1 / 33

BitTube : Case Study of a Web-based Peer-Assisted Video-on-Demand ( VoD ) System

BitTube : Case Study of a Web-based Peer-Assisted Video-on-Demand ( VoD ) System. Bo Liu, Bin Chang, Ben Gotow, Yi Cui, Yuan Xue V anderbilt A dvanced Net work and S ystems Group Vanderbilt University, USA Presenter: Bin Chang, YouTube Inc. Outline. Motivation Internet Video

Olivia
Télécharger la présentation

BitTube : Case Study of a Web-based Peer-Assisted Video-on-Demand ( VoD ) System

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. BitTube: Case Study of a Web-based Peer-Assisted Video-on-Demand (VoD) System Bo Liu, Bin Chang, Ben Gotow, Yi Cui, Yuan Xue Vanderbilt Advanced Network and Systems Group Vanderbilt University, USA Presenter: Bin Chang, YouTube Inc.

  2. Outline • Motivation • Internet Video • Can P2P Help? • On-demand P2P Streaming • BitTube Design • BitTube Implementation • Results

  3. Motivation • Internet Video • Video streams served increased 38.8% in 2006 to 24.92 billion (Source: AccuStream iMedia Research) • 53 web-video startup in 2006 (US), $521M VC funding (Source: DowJones VentureOne) • Major studio goes into Web Video • $410M video ads revenue in 2006 and grow by 89% in 2007 (0.6% of $74B TV ad market, Internet ads $16.4B in 2006, expect to grow 19% in 2007) [source: Emarketer] • Operating Cost • Hosting, storage, infrastructure, management… • Bandwidth Subscription

  4. Can P2P Help? • Successful Track Record in Broadcasting and Live Streaming • Challenges in On-demand Streaming • Peer Aggregation: a number of peers access the same media file simultaneously • Key Factors Enabling Peer Aggregation • High Access Rate • Skewed File Popularity • Short Inter-arrival Time • Concentrated Requests • Long Online Duration • Sustained Peer Mass

  5. On-demand P2P Streaming • Theoretical Study Request interarrival time = Buffer Length Request Interarrival Time = Buffer Length  request rate T Movie Length S Playback Time W Buffer Length Sequential Access Random Access Yi Cui, Baochun Li, and Klara Nahrstedt, oStream: Asynchronous Streaming Multicast in Application-Layer Overlay Networks, IEEE Journal on Selected Areas in Communications, Special Issue on Recent Advances in Service Overlays, 2004.

  6. On-demand P2P Streaming • Trace Study • Traces from the on-demand service of MSN Video • 9-month period: April – December 2006 • 520M streaming requests • 59,000 unique videos • With the aid of P2P streaming (greedy prefetch policy) • Server Bandwidth Reduction from 2.20Gbps to 79 Mbps on December 2006 • Studio Produced Content • Bitrate: 200 kbps in April 2006, 320 kbps in December 2006 • Duration: 5 to 60 minutes Cheng Huang, Jin Li, and Keith Ross, Can Internet Video-on-Demand Be Profitable?, ACM SIGCOMM, August 2007, Kyoto, Japan

  7. Peer-Assisted On-demand P2P Streaming • P2P solutions can greatly reduce server load, hence the bandwidth subscription budget • Proved by theoretical analysis and trace study • Commercial Deployment: PPLive • Server load reduction 91.7% on May 12, 2008 • However, P2P cannot replace server, but only complement it • Skewed Video Popularity • Unpopular videos are more likely to be only possessed by the server • Peer Population Fluctuation • At the beginning, a new video is only available the server, then distributed to peers one by one • Alternative Viewpoint • Server can be regarded as super peer or seed

  8. Outline • Motivation • BitTube Design • Design Rationales • Architecture • User Request Flow • BitTube Implemetation • Results

  9. Design Rationales • Minimum Interruption to Current Internet Video Infrastructure • Web-based platform should stay uninterrupted • Massive Video Archive • Adopt the most commonly-accepted P2P Technique • Our choice: BitTorrent • Retain usability of the current Internet video service • Click-to-view

  10. BitTube Architecture User Server Request Web Browser Video Server Flash Player video video Request BitTube Desktop Other Peers video Local HTTP Module Tracker Downloading Stub P2P Module Client/Server Module BitTorrent Tracker Peer List

