On the Predictive Properties of Performance Models Derived through Input-Output Relationships

Building an analytical performance model is a challenge when little is known about the functionality and behavior of the system being modeled and/or when obtaining model parameters through measurements is difficult. This paper addresses this problem by presenting an approach that derives analytic model parameters by observing the input-output relationships of a real system. More specifically, input (i.e., arrival rates for each job class) and output (i.e., average response time for each job class) measurements are used to estimate the per-class service demands and number of servers for a Queuing Network model of the system. This model, called the computed model (CM), provides the same output values for the same input values used to derive the CM. The important question is whether the CM has predictive power, i.e., can the CM predict the output values that would be observed in the real system for different values of the input? The CM’s parameters are obtained by solving a non-linear optimization problem. The paper shows through experiments that the CM is relatively robust and has predictive power over a range of input values.