Shibata, N., Yasumoto, K., and Mori, M.: P2P Video Broadcast based on Per-Peer Transcoding and its Evaluation on PlanetLab, Proc. of 19th IASTED Int'l. Conf. on Parallel and Distributed Computing and Systems (PDCS2007), (November 2007). http://ito-lab.naist.jp/themes/pdffiles/071121.shibata.pdcs2007.pdf

1. 1
P2P video broadcast based on per-peer transcoding
and its evaluation on PlanetLab
Naoki Shibata, † Keiichi Yasumoto, Masaaki Mori
Shiga University, †Nara Institute of Sci. and Tech.

2. 2
Motivation
 Watching TV on various devices
 Screen resolution of mobile phone : 96x64 ~
 Screen resolution of Plasma TV : ~ 1920x1080
 Video delivery method for wide variety of devices
 Screen resolution
 Computing power
 Available bandwidth to Internet
 Popularization of P2P video delivery
 Joost
 Zattoo

3. 3
Overview of this presentation
 Improvement to our previously proposed video
delivery method named MTcast
Features of (previous) MTcast
 Video delivery method based on P2P video streaming
 Serves requests with different video qualities
 Scalable with number of users
New improvement
 Reduced backbone bandwidth for further scalability
 Evaluation of performance on PlanetLab
 Implementation in Java language
 Evaluation in PlanetLab environment

4. 4
Outline
 Background
 Related works
 MTcast overview
 Implementation overview
 Evaluation

5. 5
Multiversion method
 Minimum delay
 No need of transcoding
 Low user satisfaction : # of served video qualities = # of versions
 High network load
500k
G. Conklin, G. Greenbaum, K. Lillevold and A. Lippman :
``Video Coding for Streaming Media Delivery on the Internet,’’
IEEE Transactions on Circuits and Systems for Video Technology, 11(3), 2001.
300k
…
request 200k
deliver 300k
request 400k
deliver 500k

6. 6
Online transcoding method
S. Jacobs and A. Eleftheriadis :
``Streaming Video using Dynamic Rate Shaping and TCP Flow Control,’’
Visual Communication and Image Representation Journal, 1998.
1000k
 Higher user satisfaction
 Additional cost for proxies
 # of qualities is restricted by capacity of proxies
Server Proxies
1000k
1000k
300k
500k
700kTranscode

7. 7
Layered multicast method
 Low computation load on server
 User satisfaction depends on # of layers
 Limitation on # of layers … High CPU usage to decode many layers
J. Liu, B. Li and Y.-Q. Zhang :
``An End-to-End Adaptation Protocol for Layered Video Multicast Using Optimal Rate Allocation,’’
IEEE Transactions on Multimedia, 2004.
200k
Base layer
300k
2nd layer
…
200k
200k
200k
+300k
+300k+500k

8. 8
Outline
 Background
 Related works
 MTcast overview
 Implementation overview
 Evaluation

9. 9
Service provided by MTcast
 Network environment
 Wide area network across multiple domains
 Number of users
 500 to 100,000
 Kind of contents
 Simultaneous broadcast of video - same as TV broadcast
 New user can join to receive delivered video from the
scene currently on broadcast
 Kind of request by users
 Each user can specify bit rate of video
 We assume resolution and frame rate are decided from bit rate

10. 10
Transcode
Idea of MTcast
1000k
Server
request 800k
reply 800k
request 500k
reply 500k
request 300k
reply 300k
Transcode
Transcode

11. 11
Building transcode tree
User nodes
Bit rate request 2000k
Bit rate request 800k
Bit rate request 300k
Bit rate request 1200k
Bit rate request 1500k
User nodes
Bit rate request 2000k
Bit rate request 2000k
Bit rate request 1920k
Bit rate request 1850k
Bit rate request 1830k
Sort
Transcode tree is video delivery tree

12. 12
Building transcode tree
User nodes
Bit rate request 2000k
Bit rate request 2000k
Bit rate request 1920k
Bit rate request 1850k
Bit rate request 1830k
Bit rate request 1800k
Bit rate request 1800k
Bit rate request 1780k
• Make groups of k user nodes from the top
• Each group is called a layer
• Minimum requested bit rate for each layer
is actually delivered to the layer
Delivered bit rate
for each layer
A constant value decided
on each video broadcast

13. 13
Building transcode tree
• Put each layer at the place of nodes in a binary tree
• In the order of depth first search
• Construct a modified binary tree
2000k
1800k
1500k 1300k
1100k
900k 700k
Video server

14. 14
Advantages of building tree in this manner
 Videos in many qualities can be served
 Number of qualities = Number of layers
 Each node is required to perform only one transcoding
 Length of video delivery delay is O(log(# of nodes))
 Tolerant to node failures
2000k
1800k
1500k 1300k
1100k
900k 700k
Video server

