Parallel Gauss-Jordan Elimination on Two-Dimensional Constant Bandwidth Storage

Abstract

As commonly used in two-dimensional magnetotelluric (MT) inversion, Occam inversion method, though featured by its stable convergence and independence on initial model, is relatively poor in time performance, with Gauss-Jordan elimination as one of the biggest time-consuming parts. For symmetrical banded coefficient matrix on two-dimensional constant bandwidth storage, the application of sequential dispatch strategy to the workload on lines of work triangle in parallel algorithm will lead to load imbalance. Therefore, based on paired-dispatch, a kind of strategy is presented to solve this problem, after which the algorithm's effect on shared memory parallel system is verified and research is carried on to focus on the optimization methods of algorithm performance. Experimental results as compared with serial algorithm show that these methods have effectively improved the performance of algorithm, which contributes to a speedup of 3.72 of parallel Gauss-Jordan algorithm as a whole, and that MT Occam inversion algorithm based on this parallel algorithm also demonstrates good speedup performance.