The strategy for optimizing the performance of Gauss-Seidel iterative algorithm based on register reuse

The GS (Gauss-Seidel) iterative algorithm is widely used in large-scale scientific computing. It is indispensable to optimize the algorithm for high performance. The existing iterative space alternate tiling parallel Gauss-Seidel algorithm has improved data locality and reduced cache misses in virtue of iteration space tiling. Basing on the above, in this paper, we have done the further work in two aspects. Firstly, we repartition most of the iterative space into register tiles, and then by unrolling the loops on register tiles, we achieve two kinds of register reuse. Secondly, the method to select repartition parameters is presented, by which we can make sure the best parameters for aimed computer to optimize the performance of Gauss-Seidel iterative algorithm. The experimental results show that the executive time of the algorithm has been greatly reduced by register reuse ideal, also the method to select repartition parameters is testified to be effective and steady. © 2012 American Scientific Publishers. All Rights Reserved.