Since the last decade, computing systems turn to large scale parallel platforms composed of thousands of processors. Many actual applications run on such systems for long duration, up to several days or weeks. Recently, statistic studies about failures on high performance computing platforms emphasize that the mean time between failures may not exceed few hours. Thus, it is necessary to develop effcient strategies providing a safe and reliable completion of applications. This may be achieved through redundancy or by storing intermediate computation states on reliable external devices. Saved states are then used to restart computations from the last checkpoint. This last approach called checkpointing is one of the most popular fault tolerance technique in parallel systems.