Parallel Optimization of Stencil Computation Base on Sunway TaihuLight

Abstract

Stencil computation is a kind of memory intensive computing core widely used in image and video processing, large-scale science and engineering calculation, which has been taken as the object of performance optimization by many scientific researchers, including parallel acceleration, communication optimization, load balance, while the memory access optimization of stencil computation needs further research [1, 2]. Based on the self-developed supercomputer “Sunlight”, this paper proposes an efficient parallel optimization method for the stencil algorithm. It includes multi-level optimization accelerated by MPI + multi-core, efficient DMA transmission, mutual concealment of communication and calculation, data reuse and other parallel optimization strategies. Experiments on Sunway platform show that the strategy proposed in this paper greatly improves the efficiency of the algorithm and the performance of the algorithm.