A Traffic Simulator with Intra-node Parallelism for Designing High-performance Interconnects

Recently, dynamic control of traffic on the interconnect has been researched for performance improvement. Better load balancing can reduce congestion on each link of the interconnect, which results in the performance improvement of communication conducted by applications. From such a viewpoint, a simulator that estimates the traffic of an application on a high performance computing system, called PFSim, has been proposed and implemented. Under the current implementation, PFSim performs the estimation of traffic for a bag of simulation conditions in a sequential-processing fashion. Therefore, the execution time of the whole simulation for practical analysis of traffic on the interconnect takes a long time because a large set of simulation conditions needs to be supplied. In this paper, we present Parallel PFSim, an enhanced version of PFSim that leverages intra-node parallelism to accelerate the simulation. Specifically, we show the architecture of Parallel PFSim. The evaluation confirms that Parallel PFSim can effectively utilize multiple cores and accelerate the simulation.