Agent migration in HPC systems using FLAME

Abstract

In HPC agent based applications, the complex interaction rules would likely cause workload imbalances that negatively affect the simulation time. In addition, the imbalance problem may vary during the application execution in accordance to the behavior of the agents. Consequently, a solution to this problem should be able to dynamically balance the load. A dynamic load balancing scheme could be based on migrating agents between processing units. In this paper, we propose a modification of the agent-based simulation framework FLAME that provides the automatic generation of the routines needed to dynamically migrate agents among different computational units. However, most agent-based simulation frameworks do not include routines for migrating agents. Moreover, we demonstrate their use in a simple load balancing scheme on a specific application. © 2014 Springer-Verlag Berlin Heidelberg.