1 Can Internet Video-on-Demand be Profitable? Cheng Huang, Jin Li (Microsoft Research Redmond), Keith W. Ross (Polytechnic University) ACM SIGCOMM 2007.

1

Can Internet Video-on-Demand be Profitable?

Cheng Huang, Jin Li (Microsoft Research Redmond), Keith W. Ross (Polytechnic University)

ACM SIGCOMM 2007

2

Outlines

Motivation Trace – User demand & behavior Peer-assisted VoD

Theory Real-trace-driven simulation

Cross ISP traffic issue Conclusion

3

Motivation

Saving money for huge content providers such as MSN Video, Youtube, Yahoo Video, Google Video,…

Video quality is just acceptable

User demand +++

Video quality+++

Traffic+

ISP Charge+Client Server

User BW +

Video quality+

User BW +++

Video quality+++

Traffic++++++++

ISP Charge+++++++P2P

Traffic++

ISP Charge++

User BW ++++++

Video quality+++++++

Traffic+++

ISP Charge+++

4

P2P Architecture

Peers will assist each other and won’t consume the server BW.

Each peer have contribution to the whole system.

Throw the ball back to the ISPs The traffic does not disappear, it moved to

somewhere else.

5

Outlines

Motivation Trace – User demand & behavior Peer-assisted VoD

Theory Real-trace-driven simulation

Cross ISP traffic issue Conclusion

6

Trace Analysis

Using a trace contains 520M streaming requests and more than 59,000+ videos from Microsoft MSN Video. http://video.msn.com/

From April to December, 2006.(9 Months)

7

8

Video Popularity

The more skewed, the much better.

• The popularity distributions are quite similar. • There is indeed a high-degree of locality.• The distribution is more skewed than a Zipf distribution.

9

Download bandwidth

Use ISP download/upload pricing table Downlink distribution

to generate upload BW distribution

10

Demand v.s. Support Available upload bandwidth at clients far exceeds user demand.

11

User behavior - Churn• Users generally view large fraction of short videos.• But less than 20% of the users view more than 60% of videos larger than 30 minutes.

12

User Behavior• Fraction of sessions that start at the beginning of a video and have no interactivity is important in the success of a peer-assisted VoD. • For < 30 min videos, 80% of the session does not have interactivity.

13

Content quality revolutionThe demand and the bitrates for VoD increase rapidly.

14

Traffic Evolution

2.271.23

Quality Growth: 50%User Growth: 33%Traffic Growth: 78.5%

They believe it is likely that bitrates will increase faster than client upload B/W.

15

Contributions of this paper The first measurement study of an on-demand video streaming

system in a large scale. We present a simple theory for peer-assisted VoD. This theory

identifies 3 basic operating modes of peer-assisted VoD system. The surplus mode, the balanced mode, and the deficit mode.

For the single-video approach, we describe 3 natural prefetching policies for exploiting surplus peer upload capacity. No-prefetching, water-leveling, and greedy policy.

We use the 9 months MSN trace, which was collected for a client-server deployment, to drive simulations for peer-assisted deployments.

We explore the impact of peer-assisted VoD on ISPs.

16

Outlines

Motivation Trace – User demand & behavior Peer-assisted VoD

Theory Real-trace-driven simulation

Cross ISP traffic issue Conclusion

17

Peer-assisted VoD

Peer-assisted VoD Users watching the video will assist in the re-distribution of

the video to other users. There is still a server (or server farm) which stores all of the

publisher’s videos. Guarantees that users playback the video at the playback

rate without any quality degradation. The server is only active when the peers alone cannot

satisfy the demand. 2 design approaches to peer-assisted VoD

Single video: a peer only redistributes the video it is currently watching. (This paper use!!)

Multiple video: a peer can redistribute a video that it previously viewed but is currently not viewing.

Modelling – Single Video The time user remains online to see the video is The bitrate of the video is r Users arrive at the system with Poisson distribution rate M is the number of user type, where a type m user has upload

link BW wm

pm : the probability that an arrival is a type m user System’s average upload B/W of an arriving user is pm wm Expected number of type m users is pm In steady state, the average total Demand is r pm r

The average Supply is pm wm If Supply > Demand

Surplus mode, small server load If Supply < Demand

Deficit mode, VERY large server load If Supply ≈ Demand

Balanced mode, medium server load

19

Prefetch Policy

Let every peer get more video data than demand (if possible) in surplus mode. And thus can tide over deficit mode.

Peers can potentially prefetch video from each other using the peers’ surplus bandwidth.

3 prefetching policies No-prefetching

Each user downloads content at the playback rate r and does not prefetch content for future needs.

At any given instant of time, the user may be downloading from multiple peers as well as from the server.

