Effect of communication overheads on Time Warp performance

Abstract

This paper describes results of an empirical study to evaluate the effect of communications delays on the performance of the Time Warp mechanism in order to assess the effectiveness of Time Warp in distributed computing environments. An implementation of Time Warp on a collection of networked workstations is used in this study. Performance using synchronous and asynchronous message passing primitives are compared, and it is observed that Time Warp experiences much more rolled back computation when using the synchronous primitives for certain applications. Message passing is decomposed into a computation component at the sender and receiver processors, and a transmission delay component that represents the amount of time the message remains “in transit” within the network. The effect of each of these components on Time Warp performance is studied. It is observed that communications latency in distributed computing environments can significantly degrade the efficiency of Time Warp for applications containing large numbers of simulator objects with small event granularity (by increasing the amount of rolled back computation), particularly applications using “self-driving” simulator objects. However, for applications containing large grained events, communication delay appears to have little effect on rollback behavior in Time Warp.