Optimizing the Memory Access Performance of FASTEST's Sispol Routine
Michael Burger, Christian H Bischof · TUbilio (Technical University of Darmstadt) · 2014
In this article the runtime behavior of the simulation software Fastest is investigated and its performance is optimized. The main performance bottleneck of Fastest is the sipsol subroutine. An analysis shows that the memory accessing behavior is the main reason for the high runtime. As a consequence, a new data structure is developed which on the one hand increases the speed but on the other hand preserves the ability to parallelize the calculations on the data grid executed within the sipsol subroutine. The ratio between calculation and memory reads/writes is improved significantly, and the performance in certain cases doubled. Finally, approaches to efficiently implement a parallelization within the new structure are proposed.