Based On Mpi - Tv University Target Electromagnetic Scattering Of Openmp Hybrid Parallel Computing Research

The study of electromagnetic(EM) scattering from electrically-large targets isalways a hot topic and challenge as well in the field of computational electromagnetic.In order to solve the problems of EM scattering from electrically-large conductors andthin dielectric coating objects accurately and efficiently, this thesis adoptsparallelization of Multilevel Fast Multipole Algorithm based on MPI-OpenMP.Firstly, Method of Moments (MoM) and it’s fast algorithms including FastMultipole Method(FMM) and Multilevel Fast Multipole Algorithm (MLFMA) arebriefly stated. From three essential elements of electromagnetic parallelization involvinghardware architecture, software environment and parallel algorithm, high performanceparallel platform(Cluster), the software environment including MPI and OpenMP, thetheory of MLFMA and the hybrid parallel strategy of MLFMA are introduced.Secondly, based on MPI, hybrid strategy of MLFMA is adopted to parallizeMLFMA on the distributed memory architecture. What’s more, the parallelization ofaggregation phase, translation phase and disaggregation phase is fully discussed. Inparticular, how to parallelize the transitional level and computation of scattering field isstated here.Aimed at a better integration with the popular high performance parallelarchitecture, this thesis proposes a hybrid parallelization of MLFMA usingMPI-OpenMP on the symmetrical multi-processing (SMP) cluster. Taking intoconsideration of coexistence of distributed memory and shared memory of SMP cluster,we introduce MPI-OpenMP to provide inter-node and intra-node two-level parallelisms.MPI is responsible for inter-node parallelism of processes, which is coarse grain, whileOpenMP is in charge of intra-node interaction of multiprocessors by providinglightweight parallelism of threads. According to the two-level parallelisms, two-levelpartition mode is employed when parallelizing the near-field interaction and far-fieldinteraction of MLFMA. And then the MPI-OpenMP hybrid scheme is optimized.Theoretical and numerical analysis show these phenomenon: MPI-OpenMP hybridscheme integrating well with the parallel architecture, avoiding intra-node communication, depressing the load imbalance and improving the parallel efficiency.Furthermore, so as to accelerate the iteration of parallelized MLFMA, the parallelizationof preconditioners involving block diagonal(BD) and sparse approximate inverse(SAI)is discussed. In the parallelization of SAI, we repartition the boxes in the finest level interms of the computational costs in SAI construction reaching the goal of load balanceand reducing the communications.Due to the fact that the characteristics of scattering from electrically-large thindielectric coating objects is always a problem concerned and explored by scholars. Theimpedance boundary condition and generalized thin-dielectric-coating approximationwhich are often used in analyzing the dielectric coating problems are incorporated intoour previously well developed parallel MLFMA code. This incorporation willefficiently solve the problem of scattering from thin dielectric coating objects.