Approaches to large‐scale parallel self‐consistent field calculations

The availability of massively parallel computers with high computation rates but limited memory and input/output bandwidth prompts the reevaluation of appropriate solution schemes for the self‐consistent field (SCF) equations. Several algorithms are considered which exhibit between linear and quadratic convergence using various approximations to the orbital Hessian. A prototype is developed to understand the computational expense of each approach. The optimal choice is found to be a conjugate–gradient method preconditioned with a level‐shifted approximation to the orbital Hessian. This is a compromise between efficiency, stability, and low memory usage. Sample benchmarks on two parallel supercomputers are also reported. © 1995 John Wiley & Sons, Inc.