  11. Request and Data Flow Other Peers Request is formatted as http://localhost/ video URL/ torrent URL Downloading stub peer list Web Browser Local HTTP Module torrent URL P2P Module BitTorrent Tracker peer list request video video URL Flash Player Client/Server Module Video Server piece request Local HTTP process reassembles the collected pieces into the complete video file HTTP/1.1 content-range entity-header feature enables it to request a particular piece of the file

  12. Service Confederation • Required changes to existing video service • Video URL Modification • Torrent File Creation P2P Network BitTorrent Tracker

  13. Screen Shot

  14. Outline • Motivation • BitTube Design • BitTube Implementation • BitTorrent Overview • From BitTorrent to BitTube • Locality-awareness • Window-based Approach • Results

  15. BitTorrent Protocol • De facto Protocol for P2P File Downloading • First software BitTorrent • Author: Bram Cohen • Development Time: Summer 2002 • Released as open source • Many BitTorrent-compliant Softwares • Writing BitTorrent-compliant Software • Stable open source code • Interoperability with existing P2P Softwares • No overlay structure needed • Highly dynamic P2P group • Many Abandoned Sessions • Distribution of small files

  16. BitTorrent Overview • A File is divided into pieces • Torrent • Tracker • Swarm • Key Functions • Tracker • Neighbor selection • Choker • Optimistic Unchoking • Piece Picker • Rarest-First Policy

  17. BitTorrent: Piece Picker Downloaded piece • Rarest-first Policy • Executed among unchoked peers • Piece with the minimum interest value is chosen • Interest value: the number of peers possessing this piece • Tie is broken arbitrarily if multiple pieces with the minimum interest value exist • Distribute the rarest pieces • Help promote the piece diversity of all peers Video File 1 2 1 3 4 2 1 3 2 Undownloaded piece Piece chosen by the policy 1 Interest value

  18. BitTorrent to BitTube • BitTorrent Limitations • Lack of Locality-awareness • Excessive amount of inter-ISP traffic • A main reason why many ISPs block or throttle BitTorrent traffic • Unnecessary Long-haul end-to-end connections • Certain pieces can often be retrieved from near-by peers • No Support for Video Streaming • Rarest-first policy disregards positions of pieces in video playback • BitTube implementations • Embedding locality-awareness into key operations of BitTorrent • Tracker, choker, and piece picker • A window-based approach to support “viewing-while-downloading” feature • Bitos, BASS, Toast, BitTorrent DNA

  19. Can BitTube Run Inside a Browser? • BitTube as a standalone software • Language: C++ • Binary code: 70 KB • Lifetime: as long as the user does not close it • BitTube embedded in browsers • Firefox Extension • Lifetime: as long as the browser runs • Other Options • Active X • User Confirmation Required • Java Applet • Redeveloping with Java • Lifetime: as long as the user stays in the website • Invisible Frame: A trick to make BitTube run longer inside a browser

  20. Piece Picker Locality Downloaded piece • Locality-first Policy • Piece with the minimum distance value is chosen • Distance value: the average distance of all peers possessing this piece • Distance value is passed down in the peer list from the tracker • Tie is broken arbitrarily Video File 2 1 3 3 1 1 2 3 1 Undownloaded piece Piece chosen by the policy 2 Distance value

  21. Window-based Piece Picker Downloaded piece • Adapting BitTorrent to video streaming • Piece selection restrained within the playback window • Enable Viewing while downloading • How to advance the playback window • Automatically pushed forward when the leftmost piece is downloaded • If the advancement falls behind the playback • The left-most piece is downloaded by the client-server thread • Streaming rate information is needed Playback Window Undownloaded piece Video File 2 1 3 3 1 1 2 3 Piece chosen by the policy Playback Window 1 Interest value 2 Distance value Video File 1 2 1 3 4 2 1 3

  22. Outline • Motivation • BitTube Design • BitTorrent Implementation • Results • Experimental Setup • Server load reduction • User Experience • Locality-Related Results • Impact to ISPs

