Analysis of Adaptive Streaming for Hybrid CDN/P2P Live Video Systems
1. ANALYSIS OF ADAPTIVE STREAMING
FOR HYBRID CDN/P2P LIVE VIDEO
SYSTEMS
Ahmed Mansy and Mostafa Ammar
School of CS, GIT
Presented by Tangkai
2. ABOUT THE AUTHOR
Ahmed Mansy
PhD Student
scalable adaptive video streaming over the Internet.
message ferry routing in Disruption Tolerant
Networks (DTNs).
Mostafa Ammar
Regents’ Professor & Associate Chair
General Interest: Computer Network Architectures
and Protocols.
Current Specific Interests: Overlay
Networks, Network Virtualization, Mobile Wirless
Networks, Disruption Tolerant Networks.
3. OUTLINE
Introduction
System Description
Single Rate System Model
Adaptive Hybrid Live Video Streaming
Analysis Validation
Illustrative Case Study
4. INTRODUCTION
Video ~ dominate traffic of the Internet.
33% in 2010 ~ 57% in 2014 (expected)
Streaming stored or live video exclude P2P sharing
CDN ~ pillar of the video distribution
Aim: delay and throughput
CDN -> edge server
CDN + adaptive streaming => DASH
5. INTRODUCTION
P2P streaming
280 PT/month in 2009
P2P + adaptive streaming => layered streaming
Cons:
Complicated (design)
High processing power (client)
Not attractive for commercial use
Pros:
Cost-efficiency
CDN/P2P Hybrid System
6. RELATED WORK
Previous works[8][11] on designing such system
LiveSky: operational commercial sys
10m users
1st work study adaptive streaming in CDN/P2P hybrid sys
[8] C. Huang, J. Li, , and K. Ross, “Can internet video-on-demand be profitable?” in Sigcomm, 2007.
[11] Hao Yin and Xuening Liu and Tongyu Zhan and Vyas Sekar and Feng Qiu and Chuan Lin and Hui Zhang
and and Bo Li, “Design and Deployment of a Hybrid CDN-P2P System for Live Video Streaming: Experience
with LiveSky,” in Multimedia, 2009.
7. IDENTIFY THE PROBLEM
Assumption
Static in client: no switch/wired ap/constant bw
Dynamic in process: departure and arrival
Bitrate adaption strategy
Linear optimization problem to obtain best suitable bitrate
CDN/P2P mode switch rule
Stochastic fluid model to obtain lower bound of num of user
Interaction between two decision and how they affect each
other
8. OUTLINE
Introduction
System Description
Single Rate System Model
Adaptive Hybrid Live Video Streaming
Analysis Validation
Illustrative Case Study
9. DNS REDIRECTION
Typical
DNS
Lookup
4. Root
DNS
Server
9.
Perfor
ming
cache
11. DNS REDIRECTION
[16]
[16] A.-J. Su, D. Choffnes, A. Kuzmanovic, and F. Bustamante, “Drafting behind akamai,” in
SIGCOMM, 2006.
12. OUTLINE
Introduction
System Description
Single Rate System Model
Adaptive Hybrid Live Video Streaming
Analysis Validation
Illustrative Case Study
13. SINGLE RATE SYSTEM MODEL
Definition
Seeder/leecher
Directly connected to CDN
Unconstrained/constrained
Unlimited number of connections to other peers
Churnless/churn
Fixed number of client
Assumption
Upload rate of all seeder or leecher are the same
(l ) (s)
ui ul uj us
14. SINGLE RATE SYSTEM MODEL
Unconstrained churnless system
To support r, at least ns seeder
nsu s nl u l nl r ul
r ns
nl nl
15. SINGLE RATE SYSTEM MODEL
Unconstrained churn system
Assumption:
User arrival follows Poisson process with rate λ[19]
User stay in sys for a period of time follows general probability
distribution with mean 1/γ
Churn happens in leech node only
Total number of user in system N(t) ~ Poisson distribution
with rate ρ= λ/γ
Simple admission policy
[19] K. Sripanidkulchai, B. Maggs, and H. Zhang, “An analysis of live streaming workloads on the
internet,” in Internet Measurement Conference (IMC), 2004.
16. SINGLE RATE SYSTEM MODEL
Formulation
Poisson distribution(large ρ) -> Gaussian distribution
Low bound
17. SINGLE RATE SYSTEM MODEL
Constrained churnless system
Def
Sin number of incoming connection a seeder can accept.(s<-l)
Yin number of incoming connection a leecher can accept.(l<-l)
Yout number of connection leecher can initiate. (l->l+s)
η as the efficiency of the P2P protocol.
