Accelerating parallel agent-based epidemiological simulations

Abstract

Background: Simulations play a central role in epidemiological analysis and design of prophylactic measures. Spatially explicit, agent-based models provide temporo-geospatial information that cannot be obtained from traditional equation-based and individual-based epidemic models. Since, simulation of large agent-based models is time consuming, optimistically synchronized parallel simulation holds considerable promise to significantly decrease simulation execution times. Problem: Realizing efficient and scalable optimistic parallel simulations on modern distributed memory supercomputers is a challenge due to the spatially-explicit nature of agent-based models. Specifically, conceptual movement of agents results in large number of inter-process messages which significantly increase synchronization overheads and degrades overall performance. Proposed solution: To reduce inter-process messages, this paper proposes and experimentally evaluates two approaches involving single and multiple active-proxy agents. The Single Active Proxy (SAP) approach essentially accomplishes logical process migration (without any support from underlying simulation kernel) reflecting conceptual movement of the agents. The Multiple Active Proxy (MAP) approach improves upon SAP by utilizing multiple agents at boundaries between processes to further reduce inter-process messages thereby improving scalability and performance. The experiments conducted using a range of models indicate that SAP provides 200% improvement over the base case and MAP provides 15% to 25% improvement over SAP depending on the model.