PARALLELIZATION OF PSEUDOSPECTRAL ALGORITHMS FOR ELECTRONIC STRUCTURE CALCULATIONS

ELECTRONIC, STRUCTURE PROGRAMS ARE BECOMING AN INCREASINGLY POWERFUL TOOL FOR SOLVING CHEMICAL AND BIOLOGICAL PROBLEMS WITHOUT COSTLY EXPERIMENTS, BUT IN ORDER TO USE CURRENT AND FUTURE COMPUTER HARDWARE TO MAXIMAL ADVANTAGE, THEY MUST BE PARALLELIZED EFFICIENTLY. THE PSEUDOSPECTRAL (PS) APPROACH UTILIZED IN THE PREVIOUSLY DEVELOPED PS-GVB SOFTWARE PACKAGE IS GRID-BASED, PERMITTING A WELL-BALANCED DISTRIBUTION OF TASKS FOR EACH REGION OF THE GRID, WHICH WILL LEAD TO EFFECTIVE PARALLELIZATION. THE RESULTING INCREASE IN SPEED WILL MAKE PS-GVB EVEN MORE USEFUL FOR AB INITIO PREDICTIONS OF ELECTRONIC STRUCTURE OF LARGE MOLECULES. PHASE I WILL PARALLELIZE THE MOST COMPUTATIONALLY INTENSIVE PORTIONS OF PS-GVB: THE ASSEMBLY OF THE LEAST-SQUARES OPERATOR, THE FOCK MATRIX, THE GVB HAMILTONIANS, AND THE ANCZOS-TYPE EIGENSOLVER USED FOR UPDATING THE WAVEFUNCTION DURING THE SELF-CONSISTENT FIELD PROCESS. THE RESULTING PROGRAM MODIFICATIONS WILL BE TESTED USING PARALLEL VIRTUAL MACHINE (PVM) SOFTWARE ON WORKSTATION CLUSTERS. THE ULTIMATE PHASE III GOAL OF THE PARALLELIZATION EFFORT IS A MULTIFUNCTIONAL SOFTWARE PACKAGE THAT DISTRIBUTES ALL TASKS EFFECTIVELY WHEN RUN ON A WIDE VARIETY OF MASSIVELY PARALLEL PROCESSORS AND CLUSTERS.