RESEARCH AND DEVELOPMENT OF A PORTABLE PARALLEL LIBRARY OF BASIC LINEAR ALGEBRA SUBPROGRAMS

WE PROPOSE TO DEVELOP A PORTABLE, PARALLEL LIBRARY OF HIGH-LEVEL BASIC LINEAR ALGEBRAIC SUBPROGRAMS FOR USE BY THE NATIONAL SCIENTIFIC COMMUNITY. SUCH A PORTABLE PARLLEL LIBRARY OF ROUTINES WILL AID GREATLY IN THE DISSEMINATION AND WIDESPREAD USE OF ADVANCED LINEAR ALGEBRA ALGORITHMS ON A VARIETY OF PARALLEL COMPUTERS AND COMPUTER NETWORKS. SUCH ALGORITHMS OFTEN FORM THE COMPUTER-INTENSIVE KERNELS OF LARGE COMPUTATIONS. THE BENEFITS OF OUR EFFORT INCLUDE: 1) OBVIATING THE NEED FOR SEPARATE DEVELOPMENT EFFORTS OF PARALLEL BLAS ROUTINES FOR EACH PARALLEL COMPUTER VENDOR, 2) HARNESSING VAST AMOUNTS OF UNTAPPED PARALLEL NETWORK COMPUTING CYCLES (BY USING NETWORK NODES WHEN THEY ARE NOT BEING USED), AND 3) DEVELOPING PARALLEL PROGRAMMING METHODOLOGY TO USE AS THE BASIS FOR OTHER PORTABLE PARALLEL SUBROUTINE LIBRARIES. THIS EFFORT WILL COMPLEMENT THE EFFORTS OF OTHER RESEARCH GROUPS TOWARD LINEAR ALGEBRA ALGORITHM DEVELOPMENT. THE EFFORT WILL INCLUDE A MACHINE-INDEPENDENT PARALLEL IMPLEMENTATION OF THE LEVEL 3 BLAS, AS WELL AS ALGORITHMIC AND PROGRAM MODIFICATIONS OF THE LAPACK PROGRAMS TO ALLOW IT TO OPERATE IN PARALLEL IN A MACHINE-INDEPENDENT MANNER. THE DEVELOPMENT EFFORT WILL BE BASED ON THE LINDA PARALLEL COMPUTING CONSTRUCT, WHICH AUGMENTS THE C PROGRAMMING LANGUAGE TO PROVIDE A UNIFORM, MACHINEINDEPENDENT PARALLEL PROGRAMMING LANGUAGE.