Agent-Based Simulation Performance
Rigorous bottleneck characterization of large-scale ABM workloads using DAMOV and hardware profiling
Agent-based modeling (ABM) is a bottom-up method for studying complex systems. We model thousands, millions, even billions of individual agents—each one simple, with its own properties and behavior, each interacting with its neighbors. From these local interactions, global complexity emerges.
Presented as a poster at ACACES 2026 (HiPEAC Summer School, Fiuggi, Italy) (Sokoli & Yaglikci, 2026).
Why agent-based simulations matter
The ambition is growing: we do not want to simulate just thousands or millions of agents, but billions—the world we live in, at the speed it actually operates, at large scale and in real time.
The EU recently committed to phasing out animal testing across 15 legislative domains, yet its own funding falls short of that ambition—pointing directly to agent-based simulations as an alternative.
Versatile applications
Agent-based simulations are general. The approach applies everywhere:
Research goal
To reach real-time, billion-agent simulation, we must first understand where performance breaks down. We started simulating real-world use cases to benchmark them, determine whether bottlenecks lie in software (data structures, architecture) or demand hardware acceleration, and guide future co-design.
To our knowledge, this is the first work that rigorously characterizes where agent-based simulations bottleneck at scale—a central contribution of this project.
Methodology
We profiled BioDynaMo, the state-of-the-art agent-based simulation framework developed at CERN, using an epidemiology use case:
- Scale: 10M to 1B agents on Google Cloud C4 VMs (16 cores each)
- Characterization: DAMOV methodology for workload analysis
- Profiling:
perfat 1 Hz to capture hardware metrics in real time during simulation
Related frameworks & code
- BioDynaMo — state-of-the-art ABM framework (CERN)
- ACACES 2026 profiling study — experiment scripts and analysis
- CARTopiaX — scalable simulation framework
This project bridges agent-based modeling research with computer architecture and performance engineering.