Since many years, the complexity of grid middlewares never ceased to increase. Thanks to the adoption of SOA (Services Oriented Architectures), they allows to design highly complex, high level, and naturally distributed software architectures. The actors (administrators, developers and users) are thus confronted to many problems related to the distributed nature of resources : How to select one replicated resource to fulfill one given task? How to deploy one or several resources, or even a complete software architecture? How to compose several resources ? etc.

Solving these problems is difficult as actors are working at the highest level, from which the infrastructure is fully abstracted. Another difficulty is due to the diversity of resources (from simple files to complex web services implying communications as well as computations), to the diversity of goals (performances, load balancing, response quality, financial aspects\ldots ), to the infrastructure performances and to the diversity of problem (selection, deployment, composition\ldots ). However, such solving is crucial as it conditions the platform performances as well as the quantity of actors work.

Our proposal is a Globus web services called the Network Distance Service (NDS). NDS allows to model intuitively a large variety of problems as graphs thanks to a notion of distance adaptable to resources and goals and adaptative to infrastructure performances. Classical graph theory algorithms are then used to solve the problems : a shortest-path algorithm is used to solve selection and composition problems ; a clustering algorithm is used to solve deployment problem. \\
Two genuine algorithms have also been developed thanks to this approach : FReDi, in order to drive dynamically some replicas, and MRKM, in order to deploy a whole software architecture.\\
Finally, one data warehouse have been designed in order to improve the analysis capabilities of monitoring systems.

Our approach is illustrated by the GGM software architecture (http://liris.cnrs.fr/PROJETS/ggm), its different target infrastructures and the different distribution problem encountered. Large scale experiments have been conducted on the Grid5000 platform and shows that optimal solutions have been computed in very short solving times and with limited user work.