Scalable Time Warp on Blue Gene Supercomputers

Abstract

n this paper we illustrate scalable parallel performance for the TimeWarp synchronization protocol on the L and P variants of the IBM BlueGene supercomputer. Scalable Time Warp performance for models that communicate a large percentage of the event population over the network has not been shown on more than a handful of processors. We present our design for a robust performing Time Warp simulator over a variety of communication loads, and extremely large processor counts-- up to 131,072. For the PHOLD benchmark model using 65,536 processors, our Time Warp simulator produces a peak committed event-rate of 12.26 billion events per second at 10\% remote events and 4 billion events per second at 100\% remote events, the largest ever reported. Additionally, for the Transmission Line Matrix (TLM) model which approximates Maxwell's equations for electromagnetic wave propagation, we report a committed event-rate in excess of 100 million on 5,000 processors with 200 million grid-LPs. The TLM model is particularly challenging given the bursty and cubic growth in event generation. Overall, these performance results indicate that scalable Time Warp performance is obtainable on high-processor counts over a wide variety of event scheduling behaviors and not limited to relatively low, non-bursty rates of off-processor communications.