Performance estimation of distributed computer systems
John A. Morse · 1988
Many computer systems encountered in the real world are best modeled as sets of independent sequential whose interaction is limited to synchronized message exchange. Such communicating sequential processes are often non-deterministic--their behavior can only be described in probabilistic terms. This thesis presents analytic methods for predicting the performance of these types of systems. The method uses an adaptation of C. A. R. Hoare's CSP language as a specification language. A Petri net is constructed from the CSP specification to capture the concurrency and synchronization constraints of the system. By doing reachability analysis on the Petri net, a probabilistic grammar is derived which describes the possible sequences of messages that could arise from execution of the system under study. Analysis of the probabilistic grammar predicts the mean and standard deviation of the total message traffic generated, and also predicts the mean number of occurrences of each message type. Several unique aspects of this approach are described, including transformation and simplification techniques that help avoid the state explosion problem inherent in this type of modeling. Specific examples drawn from both software and hardware systems are presented that validate the modeling technique.