Design and implementation strategy of a parallel agent-based Schelling model

•The Schelling model is parallelized for its application in the real-world.•The parallel model will be an example of how HPC can contribute to geography.•A performance experiment was conducted with different spatial decomposition methods.•Uneven distribution of agents leads to performance differences.•Two-stage all-to-all communication between subdomains was used to move agents.

This paper illustrates the detailed design and implementation strategy of a high-performance, parallel agent-based Schelling model. There are four key issues to implement a parallel agent-based Schelling model, i.e., the ability to handle (1) the complexity of real data processing; (2) spatial data decomposition; (3) the preference and tolerance rates for multiple ethnic groups; and (4) the communication between agents. A new parallel algorithm that addresses these issues while retaining the fundamental concepts of the original sequential Schelling model was built. The parallel Schelling model was applied to the County of San Diego’s one million household agents and one million housing unit agents using land use data and 2000 Decennial Census data. Experiments using varying numbers of CPU cores demonstrate that the computational performance of the parallel Schelling model positively correlates with increased numbers of CPU cores. This paper focuses on the implementation of a parallel agent-based Schelling model. The Schelling model is parallelized to illustrate how high-performance computing can elicit better understanding of real-world geospatial phenomena. By leveraging distributed computing resources in geospatial cyberinfrastructure, the parallel Schelling model has the potential to become a crucial component in spatial decision support systems and to help policy makers understand the key trends in urban population, urban growth, and residential segregation.