CoolStreaming - Network and System Laboratory

Download Report

Transcript CoolStreaming - Network and System Laboratory

CoolStreaming/DONet: A Data-driven
Overlay Network for Peer-to-Peer
Live Media Streaming
Authors: Xinyan Zhang., Jiangchuan Liu, Bo
Li,
and Tak-Shing Peter Yum.
Presenter: Elaine
1
Outline






Introduction
Related Work
Design and Optimization of Donet
Planet-Based Performance Evaluation
CoolStreaming
Conclusion and Future works
2

Multimedia applications



IP multicast is the best way



Live media streaming from a source to a large
population of users.
NetTV, news broadcast
The lack of incentives to install multicast routers
Political and practical issues
Application-level solution: Overlay network

Built on the unicast tunnels across cooperative
participating users.
3

Features of overlay network with streaming application.


Easily crash/leave
Streaming application



Many overlay construction algorithms advocate a tree
structure for data delivering.



High bandwidth
Continuity
highly vulnerable.
Restriction of the flow directions
Proposed data-centric overlay network



No predetermined roles
Self-organization
Data availability guides the flow directions.
4
Outline






Introduction
Related Work
Design and Optimization of Donet
Planet-Based Performance Evaluation
CoolStreaming
Conclusion and Future works
5
A brief overview of the existing overlay
streaming protocols in the pure p2p paradigm

Tree-based protocol



An internal node with high load leaving or crashing
often causes underflow in a large population of
descendants.
Unbalanced load
Repairing mechanisms are complex.
6

Gossip-based protocol
 A node sends a newly generated message to a
set of randomly selected nodes, iteratively.
 Cause redundancy especially for media
streaming application.
 Partially used for membership management
and data delivery in Donet
7
Outline






Introduction
Related Work
Design and Optimization of Donet
Planet-Based Performance Evaluation
CoolStreaming
Conclusion and Future works
8

Three key modules



Membership manager
Partnership manager
Scheduler
9

Interaction of modules
A.
Node join and Membership Management
Each Donet node has a unique id and Mcache.
 How to create and update the mcache.






Each node periodically generates a membership message
to announce its existence
SCAM(Gossip) to distribute the mcache messages..
Message format: <seq num, id,num partner, time to
live>
Entry Message format: <seq num, id,num partner,
time to live,last_update_time>
Node join algorithm
10
B.
Buffer Map Representation and Exchange



A video stream is divided into segments of uniform
length
Buffer Map: represent the availability of the segments
Continuously exchange BM with partners, and
schedules which segment is to be fetched from which
partner accordingly.
11
12
D.
Failure Recovery and Partnership Refinement
Failure Recovery



Graceful departure
Node failure
Partnership Refinement



Periodically establish new partnerships with nodes randomly
selected from its mCache.
Node i use function max{¯si,j , ¯sj,i} to score its partners
¯si,j: the average number of segments that node i
retrieved
from node j per unit time.
¯sj,i: the average number of segments that node j
retrieved
from node j per unit time.
13
Outline






Introduction
Related Work
Design and Optimization of Donet
Planet-Based Performance Evaluation
CoolStreaming
Conclusion and Future works
14
A.
Design of the Experiment System





Almost all active nodes over PlanetLab,200~300
Experiment May to June, 2004
Each active PlanetLab node runs a copy of the prototyped
program,acting a s a DONet node.
Origin node: Located in the United States
Monitoring node: In Hong Kong
15
B.
Performance under Stable Environment




All nodes persist in the lifetime of the streaming(120min)
Control overhead(left)
Scalability(left)
Playback continuity(right).
16
C.
Performance under Dynamic Environment


Dynamic node joining , leaving, and failure.
Following the ON/OFF model.
17
D.
Comparison with the Tree-based Overlay



Metrics: End-to-End delay / Continuity
The number of partners is set to 4 in Donet.
The degree of each tree node is limited to 3,
except root(4).


Different capacity and bandwidth cause some
children move to lower levels.
Fig 11/fig12/fig 13
18

A full and balanced 3-ary tree of 231 nodes
has a height of 5 only.
19
20

Summary and Caveats




Performance of Donet is acceptable
Overhead is low(1% of the video traffic)
Scalability
Compared to tree-based overly



Continuity is much better under highly dynamic environment.
End-to-end delay is comparable.
Caveats


Scalability
Representability
21
Outline






Introduction
Related Work
Design and Optimization of Donet
Planet-Based Performance Evaluation
CoolStreaming
Conclusion and Future works
22


A public Internet based DONet package, called
CoolStreaming.
Released the first version (v.0.9) on May 30, 2004.
CoolStreaming v.0.9
23
Two interesting facts:


The current Internet has enough available
bandwidth to support TV-quality
streaming(>=450 Kbps)
The larger the data-driven overlay is, the
better the streaming quality it delivers.
24
Conclusion






Proposed DONet
Simple overlay structure (according to the data
availability)
Scalable membership and partnership management
algorithm
Good playback quality under formidable network
condition
Compared to tree-based overlay,it achieves much more
continuous streaming with comparable delay.
Released CoolStreaming v.0.9
25