Latency-driven BitTorrent

Master's thesis

In recent years BitTorrent has become a notorious contributor to Internet traffic. Not only is BitTorrent responsible for over one third of all Internet traffic, but an immoderate amount of it is expensive cross-ISP or even inter-continental traffic. Much of BitTorrent's long-distance traffic is due to its random selection of peers, which can cause connected peers to be at very different locations. This causes inefficient network usage and harms client-perceived performance.

In this paper we present the design and evaluation of latency-driven BitTorrent, our approach to bias BitTorrent communication towards nearby peers. Unlike previous approaches we do this without requiring any additional infrastructure or patches to client software. A small number of cooperating BitTorrent trackers can easily deploy our system. Evaluating our approach through simulation and Planet-Lab deployment we find average and median reductions in download time of up to 25%. We find we can maintain or improve our gains in a rapidly growing network by controlling the peer sample size. We also show we reduce traffic cost with 12% less traffic going over global (transit) networks and more traffic going over short network paths.

MSc thesis

Zero-day reconciliation of BitTorrent users with their ISPs

BitTorrent users and consumer ISPs are often pictured as having opposite interests, with end-users aggressively trying to improve their download times, while ISPs throttle this traffic to reduce their costs. However, inefficiencies in both download time and quantity of long-distance traffic originate in BitTorrent randomly selecting peers to interact with. We show that biasing the link selection allows one to reduce both median download times by up to 32% and long-distance traffic by up to 16%. This optimization can be deployed by modifying only the BitTorrent trackers. No external infrastructure nor specialized client-side software deployment is necessary, thereby facilitating the adoption of our technique.

Software

• Tracker (messy, experimental version)
• Landmark
• Simulator
• Simulator dataset (200 peers, May 2008)