1. CING-YU CHU 2012.08.06 INFOCOM 2012

2. Outline  Introduction  Measurement  Measurement Results  Modeling Skype Behaviors  Analysis on TCP-friendly

3. Motivation  Skype VoIP service is well studied while video service is not  Skype video service consumes more bandwidth Up to 950 kbps  Imperative for network providers and network researchers

4. Key Questions  Q1: How does a Skype video call adapt its sending rate, video rate and quality under different network conditions?  Q2: Are Skype video calls friendly to TCP flows when they compete for network resources?

5. Methodology  Measurement Black-box approach Different network setting with ○ configurable packet loss, ○ propagation delay ○ available bandwidth Aim to measure ○ sending rate ○ throughput ○ RTT ○ video bit rate ○ frame rate

6. Methodology  Modeling Rate control model FEC model Video quality model  Analysis User back-off ○ User-level rate control scheme TCP-friendliness

7. Contribution  Measures Skype’s stationary behaviors of video calls Sending rate is insensitive to packet loss when PLR < 10% Utilization of the available bandwidth is around 80% Overly aggressive FEC scheme, 4.5 times the PLR

8. Contribution  Derive various models to verify User back-offs react fast to congestion Skype video calls are TCP-friendly ○ Due to quality-driven user back-offs

9. Measurement Test-bed

10. Experiment Design  TV news video sequence “Akiyo” From Joint Video Team (JVT) Head and shoulder movements  Virtual video camera tool  Data collection TCP-dump Skype technical reports

11. Skype Video Call  On2 video codec Video quantization step Video resolution Number of frames per seconds (FPS)

12. Measurements Results  Impact of Packet Loss Available Bandwidth Propagation Delay

13. Impact of Packet Loss  PLR varies from 0% to 12%  Propagation delay: 50ms  3 available bandwidth settings 250 kbps 750 kbps 1000 kbps

14. Impact of Packet Loss  Forward Error Correction  Two states PLR NORM state PLR >= 10% => CONS state  Different from TCP congestion control scheme

15. Impact of Available bandwidth  Available Bandwidth varies from 50 kbps to 1000kbps  Two PLRs: 2% and 10%  Propagation delay: 50ms

16. Impact of Propagation Delay  Propagation delay varies from 50ms to 2000ms  Available bandwidth: 500 kbps  PLR: 0%

17. Modeling  Sending rate  Video rate  Video Quality

18. Sending Rate Model  NORM and CONS states  25 scenarios with PLR: 0% to 12% Available bandwidth: 50 kbps to 1000 kbps  γ = 0.77, μ = -10.8 and δ = 21

19. Video Rate Model  FEC ratio  FEC ratio model Ψ = 0.15 and ω = 4.5

20. Video Quality Model  ITU-T Recommendation G.1070 frame rate video rate

21. Video Quality Model  Frame rate model

22. Video Quality Model  a = 1.431, b = 0.02228, c = 3.759, d = 184.1, e = 1.161, h = 1.446 and g = 0.03881

23. Model Validation  Co-current UDP traffic (from iPerf) 0 kbps to 600 kbps  Link capacity: 700 kbps  Propagation delay: 50ms  Available bandwidth for Skype

24. Model Validation Pearson Correlation Coefficient Sending rateVideo rateFrame rate 0.98980.98310.9545

25. Analysis  Q1: How Skype video call users respond to quality degradation resulted from network impairments? How effective user back-offs are as a user-level rate control scheme?  Q2: What is the performance of a Skype video call when it competes with other Skype calls and TCP flows? Is Skype video call TCP-friendly?

26. Network Model  LTE wireless network Multiple TCP and Skype users M/M/1/K queue using drop-tail Downlink: 100 Mbps, uplink: 50 Mbps

27. TCP Model  Reacts to packet loss and RTT p = p q + p c t = t q + t c

28. User-level Rate Control  Video drop-off probability  Number of active users  Effective traffic generated by all users

29. User-level Rate Control  Average traffic of each user Expected Skype sending rate

30. Responsiveness to Loss

31. Responsiveness to Delay

32. Competition with TCP  p c = 2% and t c = 50ms  N T = # of TCP users, N S = # of Skype users  Aggregate traffic  For each Skype user

33. Competition with TCP  Scale-up factor: k

34. Conclusion  Measures Skype video traffic Shows that Skype is robust against mild packet loss and propagation delay Skype can efficiently utilize available bandwidth  Models Skype video behaviors  Shows that Skype video is indeed TCP- friendly Based on user back-off rate control scheme

35. Q & A