1. Multiple Sender Distributed Video Streaming Nguyen, Zakhor IEEE Transactions on Multimedia April 2004

2. Agenda  Motivation  Proposed Approach  Rate Allocation Algorithm  Packet Partition Algorithm  Experimental Results  Conclusion

3. Motivation  Addresses the problem of video streaming over best effort, packet- switched networks.  Streaming of video requires high bit rates.  Packet loss and delay due to network congestion.

4. Aim of the proposed approach  Increase throughput in order to meet the high bit rate demands.  Reduce the probability of packet loss.

5. Proposed approach  Simultaneous video streaming from multiple senders to a single receiver.  Employs a receiver driven protocol  The distributed streaming protocol consists of a Rate Allocation Algorithm (RAA) and a Packet Partition Algorithm (PPA)

6. Assumption  The bandwidth bottleneck is not at the last hop.

7. Rate Allocation Algorithm (RAA)  Run at the receiver.  Determines the optimal sending rate for each sender.  Uses information about available network bandwidth, channel characteristics.  Receiver can also redistribute rates among existing senders.

8. Packet Partition Algorithm (PPA)  Run at the senders.  Ensures that every packet is sent by one and only one sender.

9. The distributed streaming framework

10. RAA : A closer look  Goal of RAA is to “determine how to split the total video rate among M senders in order to minimize the probability of packet loss.”  Used in conjunction with Forward Error Correction (FEC).

11. FEC  Used at the receiver side for recovering lost packets.  A FEC block consists of a number of packets.  N packets in a block, K data packets, N-K redundant packets.  For recovery, any K packets in a block must be received.

12. RAA (contd.)  RAA used with FEC reduces packet loss as compared to single route streaming.

13. RAA (contd.)  Notations: N : Total no. of packets in FEC block K: No. of data packets in FEC block N A : Number of packets sent by sender A in a FEC block N B : Number of packets sent by sender B in a FEC block

14. Rate Allocation Formula  We seek to minimize the probability of irrecoverable loss given by N A + N B j C(K,N 0,N 1 )= ∑ ∑ P(A,i, N A )P(B,j-i,N B ) j =N–K+1 i=0

15. Packet Partition Algorithm (PPA)  Goal of the PPA is to “determine which packets should be sent by which senders in order to prevent duplicate packets and minimize start up delay.”

16. Minimization of start-up delay in PPA  In Kazaa, senders send contiguous blocks of data. Duplication of packets cannot occur.  In PPA, senders send interleaved packets. Reduces start-up delay, but packets may be duplicated.

17. Avoiding Packet Duplication in PPA  Each sender independently arrives at the same decision as to which packet to send next.  All senders use the same information to choose the next packet to send.  This information is sent to all senders by the receiver.

18. How PPA works  The receiver sends identical control packets to all senders.  The control packet contains RTT of all senders. Sending rate for all senders. Starting sequence number.

19. How PPA works (contd.)  Each of the senders uses the information in the control packet to compute the estimated arrival time for packet k  The time difference between the arrival and playback time of packet k is then computed.  The sender that maximizes this time difference is chosen to send packet k

20. Illustration of how PPA works

21. Choosing the starting sequence no.

22. Irrecoverable loss probability for various FEC levels as a function of bad times

23. Optimal sending rate for route A as a function of bad times for route B

24. Irrecoverable loss probability ratio as a function of average bad time of route B

25. Irrecoverable probability as a function of sending rate of sender A

26. Number of lost packets per FEC block for a single sender

27. No. of lost packets per FEC block for two senders

28. Throughput of two senders

29. Sequence No. difference of two consecutive received packets

30. Conclusion  Inaccuracy in parameter estimation can skew the results.  Cannot protect against packet loss in case of coinciding congestion intervals on different routes.