A large part of today's most popular applications are data-intensive; the data volume they process is continuously growing. Specialized abstractions like Google's MapReduce and Pig-Latin were developed to efficiently manage the workloads of data-intensive applications. These models propose high-level data processing frameworks intended to hide the details of parallelization from the user. Such frameworks rely on storing huge objects and target high performance by optimizing the parallel execution of the computation. The purpose of this PhD is to provide efficient storage for the MapReduce framework and the applications it was designed for. The research conducted so far, concerned the storage layer this type of applications require. To meet these requirements we rely on BlobSeer, a system for managing massive data in a large-scale distributed context.