Parallel applications allow to execute many tasks of the same application, on many processors in the same time. These applications were first executed on supercomputers, and then on clusters. Today, this last infrastructure is evolving towards grids. Grids, in this thesis, are defined as an interconnection of clusters by a long-distance network. Most of MPI applications are written with the MPI standard, which describe how processes communicate, by message passing. Initially design for clusters, MPI is still used to program parallel applications in grids. While different problems have been resolved for communications in clusters, the long-distance network of the grid raises many problems. First, MPI messages are transmitted re-liably on the long-distance network via the TCP protocol. However, TCP is based on a data transfer by flows ; thus, it is not adapted to MPI communications. Secondly, the high latency of the long-distance network implied costly communications and costly retransmissions of lost packets. Finally, the available throughput on this access link to the network is generally lower than the sum of the throughput if all nodes want to communicate on this link, at the same time. This is creating congestion, both within the application and both between the other applications. It becomes necessary to manage this bottleneck. The main objective of this thesis is to study in detail the interactions between the parallel applications and the transport layer in the long-distance network of computing grids, and then to solve these problems.
