System/Application Designs, Optimizations and Implementations on Overlay Networks

An overlay network is a layer of virtual network topology on top of the physical network, which directly interfaces to users. With the rapid advancement of Internet and computing technology, much more aggregate information and computing resources are available from clients or peers than from a limited number of centralized servers. Overlay networks provide us with the following advantages and opportunities to better utilize the increasingly growing Internet information and resources. (1) Overlay networks allow both networking developers and application users to easily design and implement their own communication environment and protocols on top of the Internet, such as data routing and file sharing management. (2) Data routing in overlay networks can be very flexible, quickly detecting and avoiding network congestions by adaptively selecting paths based on different metrics, such as probed latency. (3) The end-nodes in overlay networks are highly connected to each other due to flexible routing. As long as the physical network connections exist, one end-node can always communicate to another end-node via overlay networks. Thus, scalability and robustness in overlay networks are two attractive features. (4) The high connectivity of increasingly more end-nodes to join overlay networks enables effective sharing of a huge amount of information and resources available in the Internet.

Typical overlay networks include multicast overlays, peer-to-peer overlays (e.g. Gnutella and Kazaa), parallel file downloading overlays (e.g. BitTorrent and eDonkey), routing overlays (e.g. skype for VoIP).

Overlay networks also create several challenges and problems for us to do research. First, overlay networks not only have no controls of physical networks, but also lack critical physical network information. Second, because of the indirect or even mis-communications between overlay and underlay networks, in practice, inefficient usage network resources are quite often in many overlay applications, such as mismatch between overlay and underlay topology, inaccurate probing results among end-to-end nodes due to network dynamics, generating a large amount of redundant messages, and others. Third, since the overlay networks are open to all kinds of Internet users, security and privacy issues can be quite serious. Fourth, overlay networks are highly decentralized, thus they are likely to have weak ability for resource coordinations. Finally, fairness of resource sharing and collaborations among end-nodes in overlay networks are two critical issues that have not been well addressed.

We are conducting research to address several issues of performance, reliability, privacy, and coordinations in overlay networks. We focus on structured and unstructured P2P overlays, routing overlays for VoIP, and parallel file downloading overlays.


Routing Overlays for VoIP

VoIP applications on overlays have become a common and cost-effective practice for many users. Having conducted intensive experiments to measure Skype's routing patterns, we conclude that (1) peer relays can significntly improve the VoIP quality in overlay networks; and (2) the peer node selection method for VoIP relays in skype is not efficient, lacking critical routing information of underlay networks. We propose an AS-aware overlay routing protocol, called ASAP, to optimize the VoIP performance and to improve Internet bandwidth utilization.

Representative Publications:


Measurement, Modeling, and Improvement of BitTorrent Overlays

Our measurements and analysis of BitTorrent traffic provide several new findings regarding the limitations of BitTorrent systems: (1) Due to the exponentially decreasing peer arrival rate in reality, service availability in such systems becomes poor quickly, after which it is difficult for the file to be located and downloaded. (2) Client performance in the BitTorrent-like systems is unstable, and fluctuates widely with the peer population. (3) Existing systems could provide unfair services to peers, where peers with high downloading speed tend to download more and upload less. Motivated by our findings, we further build a graph based multi-torrent model to study inter-torrent collaboration. Our model quantitatively provides strong motivation for inter-torrent collaboration instead of directly stimulating seeds to stay longer. We also propose a system to show the feasibility of multi-torrent collaboration.

Representative Publications:


Efficient File Searching in Unstructured P2P Overlays

Flooding is a fundamental search method in unstructured P2P systems. It has the advantage of low latency and high resilience. However, the searching cost is high by generating a huge amount of network traffic. We have prposed several efficient searching methods with low cost, while retaining the advantage of flooding.

Representative Publications:


Anonymity Support in P2P Overlays

We have proposed several anonymity protocols in P2P overlays to protect peers' privacy. Some protocols also utilize a third party, such as a supernode or a proxy, which will also ensure some security concerns in P2P communications.

Representative Publications:


Topology-aware Routing and Data Consistency in P2P Overlays

We address the problem of topology mismatching between the P2P overlay and physical networks by proposing several efficient and low cost methods to adaptively construct overlay networks aiming at approximately matching the networks at the two layers.

Representative Publications:


Building Applications on Overlay Networks and Applying P2P Principles

We have proposed a P2P assisted proxy for delivering streaming media. With a support of a DHT (structured P2P), the streaming proxy becomes scalable and its delivery quality is high and the cost is low. Another proposal is to build a browser-aware Internet caching system to significantly enlarge the effective proxy size, and to improve its effectiveness.

We have also applied a P2P collaboration principle to improve wireless network performance and energy efficiency.

Representative Publications: