Prediction of Swarm Popularity in Peer-to-Peer Networks

Abstract

Recent years have witnessed the great popularity of adopting BitTorrent(BT)- like peer-to-peer(P2P) designs in Internet Application. Different from the classic client-server solutions, P2P significantly improves the capacity, robustness, and the scalability of the system. However, it is known that P2P will also bring certain Quality-of-Service(QoS) issues. In particular, users’ service performance and availability can hardly be guaranteed. To make the matter worse, the users cannot even predict their service performance, such as the downloading rate and completion time, before joining the swarm. In this thesis, we aim to provide an option to help the users predict their service performance in BitTorrent-like P2P applications. Using the classic BT as a case study, we collect 1250 torrent files and carefully investigate their content, peer, and service performance details. Our measurement shows that the swarm popularity is related to the file type, file size, piece length, and its creation time. To better capture this relationship, we adopt different prediction methods to help the users better understand their possible downloading performance. Our evaluation shows that the neural-network-based approach has the best accuracy with minimal cost. In the future, we are aiming to further optimize the accuracy and implement the prototype of our approach in the existing open-source BT releases.