Fluid models for Internet design and simulation

Over the past decade, fluid models have been widely used, and shown to be efficient and accurate in the modeling, analysis, and design of the Internet. In literature, much of this work has focused on the design of end-host controllers and control algorithms at routers (marking functions) for the stable end-to-end operation over the Internet. However, there is a significant fraction of uncontrolled flows such as real-time video and audio flows in the Internet, and the effect of those uncontrolled flows on the design and simulation of the Internet cannot be ignored in the use of fluid models. In this thesis, we first explore the use of time-scale decomposition of the end-systems and queueing dynamics at the intermediate routers. Based on this time-scale decomposition study, for a queue-based router model we develop an equivalent fluid model that depends only on the instantaneous traffic rate. The main intuition for such a rate based model is that there exists a sufficient randomness in the Internet due to uncontrolled flows. We next study how the rate based model can be practically used for a Internet simulation/emulation with a mixture of controlled and uncontrolled flows. We address this by developing a hybrid network simulator – FluNet – which combines actual network hardware (routers and switches) at the network edge, and rate based fluid models within the network core. Second, we study the effects of these uncontrolled flows on the design choices for the end controllers and marking function. Current research has focused on the design of controllers and network algorithms with the objective of stability and convergence of the transmission rates. However, an important criterion that has not received much attention is the design of the controllers with the objective of providing QoS guarantees to real-time flows that share the links with the controlled flows. In this thesis, we study the design rules for network congestion controllers with the objective of providing QoS support for such real-time flows. Third, while much of the research based on fluid models has focused on wireline networks, a growing area of research is that of wireless multi-hop networks. The main issue in using fluid models in this context is that the MAC (media access control) is a “discrete” entity, as a result of which fluid models have not been commonly used. In this work, we first investigate fluid models for MAC and appropriate models for congestion control over multi-hop wireless networks. Based on an optimization framework with constraints that arise from the multi-hop wireless network, we propose hop-by-hop congestion control algorithms, and study their properties on the stability and peak buffer requirement. Fourth, we study a realistic MAC protocol, which leads to the appropriate fluid models used for analysis and design of networking algorithms (i.e., congestion control) over wireless multi-hop networks. In this work, we study a distributed randomized MAC protocol that converges an optimal schedule with a simple one-hop synchronized contention signaling mechanism. Furthermore, by simulation and analysis, we show that the proposed protocol adapts well to slowly-varying load/topology changes.