Shishir Nagaraja CSE 530
Acknowledgements
Slides used in today's class are due to Prof. Jian
Surveyed papers
• Service Capacity of Peer to Peer Networks (INFOCOM 2004) X. Yang, G. Veciana • The Bittorrent P2P File-sharing System: Measurements and Analysis (IPTPS 2005) J. Pouwelse, P.Garbacki, D. Epema, H. Sips • Measurements, Analysis, and Modeling of BitTorrent-like Systems (IMC 2005) L. Guo, S. Chen, Z. Xiao, E. Tan, X. Ding, and X. Zhang • Analyzing and Improving BitTorrent Performance (INFOCOM 2006) A. Bharambe, C. Herley and V Padmanabhan
Requirements
• High availability – High download speeds • No fake files • Deal with flashcrowds
Overview – system components
Web Server Tracker
Web page with link to .torrent .torrent files are distributed
t
n
e r r
o t .
C A Peer Peer [Seed] B [Leech] Downloader Peer
“US” [Leech] Overview – system components
Web Server Tracker
Web page with link to .torrent
e nc ou nn t-a Ge
C A Peer Peer [Seed] B [Leech] Downloader Peer
“US” [Leech] Overview – system components
Web Server Tracker
Web page with link to .torrent
st r li ee -p se on sp Re C A Peer Peer [Seed] B [Leech] Downloader Peer
“US” [Leech] Overview – system components
Web Server Tracker
Web page with link to .torrent
Shake-hand C A
S Peer ha ke Peer -ha [Seed] nd B [Leech] Downloader Peer
“US” [Leech] Overview – system components
Web Server Tracker
Web page with link to .torrent
pieces C A p iec es Peer Peer [Seed] B [Leech] Downloader Peer
“US” [Leech] Overview – system components
Web Server Tracker
Web page with link to .torrent
pieces C A p iec es Peer p iec Peer es [Seed] B [Leech] Downloader Peer
“US” [Leech] Overview – system components
Web Server Tracker
Web page Global registry of all with link downloaders and to .torrent seeds of a file
ce un no n st t-a r li Ge ee -p se on sp es Re piec C A p iec es Peer p iec Peer es [Seed] B [Leech] Downloader Peer
“US” [Leech] Overview – BT features
• Temporal one torrent community (no across torrents communication).
• Thousands of chunks per file – depends on the file size
• File/pieces/blocks. Content trade is piece based, while transfer is block based. Partial pieces can not be uploaded.
Overview – BT features
• Peer selection- Tit-for-Tat (download speed based)
• Central website uses mirroring to balance http access load • Further, .torrent meta-data files are distributed on .torrent servers
BT features contd.
• Moderation of content (pollution check) : – Moderated users – Unmoderated users – Moderators – all submissions are inspected; ensures integrity but introduces centralization
Research focus
• Generic network performance – Traffic volume, download/upload throughput, uplink utilization, peer behavior, service capacity, • Incentives – Fairness, peer game • Robustness – Website availability, peer workload, content integrity. • Limitation and improvement
Methodology
• Passive ways – (tracker log, torrent website monitor, traffic capture) • Active ways – Join the BT network to monitor peer performance • Simulation • Modeling
Interesting results
• YV_INFOCOM04 – Exponential growth of service capacity in initial flash crowd
– For less popular torrent (<50 peers), the performance is quite unpredictable
Interesting results
• PG_IPTPS05 – The arrival processes of peers are not Poisson as assumed by some papers (sigcomm04).
– Power law relationship btw peer speed and # peers
Interesting results
• PG_IPTPS05 – The availability of centralized supporting system is bad
– # of seeds after 10 days is not accurate predictor for content life time, one single seed can be up for long period of time.
Interesting results
• PGES_IPTPS05 – Centralized moderation seems very effective in removing fake and corrupted files. 5k of 8k torrent uploaded by 20 moderators daily.
– Donated resource for meta-data entail integrity and privacy risks without control.
Interesting results
• GCXTDZ_IMC05 – Exponentially decreasing peer arrival rate, lack of seeds make service availability poor quickly after flash crowd.
– The leecher download speed at different time stages is highly diverse.
Interesting results
• GCXTDZ_IMC05 – BT can provide unfair services. Peer contribution ratio decrease with its downloading speed increase.
– The probability for a peer to abort downloading voluntarily is independent of its speed and downloading progress.
Interesting results
• GCXTDZ_IMC05 – The life spans of torrent is btw 30-300 hours with average 8.5 days; average population of a torrent is only about 102 peers.
– Seeding time is average as 8.42 hours
Interesting results
• BHP_INFOCOM06 – The uplink utilization close to 100%
– Workload of seeds increase slowly during flash crowd
Interesting results
• BHP_INFOCOM06 – Single seed utilization outperforms that of multi- seeds
– The more concurrent upload, the less uplink utilization
Interesting results
• BHP_INFOCOM06 – For low bandwidth peers with small neighbors, random block exchange outperform LRF
– But LRF outperform random block exchange in term of # interesting connections
Interesting results
• BHP_INFOCOM06 – Block-level TFT outperform download speed based TFT in term of fairness
– Quick BW estimation help archive optimal utilization
Interesting results
• BHP_INFOCOM06 – Bandwidth-matching tracker policy further improve both utilization and fairness
Argument
• INFOCOM06 : – the simulation does not consider user control of their uplink/downlink (BitComet enable uplink speed limits) – Tradeoff of the improvement ideas – Bandwidth measure overhead is ignored – Bandwidth clustering helps but is that violation of Internet ethics by separating different capable users?
Current BT development
• DHT enabled peer search (BitComet, Bitsprite) • Decentralized torrent search (eXeem) • BT like Content delivery system (VOD, IPTV)