Locality optimizations for Jacobi iteration on distributed parallel systems

In this paper, we propose an inter-nest cache reuse optimization method for Jacobi codes. This method is easy to apply, but effective in that it enhances cache locality of the Jacobi codes while preserving their coarse grain parallelism. We compare our method to two previous locality enhancement techniques that can be used for Jacobi codes: time skewing and new tiling. We quantitatively calculate the main contributing factors to the runtime of different Jacobi codes. We also perform experiments on a PC cluster to verify our analysis. The results show that our method performs poorer than time skewing and new tiling for uniprocessor, but performs better for distributed parallel system.