2026-08-13 –, Room 3
Effective partitioning is important for the scalability of agent-based modeling (ABM) on HPC systems, but existing methods do not meet the specific challenges of complex ABMs. These usually involve heterogeneous agent types with phase-based execution, dynamic population changes, and moving agents. This talk presents ongoing research developing specialized partitioning algorithms for distributed ABMs, including a benchmark framework and the development of specialized Julia packages.
Agent-based models (ABMs) simulate complex systems through the interaction of individual agents. Scaling these simulations to millions of agents requires advanced parallel computing methods due to the highly flexible structure of ABMs and their interactions.
Compared to static computational problems such as mesh-based finite element simulations, where the computational load and communication patterns remain the same throughout execution, ABMs present a special challenge for load balancing due to their dynamic nature and the complex interactions of the agents:
ABMs can involve heterogeneous agent types that act at different phases of the simulation. Even when the overall partitioning appears good, individual agent types may be severely imbalanced across processes, leading to idle computational resources during individual phases of the execution.
Communication between agents can also take place between agents who are spatially far apart.
Dynamic population changes alter the load distribution through addition and removal of agents.
Agent movement can create hotspots and ghost zones that were not present in the initial partitioning,
Since standard partitioning algorithms do not take into account the dynamics and complex structure that are inherent in complex ABM scenarios, their results are not always satisfactory.
Talk structure:
Brief Vahana.jl introduction: A brief overview of Vahana.jl, an HPC ABM framework based on graph dynamical systems, with a focus on the new features implemented since version 1.0, as presented at JuliaCon 2023.
ABM-specific partitioning challenges: Why general-purpose partitioning approaches fail for ABMs with heterogeneous agent types, dynamic populations, and complex agent behaviors/movement patterns
Synthetic benchmarking model: Presentation of a purpose-built synthetic ABM that systematically captures the partitioning challenges specific to ABMs. This model provides a flexible environment for evaluating and comparing different load balancing strategies and different HPC ABM frameworks in general.
Outlook: Overview of our ongoing early-stage work on developing novel partitioning algorithms specifically tailored to ABM characteristics. These algorithms will be released both as integrated Vahana.jl features and as independent Julia packages.
Steffen Fürst, having studied mathematics with a focus on economics and social science, has spent the past 15 years working on various agent-based models. In the most recent 5 years, his focus has been particularly on high-performance computing environments.