In a parallel programming environment, the load sharing module - or application level scheduler - manages the execution of a program over the resources of a distributed architecture: processors and memory modules. In practice, the efficiency of the scheduling strategy highly depends on the characteristics of both application and target architecture. Consequently, to achieve code portability, it should be possible to tune the scheduling strategy. In this thesis, we propose the use of a dynamic data flow graph as the key element in order to separate the application code from its scheduling strategy. This approach allows the construction of a modular and generic software environment that enables to tune the scheduling strategy. The specificati...