Upload
rockys11
View
389
Download
0
Tags:
Embed Size (px)
Citation preview
1
Measurement and Modeling of a Large-scale Overlay for Multimedia Streaming
Long Vu, Indranil Gupta, Jin Liang, Klara Nahrstedt
University of Illinois at Urbana-ChampaignQShine 2007, Vancouver, Canada
2
Motivation
• IPTV applications have flourished (SopCast,
PPLive, PeerCast, CoolStreaming, TVUPlayer, etc.)
• IPTV growth: (MRG Inc. April 2007)
– Subscriptions: 14.3 million in 2007, 63.6 million in 2011
– Revenue: $3.6 billion in 2007, $20.3 billion in 2011
• Understanding current IPTV (P2P multimedia
streaming) applications is crucial for future designs
3
Motivation• PPLive is one of the largest deployed P2P
multimedia streaming applications
• Existing studies about PPLive:
– Network-centric: video traffic, flow rate, bandwidth utilization [X. Hei 06, A. Ali 06, Silverston 07]
– User-centric: geographic distribution, user-perceived quality [X. Hei 06]
• However, there is no study about PPLive overlay characteristics: node degree, overlay randomness, node availability, etc.
4
Study Objectives
• Evaluate overlay characteristics of PPLive
• Compare overlay characteristics PPLive with P2P file sharing overlay characteristics
• Draw conclusions for future designs and developments
5
Next…• PPLive overview
• Challenges & study methodology
• Results:– Channel size variation– Node degree– Overlay randomness– Node availability– Session length
• Compare PPLive and P2P file sharing
• Draw conclusions
6
PPLive Overview
• One of the largest deployed P2P multimedia streaming systems
• Developed in China
• Hundred thousands of simultaneous viewers
7
PPLive Episode Channels
• Most popular (number of viewers)
A Program Segment (PS)
An episode channel
Movie 1 Movie 2 Movie 3 Movie 4
PS PS PS PS PS
Day 2Day 1
Time
8
PPLive Membership Protocol
Client
An overlay
9
Challenges
• PPLive is a closed-source system
• How to develop the measurement method to measure PPLive overlay
• How to select the right metrics to measure PPLive overlay characteristics
• Many PPLive peers are unresponsive
10
Our Approaches
• Capture traffic, analyze messages
• Develop a crawler-based study methodology
• Studied Metrics:
– Channel population size
– Node degree
– Overlay randomness
– Node availability
– Node session length
11
Crawler• 10 PlanetLab geographically distributed nodes to
crawl peers (Previous works use only one machine)
• Aggregate, de-duplicate crawled peers to get the peer list
12
Operations
Snapshot collects peers in one channel
PartnerDiscovery collects partners of responsive peers
Studied
channels
Time1st Snapshot 2nd Snapshot 3rd Snapshot 4th Snapshot
10 min 10 min 10 min 10 min
13
Next…• PPLive overview
• Challenges & study methodology
• Results:– Channel size variation– Node degree– Overlay randomness– Node availability– Session length
• Compare PPLive and P2P file sharing
• Draw recommendations
14
Channel Size Varies over a day
• Peaks at noon and night
• A varies 10 times, B and C varies 2 times
• Different from P2P file sharing [Bhagwan 03]
15
Channel Size Varies over Consecutive Days
• The same channel, same program: Peaks drift• Peaks depend on time and channel content
First day Second day
16
Node Degree is Independent of Channel Size
Similar to P2P file sharing [Ripeanu 02]
Average node
degree scale-free
17
Overlay Randomness• Clustering Coefficient (CC) [Watts 98]
– for a random node x with two neighbors y and z, the CC is the probability that either y is a neighbor of z or vice versa
• Probability that two random nodes are neighbors (D)– Average degree of node / channel size
• Graph is more clustered if CC is far from D [well-known results from theory of networks and graphs]
18
Overlay Randomness May Depend on Channel Size
• Small overlay, more random
• Large overlay, more clustered
Different from P2P file sharing (small-world behavior) [Ripeanu 02, Saroiu 03]
19
Time
X
Y Y
X X
Y
10m 10m 10m
1st snapshot 2nd snapshot 3rd snapshot 4th snapshot
Node Availability (1) [Bhagwan 03]
• Number of peer pairs (X,Y):• Compute P(Y=1|X=1). E.g. P(Y=1|X=1) = 2/3• 12 hours: 72 snapshots, 24 hours: 144 snapshots
Pick 185 random peers from one snapshot
…
20
Nodes in one Snapshot Have Correlated Availability
Different from P2P file sharing [Bhagwan 03]
Correlated Availability
Nodes appearing together is likely appear together again
21
Node Availability (2)
• Compute the difference of P(Y=1|X=1) and P(Y=1)• For example: P(Y=1|X=1) = 1/3 ; P(Y=1) = 1/2
Pick 500 random peers from all peers of the whole day (144 snapshots)
Time
X
YY
X X
10m 10m 10m
1st snapshot 2nd snapshot 3rd snapshot 4th snapshot
…
22
Random Node Pairs Have Independent Availabilities
Similar to P2P file sharing [Bhagwan 03]
IndependentAvailabilities
23
PPLive Peers are Impatient
50% sessions are less than 10 minutes
Different from P2P file sharing [Saroiu 03]
24
Geometric Series Session Length
25
Comparison
26
Conclusions• Overlay characteristics of PPLive and P2P file
sharing are very different
• PPLive channel population size depends on time and channel content
• Nodes in one snapshot have correlated availability:
– route media streams (shorten startup delay)
– create sub-overlay to share content
• Simulation of P2P multimedia streaming needs to take the node availability into account
• Session Lengths are short and follow Geometric series
27
Thank you!
28
Backup slides
29
Responded Peers
• More than 50% of peers do not respond to PING messages
30
Membership Update• Experiment 1: A UIUC PPLive client attends a channel
– Turn off the UIUC client– After 10 seconds, use 10 PlanetLab crawlers to crawl that
channel, there is no IP of the UIUC client
• Experiment 2:– A UIUC PPLive client joins a PPLive channel– After 10 seconds, use 10 PlanetLab crawlers to crawl that
channel, there exists IP of the UIUC client
• Conclusion: PPLive membership update is very fast
• Unresponsive peers are not dead peers
31
Active PeersGiven an overlay (or a channel) G:
32
Channel Size Varies over a day
Coefficients match
33
Channel Size Varies over a day