Networking systems consist of network infrastructures and the end-hosts have been essential in supporting our daily communication, delivering huge amount of content and large number of services, and providing large scale distributed computing. To monitor and optimize the performance of such networking systems, or to provide flexible functionalities for the applications running on top of them, it is important to know the internal metrics of the networking systems such as link loss rates or path delays. The internal metrics are often not directly available due to the scale and complexity of the networking systems. This motivates the techniques of inference on internal metrics through available measurements. In this thesis, I investigate inference methods on networking systems from multiple aspects. In the context of mapping users to servers in content delivery networks, we show that letting user select a server that provides good performance from a set of servers that are randomly allocated to the user can lead to optimal server allocation, of which a key element is to infer the work load on the servers using the performance feedback. For network tomography, where the objective is to estimate link metrics (loss rate, delay, etc.) using end-to-end measurements, we show that the information of each end-to-end measurement can be quantified by Fisher Information and the estimation error of link metrics can be efficiently reduced if the allocation of measurements on paths is designed to maximize the overall information. Last but not least, in the context of finding the most reliable path for routing from a source to a destination in a network while minimizing the cost of exploring lossy paths, the trade-off between exploiting the best paths based on estimated loss rates and taking the risk to explore worse paths to improve the estimation is investigated, and online learning methods are developed and analyzed. The performance of the developed techniques are evaluated with simulations.