Streaming Model Computation of the FDTD Problem

摘要

The Finite Difference Time Domain (FDTD) method enables computerized simulation of the electromagnetic wave propagation. We propose a streaming model for FDTD computations oriented towards a mul-ticore processor architecture. FDTD computations are characterized by injection of small portions of data into computational nodes, processing them and returning the results into main storage. We can parallelize FDTD computations by combining the loop tiling approach and a communication mechanism based on a rotating buffers infrastructure. The described FDTD algorithm has been implemented using both of these techniques on a streaming architecture of the Cell/BE processor. The efficiency of FDTD computations has been estimated for different parameters of the assumed loop tiling algorithm and the rotating buffers mechanism.