Performance Prediction for Parallel Temperature Accelerated Dynamics

Richard J. Zamora
Seminar

Molecular dynamics (MD) is a widely used computational method in theoretical chemistry, physics, biology, and materials science. Despite its unrivaled ability to predict the dynamical evolution of interacting atoms, the method is often limited to microsecond time scales, leaving out many physical phenomena of interest. The temperature-accelerated dynamics (TAD) method overcomes this limitation by thermally accelerating the state-to-state evolution captured by MD. In an effort to modernize the TAD approach for distributed computing systems, by leveraging speculative execution, we employ a discrete event-based performance prediction framework to explore several variations of speculatively-parallel TAD (SpecTAD). By constructing a proxy application code (SpecTADSim), we provide a means to rapidly explore algorithm variations, without the need for a full-scale implementation. The investigation illuminates a nontrivial relationship between the optimal parameter set and the number of CPU resources. Furthermore, the results suggest that a majority of the available performance gains can be achieved using relatively simple modifications to the traditional TAD algorithm. The talk will emphasize the application simulator code used for this work.