Performance Modeling of Heterogeneous Distributed Applications

Heterogeneous network computing allows the development of a single complex application using a distributed network of machines; these machines may diier in terms of CPU and memory capacity and/or architecture and specialized functions. The design of eecient applications targeted to this class of systems is a complex task requiring the evaluation of diierent performance tradeoos. In this paper we present a modeling technique, based on Generalized Stochastic Petri Nets (GSPNs), for the performance analysis of heterogeneous applications. The technique identiies the features that aaect the performance of the application ; these features are modeled in isolation by GSPN submodels. The model for a given application is obtained by composing the submodels corresponding to the particular features exposed by the application. We illustrate the use of the proposed technique by modeling the CASA 3D-REACT heterogeneous application (which is one of the applications belonging to the CASA Gigabit Testbed) and evaluating its performance by analyzing the corresponding model. The model is validated by comparing the results obtained by solving the GSPN model with those obtained by measurements.