Progress Report Operational Analysis of Queueing Phenomena

·2· Introduction Operational analysts is a new method of analyzing queueing phenomena, especially the phenomena encountered by analysts of computer system performance. Operational analysis deals with data collected from a system rather than the system itself.. The data is a "behavior sequence"-an event trace of system states during an observa-lion period. Operational variables denote quantities measurable in a behavior sequence. Operational assumptions are testable in a behavior sequence. With operational analysis, we have shown that many of the results known from sto-chastic analysis of queueing systems are more general (and more robust) than has been believed. This is because these results can be applied validly to behavior sequences that would occur with probability zero in the stochastic ensembles for which we can carry out analyses. For example, operational analysis has shown that the product form soluUon for closed queueing networks may hold for deterministic systems, which helps explain why these models work so well when applied by industrial engineers to job flow in factories. Operational analysis has, more importantly, opened up new lines of research in~o the robustness of models. We believe that the most important instance of this is the erroT analysis of assumptions. ]t Is possible to measure the error between the value of a state-dependent parameter observed to be state-dependent and the value that parameter is asserted to have by a modeling assumption, and then to express the co~­ sequent error between the model's estimate of a performance metric and the true value found in the data. Suri has recently completed an error analysis to evaluate the sensitivity of the product form solution to the assumption that the service times of devices are 10ad-indepen.dent;1 he has found that model estimates of utilizations and 1 R. Sun, "RobusLness of unBlytical formulae for performance prediction in certain nonclassl-cal queueinB networks,"