A Framework for Megascale Agent Based Model Simulations on Graphics Processing Units

Abstract

Agent-based modeling is a technique for modeling dynamic systems from the bottom up. Individual elements of the system are represented computationally as agents. The system-level behaviors emerge from the micro-level interactions of the agents. Contemporary state-of-the-art agent-based modeling toolkits are essentially discrete-event simulators designed to execute serially on the Central Processing Unit (CPU). They simulate Agent-Based Models (ABMs) by executing agent actions one at a time. In addition to imposing an un-natural execution order, these toolkits have limited scalability. In this article, we investigate data-parallel computer architectures such as Graphics Processing Units (GPUs) to simulate large scale ABMs. We have developed a series of efficient, data parallel algorithms for handling environment updates, various agent interactions, agent death and replication, and gathering statistics. We present three fundamental innovations that provide unprecedented scalability. The first is a novel stochastic memory allocator which enables parallel agent replication in O(1) average time. The second is a technique for resolving precedence constraints for agent actions in parallel. The third is a method that uses specialized graphics hardware, to gather and process statistical measures. These techniques have been implemented on a modern day GPU resulting in a substantial performance increase. We believe that our system is the first ever completely GPU based agent simulation framework. Although GPUs are the focus of our current implementations, our techniques can easily be adapted to other data-parallel architectures. We have benchmarked our framework against contemporary toolkits using two popular ABMs, namely, SugarScape and StupidModel.

Cite this work

No public copy is stored for this record, so only the metadata is cited.

Stored only when both papers are in the corpus — 75 citations exist upstream.