Adaptive Message Clustering for Distributed Agent-Based Systems

Many agent-based simulation kernels rely on message passing in their core implementation. As the number of agents in a simulation increases or as the complexity of their communication expands the number of messages can increase exponentially. This is troublesome because the message content itself may be quite small, while the overhead, including message headers can dominate bandwidth and processing time. In these cases message passing becomes a bottleneck to scalability. The overhead of message exchange may saturate the network and degrade performance of the simulation. One approach to this challenge that has been investigated in related networking and simulation research centers is combining or "piggy-backing" multiple small messages together with a consolidated header. In many applications performance improves as larger, but fewer messages are sent. However, the pattern of message passing is different in the case of agent- based simulation (ABS), and this approach has not yet been explored for ABS systems. In this work we provide an overview of the design and implementation of a message piggy- backing approach for ABS systems using the SASSY platform. SASSY is a hybrid, large-scale distributed ABS system that provides an agent-based API on top of a PDES kernel. We provide a comparative performance evaluation for implementations in SASSY with a combined RMI and shared memory message passing approach. We also show performance of our new adaptive message clustering mechanism that clusters messages when advantageous and avoids clustering when the overhead of clustering dominates.