Robust network service delivery in shared resources environments

This dissertation is aimed at understanding how to design effective resource management schemes in shared resources network environments. We consider two well-known network service models: the best effort service model and the aggregate Quality of Service (QoS) model. Both models grant groups of users common access to network resources, and therefore allow traffic from distinct sources to interact in arbitrary manners. In this context, our primary interest is to understand if and when undesirable interactions among users could occur, and to devise mechanisms to make the service models robust to such interactions. The best effort service model demands that the majority of users comply with the standard TCP behavior. Non-conformant flows, when multiplexed with TCP traffic, can deprive TCP users from attaining their target resource allocation. Our main goal in this environment, is to design efficient defense schemes to protect TCP users against other forms of non-conformant traffic. Our investigation identifies that existing defense solutions perform well only in the presence of brute-force malicious behaviors, and can be defeated by more sophisticated attacking schemes. Based on such findings, we prototyped a new double horizon defense scheme, which relies on mechanisms at two different time scales (horizons) to regulate traffic. Evaluation results suggest that the implementation cost of this design is reasonable, and that it is highly effective against various types of malicious traffic while remaining transparent to TCP users. Our perspective in aggregate QoS environments is somewhat different. An aggregate QoS model explicitly allocates resources to enforce a target performance. This allocation is based on the observed overall performance of the service class, but not on the performance of individual users. To understand if and when varying traffic patterns across users can cause some of them to experience worse-than-average performance, we developed a number of explicit analytical models to evaluate the performance of individual users. This enables us to identify a set of typical cases where severe performance deviations occur, and come up with guidelines to avoid such scenarios. The resulting guidelines are subsequently incorporated into a simple tool for characterizing dangerous mixes of user traffic in practical scenarios involving QoS-sensitive traffic such as voice and video