A Parallel Simulation Methodology for Speedup and Obtaining Performance Estimates with Specific Accuracy: Experiences of its Application in Studies of Metropolitan Area Networks

Abstract

Performance measures obtained by stochastic simula tion are estimates, and one must consider the precision of the estimates before making any constructive con clusions about the investigated systems. This paper applies an automated distributed simulation method, called Spectral Analysis in Parallel Time Streams, to speed up production of performance estimates, and for run-length determination in the simulation of High- speed Metropolitan Area Networks (MANs). This method makes sequential simulators suitable for par allel execution on multiprocessors and/or networked computers. At runtime it estimates the information content of observations generated during simulation, generates a point estimate and confidence interval, and directs the run to continue until an estimate is obtained that achieves or exceeds our required level of precision. The application of this methodology for studying high-speed MANs, the speedup, intermachine communication and "warm-up" overhead, and the run lengths needed to produce estimates with a specific level of precision, are reported for each of the param eters investigated. Practical implications are discussed.