Presentation

DescriptionParallel trajectory splicing (ParSplice) is an attempt to solve the enduring challenge of simulating the evolution of materials over long time scales for complex atomistic systems. A novel version of ParSplice is introduced with features that could be useful in its scaling to exascale architectures. A two-pronged approach is used. First, latent parallelism is exploited by extending support to heterogeneous architectures, including GPUs and KNLs. Second, the efficiency of the Kinetic Monte Carlo predictor is improved, allowing enhanced parallel speculative execution. The key idea in these predictor modifications is to include statistics from higher temperature simulations. The issue of inherent uncertainty in the prediction model was addressed in order to improve the performance, as the current predictor only takes into account the previous observations to formulate the problem. The predictor was also improved by using a hybrid approach of message-passing + multi-threading. (LA-UR-17-26181)

Navigation