Watching Television Over Nationwide IP Multicast

1
Watching Television OverNationwide IP
Multicast

• Meeyoung Cha (Intern at Telefonica Research)
• with
• Pablo Rodriguez (Telefonica Research)
• Jon Crowcroft (Cambridge Univ)
• Sue Moon (KAIST)

2
Outline

• TV viewing behavior
• Architecture, trace, statistics
• Design comparison
• IP multicast, P2P, CDN
• Clustering
• Temporal, spectral (channel),behavioral patterns

3
Part1 TV Viewing Behavior

• Architecture
• Trace
• Statistics

4
Quality-assured IPTV architecture
1Gb/s
5Mb/s
5
IGMP trace snapshot
Collected over 700 DSLAMs!

• Channel switching ? a pair of IGMP leave join
  messages
• Trace summary

6
Channel holding time

• Spikes in histogram natural long-term off hours?
• Tipping point in CDF

7
Number of viewers over time

• Time-of-day effect
• 18 increase in viewing over weekends

8
Channel popularity

• Top 10 channels account for 80 viewer share
• Zipf-like popularity also shown in PPLive

9
Per-program behavior

• European Champions League, May 23 2007
• Strong membership ? potential for P2P
• P2P - storing in the box, feeding latecomers

10
Part2 Alternate Architectures

• Static vs dynamic IP multicast
• IP multicast vs CDN, P2P

11
Static vs Dynamic Multicast Trees
Source
IP router
cost 2
cost 1
DSLAM
Why? Few hundred users per DSLAM Users not
active all the time Focus skewed to few channels
STB
Dynamic
Static
12
Alternate designs for live TV
Server-based IP multicast
How do these technologies compare?
13
Example routing
TV head end
Regionalserver
IP router
cost 7
cost 3
DSLAM
STB
Locality-aware P2P
Topology-oblivious P2P
CDN
14
Assumptions for comparison

• General
• Stream in units of 30 sec
• P2P
• Streamed content stored at set-top box
• A peer serves at most one another peer
  concurrently
• Peer selection
• Locality-aware search within local DSLAM first
• Search 1000 random peers
• If no seed found, served by the server

15
Dynamic tree smart P2P win!
87 trafficlocally-served
16
Cooperative P2P IP multicast for VCR features

• Rewind to the beginning of the programs
• Streaming passed scenes for any newly joining
  channel
• Can P2P provide rewinding?
• How many peers stored content from the beginning?
• Frequent channel switching
• Unpopular programs, low usage hours
• Peers may serve only one another peer at a time
• Globally optimal scheduling across multiple
  channels
• How much locality?

17
Part3 Clustering

• Channel switching patterns
• Temporal correlation
• Channel correlation

18
1. Switching behavior (a)

• Step1 build 3-state Markov model
• Average behavior of all users
• Step2 compare how behavior of individual users
  deviate from the global pattern.

19
1. Switching behavior (b)

• Step3 threshold-based grouping
• Hard cluster
• Interpretable

15
HeavyBrowser
1.9 hr, 11 min Ch. 32 -gt 6
AverageUser
40
LightUser
0.5 hr, 11 min Ch. 5 -gt 1
View 2.8 hr/day, 12 min/session Channel browse
21, view 7
20
FocusedViewer
25
3.2 hr, 14 min Ch. 11 -gt 6
20
2. Temporal pattern (a)

• Group users based on their active times
• NMF (Nonnegative Matrix Factorization)

Figure from Brunet et al. PNAS2004
21
2. Temporal pattern (b)

• Peep into life-styles of users using NMF

Night Owls25
Early-birds25
Always-On50
22
3. Correlation across channels (a)

• Identify channels that are viewed together
• example
• Step1 calculate cosine coefficient of channels
• 155 x 155 matrix

23
3. Correlation across channels (b)

• Step2 set correlation matrix to 0-1 using
  threshold
• Only strong correlation remains
• Step3 find connected component

24
Thank you!

• ACM IMC 2007 best paper awardedI Tube, You
  Tube, Everybody Tubes Analyzing theWorld's
  Largest User Generated Content Video System

25
Backup Slides
26
Backgrounds

• Leap in IP multicast industry
• two decades ago - proposed
• now - commercial deployments with billion users!
• TV habit studies
• olden days - phone call survey, sampled monitor
• now real trace available!

27
Fun history of TV
I invented TV!

• Inventor John Logie Baird
• Born in Scotland 1888
• Timeline
• 1925 First TV, named Televisor
• 1936 BBC starts first daily broadcast (2hr)
• 1944 First color TV

John Baird, 1925
The first moving picture, 1926
28
Can we repeat the same study on cable TV?

• Backbone part doesnt change
• Last mile access is HFC (hybrid fiber coaxial),
  but not DSLAM

IGMP snoopat each CM (cable modem)
Cable Operator
Telcom Operator
IGMP snooping here!
Figure from Dischinger et al. IMC 2007
29
Average Viewing Time Per Day
30
Multicast tree restoration time

• IP Multicast up to seconds to tens of minutes
• Feasibility of IP restoration in a Tier-1
  Backbone, IEEE Network 2004
• With MPLS fast-reroute up to several seconds
• 50 ms per link, aggregated over links
• No guarantee
• Optical layer path restoration less than 1
  second
• Designing path-redundant trees NP-hard problem!
• If feasible, gives guarantee
• Multicast in optical layer, still a new
  technology

31
Alternate backbone architecture
Figure from Cha et al. IPTV workshop 2006
32
Nonnegative Matrix Factorization

• Condition
• Clustering of nonnegative sparse features
• Representation by parts
• Factor A WH
• A matrix of data
• W m x r matrix of basis vectors
• H r x n coefficient matrix
• Optimize accuracy of solution
• min A-WH F where W,H 0
• Multiplicative Update Rules, Lee-Seung 2000
  Nature
• Simple, efficient
• Guaranteed to reach global minimum using
  multiplicative update rule

33
Clustering Users Based on Time

• Problem Find dominant patterns from large trace
• Alternate Methods
• PCA (Principal component analysis)
• Showed slightly poor result due to Gaussian
  restriction!
• Signatures quite hard to interpret as they are
  eigenvectors
• SOM (Self-organizing maps)
• Known highly-dependent on initial conditions
• Entropy-based
• Time-sequence not taken into account
• Nonlinear ICA (Independent component analysis)

34
References

• IP multicast design
• Lightreading
• Case Study Resilient Backbone Design for IPTV
  Services, Workshop on IPTV over WWW, 2006
• TV viewing habits
• Neilson ratings
• A Measurement Study of a Large-Scale P2P IPTV
  system, IEEE Trans on Multimedia, 2007
• Alternate IPTV architectures
• Should Internet Service Providers Fear
  Peer-Assisted Content Distribution, ACM IMC 2005
• Clustering
• Learning the Parts of Objects with NMF, Nature
  1999
• Planet Scale Software Updates, ACM Sigcomm 2006