Performance of the Queueing Network Analyzer

Performance of the Queueing Network Analyzer By W. WHITT* (Manuscript received March 11, 1983) This paper describes the performance of the Queueing Network Analyzer (QNA), a software package developed at Bell Laboratories to calculate approximate congestion measures for networks of queues. QNA is compared with simulations and other approximations of several open networks of singleserver queues. This paper illustrates how to apply QNA and indicates the quality that can be expected from the approximations. The examples here demonstrate the importance of the variability parameters used in QNA to describe non-Poisson arrival processes and nonexponential service-time distributions. For these examples, QNA performs much better than the standard Markovian algorithm, which does not use variability parameters. The accuracy of the QNA results (e.g., the expected delays) in these examples is satisfactory for engineering purposes. I. INTRODUCTION AND SUMMARY This paper is a sequel to Whitt, 1 which described the software package called the Queueing Network Analyzer (QNA). QNA calculates approximate congestion measures for networks of queues. The first version of QNA treats open networks of multiserver queues with the first-come, first-served discipline and no capacity constraints. QNA is designed to treat non-Markovian models: The arrival processes needjiot be Poisson and the service-time distributions need not be exponential. QNA approximately characterizes other kinds of variability through variability parameters assigned to each arrival process and each service-time distribution.