Performability models and solutions

A principal goal of computing system evaluation is the measurement of the system's ability to perform. Measures such as performance, reliability, and effectiveness are often employed, but such metrics are often not suitable when evaluating systems in the increasingly important class of degradable systems. Among the measures proposed for such systems is "performability," which is simply the probability measure of the system performance variable. Classical performance and classical reliability are specialized cases of performability. To be effective, performability evaluation requires tractable techniques of solution. This dissertation concerns the modeling, calculation, and use of a system's performability. Specfically, we examine two broad classes of performability models: (1) those wherein system performance assumes values in an arbitrary, finite set, and (2) those systems where performance is continuous and identified with "reward". For the first class of models, a methodology for relating low-level behavior to high-level system performance is formalized. Behavior is characterized in terms of finite arrays of variables having finite domains. A calculus is developed for manipulating such arrays, and algorithms are described for deriving the set of low-level behaviors which result in each system performance level. The probability of each performance level (and hence the performability) is obtained by calculating the probability of the corresponding set of behaviors. Also described and illustrated is METAPHOR, a computer package implementing these algorithms. For the second class of performability models, a general method for determining the probability distribution function of the performance variable (i.e., the performability) is derived. The result is an intergal expression which can be solved either analytically or numerically. Examples of both types of solutions are given, and procedures implementing the numerical solution and included within METAPHOR are described.