Latency Prediction for P2P Overlays

Peer-to-peer (P2P) applications have become countless since they were initially proposed about ten years ago. They became popular among the general public when being used as a platform for file sharing. However, their robustness and scalability sparked interest from industry and research. Today these concepts drive some of the main businesses as back ends or migrate intensive workload away from them. Recently, Microsoft added P2P capabilities to their update mechanism. Adobe provides P2P APIs in their Flash product. Amazon’s Dynamo and Skype are two examples of commercial success in this area. One popular non-commercial example is the BitTorrent protocol. P2P is a movement “back to the roots” of the Internet. The early ARPANET was a system of equal computers providing services to each other. However, since that time the core assumptions have changed. In the first place, the Internet has changed dramatically in size. With the introduction of commercial Internet Service Providers in 1989 it became apparent that the Internet’s participants were no longer equally capable: Dedicated bandwidth-rich machines provided services to customer machines, connected via narrow-band dial-up links. Henceforth, the original P2P architecture was no longer applicable and the classical Client/Server architecture became the predominant model. With the rapid growth of the Internet even Client/Server suffered scalability issues. At the same time the consumer Internet access saw huge changes. Persistent dialup connections became common along with an increase of the offered bandwidth. Furthermore, academia introduced resilient algorithms that finally allowed the selforganisation of a large number of hosts. Hence, a “Renaissance” of the P2P paradigm began. Today Client/Server is still the predominant paradigm. Meanwhile, P2P has proven its capabilities and found adoption. Today’s P2P systems often avoid fully decentralized solutions and rather employ a hybrid approach. They employ central components to facilitate tasks such as a user management or billing. This central point creates a single point of failure, as illustrated during the recent Skype downtime of 2010. It may also constitute a bottleneck. Hence, today’s systems often use federated systems, for example, in cloud computing data centers. Despite P2P’s success, there exist only few software frameworks that support the development of truly distributed applications, such as FreePastry. Others such as OpenDHT are not maintained anymore. Often the P2P technique is tied to monolithic applications, such as the eMule client and is thus not reusable.