End-to-end techniques for network resource management
Paul Francis, Manpreet Singh · 2006
Network-based applications have evolved from simple web browsing to complex commercial transactions. This evolution has made a clear case for network QoS mechanisms, since the delays seen on the network often form the major component of an end-user's perceived delay. The key challenge in providing network QoS is the distributed and decentralized nature of the network resource. In this thesis, we have attempted to develop techniques for exposing the network as a managed resource to applications without requiring any explicit support from the network elements like routers, gateways, ISPs, etc. As part of this broad initiative, we have developed two key components: MPAT (used to control the network from end-hosts), and Netmapper (used to monitor the network from end-hosts). MPAT is the first truly scalable algorithm for providing differential services to a group of TCP flows that share the same bottleneck link. The challenge is to hold the cumulative fair share of all flows, while being fair to the background traffic. Our system can give more bandwidth to one TCP flow at the expense of lower bandwidth to another flow, in such a way that the sum total of the bandwidth that the two flows get is same as the fair share of two TCP flows. Using experiments on the real Internet, we show that our system can share congestion state across more than 100 TCP flows with throughput differentials of 95:1. We also developed a network mapping and annotation service (called Netmapper) for distributed applications sensitive to bandwidth availability, latency, or loss. The system exposes internal network state (e.g. bandwidth available at links, location of bottleneck links) using only end-to-end measurements that are TCP-friendly. Our novel scheme to identify the location of bottleneck links improves upon existing methods for bottleneck identification, which detect incorrect bottleneck link if the bottleneck is due to transient network congestion. We built Netmapper, and show, using experiments on the real Internet, that the system consumes very small amount of bandwidth, and converges even with dynamically varying network state.