Vers des supports exécutifs capables d'exploiter des machines multicors hétérogènes

Bien que les accélérateurs fassent désormais partie intégrante du calcul haute performance, les gains observés ont un impact direct sur la programmabilité, de telle sorte qu'un support proposant des abstractions portables est indispensable pour tirer pleinement partie de toute la puissance de calcul disponible de manière portable, malgré la complexité de la machine sous-jacente. Dans cette thèse, nous proposons un modèle de support exécutif offrant une interface expressive permettant notamment de répondre aux défis soulevés en termes d'ordonnancement et de gestion de données. Nous montrons la pertinence de notre approche à l'aide de la plateforme StarPU conçue à l'occasion de cette thèse.Multicore machines equipped with accelerators are becoming increasingly popular in the HighPerformance Computing ecosystem. Hybrid architectures provide significantly improved energyefficiency, so that they are likely to generalize in the Manycore era. However, the complexity introducedby these architectures has a direct impact on programmability, so that it is crucial toprovide portable abstractions in order to fully tap into the potential of these machines. Pure offloadingapproaches, that consist in running an application on regular processors while offloadingpredetermined parts of the code on accelerators, are not sufficient. The real challenge is to buildsystems where the application would be spread across the entire machine, that is, where computationwould be dynamically scheduled over the full set of available processing units.In this thesis, we thus propose a new task-based model of runtime system specifically designedto address the numerous challenges introduced by hybrid architectures, especially in terms of taskscheduling and of data management. In order to demonstrate the relevance of this model, we designedthe StarPU platform. It provides an expressive interface along with flexible task schedulingcapabilities tightly coupled to an efficient data management. Using these facilities, together witha database of auto-tuned per-task performance models, it for instance becomes straightforward todevelop efficient scheduling policies that take into account both computation and communicationcosts. We show that our task-based model is not only powerful enough to provide support forclusters, but also to scale on hybrid manycore architectures.We analyze the performance of our approach on both synthetic and real-life workloads, andshow that we obtain significant speedups and a very high efficiency on various types of multicoreplatforms enhanced with accelerators.BORDEAUX1-Bib.electronique (335229901) / SudocSudocFranceF