A program behavior model and its evaluation

We investigate a program behavior model that generates address reference strings including the interreference durations. This model views a trace as the sample path of a semi-Markov process. The transition probability and sojourn matrices of that process are statistically inferred. Two methods of evaluating the accuracy of the model are described. We experimentally evaluate the model on the shared data traces of the individual processes of a range of realistic parallel programs executing on a shared memory multiprocessor. The model generates traces that closely approximate the original trace, however, the number of model states needed for accuracy is large.<
>