) lies in higher levels of the tree, at which boxes are usually fewer than parallel processors, yielding a serious load imbalance. To solve the bottleneck, the higher levels of the tree are truncated to generate plenty of subtrees, which are distributed among processors to facilitate balancing the work load. At the coarsest level, the communication volume during translation between far-away processors is drastically reduced by adopting the far-field approximation. Therefore, the communication mainly occurs between nearby processors, which is favorable for modern computing clusters. In comparison with the parallel strategies that hybridize the spatial partitioning with the k-space partitioning, the proposed approach is more straightforward and shows good scalability.