CoolStreaming Cooperative Overlay Streaming

1
CoolStreaming - Cooperative Overlay Streaming

• Presented by Xinyan Zhang
• Supervisor Prof. Yum Tak Shing
• November 22, 2009
• Department of Information Engineering
• The Chinese University of Hong Kong

2
Outline

• Motivation
• Related Work
• CoolStreaming - Brief Introduction
• Experiments
• Conclusion Remarks

3
Motivation

• Enable large scale live broadcasting in Internet
  environment
• Capacity limitation 500Kbps, theoretical
  limitation for a 100Mbps Server is 200 concurrent
  users
• Heterogeneous network cannot guarentee the
  quality of service

4
Why Client/Server cannot scale
5
Why Collaborative Communications Work
6
Outline

• Motivation
• Related Work
• CoolStreaming - Brief Introduction
• Experiments
• Conclusion Remarks

7
Current Solution

• Current Solution
• CDN Expensive and not quite scalable for large
  number of audiences
• Application layer multicast not mature

8
Application Layer Multicast (ALM)

• Multicast is defined as sending information to
  more than one receivers at the same time.
• IP multicast is clearly the most cost-effective,
  however no scalable routing algorithm
• ALM is an attempt to replace the unsuccessfully
  deployed IP Multicast

9
Illustration of ALM
10
ALM Research Topics

• Mesh-based
• Navada
• Tree-based
• Nice

11
ALM Research Topics

• Mesh-based
• Navada
• Tree-based
• Nice

12
Bittorrent

• Bittorrent - A fast Data Distribution Protocol
• User Collaborations
• Random block fetching
• Increasing availability
• Lack
• Rate control
• Real time control
• Need many efforts to use in streaming

13
Outline

• Motivation
• Related Work
• CoolStreaming - Brief Introduction
• Experiments
• Conclusion Remarks

14
Motivation

• ALM suffers
• Frequent user join/leave
• Heterogeneous network conditions
• ALM requests
• Fast data recovery
• Low overheads
• Adaptation to various network conditions

15
Problems

• How to Control
• Scalability
• Robustness
• Overhead
• Adaptation

16
CoolStreaming

• Cooperative Overlay Streaming
• Structure
• Tree -gt Random Graph
• Gossiping based data delivery
• Each host v in the network selects some random
  host w as a communication partner and exchange
  information with each other

17
Overlay Constructions

• Similar with Gnutella
• Random Graph

18
Streaming structure

• Buffer Manager
• Scheduler
• Dispatcher

19
Performance Analysis

• Data outage
• Tree
• Distributed Dynamic Streaming

20
Performance Evaluations
21
Outline

• Motivation
• Related Work
• CoolStreaming - Brief Introduction
• Experiments
• Conclusion Remarks

22
Real Experiments

• Planet-Lab Experiments
• 200-300 hosts
• Good connectivity
• Broadcasting TV to users
• Program Euro 2004 (on going)

23
Planet-lab geographical distribution
24
Planet-Lab Brief Result 1

• 3 data outage compared with around 10 in
  improved tree algorithms (200 nodes)

25
Planet-Lab Brief Result 2

• Average data outage vs. time, 200 nodes,
  500kbps, upload rate 800kbps

26
Real Broadcasting
27
Real Broadcasting
28
Broadcasting - Facts

• Up to around 2400 concurrent users
• Bitrate 450-700kbps
• Average Packet Loss around 1 - 5 during game
  time
• Geographical distribution (approximation) 1/3
  Mainland China, 1/3 Hong Kong, 1/3 USA
• 1/3 users are in Firewall
• Heterogeneous network environment LAN, DSL,
  CABLE...

29
Broadcasting - Observations

• The more the user, likely the better the
  performance
• Self adaptation automatically converge into
  stable status (still much room to improve)
• Stable users get better services

30
Demo of streaming
31
Outline

• Motivation
• Related Work
• CoolStreaming - Brief Introduction
• Experiments
• Conclusion Remarks

32
Summary

• Experiments show
• More scalable
• More robust
• More adaptive
• Support numerous streaming users using PC
• Total bandwidth 450Mbps
• Quite amazing

33
Q A

• Thanks