Modelling Distributed Multiscale Simulation Performance: An Application to Nanocomposites

Abstract

Clay polymer nanocomposites are a new range of particle filled composite material which interact over many different length scales, ranging from the quantum mechanical level to macroscopic. Multiscale simulation is therefore an important technique to understand and, ultimately, predict the properties of the composites from their individual components. We describe two multiscale simulation scenarios in which we couple simulations running on different levels of scale: in the loosely-coupled scheme we have a unidirectional coupling of one level to the next level, while in the tightly-coupled scheme we have simulations creating multiple inputs and parameters for simulations at different levels, running concurrently. We present a performance model that predicts the multiscale efficiency of our multiscale application. Here the multiscale efficiency constitutes the fraction of runtime spent on executing the simulation codes, and not on operations facilitating the coupling between the simulations. We find that the efficiency is high (greater than 90 %) until the number of sub-simulations exceeds a critical number (> 10 in our examples).