Computational grids promise to deliver a vast computer power as transparently as the electric power grid supplies electricity. Thus, applications need to be automatically and transparently deployed on grids, in particular MPI parallel applications. However, deploying such applications on a grid is complex: the user must select compute nodes manually, launch processes on distributed, heterogeneous nodes, and configure the MPI library so it can adapt to its grid environment. Our objective is to hide the complexity of application deployment on computational grids. This paper proposes a software architecture designed to automatically deploy MPI applications on a grid. It also introduces a description and packaging model of MPI applications. Finally, the paper shows how the MPI library can be automatically configured with network topology information to optimize collective operations.