1. Playback-buffer Equalization For Streaming Media Using Stateless Transport Prioritization By Wai-tian Tan, Weidong Cui and John G. Apostolopoulos Presented by Felix Lam

2. Motivation In video streaming, playback disruption occurs when the playback buffer depletes. Due to fluctuation of available capacity, some streaming sessions may experience playback disruption while some others sharing the same bottleneck link still have plenty of buffered data.

3. Motivation Playback Buffer Equalization using Stateless Transport Prioritization (PBE-STP) Equalize the playback buffers by dynamic allocation of bandwidth to low-buffer sessions. Use packet labeling and QoS Network

4. PBE-STP: Overview and Design Streaming Server Perform streaming and congestion control (e.g. TCP) Label packets based on playback buffer occupancy

5. PBE-STP: Overview and Design Packet Classifier Assign each packet to different available types of service based on the label QoS Network Offers at least two types of services differing in delay, loss rate or both Clients Normal streaming clients

6. Evaluation of three QoS implementations 9 constant bit-rate 1.2Mbps streams Movie length: 300 seconds 6 flows start at time 0, 3 flows start at 60, 120 and 180 seconds Upon disruption, rebuffer 5-seconds data before resuming Use 3-bit TOS (Type Of Service) field for packet labeling

7. Baseline: FIFO

8. Strict Priority Packets are put into the 8 priority queues according to their labels Packets in higher priority queues are ALWAYS delivered first.

9. Premium + Best Effort Services 2 classes of priority Throughput of high priority queue is restricted to λ * For each of the eight labels b, a running average of traffic load λ b is maintained A packet is assigned to high priority if its label b <= T, where T is updated as

10. Premium + Best Effort Services λ*=3Mbps

11. Hybrid Priority Queues Combine Priority queuing and round- robin Serve K packets based on priority queuing, then serve 1 packet based on round-robin

12. Hybrid Priority Queues K=15

13. Quantitative Results Simulated with ns-2 To quantify the performance gain over FIFO queuing under different propagation delays and network load Poisson arrival of αnew streams Constant media bit rates of 100, 300, 500kbps uniformly at random Stream duration drawn randomly according to an empirical distribution Loading factor = average aggregate streaming rate / Bottleneck bandwidth

14. Performance Metrics Disruption frequency Required Initial Buffer-time (RIB) Amount of initial buffer time that would allow 95% of streams to finish without any disruption n i - # of disruptions experienced by stream I f i – duration of stream i

15. Different Network Load  = 10ms Streams started randomly at S and S’ 66 independent simulations of one hour simulated each

16. Different Network Delay

17. Streams with Mixed Delay

18. Conclusions Combine packet labeling and QoS network to reduce the probability of playback disruption of multiple streams sharing the same bottleneck Distributed algorithm for different senders without explicit coordination

19. Comments Effect of Interfering UDP and TCP traffic is not considered Fairness issue not addressed Require modification on the network infrastructure, not practical unless the bottleneck is on local network However, it shows significant incentives to perform playback buffer equalization