The role of knowledge in distributed simulation

This paper examines the suite of closed, stochastic queueing networks proposed by Nicol in order to determine how much impact additional knowledge has on the computation load and the communication load for the multiprocessor running the distributed simulation. Four different networks have been studied at three different levels of traffic intensity. For each network-load pair, four simulation models were constructed, each based on increasing use of user knowledge at a queueing node: no knowledge; service discipline knowledge; queueing discipline knowledge; and routing knowledge. In each case the number of logical process activations and the number of messages used by the distributed simulation was determined and compared to the number of activations needed in a traditional, event-listdriven simulation. In the area of computation needed, distributed simulation using a conservative synchronization algorithm permits between one half and two thirds of the computation to be eliminated when user knowledge is added to the simulation. The results presented also indicate that the differences caused by the additional knowledge are about twice as dramatic in their impact on communication load as they are on the computation load. The results from using an optimistic synchronization algorithm indicate savings of a similar magnitude, except there is a better reduction in computation than in communication. In all but one case the reduction caused by the additional knowledge is less than an order of magnitude.