The Research On Fermi-Architecture GPU Ased FDTD Methods And Related Lgorithms

In the High Performance Computing research areas, because of thehigh amount of processing cores provided by the GPU hardware, theGPU parallel computing became a hot direction in HPC. With theintroduction of Fermi GPU, the general parallel computing performanceobtained high improvement both from software and hardwareperspectives. Therefore, the research on algorithms based on Fermi GPUis of great significance for further optimizing the GPU parallel programsto obtain better performance.During the study of electromagnetic wave propagation, oscillogramof electromagnetic wave at random time from complicated environmentis required by accurately computing. In space, the distribution ofelectrical and magnetic fields follows Maxwell’s equations. However, dueto Maxwell’s equations’ feature of coupling, it’s hard to get theirnumerical solution. By making use of the Finite Different Time Domain Method (FDTD), modern computers can calculate the values of theelectromagnetic field. FDTD method processes the values in space bynumerous iterations, which would be a big challenge for computers.Considering the newest Fermi GPU, this article’s main contents andcontributions are as bellows:1, this paper, through the research on related physical background ofelectromagnetics, built the experimental model for FDTD simulation.Then, we designed and implemented the Fermi GPU based FDTDprogram in three dimensional space. This parallized FDTD program canobtain more than30times speedup than the traditional serial program.2, this paper analyzed the simulation results of the FDTDexperiments on Fermi GPU. Especially, according to the disparity of theFermi GPU and traditional GPU, relevant analysis on the results’precision and accuracy and the comparison on optimization strategieswere made. From both the software and hardware perspectives, the paperanalyzed the possible reason for the difference.3, at last, this article extensively discussed the problems occurredduring the FDTD parallization from the features of Fermi GPU, such asthe superiority and inferiority while using the shared memory as theoptimization method, and how to choose appropriate optimizationstrategy according to different algorithms. By the categorization of thedata (dense and sparse) from the algorithms’ input, this paper derived that on Fermi GPU the algorithms with different input data should choosedifferent optimization strategy so as to get satisfying performancepromotion.