Modern grid platforms present a much more complex interconnection topology than classical super-computers or clusters. Information about this topology is critical to network-aware applications, such as a grid scheduler optimizing communication times, or such as data replication managers. This paper presents a theoretical framework and the corresponding tool for automatically discovery of network topology. The goal is not to discover the physical layout of machines interconnections, but to construct a synthetic view of the effects of the topology as perceived by an application. Among other things, this requires that performance of concurrent transfers can be assessed. Our work uncovers an original mathematical model that arises from the formalization of the topology discovery problem, for which we propose an algorithm. We prove that when the underlying graph is a constellation of trees (i.e., a set of trees whose roots are interconnected by a complete clique), our algorithm produces a valid solution. We then extend the algorithm to other cases. We present some preliminary evaluation results obtained with the SimGrid simulator.