Efficient Agent-Based Simulation Framework for Multi-Node Supercomputers

In recent years the importance of a large-scale Agent-based simulation (ABS) that can handle large complex systems is increasing. We developed a large-scale ABS framework on BlueGene, which is a multi-node supercomputer. The ABS processes the agents' communications. When the number of transmissions among the agents is large, the transmission costs seriously affect the performance of the simulation. It is possible to reduce the amount of transmission among the nodes by clustering the agents which communicate heavily with each other. Assuming that an agent is a graph node, and that a data transmission between agents is a graph edge, this problem can be formulated as a maximum-flow and minimum-cut problem. In this paper we present an efficient algorithm to find an approximate solution. Our algorithm is reliable, simple, and needs little computation. We demonstrate its beneficial effects with some experiments