PETRI-NET BASED METHODOLOGY FOR TASK-SCHEDULING ON MULTIPROCESSOR ARCHITECTURES

This paper proposes a Petri Net based methodology for task scheduling on multiprocessor architectures. The aim is to build automatically a stochastic model (markovian model or stochastic simulation model) for both performance evaluation and task scheduling. The modelling process that we use consists in the consecutive elaboration of two models: the knowledge model and the action model. The knowledge model is a formalization of the structure and the working of the system, using Petri Nets. This knowledge model is made of three parts: a logical subsystem (precedence constraints between tasks), a physical subsystem (multiprocessor architecture) and a decision subsystem (scheduling method). The action model is a translation of the previous knowledge model in a programming language or in a mathematical formalism. We propose in the paper two kinds of action models: a stochastic simulation model and a markovian model. These models are dedicated to dynamic task allocation between identical processors without communication cost between tasks. It is to be remarked that these action models ar automatically deduced from the corresponding knowledge models. Finally, we report some results obtained from these two action models. In particular, we highlight that these two independent models return identical results and so validate themselves.