Synthetic traffic generation as a tool for dynamic interconnect evaluation

Run-time reconfigurable interconnection networks can provide significant performance gains in shared-memory multiprocessor systems. However, designing such networks is hard, requiring detailed but slow execution-driven simulations, since faster methods are currently not suitable for use with dynamic network topologies. In this paper, we extend one of these methods, synthetic traffic generation, to incorporate the dynamic traffic behavior necessary to accurately determine the performance of a reconfigurable network. Our synthetic traffic flow has the same characteristics as the flow resulting from an execution driven simulation, but can be much shorter: we can gain a reduction in simulation time of up to 100x at only a limited expense in accuracy. This way, it is possible to quickly analyze the dynamic interconnect requirements of an application and evaluate various aspects of a proposed reconfigurable interconnect implementation.