Today, clusters are often interconnected by long distance networks within grids to offer a huge number of available ressources to a range of users. MPI, the standard communication library used to write parallel applications, has been implemented for clusters. Two main features of grids: long distance networks and technological heterogeneity, raise the question of MPI efficiency in grids. This report presents an evaluation of four recent MPI implementations (MPICH2, MPICH-Madeleine, OpenMPI and GridMPI) in the french research grid: Grid'5000. The comparison is based on the execution of pingpong, NAS Parallel Benchmarks and a real application in geophysics. We show that this implementations present performance differences. Executing MPI applications on the grid can be beneficial if the parameters are well tuned. The paper details the tuning required on each implementation to get the best performances.