15. 15
Recovery from node failure
No increase in number of video transcoding on each node
• Degree of tolerance of node failure depends on :
• Number of nodes in each layer
If there are many nodes in layer, it has greater tolerance of failure
• Available bandwidth on each node
• Buffered video data is played back during recovery
• Users never notice node failures

16. 17
Extension for real world usage
Each link is an overlay link
Traffic may go back and forth many times
between ASs
Precious bandwidth between ASs is consumed
Nodes in service provider A
Nodes in service provider B
Nodes in service provider C
Idea of extension
Nodes in a same AS should connect in priority
Priority of connection is decided according to
hop count and available bandwidth

17. 18
Outline
 Background
 Related works
 MTcast overview
 Implementation overview
 Evaluation

18. 19
Design policy
 Usable on PlanetLab
 Usable in many similar projects
 Easily modifiable
 Good performance, if possible
 Why not use JMF?
It’s not maintained
Huge buffering delay

19. 20
Modular design
 We designed many classes for video delivery
 Transcoder
 Transmitter and receiver to/from network
 Buffer
 etc.
 Each node is a set of instances of these classes
 Each node instantiate these classes and connects
the instances according to the command from a
central server
Workings of each node can be flexibly changed by
changing commands from the central server

20. 21
Outline
 Background
 Related works
 MTcast overview
 Implementation overview
 Evaluation

21. 23
Results of evaluation published in [9]
 Computation load of transcoding
 Measured computation load when video playback and
transcoding are simultaneously executed
 Measured on desktop PC, notebook PC and PDA
 Result : All processing can be performed in real time
 Computation load of making transcode tree
 1.5 secs of computation on Pentium 4 2.4GHz
 Time complexity： O( n log n )
 Network load : Practical if the computation node of transcode tree
has enough bandwidth
 User satisfaction
 Satisfaction degree is defined as to [3]
 Made a network with 6000 node using Inet 3.0
 Satisfaction with our method was at least 6% higher than layered
multicast method
 Satisfaction becomes better as the number of nodes increases

22. 24
Video quality degradation by transcoding
 Video quality may degrade by multiple transcoding
 We measured PSNR value when video is transcoded
in our method
 We compared :
 A video transcoded only once
 A video transcoded multiple times

23. 25
Effectiveness of bandwidth reduction(1/2)
 Compared physical hop count in transcode tree
 By our method
 By randomly selecting node to connect
 Comparison by simulation
 Number of user nodes : 1000
 333 nodes has bandwidth between 100 and 500kbps
 333 nodes has bandwidth between 2 and 5Mbps
 334 nodes has bandwidth between 10 and 20Mbps
 Result of simulation
 Hop count by the random method : 4088
 Hop count by our method : 3121
 25% reduction of hop count by our method

24. 26
Effectiveness of bandwidth reduction(2/2)
 Compared physical hop count in transcode tree
 By our method
 By randomly selecting node to connect
 Comparison on PlanetLab
 20 user nodes on 7 countries
 Result
 Random selection : hop count 343, 361, 335
 Our method : hop count 314, 280, 277
 16% reduction of hop count by our method

25. 27
Time to start up the system on PlanetLab
 Measured time to the following events since beginning
 All nodes complete establishing connection
 All nodes receive the first one byte of data
 Comparison on PlanetLab
 20 user nodes on 7 countries
 Nodes are cascaded, not connected in tree
 Result of evaluation
 Observation
 Most of time is consumed to establish connections
 All operations are performed in parallel, and thus the observed
time is the time for the slowest node to establish connection

26. 28
Time to recover from node failure on PlanetLab
 Measured following time since a node failure
 Establishing connection to a new node
 Receiving data from the new node
 Comparison on PlanetLab
 20 user nodes on 7 countries
 Nodes are cascaded, not connected in tree
 Result of evaluation
 Observation
 These are practical values
 During recovery time, buffered data is played back, and thus user never
notices node failure

27. 29
Conclusion
 Improved MTcast
 Bandwidth usage between ASs is reduced
 Made a prototype system in Java
 Evaluation on PlanetLab
 Ongoing works include
 Serving request of many parameters including picture
size, framerate and audio channels
 Further reduction of bandwidth between nodes

28. 30
Shibata, N., Yasumoto, K., and Mori, M.: P2P Video Broadcast
based on Per-Peer Transcoding and its Evaluation on
PlanetLab, Proc. of 19th IASTED International Conference on
Parallel and Distributed Computing and Systems (PDCS2007),
pp. 478-483. [ PDF]
Sun, T., Tamai, M., Yasumoto, M., Shibata, N., Ito, M. and Mori,
M.: MTcast: Robust and Efficient P2P-based Video Delivery
for Heterogeneous Users, Proceedings of 9th International
Conference on Principles of Distributed Systems (OPODIS2005),
pp. 176-190.
DOI:10.1007/11795490_15 [ PDF ]