Multiprocessor Resource Estimation Using a Stochastic Modeling Approach

A modeling approach based upon the notions of thread spawning and maximum length probability density functions is presented. By using a data dependence graph, or alternatively, an available parallelism profile, probability density functions may be derived and used as input to a queuing system model that can predict required resources.. The accuracy of the model is measured by validation using parallelism profiles under different halting criteria. The result of this work is the establishment of a modeling framework that can later be used to estimate the effects of non-zero interprocessor latencies and limitations due to a finite number of processing elements.