Network support for global-scale peer-to-peer systems

Peer-to-peer systems research has recently attracted much attention because the peer-to-peer architecture can potentially allow powerful solutions to be deployed rapidly in the Internet. However, building high performance peer-to-peer systems faces two networking challenges that are rooted in fundamental limitations of the Internet architecture.; In this dissertation, we show that despite the fundamental limitations of the Internet architecture, end system-based algorithms and services can efficiently address the networking challenges in building global-scale peer-to-peer systems. To address the heterogeneous connectivity challenge, we have designed and implemented a connectivity service called AYES. AYES unifies the identification of heterogeneous Internet hosts by using domain names and enables universal connectivity by using waypoints to relay data traffic between heterogeneous networks. To address the performance heterogeneity challenge, a naive on-line network measurement-based approach is clearly unscalable. We propose to model the Internet network distance relationship as a geometric space using a technique called Global Network Positioning (GNP). By performing experiments on measured Internet network distance data, we have discovered that the Internet network distance relationship can be accurately modelled by a low-dimensional Euclidean space. In our experiments, 7 dimensions are sufficient to attain high accuracy. Moreover, very few network paths need to be measured to construct the model. Based on the finding of our GNP study, we have designed and implemented a Network Positioning System (NPS) that enables every Internet end host to compute its network position in the Euclidean space model distributedly. Using NPS positions, accurate estimates of network distances among hosts in peer-to-peer systems can be rapidly computed without measuring the network paths directly, therefore removing a major performance bottleneck in peer-to-peer systems. Finally, using NPS to rapidly select end hosts with promising network distance performance, light-weight measurement-based techniques such as round-trip time probing, 10KB TCP probing, and bottleneck bandwidth probing can potentially be used to identify high TCP throughput end hosts. Our analyses of the performance of a large number of Internet hosts show that these techniques perform quite well, are highly complementary to each other, and can efficiently guide the selection of a high TCP performance host in peer-to-peer systems. (Abstract shortened by UMI.)...