1. Receiver-Driven Bandwidth Sharing for TCP and its Application to Video Streaming Puneet Mehra, Christophe De Vleeschouwer, and Avideh Zakhor IEEE Transactions on Multimedia, August 2005

2. Introduction Last-hop connections are often the bottleneck on the Internet Many applications run together to compete for the bandwidth resource Propose Receiver-Driven Bandwidth Sharing System (BWSS) Apply on last-hop link Allocate bandwidth according to user preferences

3. System overview Objectives: Achieve full utilization of receiver ’ s access link Satisfy user preferences (how the bottleneck bandwidth should be shared) Essential idea: constrain the throughput of certain low priority flows to provide additional bandwidth to high priority flows

4. Block diagram of bandwidth sharing system (BWSS)

5. TCP flow control system (FCS) Achieve a particular bit-rate for a given TCP flow Input: desired bit rate Output: advertised window (an integer number of packets, must be greater than zero) Goal: R: actual rate, T: target rate T ω

6. FCS algorithm

7. FCS analysis ( ) T: target rate  set ω to Proceed in an iterative bandwidth estimation to adjust advertised window T ω R

8. FCS analysis

9. Measuring flow RTT and bandwidth Calculate RTT: use TCP Timestamp option Calculate bandwidth Estimate at the end of each period ψ, ψ is set to in all experiments, εis set to 0.3 in all experiments

10. Frequency to adjust the advertisement window Frequency φ is bounded based on: Round trip time (RTT) Bandwidth estimation period (ψ) φ should be greater than RTT+ψ We set φ>RTT+3 ψ Allow enough time for the window adjustment to affect on throughput

11. Bandwidth sharing system (BWSS) Separate “ priority ” and “ weight ” Some aps require minimal bandwidth, but the more is not necessarily the better, ex: streaming Some aps don ’ t require minimal bandwidth, but the more is the better, ex: ftp Minimal rate should be provided to each connection in decreasing order of priority The remaining bandwidth should be shared proportionally to the weight

12. Target rate allocation & receiver preferences If system ’ s total bit rate Else =min(, max(0, ) )

13. Congestion Congestion in certain connection Threshold: γT  allocate bandwidth to other application or not

14. Experiment environment Buffer incoming packets to limit bandwidth Throughput: 960kbps Additional delay: 30ms

15. Experiment 1: the ideal case First 40 seconds: standard TCP Weights: 2, 1, 3, respectively No minimal rate

16. Experiment 2: Link bandwidth reduction Weight: 1, 2, respectively 40~80s: a 320 kbps UDP stream

17. Experiment 3: bandwidth redistribution Stream minimal rate: 496 kbps 40~100s: congestion, stream ↓ 160kbps

18. Experiment 4: RealVideo streaming (under standard TCP)

19. Experiment 4: RealVideo streaming (under UDP)

20. Experiment 4: RealVideo streaming (under BWSS)

21. Conclusion Investigated a bandwidth-sharing system (BWSS) for TCP connections Allow users specify preferences Don ’ t need to change network infrastructure Experiment results show that streaming with BWSS offers superior performance Drawback: BWSS may suffer starving condition