Global Sensitivity Analysis of a Large Agent-Based Model of Spatial Opinion Exchange: A Heterogeneous Multi-GPU Acceleration Approach

Sensitivity analysis is an important step in agent-based modeling of complex adaptive spatial systems to evaluate the contribution of influential variables to model response. Sensitivity analysis of agent-based models is computationally demanding, however, and this analysis tends to be intractable for large agent-based modeling. This computational challenge greatly limits our ability to investigate complex spatial dynamics using large agent-based models. The objective of this study is to gain insight into this computational issue by focusing on the sensitivity analysis of large agent-based modeling of spatial opinion exchange, accelerated using multiple graphics processing units (GPUs). We present a heterogeneous parallel computing approach based on nested parallelism for the global sensitivity analysis of the model. The agent-based opinion model is parallelized using many-core GPUs for the simulation of a large number of spatially aware and interacting agents. These agents exchange opinions for developing consensus on topics through processes of spatial neighborhood search and opinion update. Global sensitivity analysis of the opinion model is conducted using a variance-based approach, requiring numerous model runs for Monte Carlo integration. Intermodel parallelization is introduced to enable Monte Carlo runs of sensitivity analysis. We conduct global sensitivity analysis on a multi-GPU cluster. Experimental results indicate GPU-accelerated general-purpose computation provides an efficacious and feasible solution for the sensitivity analysis of large agent-based models. The heterogeneous parallel computing approach provides valuable insight into large-scale spatiotemporal problem solving by leveraging cyberinfrastructure-enabled computational capabilities.