1. Reza Rejaie AT&T Labs - Research1 Reza Rejaie AT&T Labs – Research Menlo Park, CA. http://www.research.att.com/~reza ICON 2000 In collaboration with Mark Handley(ACIRI at ICSI) Deborah Estrin (ISI/USC) Architectural Considerations for Playback of Quality Adaptive Video over the Internet

2. Reza Rejaie AT&T Labs - Research2 Motivation Rapid growth in client-server streaming over the Internet Internet applications should deploy end-to-end congestion control Streaming applications require QoS  Most of the existing solutions do not properly integrate congestion control into Internet streaming applications

3. Reza Rejaie AT&T Labs - Research3 Unicast Streaming Applications Delay-sensitive Semi-reliable Rate-based  Require QoS from the end-to- end point of view Internet Adaptation Buffer Decoder TCP Server Display Encoder Source

4. Reza Rejaie AT&T Labs - Research4 Goal of this paper To provide high level architectural insights in the design of unicast Internet streaming applications  Guidelines for Internet streaming application designers

5. Reza Rejaie AT&T Labs - Research5 Outline Design principles for Internet applications  Key components Design space for key components of Internet streaming applications An architecture Generalizing the architecture Conclusion & future directions

6. Reza Rejaie AT&T Labs - Research6 Design Principles for Internet App. Social behavior Being adaptive Recovery from losses

7. Reza Rejaie AT&T Labs - Research7 Social Behavior Network resources are shared End-systems should adjust their tx rate based on state of the network Rapidly decrease when congestion occurs Periodically increase to probe for spare capacity  End-to-end Congestion Control Congestion control should result in fairness, stability and high utilization  Congestion control is network-dependent Design Principles

8. Reza Rejaie AT&T Labs - Research8 Being Adaptive 1) Bandwidth heterogeneity 2) Shared resources  Internet applications should be able to:  operate over wide range of available bw  cope with wide & unpredictable variations in bw  Internet applications should be quality adaptive  Quality adaptation is application-specific Design Principles

9. Reza Rejaie AT&T Labs - Research9 Recovery from Loss Packets are pseudo-randomly lost in the network mainly due to congestion Internet applications should be able to recover from losses to meet required level of reliability  Error control  Error control is application-specific Design Principles

10. Reza Rejaie AT&T Labs - Research10 Key Components 1) Congestion Control(CC) 2) Quality Adaptation(QA) 3) Error Control(EC)  What is the design space for each key component in the context of Internet streaming applications?

11. Reza Rejaie AT&T Labs - Research11 Congestion Control Challenge: to design an effective Increase, Decrease algorithm? Goal: to be network-friendly if TCP traffic dominates => TCP-friendly Additive Inc., Multiplicative Dec. Slope of increase & decreasing factor Some existing TCP-friendly CC schemes: SCP, LDA, RAP, TFRC  Performing CC results in unpredictable and potentially wide variations in bandwidth Design Space

12. Reza Rejaie AT&T Labs - Research12 Mechanisms to Adjust Quality Adaptive encoding CPU-intensive Switching between multiple encoding High storage requirement Layered encoding Inter-layer decoding dependency Suited for multimedia proxy caching Design Space/Quality Adaptation

13. Reza Rejaie AT&T Labs - Research13 Quality Adaptation Adjust delivered quality with long-term changes in available bandwidth receiver buffering only absorbs short-term changes in bandwidth Challenge: frequent changes in delivered quality degrades perceptual quality encoding-specific More buffering allows smoother variations in delivered quality Buffering strategy? QA is the least understood aspect of streaming applications Design Space

14. Reza Rejaie AT&T Labs - Research14 Error Control Streaming applications require quality instead of complete reliability Techniques for loss repair are well understood Appropriate EC mechanism depends on: Required level of reliability by encoding Accepted delay for loss recovery Expected/measured loss pattern EC & QA should share the available bw in order to maximize delivered quality  Encoding-specific Design Space

15. Reza Rejaie AT&T Labs - Research15 Buffer Manager Archive Error Control Quality Adaptation Transmission Buffer Cong. Control Acker Decoder Playback Buffer Internet Server Client Adaptation Buffer Data path Control path Buffer Manager End-to-End Architecture

16. Reza Rejaie AT&T Labs - Research16 Generalizing the Architecture Architecture is generic. Knowing: 1) Designated func. to each component 2) Design space for each component 3) Effect of each design choice on other components Replace algorithms in each module Address implications on other modules Examples: Use a new CC mechanism => Adjust details of QA Change EC mechanism => Adjust bw sharing between QA & EC

17. Reza Rejaie AT&T Labs - Research17 Natural Design Steps Select a network-friendly CC scheme Select an EC mechanism that satisfies application requirements Design a QA mechanism that bridges the gap between app. requirements & CC maximizes perceptual quality for a given encoding Design a bandwidth sharing strategy for QA and EC mechanisms

18. Reza Rejaie AT&T Labs - Research18 Conclusion High level architectural insights in design of Internet streaming app. Key components => designated func. Design space of each component Inter-dependency among key components Set of high level design guidelines for streaming applications Applicable to live but non-interactive streaming applications

19. Reza Rejaie AT&T Labs - Research19 Future Directions Prototyping a sample architecture MPEG4 encoding Layered QA TCP-friendly Retx-based error control Extending the arch to include multimedia proxy caches (mcaches) To deliver high quality streams in a large scale Utility-based QA and More details: http://www.research.att.com/~reza