Interval-based performance analysis of computing systems

Analytic performance models are often used for predicting the performance of computing systems. Existing models accept single valued parameters as input and produce single valued performance measures as outputs. This research proposes to associate intervals or ranges of values with performance measures and key system parameters. Such an approach is appropriate when exact parameter values are unknown but approximate ranges for parameters may be estimated. Conventional arithmetic cannot handle intervals and interval arithmetic-based techniques are required. The paper reports on the feasibility of application of interval arithmetic in the solution of existing well-known models of computing systems. One of the problems with using interval arithmetic is the potential loosening in the interval for the model output: the computed interval may be wider than the actual interval. A computational method based on the notion of interval splitting is introduced in this paper for controlling this problem. The technique is found to be effective in the context of a number of models.>