1. Adaptive Multi-source Streaming in Heterogeneous Peer-to-peer Network Vikash Agarwa; Reza Rejaie Twelfth Annual Multimedia Computing and Networking (MMCN ’05)

2. Agenda Introduction Assumptions and Target Environment PAL –Sliding Window –Quality Adaptation –Packet Assignment Performance Evaluation Conclusion

3. Introduction P2P streaming requires –Overlay construction –Content Delivery Challenges –The available bandwidth from each sender is not known a priori and could significantly change during a session –Different senders have different bandwidth and RTT, or even leave

4. Introduction Multi-source MDC : Commonly used separate layer for each sender Drawback: 1.Available bandwidth may be too small 2.Available bandwidth may be too large 3.Unpredictable variation needs complicates inter-layer synchronization

5. Introduction Propose P2P Adaptive Layered Streaming (PALS) –Receiver driven –Heterogeneous senders –Periodically determine the target quality (i.e. the number of layers) that can be streamed –Relatively low amount of receiver buffering –Can be used in non-interactive streaming application

6. Assumptions & Target Environment 1.Request a single video from group of senders 2.UDP connections and TCP- friendly congestion control 3.Sender differs from bandwidth and RTT and bottlenecks 4.Assume provided information of senders (PRO) 5.Assume layered coded 6.Assume all layers have same constant bit rate, C.

7. PAL Mechanism 1.Receiver monitors EWMA bandwidth from each senders 2.Sliding Window (SW) a)T ewma = T j ewma b)Predict total number of incoming packet : K = (Tewma*∆)/pktsze 3.Quality Adaptation (QA) a)Number of active layer that can be played b)Packet from each layer to be request in future

8. PAL Mechanism 4.Packet Assignment (PA) a)Divides selected packets into disjoint subnet b)Sends a separate request to each sender c)Each sender maintains a single list of pending packets for each receiver peer

9. Sliding Window Δ : buffering window = 5*SRTT δ : minimum threshold. If exceeds -> slide window

10. Sliding Window Drawback : 1.Duplicate request -> low 2.Overestimating bandwidth -> overwriting request, packet ordering 3.Underestimating bandwidth -> idle, further request (window overlap)

11. Quality Adaptation Invoke once per window 1.If (T ewma >= n * C), utilize the excess bandwidth to request future packet BUF add : Total buffer data before adding a layer 2.If (T ewma < n * C), drain the buffer~~, drop top layer

12. Packet Assignment Weighted round-robin Partially Available content : two pass –Treat all senders have the entire content –Per-packet assignment

13. Performance Evaluation Window overlap = 20% N c = 5 C = 80KBytes/sec δ=2*SRTT max

14. Quality comparing single and multiple layer static streaming

15. Duplicate packet Low duplicate packets – 8% Duplicates decrease as window size increase Loss remains below 5%

16. Dynamic of bandwidth variation

17. Conclusion Multiple-source MDC streaming –Sliding Window uses for predicting number of incoming packets –Quality Adaptation uses for balance between the quality and the potential loss –Packet Assignment deals with the packet distribution among the senders