Automating parallel simulation using parallel time streams

Abstract

This paper describes a package for parallel steady-state stochastic simulation that was designed to overcome problems caused by long simulation times experienced in our ongoing research in performance evaluation of high-speed and integrated-services communication networks, while maintaining basic statistical rigors of proper analysis of simulation output data. The package, named AKAROA, accepts ordinary (nonparallel) simulation programs, and alll further stages of stochastic simulation should be transparent for users. The package employs a new method of sequential estimation for the Multiple-Replications-in-Parallel scenario. All basic functions, including the transformation of originally nonparallel simulators into ones suitable for parallel execution, control of the precision of estimates, and stopping of parallel simulation processes when the required precision of the overall steady-state estimates is achieved, are automated. The package can be used on multiprocessor systems and/or heterogeneous computer networks, involving an arbitrary number of processors. The design issues, architecture, and implementation of AKAROA, as well as the results of its preliminary performance studies are presented