Numerical Performance of Matrix Inversion with Block Pivoting.

摘要

An experiment with matrix inversion using block pivots is presented. Large scale matrix computations can often be performed more efficiently by use of partitioning. Such matrix manipulation lends itself to paged or cache memory systems since computation is staged to be completely performed in local blocks of controllable size. On other systems retrieval overhead can be balanced with computation for 'in-memory/out-of-memory' applications. Parallelism in such schema leads to efficient utilization of some multiple processor environments. Timing results indicate, however, that choice of block size should not necessarily be dictated by hardware page size for most efficient operation and that classical methods of estimating computation times are not always adequate.