Scaleable parallel model development and performance analysis for MIMD molecular dynamics simulations

Abstract

A scaleable parallel model developed specifically for molecular dynamics simulations is described. The description encompasses both hardware and software. It is mapped onto a transputer array, programmed, and used to verify concepts and gather experimental data. Simulations are run with sample sizes ranging from 1024 to 42,546 particles on arrays from 4 to 128 processors. Performance data is generated by fixing the sample size, changing the number of node processors used to run the simulation, and measuring execution speeds. The parallel implementation is dissected and used to develop a detailed mathematical performance model. The methodology for developing the mathematical model is described. Performance equations are derived for all computation and communication routines in the code, and combined to form a modular system performance estimator. Accuracy of the mathematical model is verified by comparing the measured baseline times against a set of performance estimation curves generated by the mathematical formulation. The computation and communication constraints to faster execution are identified and discussed, including the issues of load balancing, data partitioning strategy, and upper and lower limits on the number of processors that should be used for a simulation. Algorithm and architectural improvements are described, and their impact on performance predicted using the analytical tools developed.