Characterization of BitTorrent Swarms and their Distribution in the - - PowerPoint PPT Presentation

Institute of Computer Science Chair of Communication Networks

• Prof. Dr.-Ing. P. Tran-Gia

Characterization of BitTorrent Swarms and their Distribution in the Internet

Tobias Hoßfeld, Frank Lehrieder, David Hock, Simon Oechsner University of Würzburg, Germany Zoran Despotovic, Wolfgang Kellerer, Maximilian Michel DoCoMo Communication Laboratories Europe GmbH, Germany

Characterization of BitTorrent Swarms and their Distribution in the Internet Frank Lehrieder 2

Agenda

 Introduction

• BitTorrent-like P2P networks
• Aim: characterization of real-life BitTorrent swarms
• Methodology and data sets

 Measurements

• Swarm sizes
• AS clustering of peers
• Traffic of BitTorrent swarms

 Characterization: distribution of peers over ASs  Conclusion

Characterization of BitTorrent Swarms and their Distribution in the Internet Frank Lehrieder 3

BitTorrent-like P2P Networks

 In wide use for user-assisted content distribution, mostly file-sharing  Responsible for a large fraction of today’s traffic in the Internet  Example network:

Tracker: Index server, knows addresses

• f all peers in the swarm

Swarm: Set of all peers exchanging the same file Seed: Peer which has the complete file, uploads only Leecher: Peer which does not have the complete file, uploads and downloads data Transfer of data chunks: File is divided in chunks of 512 KB

Characterization of BitTorrent Swarms and their Distribution in the Internet Frank Lehrieder 4

 Major research topic:

Application layer traffic optimization (ALTO) for BitTorrent networks

 Performance evaluation difficult

• Crucial impact of evaluation scenarios
• Slightly modified mechanisms

lead to different results

 What is the nature of real-life BitTorrent swarms in the Internet?

• Distribution of peers over swarms
• Distribution of peers over ASs
• Exploitation potential for ALTO mechanisms
• Time dynamics, file sizes, content, …

Aim: Characterization of Real-Life BitTorrent Swarms

Autonomous Systems (AS) Peers

“the Internet”

Characterization of BitTorrent Swarms and their Distribution in the Internet Frank Lehrieder 5

Available Data Sets

Characterization of BitTorrent Swarms and their Distribution in the Internet Frank Lehrieder 6

Swarm Sizes

 Almost all swarms have less than 100 peers (exception: Pop.)  Maximum swarm sizes are by far larger than the mean value  The fraction of swarms containing 80% of the peers (p80) is

roughly 0.2 for most of the data sets.

ID Mean Max. p80 Mov. 25.46 20079 0.13 TV. 15.53 7276 0.17 Mus. 9.76 3813 0.25 KPi. 11.12 72988 0.18 KMi. 6.99 763 0.45 KDe. 9.73 1883 0.31 Pop. 691.14 30961 0.45 24h. 146.68 19748 0.12

Mean and max. nr. of peers/swarm, fraction

• f swarms containing 80% of all peers (p80)

Characterization of BitTorrent Swarms and their Distribution in the Internet Frank Lehrieder 7

AS Clustering of Peers

 To which degree are peers of a swarm clustered in their ASs?

• “-clustered” peers have at least (-1) other peers in the same AS
• AS clustering  of swarm s:

 = #(-clustered peers) / swarm size

 Most swarms have a very low

fraction or even no peers at all clustered in their ASs

• Only 4% of the music

swarms have an AS with 5 or more peers

• Only 12% of the movie

swarms have an AS with 5 or more peers

3 = 3/5 4 = 0

Example swarm

Characterization of BitTorrent Swarms and their Distribution in the Internet Frank Lehrieder 8

Traffic of BitTorrent Swarms

 Two simple approximations for the traffic of a swarm

• “with file sizes”: traffic is proportional to (swarm size * file size)
• “w/o file sizes”: traffic of a swarm is proportional to swarm size

 80-90% of the traffic are

• wed to 20% of the swarms

(pareto principle)

 “Potentially local traffic”=

traffic of a swarm * 2

 ALTO mechanisms useful

• nly in the top 20% of the

swarms

Characterization of BitTorrent Swarms and their Distribution in the Internet Frank Lehrieder 9

Characterizing the Distribution of Peers over ASs

 Intention

• Input for performance evaluation
• Real-life distribution of peers
• ver ASs within a swarm

 For every swarms s spread over n ASs

• Assign AS ids k{1,…,n} to ASs

with decreasing nr. of peers

• Fs(k): fraction of peers in s that belong to AS with id k

 Average Fs(k) of all swarms s: F(k) (=> dark blue bars)  Fit F(k) with power-law function: P(k)=a/kb + c (=> red curve)

Characterization of BitTorrent Swarms and their Distribution in the Internet Frank Lehrieder 10

Conclusion

 Measurement study comprises swarms of

• Different index servers (piratebay, mininova, demonoid)
• Different types of content (music, movies, regional content)

 Measurement results

• Most swarms are small and cannot use ALTO mechanisms
• Most traffic (80-90%) produced by a few large swarms
• ALTO mechanisms have a high potential in these swarms
• Further results: regional swarms, timely dynamics, distribution of

peers over countries, number of peers vs. AS degree

 Characterizations of BitTorrent swarms for performance evaluations

• Distribution of peers over ASs within a swarm
• Further char’s: file sizes, number of peers, and top AS fraction

Characterization of BitTorrent Swarms and their Distribution in the Internet Frank Lehrieder 11

BACKUP

Characterization of BitTorrent Swarms and their Distribution in the Internet Frank Lehrieder 12

Distribution of Peers over ASs and Countries

 Average number of peers per AS is very small (<5) for most

swarms

 Maximum number of peers per AS is still quite small  AS affiliation not the only metric: country codes (MaxMind GeoIP)

Characterization of BitTorrent Swarms and their Distribution in the Internet Frank Lehrieder 13

Peculiarities of “Regional” Swarms

 16 example swarms considered  Swarm sharing regional content

• Spread over less ASs
• Higher top AS fraction

 Calculate distribution of peers

• ver ASs for every swarm

 Determine kurtosis of this distr.  Higher kurtosis for regional

swarms (due to concentration in large ASs)