Development of Computational Chemistry Methods for Parallel Processors


Computational chemistry aims to model the behaviour of atoms and molecules using computers. There are many parts to computational chemistry, but our specific interest is in electronic structure theory. These methods seek to solve, with minimal approximation, the fundamental equations of quantum mechanics for a system of electrons and nuclei. The methods used are complex having been developed over many decades. In part because of this historical legacy many of the computer codes that are currently in use are poorly adapted to massively parallel processors. This project aims to address this issue.


Principal Investigator

Alistair Rendell
Computer Science
Faculty of Science
ANU

Project

x32

Co-Investigator

Stephen Titmuss
Computational Molecular Biology
JCSMR
ANU

RFCD Codes

250601


Significant Achievements, Anticipated Outcomes and Future Work

Our initial work has been exploratory, assessing the parallel capability and performance of some current electronic structure codes. To this end we have considered the GAMESS (US Version), Gaussian 98, and NWChem codes. Ultimately our aim is to perform routinely simple (i.e. Hartree-Fock and density functional) energy and gradient calculations on systems containing several hundred atoms, while making more complicated perturbation theory and coupled cluster calculations routine on systems with fifty or more atoms. To this end our future work will be focused on the Gaussian code, working closely with the developers to improve the parallel performance of the various linear scaling algorithms.

 

Computational Techniques Used

Essentially, electronic structure codes use spectral methods to solve a system of differential equations. The behaviour of each electron is described by a series of basis functions that are centered on each nuclear center. The formation and manipulation of integrals involving these functions and a variety of operators form the core of all electronic structure methods. The specific codes used here are GAMESS (US Version), Gaussian 98 and NWChem. All three codes use a multiple instruction multiple data (MIMD) approach to parallelisation, but differ in the method of parallelisation (e.g. message passing, Linda, and Global Arrays) and degree to which different quantities have been distributed across the memory associated with each process. The APAC National Facility SC provides us with access to significantly more processors than would be available on any other resource.

 

Publications, Awards and External Funding

This work has been supported, in part, by funds from the APAC / ANU computational chemistry expertise program.