Experiments investigating the performance of two parallel shared-memory simulators are described. The emphasis is on real-life, complex, and heterogeneous problems.
The application chosen is the simulation of large ATM networks. Two configurations were used: the 11-switch Western Canadian regional network and the 54-switch Canadian National Test Network. The configurations were simulated for three and four different traffic load scenarios, respectively.
The two parallel simulators used are a conservative approach proposed initially by Chandy and Misra in 1979, and an optimistic one, known as TimeWarp, originally proposed by Jefferson in 1985. They are compared against a central-event-list sequential simulator. All simulators were optimized for the target application, and ways to reduce overheads associated with the parallel implementations are detailed in the text.
The authors carefully describe the simulators, the application, and the results of the experiments. The paper is a good source of information on the pros and cons of parallel simulators, in terms of overhead, scalability, and suitability of a given approach to a specific application.
Two important open problems mentioned are the scheduling and load balancing of the various processes on the available processors, when executing programs in parallel. In the experiments described, these tasks were always done statically and manually on a trial-and-error basis.