Study Of Hybrid And Computing Technology For Fast Multipole Algorithm

As we all know,in order to manufacture large-scale equipment such as ships and airplanes with better performance,it is necessary to evaluate their electromagnetic characteristics in the design process.Limited to the problems of large site and equipment,it is difficult to measure the electromagnetic characteristics of these equipment.Therefore,at present,numerical analysis of large-scale equipment such as ships and airplanes is carried out based on electromagnetic theory to obtain their electromagnetic characteristics which is one of the most important means at present.The calculation of electromagnetic scattering characteristics of aircraft and ships belongs to the electromagnetic calculation problem of electrically large-scale models.How to quickly solve these problems and obtain results that meet engineering requirements has always been a hot spot in computational electromagnetics.Limited by the time and space complexity of the algorithm,the early electromagnetic field integral equation methods such as The Method of Moments(MOM)were relatively small in solving scales.Later,someone improved The Method of Moment solving process from the mathematical and physical levels,and proposed The Multilevel Fast Multipole Algorithm(MLFMA).MLFMA reduces the amount of calculation and storage to O(Nlog N),which makes it possible to analyze and calculate the electromagnetic scattering of electrically large targets.In order to further speed up the efficiency of solving the problem,the algorithm must be improved at the computer level.In recent years,heterogeneous computing technology has developed rapidly.The GPU accelerator cards commonly used in heterogeneous computing have greatly improved their floating-point computing capabilities and storage capacity.The GPU software programming framework CUDA has also added many new features.Computing clusters are also developing in the direction of CPU-GPU heterogeneous parallelism.Therefore,it is necessary to study how to use the new features of CUDA to realize CPU-GPU heterogeneous parallel acceleration of MLFMA.In addition,it takes a long time for MLFMA to analyze the electromagnetic scattering characteristics of the super-electric large model,the use of cluster system resources,and the problem of program reliability operation are acute.It is also necessary to study MLFMA fault tolerance technology.This thesis mainly studies the following aspects.(1)Combined with the current computer development trend,in order to further improve the calculation efficiency and solution scale of MLFMA,this thesis studies the parallel algorithm of MLFMA based on the main processor and coprocessor heterogeneous platform.The research focuses on the far-field interaction in MLFMA,which includes three processes:aggregation,transfer and disaggregation.Firstly,the implementation of aggregation,transfer and disaggregation on GPU is studied to realize the load balancing and thread parallel between the process blocks in GPU.Aiming at the problem that the data transmission between CPU and GPU takes a long time and affects computing efficiency,the "GPU memory-CPU memory" off-core technology,GPU memory and CPU memory asynchronous data transmission technology is used to further optimize the overall calculation process,improve the efficiency of algorithm execution,and expand the algorithm solution.scale.For the completed heterogeneous parallel MLFMA subroutines,different models and examples are used to verify the correctness of the heterogeneous parallel MLFMA algorithm,and compared with the traditional isomorphic parallel MLFMA algorithm,it verifies the acceleration effect of the heterogeneous parallel MLFMA in this thesis.(2)In view of the unexpected interruption of the supercomputer cluster computing node failure in the parallel MLFMA iterative solution,the parallel iterative solution part is improved,and the approximate solution after each iteration of MLFMA is written to the disk to realize the interruption of the iterative solution process.Improve the reliability and robustness of the multi-layer fast multi-pole algorithm.