1. Transport of Real-Time Flows
Bernd Girod
Information Systems Laboratory
Stanford University

2. THE MEANING OF FREE SPEECH
The acquisition by eBay of Skype is a helpful reminder to the world's trillion-dollar telecoms industry that all phone calls will eventually be free . . .
. . . Ultimately—perhaps by 2010—voice may become a free internet application, with operators making money from related internet applications like IPTV . . .
[Economist, September 2005]

3. IPTV Rollout Verizon 10M households by 2009 IPTV SBC 18M households
[IEEE Spectrum, Jan. 2005]

4. Why Is Real-Time Transport Hard?
Internet is a best-effort network . . .
Congestion Insufficient rate to communicate
Packet loss Impairs perceptual quality
Delay Impairs interactivity of services;
Telephony: one way delay < 150 ms [ITU-T Rec. G.114]
Delay jitter Obstructs continuous media playout

5. Clean Slate Internet Design Research Areas
Network Architecture
Heterogeneous
Applications
Physical Layers
Security
Econo- mics &
Regu- lation

6. Outline of the Talk QoS vs. best effort Resource allocation for IPTV
Rate-distortion optimized streaming

7. How 1B Users Share the Internet
Rate r
TCP Throughput
RTT
maximum
transfer
unit
data rate
Growing congestion p
packet
loss rate
0.0001
0.001
0.01
0.1
round
trip time
[Mahdavi, Floyd, 1997]
[Floyd, Handley, Padhye, Widmer, 2000]

8. QoS vs. Best Effort Reservation-ism Best Effort-ism
Voice and video need guaranteed QoS (bandwidth, loss, delay)
Implement admission control: “Busy tone” when network is full
Best effort is fine for data applications
Best Effort-ism
Best Effort good enough for all applications
Real-time applications can be made adaptive to cope with any level of service
Overprovisioning always solves the problem, and it’s cheaper than QoS guarantees

9. Simple Model of A Shared Link
Link of capacity C is shared among k flows
Fair sharing: each flow uses data rate C/k
Homogeneous flows with same utility function u(.)
Total utility C
[Breslau, Shenker, 1998]

10. Rigid Applications u Utility u=0 below of minimum bit-rate B
Maximum total utility U=k* is achieved by admitting at most k* flows 1 B
C/k
[Breslau, Shenker, 1998]

11. Rigid Applications (cont.)
Expected loss in total utility w/o admission control
Gap DU is substantial when number of admissable flows k* is small
Gap DU usually disappears with growing capacity C  Overprovisioning can solve the problem!
[Breslau, Shenker, 1998]

12. Elastic Applications
Elastic applications: utility function u(k), such that total utility U(k)=ku(C/k) increases with k
Example: u(C/k)=1-aC/k
All flows should be admitted: best effort! u
C/k

13. Video Compression H.264 video coding for 2 different testsequences
Video is elastic application
Rate must be adapted to network throughput
How to achieve rate control for stored content or multicasting?
Utility function depends on content: should use unequal rate allocation
Good
picture
quality
Foreman
Mobile
Bad
picture
quality

14. Different Utility Functions
Example: uk(rk)=1-akrk
With rk>=0  Karush-Kuhn-Tucker conditions (“reverse water-filling”)
Better than utility-oblivious “fair” sharing
Equal-slope “Pareto condition” uk
Vilfredo Pareto rk

15. Distribution of IPTV over WLAN
5 Mbps
11 Mbps
2 Mbps
Home Media
Gateway
[courtesy: van Beek, 2004]

16. Video Streaming Over Shared Channel
Transcoder
Decoder
Transcoder
Decoder 1 1
Receiver
Transcoder
Decoder
(Multi-Channel) 2 2
Transcoder
Decoder 3 3
Controller
[Kalman, van Beek, Girod 2005]

17. Dynamically Changing RD Function
Scene Cut
Weighted average of last I-Frame, P-Frame and B-Frame used for RD estimation

18. Utility-Based Resource Allocation
Equal Time Allocation
Min-MSE Allocation
Average PSNR: 37.5 dB
Average PSNR Stream 4: 32.9 dB
Average PSNR: 37.6 dB
Average PSNR Stream 4: 35.5 dB

19. Tx Backlog for 4 Video Streams 85% WLAN Utilization
[Kalman, van Beek, Girod 2005]

20. Streaming of Stored Content
Media files are already compressed:
How can we nevertheless adapt to network?
100s to 1000s simultaneous
streams
DSL
Cable
wireless
Server
Network
Client

21. Not All Packets are Equally ImportantB P B P B P I B P B P B P … … … A A …

22. Not All Packets are Equally ImportantB P B P B P I B P B P B P … … … A A …

23. Distortion-Aware Packet Dropping
Good
Picture
quality
Distortion aware
Packet dropping
No retransmissions
QCIF Carphone
I-P-P-P-P-P
Bad
picture
quality
Oblivious
Percentage of Packets Retained [%]
[Chakareski, Girod, ICME 2004]

24. Rate-Distortion Optimized (RaDiO) Streaming
“Decide which packets to send (and when) to maximize picture quality while not exceeding an average rate” [2001]
Repeat
request
Repeat
request
Repeat
request
Request
stream
Video data
Packet schedule
Rate-distortion
preamble
Server
Network
Client

25. A Brief History of Media Streaming
Media streaming w/o congestion avoidance: “reckless driving” [1995]
TCP-friendly rate control: “Limit average rate for fair sharing with TCP” [1997]
Rate-distortion optimized packet scheduling (RaDiO): “Decide which packets to send (and when) to maximize picture quality while not exceeding an average rate” [2001]
Congestion-distortion-optimized scheduling/routing (CoDiO): “Decide which packets to send (and when) to maximize picture quality while minimizing network congestion.” [2004]

26. Congestion vs. Rate Congestion: queuing delay that packets experience
weighted by size of the packet
averaged over all packets in the network
Congestion increases nonlinearly with link bit-rate
Congestion D
[seconds]
Rate R
Rmax

27. Video Distortion with Self Congestion
Good
Picture
quality
Self congestion causes late loss
Bad
picture
quality
Bit-Rate [kbps]

28. Streaming with Last Hop Bottleneck
Random cross traffic
High bandwidth links
Video traffic
Low bandwidth
last hop
Acknowledgments

29. Delay distribution pdf Overall delay distribution
Queue length determines delay of last hop
pdf
delay C

30. Comparison RaDiO vs. CoDiO
50 %
PSNR [dB]
PSNR [dB]
Rate [kbps]
End-to-end delay [ms]
Simulations using H.263+
Rate : 10 fps
Sequence : Foreman (32kbps,32kbps)
Sequence length : 60s
Playout deadline : 600ms

31. Concluding Remarks Over-provisioning makes QoS superfluous
Elastic applications don’t need QoS
Joint rate control for access bottlenecks (e.g. IPTV, WLAN)
Media-aware congestion control (e.g. CoDiO)

32. The End