Some applications of pipelining techniques in parallel scientific computing

Abstract

In this thesis, we study the applicability of pipelining techniques to the development of parallel algorithms for scientific computation. General principles for pipelining techniques are discussed and two applications, Gram-Schmidt orthogonalization and chasing algorithms, are considered. For each application, the pipelined parallel implementation is discussed and performance analysis is presented. For Gram-Schmidt orthogonalization, we investigate pipelined parallel algorithms based on various columnwise partitioning schemes. For chasing algorithms, in addition to the pipelining, we apply block-cyclic partitioning, group message-passing techniques to enhance the performance of the pipelined parallel algorithms. The numerical results for the use of these techniques are also presented and compared.