Performance Evaluation of A Seismic Data Analysis Kernel on The KSR Multiprocessors

The paper investigates the effective performance attainable for a specific class of application programs on shared memory supercomputers. Specifically, we are to investigate how seismic data analysis applications behave on the Kendall Square Research Inc.'s KSR multiprocessors. The computational kernel of seismic computation algorithms is parallelized and its performance is analyzed. Three approaches for parallelizing the g5 kernel are analyzed: column-based, row-based, and grid-based parallelizations. All three approaches result in well balanced decompositions, but differ significantly in data locality. In general, the column-based approach has the best data locality, while the small grid-based approach has the worst. These results clearly indicate that data locality is one of the critical factors for attaining high performance for the g5 kernel. The best parallelized g5 kernel code achieves about 44% of both the KSR-1 and KSR-2 machines' peak computational performance