Presentation is loading. Please wait.

Presentation is loading. Please wait.

1 USC INFORMATION SCIENCES INSTITUTE An End-to-end Architecture for Quality- Adaptive Streaming Applications in Best- effort Networks Reza Rejaie

Similar presentations


Presentation on theme: "1 USC INFORMATION SCIENCES INSTITUTE An End-to-end Architecture for Quality- Adaptive Streaming Applications in Best- effort Networks Reza Rejaie"— Presentation transcript:

1 1 USC INFORMATION SCIENCES INSTITUTE An End-to-end Architecture for Quality- Adaptive Streaming Applications in Best- effort Networks Reza Rejaie reza@isi.edu USC/ISI http://netweb.usc.edu/reza April 13, 1999

2 2 USC INFORMATION SCIENCES INSTITUTE Motivation n Rapid growth in deployment of realtime streams(audio/video) over the Internet n TCP is inappropriate for realtime streams ¶ The Internet requires end-system to react to congestion properly and promptly Ë Streaming applications require sustained consumption rate to deliver acceptable and stable quality

3 3 USC INFORMATION SCIENCES INSTITUTE Best-effort Networks (The Internet) n Shared environment Š Bandwidth is not known a prior Š Bandwidth changes during a session Š Seemingly-random losses n TCP-based traffic dominates n End-to-end congestion control is crucial for stability, fairness & high utilization Ô End-to-end congestion control in a TCP-friendly fashion is the main requirement in the Internet

4 4 USC INFORMATION SCIENCES INSTITUTE Streaming Applications n Delay-sensitive n Semi-reliable n Rate-based Ô Require QoS from the end-to-end point of view Internet Adaptation Buffer Decoder TCP Server Display Encoder Source

5 5 USC INFORMATION SCIENCES INSTITUTE ¶ Designing an end-to-end congestion control mechanism · Delivering acceptable and stable quality while performing congestion control The Problem

6 6 USC INFORMATION SCIENCES INSTITUTE Outline Ô The End-to-end Architecture Š Congestion Control (The RAP protocol) Š Quality Adaptation n Extending the Architecture Š Multimedia Proxy Caching n Contributions n Future Directions

7 7 USC INFORMATION SCIENCES INSTITUTE Buffer Manager Archive Error Control Quality Adaptation Transmission Buffer Cong. Control Acker Decoder Playback Buffer Internet Server Client Adaptation Buffer Data path Control path The End-to-end Architecture Buffer Manager

8 8 USC INFORMATION SCIENCES INSTITUTE Outline n The End-to-end Architecture Ô Congestion Control (The RAP Protocol) Š Quality Adaptation n Extending the Architecture Š Multimedia Proxy Caching n Contributions n Future Directions

9 9 USC INFORMATION SCIENCES INSTITUTE Previous works on Congestion Ctrl. n Modified TCP Š [Jacob et al. 97], SCP[Cen et al. 98] n TCP equation Š [Mathis et al. 97], [Padhye et al. 98] n Additive Inc., Multiplicative Dec. Š LDA[Sisalem et al. 98] n NETBLT[Lixia!] Ô Challenge: TCP is a moving target

10 10 USC INFORMATION SCIENCES INSTITUTE Overview of RAP n Decision Function n Increase/Decrease Algorithm n Decision Frequency Ô Goal: to be TCP-friendly Time Rate Decision Frequency Decision Function + -- Increase/Decrease Algorithm

11 11 USC INFORMATION SCIENCES INSTITUTE Congestion Control Mechanism n Adjust the rate once per round-trip-time (RTT) n Increase the rate periodically if no congestion n Decrease the rate when congestion occurs Š Packet loss signals congestion n Cluster Loss Š Grouping losses per congestion event

12 12 USC INFORMATION SCIENCES INSTITUTE Rate Adaptation Algorithm n Coarse-grain rate adaptation Š Additive Increase, Multiplicative Decrease (AIMD) n Extensive simulations revealed: Š TCP’s behavior substantially varies with network conditions, e.g. retransmission timeout, bursty Š TCP is responsive to a transient congestion Ô AIMD only emulates window adjustment in TCP

13 13 USC INFORMATION SCIENCES INSTITUTE Rate Adaptation Algorithm(cont’d) n Fine-grain rate adaptation Š The ratio of short-term to long-term average RTT n Emulates ACK-clocking in TCP n Increase responsiveness to transient congestion

