Efficient parallel simulation of spatially-explicit agent-based epidemiological models

Abstract

Agent-based approaches enable simulation driven analysis and discovery of system-level properties using descriptive models of known behaviors of the entities constituting the system. Accordingly, a spatially-explicit agent-based ecological modeling, parallel simulation, and analysis environment called SEARUMS has been developed. However, the conservatively synchronized parallel simulation infrastructure of SEARUMS did not scale effectively. Furthermore, the initial multithreaded shared-memory design prevented utilization of resources on multiple compute nodes of a distributed memory cluster.Consequently, the simulation infrastructure of SEARUMS was redesigned to operate as a Time Warp synchronized parallel and distributed discrete event simulation (PDES) on modern distributed-memory supercomputing platforms. The new PDES environment is called SEARUMS++. The spatially-explicit nature of the models posed several challenges in achieving scalable and efficient PDES, necessitating new approaches in SEARUMS++ for: ① modeling spatial interactions and initial partitioning of agents, ② logical migration of an agent during simulation using proxy agents to reflect migratory characteristics, and ③ ghosting of agents using multiple proxy agents to handle boundary cases that occur during logical migration of agents.This article presents our optimization efforts involving new methods to address aforementioned challenges. The design of SEARUMS++ and experimental evaluation of various alternatives that were explored to achieve scalable and efficient PDES are also discussed. Our experiments indicate that SEARUMS++ provides 200% performance improvement and maintains scalability to a larger number of processors, thus enabling efficient parallel simulation of spatially-explicit agent-based epidemiological models.