Performance evaluation of N-processor Time Warp using stochastic activity networks

Abstract

The speedup obtainable with Time Warp parallel discrete-event simulation varies greatly with the characteristics of the simulation and the Time Warp implementation. Analytic studies have been done to determine the expected speedup of Time Warp, but these studies have been limited to bound analysis or analysis considering a few overheads with the others assumed negligible. The models used in these studies have often been constructed directly in terms of a Markov process, making the construction process difficult. This thesis uses stochastic activity networks to construct a model of Time Warp at a higher level, enabling the construction of a more detailed underlying Markov process. The result is an analytic model for N-processor Time Warp that considers the combined effects of limited optimism, cascaded rollback, communication cost, and rollback cost. Given these effects, solutions for performance variables such as speedup, fraction of time blocked, and channel utilization are obtained.