Abstract: Most biological phenomena have microscopic foundations yet span macroscopic length- and time-scales, necessitating multiscale computational models. Efficient simulation of these complex multiscale models on modern heterogeneous architectures poses significant challenges in scheduling and co-managing resources such as computational power, communication bottlenecks, and filesystem bandwidth. To address these challenges, we present a novel massively parallel Multiscale Machine-Learned Modeling Infrastructure (MuMMI), which combines a large length- and time-scale macro model with a high-fidelity molecular dynamics (MD) micro model using machine learning. We describe our infrastructure which is designed for high scalability, efficiency, robustness, portability, and fault tolerance on heterogeneous resources. We demonstrate MuMMI conducting the largest-of-its-kind simulation to investigate the dynamics of KRAS proteins in cancer initiation. Concurrently running up to 36,000 jobs on 16,000 GPUs and 176,000 CPU cores, we executed 120,000 MD simulations surpassing an aggregate simulation time of 200 milliseconds, orders of magnitude greater than comparable studies.
Back to Technical Papers Archive Listing