On learning how to predict

To meet performance guarantees, we should consider building computer systems more like queueing models we do understand, rather than trying to build complex queueing models of systems we do not understand. This approach is already familiar to software engineers, who use models to constrain software development so that the results are more reliable and understandable. Four real examples illustrate how a performance model can guide a system's design. The need for experimentation with models is emphasized.