During the last years, the proliferation of modern devices capable of capturing context information through various sensors has triggered the blossom of context-aware systems, which automatically adapt their behaviour based on the detected context. For many emerging context-aware applications, context may include a huge amount of entities possibly dispersed geographically over a wide area. In such large-scale scenarios, the efficient processing of context information becomes a challenging task. In this dissertation, we are going to focus on the problem of the efficient processing of context information. In particular, we will consider the problem of deriving high-level context information, also referred to as situation in the literature, from sensor data streams captured by a large set of geographically distributed sensors. First, we present the architecture of a distributed system that uses reasoning algorithms to detect situations in an overlay network of data stream processing operators. Then we are going to introduce our strategies for the optimal distribution of data processing between processing nodes in order to save network resources, by optimizing for bandwidth-delay product, and fulfill given QoS requirements, such as end-to-end latency constraints. To this end, we formulate three (constrained) optimization problems, which search for an optimal placement of operators onto physical hosts with respect to different application constraints. The proposed algorithms are executed in a distributed way, by using local knowledge of the system. Our evaluation shows that our algorithms achieve good approximations of the optimal solutions, while inducing limited communication overhead.