Parallel implementation of a mesh-based density functional electronic structure code

We describe the implementation of the mesh-based first-principles density functional code DMol on nCUBE 2 parallel computers. The numerical mesh nature of DMol makes it naturally suited for a massively parallel computational environment. Our parallelization strategy consists of a domain decomposition of mesh points. This evenly distributes mesh points to all available processors and leads to a substantial computational speedup with limited communication overhead and good node balancing. To achieve better performance and circumvent memory storage limitations, the torus wrap method is used to distribute both the Hamiltonian and overlap matrices, and a parallel matrix diagonalization routine is employed to calculate eigenvalues and eigenvectors. Benchmark calculations on a 128-node nCUBE 2 are presented. © 1995 by John Wiley & Sons, Inc.