Research On RCS Computing Method Based On Parallel Multilevel Fast Multipole Algorithm

Radar Cross Section(RCS)is a pivotal parameter to describe the electromagnetic scattering properties of the target objects.How to efficiently obtain the RCS of targets,especially to obtain the RCS of complex shaped targets and electrically large targets,has been a hot topic worthy of deep reasearch.To meet the computational needs of complex scattering problems,it has been of great practical value to build a computing system which is competent for RCS computing tasks.Since the Method of Moments(MoM)suffers from temporal and spatial complexity of O(N~2),the Multilevel Fast Multipole Algorithm(MLFMA)with O(N)spatial complexity and O(NlogN)temporal complexity has become an optional solution for implementing an efficient RCS computing system.However,due to the limitation of computing power and memory capacity of a single computer,MLFMA implementated on a single process dose not have the function of solving scattering problems up to millions of unknows.With the advances in computational power nowadays rely on the incorporation of multi-core CPUs and along with the reductions in the cost of commodity computers,it provides a good hardware environment for building the RCS computing system based on parallel MLFMA.The RCS computing system that takes computational efficiency as an important indicator requires the support of a high-performance parallel MLFMA computing method.For the sake of the parallel efficiency of parallel MLFMA,the near-field interaction impedance matrix is allocated to each process according to the load balancing strategy based on a greedy algorithm to achieve efficient near-field interaction calculations;In order to accelerate the computation of the far-field interaction of MLFMA,a distributed oct-tree based on the Morton key technology to store a large number of non-empty group partitions generated in large-scale scattering problems is considered in this paper,and a distributed layer-transition layer-shared layer hybrid partition strategy to distribute the non-empty groups and plane waves in each layer of the oct-tree is implemented.The stage of aggregation,translation and disaggregation with low communication costs based on MPI protocol is well-designed in this paper.The relevant numerical experiments has illustrated that the parallel MLFMA computing method implemented in this paper performs 13 times speedup in solving a large-scale scattering problem more than 8 million unknowns,which has broken through the bottleneck of stand-alone computing,the CPU times has been shortened from 6 hours and 20 minutes to 28 minutes.