Distributed Peer Discovery in Large-Scale P2P Streaming Systems : Addressing Practical Problems of P2P Deployments on the Open Internet

Sammanfattning: Peer-to-peer (P2P) techniques allow users with limited resources to distribute content to a potentially large audience by turning passive clients into peers. Peers can self-organize to distribute content to each other, increasing the scalability of the system and decreasing the publisher’s costs, compared to a publisher distributing the data himself using a content delivery network (CDN) or his own servers.Peer discovery is the mechanism that peers use to find each other. Peer discovery is a critical component of any P2P-based system, because P2P networks are dynamic by nature. That is, peers constantly join and leave the network and each individual peer is assumed to be unreliable. This thesis addresses practical issues in distributed peer discovery mech- anisms in the context of three different large-scale P2P streaming systems: a (1) BitTorrent-based streaming system, (2) Spotify, and (3) our own mobile P2P streaming system based on the upcoming Peer-to-peer Streaming Protocol (PPSP) Internet standard.We dramatically improve peer discovery performance in BitTorrent’s Mainline DHT, the largest distributed hash table (DHT) overlay on the open Internet. Our implementation’s median lookup latency is an order of magnitude lower than the best performing measurement reported in the literature and does not exhibit a long tail of high-latency lookups, which is critical for P2P streaming applications.We have achieved these results by studying how connectivity artifacts on the underlying network —probably caused by network address translation (NAT) gateways— affect the DHT overlay. Our measurements of more than three million nodes reveal that connectivity artifacts are widespread and can severely degrade DHT performance.This thesis also addresses the practical issues of integrating mobile devices into P2P streaming systems. In particular, we enable P2P on Spotify’s Android app, study how distributed peer discovery affects energy consumption, and implement and evaluate backwards-compatible modifications which dramatically reduce energy consumption on 3G.Then, we build the first complete system that not only is capable of streaming content to mobile devices but also allows them to publish content directly into the P2P system, even when they are behind a NAT gateway, with minimal impact on their battery and data usage.While our preferred approach is implementing backwards-compatible modifications, we also propose and analyze backwards-incompatible ones. The former allow us to evaluate them in the existing large-scale systems and allow developers to deploy our modifications into the actual system. The latter free us to propose deeper changes. In particular, we propose (1) a DHT-based peer discovery mechanism that improves scalability and introduces localityawareness, and (2) modifications on Spotify’s gossip-like peer discovery to better accommodate mobile devices