Predicting the performance of reconfigurable optical interconnects in distributed shared-memory systems

New advances in reconfigurable optical interconnect technologies will allow the fabrication of low-cost, fast and run-time adaptable networks for connecting processors and memory modules in large distributed shared-memory multiprocessor machines. Since the switching times of these components are typically high compared to the memory access time, reconfiguration exploits low frequency dynamics in the network traffic patterns. These are however not easily reproduced using statistical traffic generation, a tool commonly used when doing a fast design space exploration. In this paper, we present a technique that can predict network performance, based on the traffic patterns obtained from simulating the execution of real benchmark applications, but without the need to perform these slow full-system simulations for every parameter set of interest. This again allows for a quick comparison of different network implementations with good relative accuracy, narrowing down the design space for more detailed examination.