Assume that each user views the video without gaps. Water-leveling prefetching

Peer only prefetches from peers arrive before it and have sufficient upload bandwidth, and demand is depend on the user buffer level.

Make all the peers to have the same buffer levels of prefetched content. Greedy prefetching

Each user simply dedicates its remaining upload BW to the next user right after itself. For each user i, donate it’s upload BW and aggregated BW to user i+1

[4] C. Huang, J. Li, and K. W. Ross, “Peer-Assisted VoD: Making Internet Video Distribution Cheap,” IPTPS, Bellevue, WA, Feb. 2007.

20

Outlines

Motivation Trace – User demand & behavior Peer-assisted VoD

Theory Real-trace-driven simulation

Cross ISP traffic issue Conclusion

21

Methodology

Discrete-event simulator. Driven by 9 months of MSN Video trace. 2 videos

Gold stream: the most popular video, popular for a few days.

Silver stream: the second most popular video, popular for a month.

Focus on the balanced mode. Greedy prefetching. no P2P: the resources used by the pure client-

server deployment.

22

Simulation: Non-early-departure Trace

•P2P deployment at the current quality level, typically no server resources are needed. Some resources needed when few concurrent users.

23

Simulation: Early departure (No interaction) When video length > 30mins, 80%+ users

don’t finish the whole video.

• Even with early departures peer-assistance can provide a good improvement. • Prefetching continues to provide improvements over non-prefetching.

Table 4: Server rates (in Mbps) under different system modes with early departures. April 2006.

24

Simulation: Full Trace

How to deal with buffer holes? As user may skip part of the video.

2 strategies Conservative: assume that user upload BW=0

after the first interaction. Optimistic: ignore all interactions, there is no hole

in the user’s buffer.

25

Results of full trace simulation (1/2)

• Due to interactivity, a user might have holes in its buffer.•The actual performance will lie between these two bounds.

26

Results of full trace simulation (2/2)

Cost Reduction

With peer-assisted VoD, server BW reduction from 2.2Gbps to 79.4Mbps on Dec. 2006.

28

Outlines

Motivation Trace – User demand & behavior Peer-assisted VoD

Theory Real-trace-driven simulation

Cross ISP traffic issue Conclusion

29

ISP-unfriendly P2P VoD

ISPs, based on business relations, will form economic entities. 3 relationships between ISPs:

1) transit relationship (also called customer-provider) one ISP purchasing Internet access from another ISP and paying

for the bandwidth usage. 2) sibling relationship

the interconnection among several ISPs belonging to the same organization.

3) peering relationship ISPs pairing with each other. Peering ISPs can exchange traffic

directly, which would otherwise have to go through their providers. Traffic do not pass through the boundary won’t be charged.

ISP-unfriendly P2P will cause large amount of traffic.

30

Simulation results of unfriendly P2P

Most P2P VoD crosses ISP boundaries.

31

Simulation results of friendly P2P Peers lies in different economic entities do

not assist each other.

•Silver stream single video, 5000 distinct video distributions.•Top 10 more popular videos among the 12000 in traces.• When an entity contains few peers, the sharing becomes more difficult as well, and the server bandwidth is increased accordingly.

Table 8: Server bandwidth (in Mbps) in an ISP-optimized scenario.

Conclusion

Peer-assisted VoD, with the proper prefetching policy, can dramatically reduce server bandwidth costs.

Peer-assisted VoD can be both server and

ISP friendly.

33

References

[3] B. Cheng, X. Liu, Z. Zhang, and H. Jin, “A Measurement Study of a Peer-to-Peer Video-on-Demand System,” IPTPS, Bellevue, WA, Feb. 2007.

[4] C. Huang, J. Li, and K. W. Ross, “Peer-Assisted VoD: Making Internet Video Distribution Cheap,” IPTPS, Bellevue, WA, Feb. 2007.

[19] A. Al Hamra, E. W. Biersack, and G. Urvoy-Keller, “A Pull-based Approach for a VoD Service in P2P Networks,” IEEE HSNMC, Toulouse, France, Jul. 2004.

[20] Y. Cui, B. Li, and K. Nahrstedt, “oStream: Asynchronous Streaming Multicast in Application-Layer Overlay Networks,” IEEE JSAC, 22(1), 2004.

[21] J. Li, Y. Cui, and B. Chang, “PeerStreaming: Design and Implementation of an On-Demand Distributed Streaming System with DRM Capabilities,” Multimedia Systems Journal, 2007.

[22] S. Annapureddy, C. Gkantsidis, P. R. Rodriguez, and L. Massoulie, “Providing Video-on-Demand Using Peer-to-Peer Networks,” Microsoft Research Technical Report, MSR-TR-2005-147, Oct. 2005.