Key Techniques And Applications Of Parallel FDTD Based On Supercomputer Platforms

Electromagnetic field and electromagnetic wave is playing an increasingly important role in life. Numerical calculation has the advantage of good accuracy, high efficiency, flexibility, convenience and so on, so it becomes the modern method of electromagnetic characteristics analysis and design, and is playing an increasingly important role. Finite-difference time-domain (FDTD) method, as one of the three classical numerical methods of electromagnetic field, has its own unique advantages. For electrically large and complex (small) structure simulation, huge memory resources and computation time consumption has become limiting factors of the algorithm. With the rapid development of today’s computer hardware and software, a large computer cluster provides the probability for solution to large-scale complex electromagnetic problem. To take advantage of the stronger computing ability platform to solve problems of large scale electromagnetic field becomes one of the hotspots and challenges in the field of computational electromagnetics. As an emerging domestically-made computer in large computer cluster, the compiler environment is aimed at independent development of the domestic CPU architecture, and there are big differences with general computer platform environment. Meanwhile making electromagnetic field research work in the pure domestically-made supercomputer platform has a higher security, so the transplantation and test for autonomous development parallel FDTD algorithm in the domestically-made computer is particularly important.In this thesis, based on domestic supercomputer platforms, several key techniques of the parallel FDTD method and their applications in electromagnetic problems are studied in depth and in detail, the work of the author is mainly focused on:(1). Based on characteristics of the FDTD method, Descartes (Cartesian) topology is used to describe the logical relationship between the sub-domains in the whole computation space, and MPI-based parallel FDTD method is realized. A communication model of affecting parallel FDTD on parallel performance is established. The theoretical guidelines of virtual topology are presented on how to obtain the higher parallel performance. Meanwhile a parallel automatic modeling method is presented to transfer the triangular models into FDTD cubic models, so that the core program is not corresponding with models and the speed of modeling is greatly improved, compared to serial modeling technology. When the problem size is larger, the efficiency can be improved by more than 60%.(2). Under the guidance of the guidelines above, a large number of tests for parallel FDTD virtual topology optimization have been carried out on two general platforms, National Supercomputer Center in Tianjin and National Supercomputing Center in Shenzhen. Through the result analysis on Tianjin center, it is concluded that the better performance of parallel algorithm can be obtained in the virtual topology with less across-nodes communication. And through the results analysis on Shenzhen center, it is concluded that the better performance of parallel algorithm can be obtained in the virtual topology with less main communication task. Summarizing the results on two general platforms and the theoretical rules from the communication model, the general rules of optimal virtual topology are given on general platforms. That is:a. Select MPI virtual topology scheme to make the total communication the smallest. b. When the total communication is the same, set the topology as with less crossing-node communication. c. When the amount of crossing-node communication of different topologies is approximately the same, select the topology with a more balanced communication load. Based on the rules, the algorithm can obtain higher parallel performance.(3). Considering the difference between pure domestically-made CPU computer and general computer platform, optimal virtual topology rules on pure domestically-made platform are studied on how to obtain higher parallel performance. Restricted to domestic CPU memory, the entire tests are divided into several parts. Through the analysis of test results, the rules of optimal virtual topology are proposed, which are basic consistent with that on general platform.(4). On the basis of characteristics of domestic heterogeneous platform, parallel FDTD method is transplanted. For the unique programming way on domestic heterogeneous platforms, the parallel strategy of parallel FDTD method is optimized. Several examples are tested, and the acceleration performance of the algorithm is obtained.(5). The parallel performance of FDTD algorithm was tested on the several different test platforms above. The algorithm is implemented on pure domestically-made CPU computer platforms using maximum number of CPU cores of 10000 and parallel efficiency reach up to 67%,300000 cores on general platform and parallel efficiency reach up to 80%. The maximum number of cores using broke 4 million cores on domestic heterogeneous platform and the Yee grid is about 1 thousand billion. Moreover, according to the search report, one can see that the parallel scale of this title is the largest in the world.(6). Several typical electromagnetic applications have been simulated on different supercomputer platforms, including the antenna radiation characteristics analysis, the scattering characteristics of carrier aircraft platform, disturbance of airborne antenna pattern analysis and the near field aircraft around analysis when lightning attack it, etc.(7). A parallel FDTD software is formed for secondary development using GiD software. Through the software, several integration processes can be realized from modeling, parameter setting, media assignment and MESH subdivision, to calculate and post-processing display, etc.