Déploiement adaptatif d'applications parallèles : Algorithme de vol de travail appliqué au déploiement d'applications parallèles sur des grappes de grande taille

National audienceThis article presents a library named Taktuk for the deployment of applications on large sized clusters (thousands of nodes). This library is conceived for the development of interactive tools and thus has to complete a whole deployment in the shortest time. With this objective in mind, we show that the behavior of standard communication tools that we use can be modelized as a classical communication and we deduce from this a theoretical algorithm for optimal deployment. We then present our implementation choices that take into account the possible heterogeneity of the execution platform and the uncertainties about the value of some parameters. The evaluation of our tool highlights its near optimality and its adaptability that are essentially due to the work-stealing algorithm which is the heart of our system