A Study Of Utility Maximization Based Network Bandwidth Allocation

The resource allocation is a key issue in every distributed system. In such systems, the overall amount of resource is always less than its demand, thus resource sharing is evitable. Sharing means the resource should be allocated. How to allocate resource reasonably is the primary problem which will be addressed in this thesis.The thesis will discuss the problem of utility-maximization based bandwidth allocation. The content includes resource allocation model in microeconomics, relevant microeconomical concepts and interpretation of various TCP protocols and AQM algorithms in the context of microeconomics. Next, we'll study in detail two variants of classic utility-maximization based bandwidth allocation model.Firstly, a router level bandwidth allocation model is proposed by combining the theory of generalized matrix inverse and utility-maximization based bandwidth allocation. In this model, the bandwidth allocation of each flow is determined at the routers, TCP protocols do not need to adjust each flow's bandwidth allocation. In addition the model has better ability to avoid the network being in a congested state and can isolate malicious flows and non-elastic flows to protect various elastic flows. In the thesis, we'll prove that this model is equivalent to the classic utility-maximization based bandwidth allocation model in that they all aim at maximizing the aggregate utility and the aggregate utility and the allocation are the same under these two models.Secondly, the thesis will study an extension model of the utility-maximization based bandwidth allocation. The aim is to improve the fairness of bandwidth allocation of TCP protocols; especially the bais of TCP Reno against long round-trip-time flows. In this model, every link in the network is associated with a fairness index constraint. The goal is to limit the differences among the bandwidth allocation of the flows traversing the same link. We'll study the fairness-efficiency tradeoff in this model and propose solution to attain good fairness and efficiency performance. Also a distributed algorithm that realizes this extension model is proposed. Based on these theoretical results, the model is implemented in the environment of TCP Reno-RED feedback system by using the technique of differential dropping. Simulation results show that using this extended model, the bias of TCP Reno against the long round-trip-time flows is relieved to a great extent and the resulting bandwidth allocation is more reasonable.