Sizing and capacity planning are key issues that must be addressed by anyone wanting to ensure a distributed system will sustain an expected workload. Solutions typically consist in either benchmarking,or modeling and simulating the target system. However, full-scale benchmarking may be too costly and almost impossible, while the granularity of modeling is often limited by the huge complexity and the lack of information about the system. To extract a model for this kind of system, we propose a methodology that combines both solutions by first identifying a middle-grain model made of interconnected black boxes, and then to separately characterize the performance and resource consumption of these black boxes. Then, we present two important issues : saturation and stability, that are key to system capacity characterization. To experiment our methodology, we propose a component-based supporting architecture, introducing control theory issues in a general approach to autonomic computing infrastructures.