A Distributed Discrete Event Simulation Framework for Timed Petri Net Models

Parallel and distributed simulation techniques are consolidating as a potentially eeective way of improving elapsed time for discrete event simulation experiments of large system models. The common aim to all the approaches is to divide a single simulation program into logical processes (LPs) to be executed concurrently on individual processing nodes of a parallel computer. Two issues are critical in this concern: First there is the problem of deening protocols that allow an eecient, distributed handling of local event simulations in the LPs while at the same time maintaining the proper causality among events, and secondly the problem of partitioning of large simulation models into a number of cooperating simulation LPs an mapping them onto processing nodes. Most of the literature so far has concentrated on queueing networks as simulation models, trying to tackle these problems. In this work timed Petri nets will be studied as speciications of simulation models, and both the development (or adaptation) of distributed simulation protocols, as well as the partitioning of large Petri net models will be studied in order to optimize the elapsed (parallel) simulation time. Distributed simulation mechanisms based on the two classical approaches (conservative, optimistic) for timed transition Petri nets (TTPNs) are systematically developed. The proposed partitioning is based on characteristics and mathematical properties of the structure of the Petri net model that can be eeectively computed oo-line before the start of the simulation experiment. These properties statically identify portions of models that potentially yield concurrent events and portions that have no chance to yield useful parallelism. Based on this information an automated partitioning into minimum grain size LPs is possible, and rules for packing grains can be given in order to nd an optimum balance of communication and computation requirements of the distributed simulation to t to dedicated multiprocessor systems. The distributed simulation framework has been implemented on three multiprocessor hardware platforms (Intel iPSC/860, Transputer and Sequent Balance). Speedup characteristics of sample TTPNs on these machines are compared, and the performance innuences are categorized according to empirical observations. This work demonstrates, that distributed simulation on a real multipro-cessor system can in fact gain speedup over the sequential simulation { this can be achieved even for very small scale simulation models.