Computer Science

Oral Comprehensive Defense Details

Investigating the Mutual Impact of the P2P Overlay and the AS-level Underlay

Abstract

In parallel with the growth of the Internet's infrastructure, peer-to-peer class of applications emerged, in which participating users (peers) connect together and form an overlay network to assist each other towards a common goal. P2P concept brings about a self-scalable, low-cost, user-centric structure for a variety of applications, mainly content distribution. P2P applications quickly became popular enough to be responsible for up to 70\% of the Internet traffic according to some reports. Their popularity and unique features make the study of P2P applications an important research subject. Capturing perfectly accurate and complete pictures of these systems is infeasible and therefore the researchers have tried to provide representative pictures to study the features and shortcomings of different P2P applications. The increasing amount and irregular pattern of traffic generated by this class of applications have led to a growing concern among ISPs who carried the traffic relayed by their customers running P2P applications. To control the increasing traffic cost, in many cases ISPs started throttling or even blocking P2P traffic while the P2P developers and the users have been trying to evade the imposed limitations. Several research works have targeted the problem of making P2P applications ISP-friendly by trying to reduce the costly traffic via reshaping the overlay. However, no previous work has provided a global picture of the impact (imposed traffic pattern) of the P2P applications on the network. The key question that is the core of this dissertation is the following: What is the impact of a P2P overlay on the underlying network? We try to answer this key question in three parts: (i) Measurement and characterization of P2P/overlay applications, (ii) Capturing and characterizing the Internet?s AS-level topology, and (iii) Characterizing the impact of a P2P/overlay application on the underlying AS-level network. First we monitor the popular Gnutella P2P application over a 15-month period in which the user population quadrupled, capturing the main peer-level properties and their evolution. Next, in order to capture the peer properties of very large scale P2P networks, we propose RDS sampling technique, as an alternative to taking full snapshots. After investigating the overlay and peer properties, we focus on the data exchange in P2P applications and we perform a case study on BitTorrent in which we use BitTorrent tracker logs to capture peer performance and find the root causes of the observed performance for each individual peer. Next, we switch to AS-level topology and we propose a novel technique for capturing geographic footprint of ASes by mapping the location of their users participating in popular P2P applications. We use KDE technique to provide an estimate of the user density, and we use the derived user density to locate likely PoP locations of each AS. Finally, we propose a method for characterizing the traffic load of a P2P application on the AS-level underlay. We capture Gnutella snapshots using the tools and techniques presented earlier, and group the participating peers in each snapshot according to their corresponding ASes. We perform BGP simulation, in order to capture paths connecting each pair of ASes and then calculate the traffic load on each AS and each inter-AS link. This dissertation includes my previously published and my co-authored material.