PARALLEL LANCZOS ALGORITHMS FOR NOVEL ARCHITECTURE COMPUTERS - A COMPARATIVE-EVALUATION

The current availability of a wide range of novel architecture compute rs has stimulated a quest for efficient algorithms for the solution of the commonly occurring numerical mathematical problems which can expl oit the facilities provided by these machines. When only single-proces sor machines were available the most efficient algorithm for one compu ter was, in general, the most efficient for another since all shared t he same basic architecture. In this paper the suitability of a modific ation of the Lanczos algorithm for the computation of partial eigensol utions of real symmetric matrices on a variety of parellel computers i s assesses. The modification incorporates an extension to the classica l Lanczos algorithm which permits an accurate determination of converg ence and which may be efficiently implemented on vector, SIMD and MIMD computers are assesed and are critically compared with those of the u nextended method. Some conclusions are drawn about the effectiveness o f the various architectures for the execution of the algorithm.