The impact of certain parameter estimation errors in queueing network models

The effect that parameter estimation errors have on performance in closed product form queueing networks is investigated. In particular, the effects of errors in the relative utilization estimates of the servers are analyzed. It is shown that in single class load independent networks, the resulting errors in throughput and utilizations are of approximately the same percentage as the errors in the relative utilization estimates. This result does not hold in networks with load dependent servers or multiple customer classes. The percentage errors in mean queue length depend upon the degree of multiprogramming in the network. Errors in mean queue lengths can become unbounded as the degree of multiprogramming becomes unbounded. Implications of these results to computer system modeling are discussed.