Abstract : Performances of programs on distributed memory parallel machines are highly dependent of the efficiency of interprocessor communications. Parallel programming environments often offer poor support for high level communication models. This thesis deals with high level group communications in such architectures. This problem has two main aspects : first, correct and efficient handling of message exchanges for diffusion between processors, second, construction of group communication protocols for processes. We first address the problem of correct message passing within diffusion exchanges on top of efficient and correct routing functions. The proposed solution is general and is independent of the representation of the routing function, and it is independent from the topology of the interconnection network. Moreover constant memory space is required to prevent deadlocks in general networks. Futhermore, using interval labelling to represent the routing function, we show how our diffusion method can be integrated in a routing processor. Secondly, we consider the construction of high level diffusion protocols (synchronous, asynchronous), and propose a diffusing virtual machine for the parallel generic kernel ParX developped by our group. Our virtual machine is based on two diffusion protocols which can be used to correctly and efficiently build most of existing global communication schemes. The protocol interface offers primitives for coherent management of process groups, message exchanges and control (broadcasting and scattering). Our virtual machine is designed to be a minimal support to efficiently and correctly implement different existing group communication interfaces and libraries (PVM, MPI, etc.). The set of proposed solutions has been integrated in the prototype version of the ParX kernel, and some performances evaluated.