Distributed service architectures are mandatory to handle the platform scale and dynamicity hindering the development of grid and P2P applications. These large-scaled distributed applications are difficult to design, develop and tune because of both theoretical and practical issues. This paper presents the GRAS framework that allows developers to first implement and experiment with such an infrastructure in simulation, benefiting from a controlled environment. The infrastructure can then be deployed in-situ without code modification. We detail our design goals, and contrast them with the state of the art. We study the exchange of a message (from the Pastry protocol) using either GRAS or several other solutions. We quantify both the code complexity and the performance and find that GRAS performs better according to both metrics.