The treatment of sparse numerical problems on large scale systems is often reduced to that of their kernels. For reasons of efficiency in time and space, specific compressing formats are used for storing the matrices of these problems. Most of sparse scientific computations are led to linear algebra problems. Here two fundamental problems are often considered: linear systems resolution and eigenvalue computation. In this thesis, we address the study of distribution of computations performed in iterative methods to solve such problems. The sparse matrix-vector product (SMVP) constitutes a basic kernel in such iterative methods. Thus, our problem reduces to the study of SMVP distribution on large scale distributed systems. Three phases are required for achieving our objectives (i)- pre-processing, (ii)- processing and (iii)- postprocessing. In phase 1, we first process the optimization of four versions of the SMVP algorithm corresponding to four specific matrix compressing formats; we study then their performances on sequential target machines. In addition, we focus on the study of load balancing in the procedure of data distribution (i.e. the sparse matrix rows) on a large scale distributed system. Concerning the processing phase, it consists in validating our study by a series of experimentations achieved on a volunteer distributed system that we installed through using XtremWeb-CH middleware. As to the post-processing phase, it consists in interpreting the experimental results previously obtained in order to deduce adequate conclusions
