Relaxation of Synchronization Constraints in Parallel Simulation of ATM Network

Parallel simulation using conservative or optimistic approach does not exhibit good performance when applied to large ATM networks. We introduce a statistically correct approach based on the observation that some performance measurements can be obtained without the total time ordering established by conventional approaches. In our approach events can be processed in partial time order. Upon reception of an outdated message, a process does not attempt to cancel the bad computations it has done but uses local estimations instead. To demonstrate our approach a parallel simulation of an ATM network is performed applying this idea. We focus on the mean buuer length of each switch outputs and found out that the results are very close to those of a correct simulation. The statistically correct approach reduces the simulation cost while keeping a reasonable level of accuracy.