Parallelization in quantum chemistry: The MNDO code

Abstract

This paper describes the techniques and methodologies employed during parallelization of the semi-empirical quantum-mechanical (SEQM) code MNDO, with the specific requirement that the program run efficiently on a range of distributed memory parallel platforms. We present two approaches which reflect the different ways in which this type of code are used. The central problem for one of these concerns the parallelization of a series of matrix operations of the type common to many computational applications (matrix-matrix multiplication and matrix diagonalisation). In particular, we provide one possible parallel solution which tackles the core double matrix multiplication. We discuss the preliminary results of our prototype port and indicate to what extent the parallel code can fulfill the requirements of industrial users

GPU Linear Algebra Libraries and GPGPU Programming for Accelerating MOPAC Semiempirical Quantum Chemistry Calculations

“…In contrast to the strategy adopted in Lin’s thesis, however, and according to the approach adopted in the MNDO program, − , we have performed the two multiplications shown in eq , by taking advantage of the improved performances of linear algebra libraries developed for GPUs and equivalent libraries for multithreaded CPUs. So, we have decided to use matrix multiplication subroutines for both architectures (cublasDGEMM and DGEMM) in this step (see eq ).…”

“…Moreover, a distributed-memory parallelization of MNDO code was also performed using message-passing, both at the coarse-grained and fine-grained level. − …”