Parallel matrix inversion on a subcube-grid (Q1208521)

The authors propose a new medium-gain parallel algorithm for computing a matrix inverse on a hypercube multiprocessor. It is shown that the subcube-grid topology can be used to implement the Gauss-Jordan inversion with column interchanges. How the submatrix partitions can be exploited in the design of a variety of matrix algorithms for hypercubes is described. Numerical experiments were performed on a 128-processor Intel iPSC/860 hypercube multiprocessor. Significant improvements in speedup and efficiency are reported.