Probability leecher can find new content in other leechers.
d as the average download rate for any leecher
18. SINGLE RATE SYSTEM MODEL
= average num of seeder
connected to each leecher.
Average leecher download rate is not directly related to the
constraints of the system Sin/Yin.
only difference is η with unconstrained churnless sys.
19. SINGLE RATE SYSTEM MODEL
Constrained system with churn
Estimation -> bound
N ~ Gaussian dist( ), (1 − α) confidence interval
,
20. SINGLE RATE SYSTEM MODEL
is inversely proportional to ρ which means that the higher
client arrival rates λ and the longer clients stay in the system
1/γ, the lower becomes.
High guarantee of number of seeder
22. OUTLINE
Introduction
System Description
Single Rate System Model
Adaptive Hybrid Live Video Streaming
Analysis Validation
Illustrative Case Study
23. ADAPTIVE HYBRID LIVE VIDEO STREAMING
Problem
Which clients should be downgraded to streams of lower
bitrates?
What should these new lower bitrates be?
How to get an optimal allocation of bitrates to clients while
minimizing client downgrading?
Does the adaptive solution always exist?
Object
client dissatisfaction: difference between bitrate it requested
and it actually received
Minimize total client dissatisfaction over all clients.
24. ADAPTIVE HYBRID LIVE VIDEO STREAMING
Unconstrained churnless system
Def:
Bitrates provided by the CDN r1 > r2 > ... > rR
Define xij as the fraction of clients that request bitrate ri but receive
bitrate rj
25. ADAPTIVE HYBRID LIVE VIDEO STREAMING
Linear Optimization problem has a solution. values for xij
and ns i
ns the number of seeders that should receive video of bitrate ri
i
from the proxy.
ns =0
i
bitrate ri will not be supported by the server
no clients requested bitrate ri
some clients requested ri but the server decided not to deliver it and
downgraded these clients to lower bitrates
ns >0
i
does not necessarily mean some clients requested bitrate ri
it could mean that no clients requested rate ri but the server chose to
downgrade some of the clients
xij randomly choose fraction of leecher requested ri and delivered rj
26. ADAPTIVE HYBRID LIVE VIDEO STREAMING
Unconstrained churn system
client will request a video stream of bitrate with probability
where λ is the general client arrival rate
number of clients of bitrate at any time in the system
becomes a Poisson random variable with an average
Non-linear optimization problem. Use a linear approximation
27. ADAPTIVE HYBRID LIVE VIDEO STREAMING
Constrained churnless system
Constrained churn system
28. ADAPTIVE HYBRID LIVE VIDEO STREAMING
CDN adaptive live streaming
Ce r 1
guarantees with confidence (1 − α) that edge server capacity will
be sufficient for providing bitrate r to arriving clients with rate ρ.
29. ADAPTIVE HYBRID LIVE VIDEO STREAMING
CDN v.s. Hybrid Performance
Churnless
Linear optimzation problem -> xij
Churn
approximation
30. OUTLINE
Introduction
System Description
Single Rate System Model
Adaptive Hybrid Live Video Streaming
Analysis Validation
Illustrative Case Study
31. ANALYSIS VALIDATION
Validate single bitrate streaming only
On BitTorrent
Tracker: proxy
Seeder: download torrent and video files
Leecher: download torrent
Parameter
10s chuck
Us/Ul 350kbps/500kbps
ρ 100~400 clients/hour
γ ~ mixed-exponential distribution PDF
Sin = 20, Yin = 10
32. ANALYSIS VALIDATION
Solid line means enough seeder to support bitrate
Fig 4(a) – Fig2(b)
34. OUTLINE
Introduction
System Description
Single Rate System Model
Adaptive Hybrid Live Video Streaming
Analysis Validation
Illustrative Case Study
35. ILLUSTRATIVE CASE STUDY
Metric
Inter-client fairness
Request and actually received
Saving in CDN server capacity
Profile
low/uniform/high (for bitrate)
36. ILLUSTRATIVE CASE STUDY
Inter-client fairness
Single bitrate manner
Downgrade for all if overloaded.
Adaptive: fairness drop
Single bitrate
Start at lower than 100%/Constant/even better
37. ILLUSTRATIVE CASE STUDY
Capacity saving
Fairness->100%
Saving is less in high profile: asymmetric bw(US/China)