14 14 USC INFORMATION SCIENCES INSTITUTE Coarse vs fine grain RAP fig Impact of fine-grain rate adaptation

15 15 USC INFORMATION SCIENCES INSTITUTE n RAP against Tahoe, Reno, NewReno & SACK n Inter-dependency among parameters n Config. parameters : Š Bandwidth per flow Š RTT Š Number of flows RAP Sinks TCP Sinks RAP Traffic SW TCP Traffic SW RAP Sources TCP Sources RAP Simulation Fairness Ratio = Avg. RAP BW Avg. TCP BW

16 16 USC INFORMATION SCIENCES INSTITUTE Fairness ratio across the parameter space without F.G. adaptation

17 17 USC INFORMATION SCIENCES INSTITUTE Fairness ratio across the parameter space with F.G. adaptation

18 18 USC INFORMATION SCIENCES INSTITUTE Impact of RED switches on Fairness ratio

19 19 USC INFORMATION SCIENCES INSTITUTE Summary of RAP Simulations n RAP achieves TCP-friendliness over a wide range Š Fine grain rate adaptation extends inter-protocol fairness to a wider range n Occasional unfairness against TCP traffic is mainly due to divergence of TCP congestion control from AIMD Š Pronounced more clearly for Reno and Tahoe Š The bigger TCP’s congestion window, the closer its behavior to AIMD n RED gateways can improve inter-protocol sharing Š Depending on how well RED is configured Ô RAP is a TCP-friendly congestion controlled UDP

20 20 USC INFORMATION SCIENCES INSTITUTE Outline n The End-to-end Architecture Š Congestion Control (The RAP protocol) Ô Quality Adaptation n Extending the Architecture Š Multimedia Proxy Caching n Contributions n Future Directions

21 21 USC INFORMATION SCIENCES INSTITUTE Buffer Manager Archive Error Control Quality Adaptation Transmission Buffer Cong. Control Acker Decoder Playback Buffer Internet Server Client Adaptation Buffer Data path Control path Quality Adaptation Buffer Manager

22 22 USC INFORMATION SCIENCES INSTITUTE The Problem · Delivering acceptable and stable quality while performing congestion control n Seemingly random losses result in random & potentially wide variations in bandwidth n Streaming applications are rate-based

23 23 USC INFORMATION SCIENCES INSTITUTE Role of Quality Adaptation n Buffering only absorb short-term variations n Long-lived session could result in buffer overflow or underflow n Quality Adaptation is complementary for buffering Ô Adjust the quality with long-term variations in bandwidth BW(t) Time

24 24 USC INFORMATION SCIENCES INSTITUTE n Adaptive encoding [Ortega 95, Tan 98] Š CPU-intensive n Switching between multiple encoding Š High storage requirement n Layered encoding[McCanne 96, Lee 98] Š Inter-layer decoding dependency Ô When/How much to adjust the quality? Mechanisms to Adjust Quality

25 25 USC INFORMATION SCIENCES INSTITUTE n Assumptions Š AIMD variations in bandwidth(rate) Š Linear layered encoding n Constraint Š Obeying congestion controlled rate limit n Goal Š To control the level of smoothing Assumptions & Goals

26 26 USC INFORMATION SCIENCES INSTITUTE Layered Quality Adaptation Layer 2 Layer 1 Layer 0 + bw (t) 2 1 0 Internet Decoder Time(sec) BW(t) bw (t) 1 0 C C C Display Linear layered stream buf 0 1 2 bw (t) 2 Quality Adaptation Consumption rate C Time(msec) BW(t) ac b Filling Phase Draining Phase C C

27 27 USC INFORMATION SCIENCES INSTITUTE Buffering Tradeoff n Each buffering layer can only contribute at most C(bps) Š Buffering for more layers provides higher stability bw (t) 1 0 C C C buf 0 1 bw (t) 2 Time BW(t) nC buf 2 C C C C n Buffered data for a dropped layer is useless for recovery Š Buffering for lower layers is more efficient Ô What is the optimal buffer distribution for a single back-off scenario?

