Experimental analyses of data distribution on data center networks

In recent years, operators of large data centers have begun to use BitTorrent to distribute files to large numbers of machines within the data center. The rationale for this trend is clear and well-motivated: BitTorrent is easy to use and enables the scalable distribution of large files to many machines. However, peer-to-peer data distribution applications like BitTorrent are highly configurable, and parameter tuning can have a substantive impact on performance and efficiency. In this work, we use an experimental approach to study the impact of peer-to-peer configuration choices in data center networks. This understanding enables us to improve the performance of Murder, a popular BitTorrent variant used in data centers, by a factor of 5. Moreover, we show that LANTorrent, an alternative tool that uses chain-based distribution rather than swarms, is 12 times faster than Murder.