Exploiting model uncertainty to improve the scalability of long-time simulations using Parallel Trajectory Splicing

Abstract

We consider parallel trajectory splicing (ParSplice), a specialized molecular dynamics method that extends simulation timescales through a parallel-in-time strategy, enabling parallel speedups proportional to the number of worker-processes deployed. In practice, the ability for ParSplice to scale significantly improves when it is possible to predict the future evolution of the atomistic trajectory. We propose improved predictive statistical models that are built on-the-fly in order to maximize computational efficiency. By imposing physical constraints and explicitly considering uncertainties in model estimation we show a significant improvement in the scalability of ParSplice, and hence a corresponding increase in the timescales that can be reached by direct simulation.

Resource allocation for task-level speculative scientific applications: a proof of concept using Parallel Trajectory Splicing

“…In order to assess the potential performance gains accessible with this new approach without the considerable time investment required to rewrite the ParSplice production code, we instead chose to make use of a simulator; a strategy which has proved beneficial 48,49 in developing Accelerated Molecular Dynamics 50 methods, to which ParSplice belongs. The ParSplice Simulator (ParSpliceSIM) was designed to directly mirror the logic of the actual ParSplice code with the exception that dynamics are generated from a user-specified Markov Chain that can be used to statistically sample segment endpoints, rather than using computationally expensive MD calculation 51 . ParSpliceSIM is therefore computationally light, simple, and runs in serial.…”