28 28 USC INFORMATION SCIENCES INSTITUTE n Optimal buffer state depends on time of the back-off n Draining pattern depends on the buffer state n Back-off occurs randomly Ô Keep the buffer state as close to the optimal as possible during the filling phase Time BW(t) Draining Phase 4C C C Buf. data & Filling Phase Optimal Inter-layer Buffer Allocation Buf. data & BW share of L0 BW share of L1 BW share of L2

29 29 USC INFORMATION SCIENCES INSTITUTE n Add a layer when buffering is sufficient for a single back-off n Drop a layer when buffering is insufficient for recovery n Random losses could result in frequent add and drop Š unstable quality Ô Conservative adding results in smooth changes in quality Time Adding & Dropping BW(t) Draining Phase nC Draining Phase (n-1)C Buf. data for L0 Buf. data for L1 Buf. data for L2

30 30 USC INFORMATION SCIENCES INSTITUTE n Conservative adding Š When average bandwidth is sufficient Š When sufficient buffering for K back-offs n Buffer constraint is preferred and sufficient Š Directly relate time of adding to the buffer state Š Effectively utilizes the available bandwidth n K is a smoothing factor Š Short-term quality vs long-term smoothing Smoothing

31 31 USC INFORMATION SCIENCES INSTITUTE Proper Buf. State recovery from 1 backoff Proper Buf. State recovery from 2 backoffs Proper Buf. State recovery from K backoffs Add a Layer Drop a Layer Filling Smooth Filling & Draining Draining

32 32 USC INFORMATION SCIENCES INSTITUTE KB/s C = 10 Buf. L3 (KB) 9.5 Buf. L2 (KB) Buf. L1 (KB) Buf. L0 (KB) 17.5 Buf. L3 (KB) Buf. L2 (KB) Buf. L1 (KB) Buf. L0 (KB) 17.5 KB/s C = 10 40 Time(sec) Effect of smoothing factor (K = 2) (K = 4) 40 Time(sec) TX rate & Quality TX rate & Quality

33 33 USC INFORMATION SCIENCES INSTITUTE Buf. L3 (KB) C = 10 17.5 KB 90 Time(sec)3060 (K = 4) 90 Time(sec) 3060 Adapting to network load KB/s TX rate & Quality

34 34 USC INFORMATION SCIENCES INSTITUTE No of Dropped Layers

35 35 USC INFORMATION SCIENCES INSTITUTE Summary of the QA results n Quality adaptation mechanism can efficiently control the quality n Smoothing factor allows the server to trade short-term improvement with long-term smoothing n Buffer requirement is low Ô Deploying for live but non-interactive sessions!

36 36 USC INFORMATION SCIENCES INSTITUTE n Delivered quality is limited to the average bandwidth between the server and client n Solutions: Š Mirror servers Š Multimedia proxy caching Limitation of the E2E Approach Server Client Internet Client Time L 0 L 1 L 2 L 3 L 4 Quality( layer )

37 37 USC INFORMATION SCIENCES INSTITUTE Outline n The End-to-end Architecture Š Congestion Control (The RAP protocol) Š Quality Adaptation n Extending the Architecture Ô Multimedia Proxy Caching n Contributions n Future Directions

38 38 USC INFORMATION SCIENCES INSTITUTE Server n Assumptions Š Proxy can perform: –End-to-end congestion ctrl –Quality Adaptation n Goals Š Improve delivered quality Š Low-latency VCR-functions Š Natural benefits of caching Proxy Internet Multimedia Proxy Caching Client

39 39 USC INFORMATION SCIENCES INSTITUTE Challenge n Cached streams have variable quality Ô Layered organization provides opportunity for adjusting the quality Time L 0 L 1 L 2 L 3 L 4 Quality ( layer ) Stored stream Played back stream

40 40 USC INFORMATION SCIENCES INSTITUTE Issues n Delivery procedure Š Relaying on a cache miss Š Pre-fetching on a cache hit n Replacement algorithm Š Determining popularity Š Replacement pattern

41 41 USC INFORMATION SCIENCES INSTITUTE Cache Miss Scenario n Stream is located at the original server n Playback from the server through the proxy n Proxy intercepts and caches the stream n No benefit in a miss scenario Server Internet Proxy Client

42 42 USC INFORMATION SCIENCES INSTITUTE Cache Hit Scenario n Playback from the proxy cache Š Lower latency Š May have better quality! n Available bandwidth allows: Š Lower quality playback Š Higher quality playback Server Proxy Internet Client

