A mapping method for distributed heterogeneous systems

We present a method to characterize a distributed program for evaluating processes to processors allocation strategies in a heterogeneous computing system. The method is based on a Generalized Stochastic Petri Net model of the given hardware/software connguration. The model is used to compute various performance indices, depending on the nature of the application. Model analysis is carried out either analytically or on a simulation basis. The proposed technique is automated to a large extent. Distributed heterogeneous computing seems to be an eeective approach to supercom-puting, and/or to improve utilization of available hardware 1]. A Distributed Heterogeneous System (DHS) is a collection of diierent machines, usually interconnected via local or wide-area networks, which can be programmed on behalf of a single application, as if it were a global \virtual" computational resource. EEcient programming tools have been developed for this purpose in the last years 2], and industrial standards are being developed 3]. Nevertheless distribution and heterogeneity rise a number of problems and issues 4]. Among them, optimal resource management is one of the most diicult to solve; indeed allocation problems are known to be generally intractable, and scheduling and task allocation theory for such systems are rapidly expanding research areas 5{7]. In a DHS the goal of the task allocation problem is usually to determine the hard-ware/software connguration which minimizes the global execution cost of an application. The parameters that characterize the application and the hardware connguration are diicult to compute analytically, especially if the distributed program behavior is non-deterministic, or if it in turns depends on the allocation scheme. For this reason, modeling and simulation tools are needed as an aid in the design and evaluation of a concurrent application. In this paper we present a method to compute processing and communication parameters , as well as overall performance indices of a given hardware/software distributed