The problem studied in this thesis is the service discovery on platforms distributed at large scale, a service being a computing service (software components, scientific computing libraries, or binaries) offered with some characteristics and a performance level related to the hardware supporting it. Traditional approaches, designed for reliable and small scale environments, rely upon centralized solutions, unable to scale well in geographically distributed unreliable platforms. Our contribution centers around three main parts. 1) We propose a novel approach called DLPT (Distributed Lexicographic Placement Table), whose design is inspired by peer-to-peer systems. It calls upon an indexing system structured as a prefix tree. This structure supports multi-attribute range queries. 2) We study the mapping of nodes of this tree onto heterogeneous processors of the dynamic underlying network. We propose and adapt some load balancing heuristics for this kind of architectures. 3) Our architecture, targeted for platforms within which processors are unreliable and constantly joining and leaving the network, requires fault-tolerance mechanisms. Replication, usually used, is costly and unable to manage transient faults. We propose alternative best-effort mechanisms based on the self-stabilization theory for the construction and maintenance of prefix trees in a peer-to-peer environment. Among the mechanisms provided, one is proven to be snap-stabilizing. This means that the tree is rebuilt in an optimal time after an arbitrary number of faults. This approach is written in a coarse grain communication model and assumes several restrictions on initial topology handled, making it hard to implement on real platforms. To address these drawbacks, another self-stabilizing protocol is given for actual message-passing environments. Finally, we present a software prototype of this architecture and its first promising experiments on the Grid'5000 platform.