Abstract : Massivelly parallel computers with thousand and more processors are considered one of today's promising technologies to achieve high performance computing. Such large-scale multiprocessors machines are usually organized as sets of nodes, where each node has its own processor and local memory, connected by some interconnection network. As generally nodes do not share memory, they communicate by passing messages through the network. In this dissertation, we adress the problem of messages routing in massivelly parallel computers. The stress is put on scalable algorithms which require an amount of resources independent of the network size and shape. Through the example of the supernode architecture (dynamically reconfigurable networks of transputers) we show that the complexity of handling the message exchanges by dynamically connecting processors is high in large scale machines. Our study focusses then on the problem of deadlock-free routing in non regular networks and we propose a novel algorithm. Recently the trend in parallel computer architectures is to offer hardware support for handling messages exchanges within nodes. To efficiently achieve this objective for massivelly parallel computers, new methods for compacting routing information on a node are required. A technique well suited for non regular networks is the interval routing introduced by Santoro and Khatib. For this kind of methods we propose deadlock-free solutions for k-ary ncubes and general networks. For the k-ary ncube our method gives moreover nearly optimal paths. Finally, we propose an original extension for interval labelling which needs routing tables of size $O(d^2)$ (where $d$ is the number of neighbors) for a node. This extension allows to represent more routing functions than the original interval labelling.