Abstract: Large Grids are build by aggregating smaller parallel machines through a public long-distance interconnection network (such as the Internet). Therefore, their structure is intrinsically hierarchical. Each level of the network hierarchy gives performances which differ from the other levels in terms of latency and bandwidth. MPI is the de facto standard for programming parallel machines, therefore an attractive solution for programming parallel applications on this kind of grids. However, because of the aforementioned differences of communication performances, the application continuously communicates back and forth between clusters, with a significant impact on performances. In this report, we present an extension of the information provided by the run-time environment of an MPI library, a set of efficient collective operations for grids and a methodology to organize communication patterns within applications with respect to the underlying physical topology, and implement it in a geophysics application.