Congestion modeling for reconfigurable inter-processor networks

In this paper, we attempt to model congestion on a reconfigurable multi-processor communication network. This reconfigurable network adapts its topology at given intervals to the properties of the network traffic, which may alter over time. Using our congestion model, one can quickly estimate packet latency for a given set of network parameters. This allows a network designer to do design-space explorations without having to resort to detailed, slow simulations.The model is derived by viewing the network as an interconnected set of queues and servers. For each reconfiguration interval and each network link, we construct inter-arrival and service time distributions. Queuing theory can now give us the average waiting time on a single link. Combining waiting times per link over a network path yields the congestion experienced by each packet.After presenting the model itself, we show simulation results and determine the accuracy of our model. We analyze the assumptions made by the model, their effect on the accuracy, and propose some possible future improvements.