Summary: | 博士 === 國立交通大學 === 資訊工程系 === 90 === Nowadays the Internet only provides one service class, the best-effort service, which does not guarantee any timeliness or transmission rate. With the transition to a commercial infrastructure, there is an increasing need to provide better service quality in the Internet.
In this dissertation, we present four efficient approaches for supporting scalable quality of service (QoS) over the Internet. The proposed approaches aim at achieving weighted fair rate allocations. More specifically, each flow is assigned a weight which determines the service rate it will receive. The proposed approaches follow the design philosophy of the differentiated services (Diffserv) model; that is, keeping the core network as simple as possible for good scalability.
The first approach addresses QoS issues from the aspect of end-hosts. Similar to the current Internet architecture, the traffic control is mainly accomplished through end-host algorithms. Routers in the network do not have to provide any particular support for the service quality. Instead of using a reactive flow control scheme commonly used in end-to-end protocols, the proposed approach uses a proactive flow control scheme. By keeping a certain amount of extra packets in proportion to its weight for each flow in the network, the proposed approach is able to achieve weighted fair rate allocations with a simple and efficient distributed algorithm. We evaluate the performance of the proposed approach through both simulations and experiments on Linux.
The main advantage of the first approach is the simplicity, which eases the deployment. However, since it is difficult to ask every user to follow the same flow control rule, the first approach cannot guarantee the throughput of a well-behaved user if any ill-behaved traffic is present. Therefore, we extend the first approach to an edge-to-edge flow control algorithm, in which the packets of each flow are queued at ingress edge routers before they can be forwarded to the core network. In contrast to the window-based flow control used in the first approach, rate-based flow control is used in the second one. With the proposed approach, core routers do not have to provide any particular supports. Therefore, this approach can be considered as an extreme case for realization of the Diffserv model.
In the third approach, we propose the so-called Multi-Level Fair Queueing (MLFQ) algorithm, in which both edge routers and core routers are required to be modified. As compared with existing approaches, such as Core-Stateless Fair Queueing (CSFQ) and Rainbow Fair Queueing (RFQ), MLFQ achieves better fairness of rate allocations and higher throughput. Moreover, it supports layer-encoded applications. In particular, it does not incur extra implementation complexity.
Finally, a closed-loop flow control based on the idea of virtual flow is proposed. This approach avoids the potential bandwidth waste in Fair Queueing algorithms. Through simulations, we show that it significantly improves the network throughput. We also give an analytical argument for the convergence of the proposed approach. For a single bottleneck configuration, we prove that the system is guaranteed to achieve weighted fair rate allocations in O(log N) cycles, where N is the number of flows.
|