Research Of Parallel FDTD Algorithm Based On LAN

The Finite difference time domain (FDTD) method has been widely used inComputational Electromagnetics. Because the FDTD algorithm expression is simpleand effective, it is suitable for resolve the electromagnetic problems and achievevisual computing which require recurrencing the numerical value ofElectromagnetic Fields. When calculating complex geometrical structure andguaranteeing the computational accuracy and stability conditions, increasing thenumber of FDTD grid is necessary, as result of that computer memory andcomputing time is so huge that computers in general can not undertake thesecomputing tasks, which is the main ottleneck in FDTD algorithm application. Withthe rapid development of computer hardware, nowadays the computers with parallelcomputing ability and application are more and more extensive. To improve theperformance of parallel computation techniques based on the LAN (Local AreaNetwork) using in Computational Electromagnetics, this dissertation studies how tosolve the subdivision FDTD computational process. Meanwhile, the FDTDalgorithm itself has the parallelism as calculating. In the whole FDTD program,calling the MPI library realizes the electromagnetic field data communication,operation and and collection from each compute nodes in the parallel computersystem and achieves data output in the host node, and then optimize the FDTDparallel algorithm and test the performance of parallel computation.This dissertation studies and complete the parallel computing cluster usingcomputers connected together in LAN in consideration of current Microwave Labhardware conditions. This parallel computing cluster and the whole FDTD parallelalgorithm basic content including three parts: the parallel hardware platform, theparallel software platform, the parallel programming mode. The parallel hardwareplatform contains computer nodes based on the laboratory's local area network anda router. The parallel software platform contains two parts: running environment ofthe parallel software based on MPI library and the parallel software developmentenvironment. The parallel programming mode is based on the MPI environment andmaster and slave mode, which breaks down the parallel FDTD program into a hostprogram and a slave program, and the programmers need writing host and slave parallel calculator program codes. In such a manner, the parallel FDTD program ismore suitable for current the hardware environment and network conditions, inaddition enhancing the stability and expansibility of the parallel computing clustersystem which is appropriate for improvement latterly.The major achievements of this dissertation are as follows: bringing up thestrategy of dynamic subdivision of FDTD computational domain, which distributesthe parallel FDTD calculating tasks to each compute node according to its capacity,writing this function program code with C programming language. These traditionaloptimization methods, such as changing the parallel granularity or using theOpenMP functions and MPI library achieving the concept of two levelparallelization on parallel cluster, all of that can not go to the root of the problemthat achieving the computer resource optimization. The procedures of strategy ofdynamic subdivision of FDTD computational domain are that tests every computingnode performance in the parallel computing system, and according to thenormalization performance index which comes from the slave nodes, the mastercomputer calculates the suitable computing predicted and sends the computerinstruction to each nodes with MPI library. The cluster system will test thisnormalization performance index before executing the parallel program code.The other achievement of this dissertation is that the pointer construction takesthe place of the structure of arrays in the parallel FDTD program data structure.Traditional method of declaring an array is clear and simple, at the same time couldreduce the parallel programming difficulty. On the other hand there are two maindisadvantages of that: the process of declaring an array memory space andinitialization waste of calculation resources too large and the arrays have manydisadvantaged in flexibility of using. Declaring the pointer construction couldremedy the disadvantage of these disadvantages, and using the pointer constructionaccurately and neatly in development of the parallel FDTD computing programscould reduce the complex rate, in the meantime, the parallel calculator programscould be greater simplicity and more compact and efficient.