Dissertation Defense: Peng Yang

Abstract: The Internet has become increasingly heterogeneous as new technologies are deployed at various layers in the Internet protocol stack, from new peer-to-peer (P2P) technologies and cloud technologies in the application layer, to various transmission control protocols (TCP) in the transport layer, and to new wireless access technologies in the media access control and physical layers. Thus, the Internet has become very complex and very diversified in terms of formation, control and usage. The heterogeneous Internet greatly complicates the Internet traffic control. Firstly, two seemingly unrelated Internet traffic control mechanisms can step on each other in an unexpected way, which results in performance degradation for both mechanisms. Secondly, different entities perform the Internet traffic control mechanisms based on uncoordinated and inaccurate local observations, which may severely damage the interaction among the Internet traffic control mechanisms performed by different entities. Thirdly, the lack of deployment information of some Internet traffic control mechanisms can hinder us from optimally designing other Internet traffic control mechanisms, since they are dependent on each other. In this thesis, we tackle the complexity of the Internet traffic control for two important types of the Internet traffic: P2P traffic and TCP traffic generated by the P2P applications and the TCP protocols, respectively.

The P2P applications have been causing great trouble for the Internet for a long time. Firstly, P2P traffic control mechanisms may adversely step on the traffic management conducted by the Internet service providers (ISPs) and increase their operating cost. Secondly, the traffic generated by the P2P applications is not well coordinated and controlled, and thus results in performance degradation of the P2P applications. In this thesis, we propose two P2P application designs to reduce the operating cost of the ISPs, and to increase the performance of the P2P applications. The proposed designs require that the ISPs provide guidance information to the P2P traffic control. Thus, we refer to them as ISP-guided P2P application designs. The ISP guidance information helps the coordination between the P2P traffic control mechanisms and the ISP traffic management, and thus can reduce the operating cost of ISPs. The proposed designs also generate well-coordinated P2P traffic, and thus can increase the performance of the P2P applications.

Compared with the P2P applications, the TCP protocols have a more profound impact on the Internet. Decades ago, the Internet was dominated by one TCP protocol, called RENO. This situation has changed drastically recently, as dozens of new TCP protocols were proposed and deployed in the Internet. However, to date, we do not fully understand the impact of the heterogeneous TCP protocols on other Internet traffic control mechanisms and on the Internet as a whole. A fundamental reason is the lack of the deployment information of various TCP protocols. This motivated us to do the TCP protocol census, i.e., identifying the TCP protocols used by the Internet hosts. We conducted large-scale Internet experiments, and found that today’s Internet has already become highly heterogeneous in terms of TCP protocols. This result means that many Internet traffic control mechanisms designed for the RENO protocol need to be re-designed. Otherwise, they cannot operate optimally. We pick one such influential Internet traffic control mechanism, called router buffer sizing, which is to set an appropriate router buffer size so that the utilization of the Internet links and the performance of TCP flows are optimized. Different from many traditional and recent studies on the router buffer sizing problem that still assume that the Internet is dominated by the RENO protocol, we study the router buffer sizing problem under the heterogeneous TCP protocols. We found that those router buffer sizing rules are not applicable anymore. They either overprovison the router buffer and result in the bufferbloat problem where the excessive buffer lowers the performance of TCP flows, or cannot set the optimal router buffer size for the heterogeneous TCP protocols. To this end, we propose a new rule to set the router buffer size. Evaluations show that our rule can optimize the utilization of the Internet links and the performance of TCP flows under the heterogeneous TCP protocols in the current Internet.