43 43 USC INFORMATION SCIENCES INSTITUTE Lower quality playback n Missing pieces of the active layers are pre- fetched on-demand n Required pieces are identified by QA n Results in smoothing Time L 0 L 1 L 2 L 3 L 4 Quality ( no. active layers ) Pre-fetched data Stored stream Played back stream

44 44 USC INFORMATION SCIENCES INSTITUTE n Pre-fetch higher layers on-demand n Pre-fetched data is always cached n Must pre-fetch a missing piece before its playback time n Tradeoff Time L 0 L 1 L 2 L 3 L 4 Quality ( no. active layers ) Pre-fetched data Stored stream Played back Stream Higher quality playback

45 45 USC INFORMATION SCIENCES INSTITUTE Replacement Algorithm n Goal: converge the cache state to optimal Š Average quality of a cached stream depends on –popularity –average bandwidth between proxy and recent interested clients Š Variation in quality inversely depends on –popularity Server Proxy Internet Client

46 46 USC INFORMATION SCIENCES INSTITUTE n Number of hits during an interval n User’s level of interest (including VCR- functions) n Potential value of a layer for quality adaptation Š Calculate whit on a per-layer basis n Layered encoding guarantees monotonically decrease in popularity of layers Popularity whit = PlaybackTime(sec) / StreamLength(sec)

47 47 USC INFORMATION SCIENCES INSTITUTE n Multi-valued replacement decision for multimedia object n Coarse-grain flushing Š on a per-layer basis n Fine-grain flushing Š on a per-segment basis Fine-grain Coarse-grain Cached segment Replacement Pattern Time Quality(Layer)

48 48 USC INFORMATION SCIENCES INSTITUTE Summary of Multimedia Caching n Exploited characteristics of multimedia objs n Proxy caching mechanism for multimedia streams Š Pre-fetching Š Replacement algorithm Ô Adaptively converges state of the cache to the optimal

49 49 USC INFORMATION SCIENCES INSTITUTE Contributions n End-to-end architecture for delivery of quality-adaptive multimedia streams n RAP, a TCP-friendly cong. ctrl mechanism over a wide range of network conditions n Quality adaptation mechanism that adjusts the delivered quality with a desired degree of smoothing n Proxy caching mechanism for multimedia streams to effectively improve the delivered quality of popular streams

50 50 USC INFORMATION SCIENCES INSTITUTE Future Directions n End-to-end Congestion Control Š RAP’s behavior in the presence web-like traffic Š Emulating timer-driven regime TCP Š Bi-directional RAP connections, Reverse ns forward path congestion control Š Experiments over CAIRN & the Internet Š Integration of RAP and congestion manager Š Adopting RAP into class-based QoS Š Using RAP for multicast congestion control Š Congestion control over wireless networks

51 51 USC INFORMATION SCIENCES INSTITUTE Future Directions(cont’d) n Quality Adaptation Š Extending to other rate adaptation mechanisms n Multimedia Proxy Caching Š Other replacement patterns & popularity functions(e.g. chunk-based) n Traffic Measurement and Characterization Š Imiprical evaluation of streaming applications

52 52 USC INFORMATION SCIENCES INSTITUTE An End-to-end Architecture for Quality- Adaptive Streaming Applications in Best- effort Networks Reza Rejaie reza@isi.edu USC/ISI http://netweb.usc.edu/reza April 7, 1999

53 53 USC INFORMATION SCIENCES INSTITUTE Thank you Reza Rejaie http://netweb.usc.edu/reza reza@isi.edu

54 54 USC INFORMATION SCIENCES INSTITUTE Archive Media Server Internet TCP Traffic TCP Traffic Target Environment

55 55 USC INFORMATION SCIENCES INSTITUTE Optimal Buffer Allocation Scenario 1Scenario 2Scenario 3 Backoff 2 Optimal buffer state is not unique S1 and S2 are extreme cases S1 requires more buffering layers S2 requires more buffer share per layer Buffer allocation for S1 can recover from S2 but not vice versa


Download ppt "1 USC INFORMATION SCIENCES INSTITUTE An End-to-end Architecture for Quality- Adaptive Streaming Applications in Best- effort Networks Reza Rejaie"

Similar presentations


Ads by Google