Linear algebra software on a vector computer (Q1118969)

This paper reports on the performance and accuracy of common library subroutines for a range of numerical linear algebra algorithms implemented on a vector computer. Many large scale computations require the solution of subproblems which can be formulated in terms of vector operations, and vector computers (CRAY, CYBER, FUJITSU, IBM) are designed to implement such operations efficiently. To attain this aim user software must be written in vector mode. The main software libraries [cf. \textit{Th. J. Aird}, in W. R. Cowell (ed), Sources and Development of Mathematical Software, 264-301, Prentice-Hall, NJ (1984)] which are considered are SCILIB, which is CRAY specific and based on BLAS, LINPACK and EISPACK, and the software-house libraries ISML and NAG. Limited tests with ZAMLIB, from the authors' institute, are also included. Most of the tests were performed on a CRAY X-MP/22, with special attention being paid to large-order test matrices. Special attention was paid to accuracy and efficiency as measured by CPU-time. The Nag library (Mk 11) made better use of the vector processing facilities of the CRAY than did IMSL (Ed 9.2). However first tests with IMSL (Ed 10) indicated that this library was taking good advantage of the architecture of vector computers.