Scaling an Agent-Based Epidemic Simulation on State-Level Digital Twins
Joy Kitson, University of Maryland, College Park
Preventing the spread of infectious diseases requires implementing interventions at different levels of government and evaluating the potential impact and efficacy of those preemptive measures. Agent-based modeling can be used for detailed studies of epidemic diffusion and the impact of potential interventions, through their ability to capture the complex emergent behavior which results from simple rules for agent behavior. We present refinements to Loimos, a highly parallel simulator of infectious diseases spread in realistic, high resolution “digital twin” populations. We present several performance optimizations implemented in Loimos, namely a geographically based static load balancing scheme, and an efficient scheme for computing exposures between susceptible and infectious agents. We also provide detailed performance analysis demonstrating the benefits of these optimizations. We additional validate Loimos against an established simulator, EpiHiper, on full state-level digital twin populations. We demonstrate that our implementation of Loimos is able to scale to large core counts on Perlmutter at NERSC. In particular, Loimos is able to simulate a realistic interaction network for the state of Michigan in an average of 0.17 seconds per simulation day when executed on 4096 cores on Perlmutter, processing around 268 trillion interactions (person-person edges) in about 34 seconds for an average of 7.95 trillion traversed edges per second (TEPS).