Threshold pivoting for dense LU factorization on distributed memory multiprocessors

The cost of pivoting in LU factorization has become non-negligible on MIMD computers due to the acceleration of floating point arithmetic and comparatively slow communications. This study addresses the LU factorization of matrices stored by row. The importance of efficiently broadcasting pivot data is stressed. Multicasting is found preferable in this respect to broadcasting along minimum spanning trees. Threshold pivoting is shown to effectively reduce the number of messages while preserving a good load balance.