A Comparison of Two Methods for Advancing Time In Parallel Discrete Event Simulation

Anthony P. Galluscio, John T. Douglass
Brain A. Malloy and A. Joe Turner
[email protected]
Department of Computer Science
Clemson University
Clemson, SC 29634

Abstract
In this paper, we compare the design and implementation of a parallel simulation of a traffic flow network using two different approaches: event-driven and time-driven. We begin by designing an efficient event-driven approach to model the traffic network; our design matches a time-driven model of the same traffic network[5]. Our experiments with the sequential implementation of the two approaches correlates with previous research [12]. Exploiting the look-ahead in the event-driven model, we design a conservative parallel implementation of the traffic flow problem where we obtain a maximum speedup of 9.27 using 16 Sun workstations. This speedup is appreciable since our parallel architecture is parallel virtual machine (PVM)[8], not known for fast communication, and we use wall-clock time as a measure of execution speed. We show that appreciable speedup can be achieved in parallelizing either the event-driven or time-driven approach. We also show that speedup is a misleading metric when used to compare the parallelizability of the two approaches. Parallel performance, as measured by speedup, may be better when the sequential performance is poor. For example, the time-driven implementation achieved better speedup (3.21 to 3.56) than the event-driven implementation (0.59 to 0.97) for few cars in the system; however the sequential time-driven implementation required longer to execute than the event-driven implementation for few cars in the system. Similarly, for many cars in the system, the event-driven implementation achieved better speedup (9.01 to 9.27) than the time-driven implementation (5.99 to 9.12).

keywords: modeling, event-driven, time-driven, methodology, conservative parallel distributed simulation.