  23. Experimental Setup • PlanetLab testbed • BitTube deployed on 200 nodes • Two PC machines at VANETS lab • video server and tracker • Intel Xeon 2.80GHz CPU, RedHat Linux • 10 minutes seeding time • Piece-level traces recorded at each PlanetLab node • Sequence number, receiving time, IP of source peer • Test Video File • Flash File downloaded from YouTube • Requested 53165 times from 18:18:33 to 22:31:59 on September 5, 2007 • Average 1.5 requests per second • 59MBytes and lasts 28 minutes 28 seconds • Original file is 7.5 Mbytes and lasts 3 minues 10 seconds

  24. Experimental Setup • What results do we want to obtain? • Server Load • User Experience • Impact on ISPs • Peer Contributions • Interplaying of Design Goals • Original design goals in BitTorrent • Locality-awareness • Window-based

  25. Server Load • Server Load Reduction • Average time length per piece • 4.25 seconds • BitTorrent (unlimited window size) • 91.8% • Window-based approach (window size = 20 pieces) • Rarest-first policy • 94.5% • Piece-picker Locality • 89.6% • Choker Locality • 85.3% • Tracker Locality • 83.5%

  26. Server Load per Peer Piece picker locality (window size = 20 pieces) Rarest-first Policy (window size = 20 pieces) BitTorrent (unlimited window size) • The more self-supporting peers, the greater the server load reduction

  27. User Experience: Interruption • Aggregate Interruption Time • Interruption stage is triggered if any piece is missing in playback • Aggregate interruption time is the summation of time lengths of all interruptions experienced by the peer • 80% peers have no interruptions • 10% peers have interruptions less than 100 seconds BitTorrent (unlimited window size) Choker Locality (window size = 20 pieces) Picker Locality (window size = 20 pieces) Tracker Locality (window size = 20 pieces)

  28. Locality-Related Results • Minimum AS Hop Strategy • Each peer finds from previously-joined peers the one with the minimum AS hop as its parent • Result: a minimum spanning tree without degree constraint • Optimal P2P Strategy to minimize the total number of AS hops • Serve as the baseline strategy against which other realistic strategies are evaluated ISP A ISP B ISP C 1 5 3 8 7 4 6 2

  29. AS Hop Count BitTorrent (unlimited window size) • Average AS Hop Count • Average value of AS hop counts traveled by all pieces download by each peer • Minimum AS Hop Strategy • Half the peers cause 0 AS hop count • Tracker Locality has the shortest AS hop count Choker Locality (window size = 20 pieces) Picker Locality (window size = 20 pieces) Tracker Locality (window size = 20 pieces) Minimum AS Hop

  30. Redundancy • Redundancy • Number of times a piece has to enter an ISP until all peers in the ISP finish their downloading • Lowest value is 1 • Highest value is the number of peers in the ISP • Normalized Redundancy • Redundancy normalized by number of peers • Less than 80 ISPs are involved BitTorrent (unlimited window size) Choker Locality (window size = 20 pieces) Picker Locality (window size = 20 pieces) Tracker Locality (window size = 20 pieces) Minimum AS Hop

  31. Peer Contribution • Sorted list of peers based on the number of bytes they have uploaded during the experiment • Tracker Locality has the most uneven distribution of peer contribution • Original BitTorrent and choker locality give the most even distribution of peer contribution • Minimum AS hop strategy only makes a few peers contribute • Unevenness of peer contribution is obvious across all policies BitTorrent (unlimited window size) Choker Locality (window size = 20 pieces) Picker Locality (window size = 20 pieces) Tracker Locality (window size = 20 pieces) Minimum AS Hop

  32. Number of Supplying Peers • Sorted list of peers based on the number of supplying peers they have got pieces from during the downloading • Tracker Locality allows the least number of supplying peers • Original BitTorrent allows the most number of supplying peers • Minimum AS hop always allow only one supplying peer BitTorrent (unlimited window size) Choker Locality (window size = 20 pieces) Picker Locality (window size = 20 pieces) Tracker Locality (window size = 20 pieces)

  33. Summary • A P2P solution to complement the current Internet video service • Motivation: saving bandwidth cost • Design Considerations: Minimum interruption to existing infrastructure • PlanetLab experiment shows great server load reduction without sacrificing user experience • Accommodating BitTorrent to locality-awareness and streaming applications • Less randomness are likely to shrink server load saving and cause more uneven peer contribution • But the cost saving is still significant • Worst performance: 83.5% saving (tracker locality) • Best performance: 94.5% saving (window-based rarest-first policy)

More Related