Parallel simulation of Markovian queueing networks

Abstract

Consider simulating a large queueing network, i.e., one with many queues and jobs, on a parallel computer. Suppose the network is partitioned so that each processor is assigned a set of queues to simulate. The basic difficulty in parallel simulation is synchronizing the simulation clocks on each of the processors so that events appear to get executed in the proper order. If the network possesses a Markovian structure, then algorithms described by the authors (1993) can be used to simplify the clock synchronization problem. For appropriately structured queueing networks, these algorithms are quite efficient and large speedups have been obtained; in one example a speedup of 220 was measured on 256 processors of the Intel Touchstone Delta computer.<>