A group of researchers have developed a system to make accurate predictions on the network performance between two arbitrary computers on the Internet. In the paper (Google HTML version), they define iPlane, a system that, in addition to doing simple tests on round-trip latency and throughput, accounts for the actual network topology. Such a system has many possible uses for P2P networks and distributed services. For example, the Coral Content Distribution Network could use this data to make sure that each client trying to use the service gets routed to the node that gives the best connection. One of the people involved with the paper, Tomas Isdal, wrote his master’s thesis on the use of the iPlane system to enhance the performance of BitTorrent. Normally, a BitTorrent tracker will only return a random subset of the peers it knows about when queried by a peer. The modified version returned a list of peers that was half chosen at random and half peers thought to provide the best throughput by the iPlane system. This resulted in a dramatic increase in performance: the time it took for 80% of the peers to complete their downloads was reduced by about 20-25%.

This system relates to the graph theory we have covered in class because it involves identifying clustering in a (very large) graph of interconnected routers and clients/servers. Such knowledge could be very useful on a network like Cornell’s because of the incredible amount of intra-campus bandwidth available compared to the connections to outside campus. If Cornell was running a Coral node, and many people on campus were trying to acquire a file using the Coral network, they would all be directed to the one located on campus, which would fetch the file only once from the greater internet and then serve it to all the computers on campus that had requested it.

BitTorrent itself also constitutes a kind of trading network. The rules are less rigid, but the algorithms used to choose who to upload to depend heavily on who is most willing to upload to the peer. In essence, while nodes are not restricted nearly as much in the number of nodes they can trade with and “freebies” are occasionally given out, nodes will consistently choose to trade with the node that gives them the best deal: in this case, the fastest download.