With data volumes increasing at a high rate and the emergence of highly scalable infrastructures (cloud computing, petascale computing), distributed management of data becomes a crucial issue that faces many challenges. This thesis brings several contributions in order to address such challenges. First, it proposes a set of principles for designing highly scalable distributed storage systems that are optimized for heavy data access concurrency. In particular, it highlights the potentially large benefits of using versioning in this context. Second, based on these principles, it introduces a series of distributed data and metadata management algorithms that enable a high throughput under concurrency. Third, it shows how to efficiently implement these algorithms in practice, dealing with key issues such as high-performance parallel transfers, efficient maintainance of distributed data structures, fault tolerance, etc. These results are used to build BlobSeer, an experimental prototype that is used to demonstrate both the theoretical benefits of the approach in synthetic benchmarks, as well as the practical benefits in real-life, applicative scenarios: as a storage backend for MapReduce applications, as a storage backend for deployment and snapshotting of virtual machine images in clouds, as a quality-of-service enabled data storage service for cloud applications. Extensive experimentations on the Grid'5000 testbed show that BlobSeer remains scalable and sustains a high throughput even under heavy access concurrency, outperforming by a large margin several state-of-art approaches.

Les applications data-intensive sont largement utilisées au sein de domaines diverses dans le but d'extraire et de traiter des informations, de concevoir des systèmes complexes, d'effectuer des simulations de modèles réels, etc. Ces applications posent des défis complexes tant en termes de stockage que de calcul. Dans le contexte des applications data-intensive, nous nous concentrons sur le paradigme MapReduce et ses mises en oeuvre. Introduite par Google, l'abstraction MapReduce a révolutionné la communauté intensif de données et s'est rapidement étendue à diverses domaines de recherche et de production. Une implémentation domaine publique de l'abstraction mise en avant par Google, a été fournie par Yahoo à travers du project Hadoop. Le framework Hadoop est considéré l'implémentation de référence de MapReduce et est actuellement largement utilisé à des fins diverses et sur plusieurs infrastructures. Nous proposons un système de fichiers distribué, optimisé pour des accès hautement concurrents, qui puisse servir comme couche de stockage pour des applications MapReduce. Nous avons conçu le BlobSeer File System (BSFS), basé sur BlobSeer, un service de stockage distribué, hautement efficace, facilitant le partage de données à grande échelle. Nous étudions également plusieurs aspects liés à la gestion des données intermédiaires dans des environnements MapReduce. Nous explorons les contraintes des données intermédiaires MapReduce à deux niveaux: dans le même job MapReduce et pendant l'exécution des pipelines d'applications MapReduce. Enfin, nous proposons des extensions de Hadoop, un environnement MapReduce populaire et open-source, comme par example le support de l'opération append. Ce travail inclut également l'évaluation et les résultats obtenus sur des infrastructures à grande échelle: grilles informatiques et clouds. / Data-intensive applications are nowadays, widely used in various domains to extract and process information, to design complex systems, to perform simulations of real models, etc. These applications exhibit challenging requirements in terms of both storage and computation. Specialized abstractions like Google’s MapReduce were developed to efficiently manage the workloads of data-intensive applications. The MapReduce abstraction has revolutionized the data-intensive community and has rapidly spread to various research and production areas. An open-source implementation of Google's abstraction was provided by Yahoo! through the Hadoop project. This framework is considered the reference MapReduce implementation and is currently heavily used for various purposes and on several infrastructures. To achieve high-performance MapReduce processing, we propose a concurrency-optimized file system for MapReduce Frameworks. As a starting point, we rely on BlobSeer, a framework that was designed as a solution to the challenge of efficiently storing data generated by data-intensive applications running at large scales. We have built the BlobSeer File System (BSFS), with the goal of providing high throughput under heavy concurrency to MapReduce applications. We also study several aspects related to intermediate data management in MapReduce frameworks. We investigate the requirements of MapReduce intermediate data at two levels: inside the same job, and during the execution of pipeline applications. Finally, we show how BSFS can enable extensions to the de facto MapReduce implementation, Hadoop, such as the support for the append operation. This work also comprises the evaluation and the obtained results in the context of grid and cloud environments.