In distributed memory parallel architectures, interprocess communication is one of the main effi.ciency factor for the applications. The system which is in charge of this task, i.e, the communication kernel, must integrate an essential functionality for such architectures: message routing. This functionality is ensured by a specifie component of the communication kernel : the routing kernel, whose purpose is to convey a message from a sender node to a receiver node. Messages conveying requires a routing strategy which specify the communication paths for each pair of (source,destination) processors of the interconnection network. Such routing strategy must satisfy in one hand, correction criteria and in other hand, efficiency criteria. The purpose of this thesis is the design of routing strategies for networks of processors, which satisfy at the same time, correction criteria and e:fficiency criteria. We propose a method to design routing strategies, which allows by an incrementai approach, to satisfy both of the criteria: the multi-level communication and the associated primary communication scheme. To prove the e:fficiency of the method, we apply it to a particular topology : The torus network. The several obtained routing algorithms are correct and very efficient. We propose also a technique to implement our routing method, which allows to compute the routing tables directly on the target network processors. Thus, this technique allows to obtain a self constructing system.
