Revisiting Minimization Strategies for Solving Eigenvalue Problems

Osni Marques, Doru Thom Popovici, Mauro Del Ben and Andrew Canning (Lawrence Berkeley National Laboratory, USA) Invited Talk

This presentation will discuss ongoing work on unconstrained minimization schemes for the solution of eigenvalue problems. These schemes employ a preconditioned conjugate gradient approach that avoids an explicit reorthogonalization of the trial eigenvectors, in contrast to typical iterative eigenvalue solvers, therefore reducing communications and becoming a potential alternative for the solution of large problems on massively parallel computers. We target problems related to electronic structure calculations. We show results for a set of benchmark systems with an implementation of the unconstrained minimization in first-principles materials and chemistry codes that perform electronic structure calculations based on a density functional theory (DFT) approximation to the solution of the many-body Schrödinger equation. The results show that the unconstrained formulation, together with an appropriate preconditioner, offers good convergence properties and scales well on a large number of cores. We also show results for an implementation in the commonly used plane wave formulation of DFT.