Performance is one of the key features of parallel and distributed computing systems. For that reason, a significant research effort was invested in the development of approaches in the area of performance modeling and prediction. Since many parallel applications from scientific computing use MPI communication operations to distribute or collect data, we present in this paper a novel (off-line) approach that addresses the performance prediction of MPI routines in multi-clusters platforms. The main objective of this approach is to predict accurately and efficiently the performance of a given routine. Our solution is based principally on models for point to point (P2P) MPI routines which are obtained after a short profiling procedure. Since collective communication routines are composed of P2P routines, the performance prediction of the formers is done on the basis of a rapid emulation of these routines and on an evaluation of P2P routines models. Experimental results obtained on a grid platform demonstrated the interest of the proposed approach.