BLOCK ALGORITHMS FOR TRIANGULAR LINEAR SYSTEMS

摘要

Parallel blocked versions of short-cut algorithms are presented for solving triangular systems with, possibly, multiple right hand sides. They mainly use BLAS3-routines for their implementation and require the data to be distributed through a square-block torus-wrap mapping. Numerical experiments on an INTEL Paragon XP/S show an efficiency of 50-75% for a wide range of block sizes and mesh forms. [References: 7]