Extreme-scale Dynamic Exploration of a Distributed Agent-based Model with the EMEWS Framework

Area of Science:

• Computational science
• Complex systems modeling
• Scientific workflow management

Background:

• Agent-based models (ABMs) are vital for studying complex systems across various scientific domains.
• The computational demands of large-scale ABM simulations hinder their development and validation.
• Efficiently exploring the parameter space of ABMs is essential for understanding their emergent behaviors.

Purpose of the Study:

• To introduce the Extreme-scale Model Exploration with Swift (EMEWS) framework for efficient ABM simulation ensembles.
• To demonstrate EMEWS's capability in integrating model exploration algorithms with ABMs.
• To address the computational challenges in large-scale, distributed ABM simulations.

Main Methods:

• Developed the EMEWS framework, integrating stateful tasks with many-task computing (MTC).
• Enabled composition and execution of large simulation ensembles and black-box scientific applications.
• Integrated model exploration (ME) algorithms using R and Python libraries.
• Applied EMEWS to a distributed Message Passing Interface (MPI) agent-based infectious disease model.

Main Results:

• EMEWS efficiently composes and executes large ensembles of simulations.
• The framework scales to millions of tasks, maintaining state and locality.
• Successfully integrated an active learning ME algorithm for dynamic parameter space characterization.
• Demonstrated efficient exploration of a complex infectious disease ABM's parameter space.

Conclusions:

• EMEWS provides a scalable and efficient solution for high-performance ABM workflows.
• The framework facilitates multi-language problem-solving for complex scientific applications.
• EMEWS significantly enhances the ability to explore and understand complex agent-based models.