Distributed flow control for next generation networks

In this work, we investigate the flow control problem for the next generation Internet. The flow control problem involves determining the rates at which traffic flows from various sources should be transmitted over the network. Providing efficient and distributed flow control mechanisms is very important in maximizing network throughput as well as minimizing network delays. Flow control is inherently difficult because we expect that networks will continue to become even larger and more complex (hence scalability becomes critical) and will need to support a variety of different applications with diverse performance requirements (all traffic may not even be controllable). For example; the network may send explicit data rate information to the source or may send simple binary-bit indications of the network status to the source. While the explicit-rate flow control model is widely used for theoretical analysis, in practice, it is the binary-bit scheme that has been widely deployed. One example of the binary-bit flow control is the famous Additive Increase Multiplicative Decrease (AIMD) that has been used in the most dominant protocol of the current Internet, TCP. In our work, we develop our theoretical results assuming an explicit-rate flow control model and then based on this analysis provide practical AIMD type of solutions for Internet. Unlike other works in the literature, we consider a network with both controllable and uncontrollable flows. We begin by analyzing a network with a single bottleneck link and then extend the analysis to a general network with multiple links. In contrast to other works, we focus on the queueing delays caused by the controller and provide guidelines on how to design an efficient flow control system based on rigorous queueing theoretic results. We believe that the queueing behavior is very important because it determines network performance. In fact the main objectives of flow control, i.e., high utilization, low loss rate, and fairness are all related to the queueing behavior within the network. Based on our theoretical results, we provide efficient distributed flow control mechanisms that can be implemented in the TCP context using Active Queue Management (AQM) schemes. We further analyze an AIMD network using general AQM schemes when the number of flows in the network is large. The purpose of this work is to clearly understand how AIMD can affect the queueing behavior within a network, and whether different AQM schemes can result in widely differing network performance