Migration de processus dans les systemes massivement paralleles

This thesis discusses process migration in massively parallel systems. This capability is useful to manage efficiently the resources in a system. The issues that we deal with are the migration transparency, the minimisation of the migration overheads and the suitableness of migration algorithms to the scalability of the machine architecture. To migrate a process a set of operations must be performed: freezing the process, transferring its state to the new processor and restarting the process execution. The interaction between the migrated process and other entities in the system must also be handled in order to insure correctness. For these operations we have proposed new migration algorithms which fulfil the requirements above. Our algorithms have been implemented in the ParX Kernel. The performance measurements show that supporting process migration in a massively parallel system do not inquire significant penalties in performance of the system. Besides, we have proposed a load balancing algorithm which uses the process migration facility. Compared to an algorithm relying only on the placement of the processes, the experiments carried out show that thanks to the process migration facility our algorithm offers a potential benefit when the processes execution and inter arrival times are both variable and the overhead to migrate a process is negligible compared to its execution time.Cette these traite de la migration de processus dans les systemes massivement paralleles. L'interet d'une telle fonctionnalite est de permettre a un systeme d'exploitation une gestion efficace des ressources. Les criteres de conception sont la transparence de la migration, la reduction des couts induits, et l'adequation entre les algorithmes de migration et l'extensibilite des architectures cibles. La migration d'un processus vers un nouveau processeur necessite la suspension du processus, le transfert de son contexte d'execution et la reprise de l'execution sur le nouveau processeur. De plus, les protocoles de communication et d'acces doivent etre reconsideres afin de tenir compte de la migration de processus. Pour ces differentes actions et selon les criteres de conception retenus, de nouveaux algorithmes ont ete proposes. La mise en oeuvre de ces algorithmes dans le noyau de systeme Parallele ParX nous a permis de montrer que la realisation d'un mecanisme de migration de processus dans un systeme massivement parallele peut s'effectuer sans penaliser les performances du systeme de facon significative. Enfin, nous avons propose un algorithme de repartition de charge qui utilise la migration de processus. Compare a un algorithme uniquement fonde sur le placement des processus, les experimentations effectuees montrent que notre algorithme ameliore les temps de reponse du systeme grace a la migration de processus. Cette amelioration est obtenue lorsque les temps d'execution et d'inter-creation des processus sont variables et le cout de migration negligeable par rapport a la duree d'execution des processus