Computational Chemistry on Workstation Clusters: Parallel Programming for Molecular Dynamics and Neural Networks

Abstract

The use of RISC workstation clusters to obtain supercomputer level performance with the use of coursegrained message passing parallelism is described. Two types of computational chemistry applications are discussed: molecular dynamics and neural networks.

Several molecular dynamics programs have been parallelized and have shown very good improvements in computational speed. For a 14000 atom Myoglobin dynamics calculation, we have obtained a 6 times acceleration with the program CHARMM on a eight node HP735 cluster using FDDI (Fiber Data Distributed Interface). A feed forward backpropagation neural network was developed that uses a parallelized ‘molecular dynamics-like’ algorithm. It shows good convergence behavior and very efficient parallel speedup when applied to the problem of protein secondary structure prediction.

Key Words:

• Molecular dynamics
• CHARMM
• neural networks
• parallel