Linear-scaling DFT-pseudopotential calculations on parallel computers

A parallel implementation of linear-scaling first-principles total-energy calculations is presented. The theoretical basis is density functional theory (DFT) and the pseudopotential approximation. The linear-scaling method is the one due to the present authors, but the parallelisation techniques are also relevant to other linear-scaling DFT methods. The theoretical and computational framework of the linear-scaling method is summarised, in order to identify the main classes of computer operation required. For each class of operation, the issues involved in distributing tasks and data between processors are discussed and a solution is proposed. Practical tests of the proposed implementation on a Gray T3D are presented, and it is shown that the scaling with respect to both the number of atoms and the number of processors is excellent for systems containing up to over 6000 atoms.