1. Allocation of Layer Bandwidths and FECs for Video Multicast Over Wired and Wireless Networks T.-W. Angus Lee, S.-H. Gary Chan, Qian Zhang, Wen-Wu Zhu, and Ya-Qin Zhang IEEE Trans. on CSVT. VOL. 12, NO. 12, DEC 2002

2. Outline Introduction to “Layered Multicast”. Introduction to “Layered Multicast”. Architecture of the video system Architecture of the video system Wired vs. wireless clients. Wired vs. wireless clients. Using a “transcoding” gateway or not. Using a “transcoding” gateway or not. Recovery of packet-loss Recovery of packet-loss Feedback or forward-error correction code(FEC). Feedback or forward-error correction code(FEC).

3. Architecture of the video system.

4. Wired vs. wireless networks In wired networks In wired networks Packets are dropped mainly due to congestion at the routers. Packets are dropped mainly due to congestion at the routers. Packet-level FEC in the form of parity packets are used to recover packet loss. Packet-level FEC in the form of parity packets are used to recover packet loss. In the wireless hop In the wireless hop Packets are often lost due to random bit errors caused by fading of multipath effect. Packets are often lost due to random bit errors caused by fading of multipath effect. Byte-level FECs in the form of parity bytes are used to recover bit error. Byte-level FECs in the form of parity bytes are used to recover bit error.

5. Problems to be solved Given the heterogeneous error and bandwidth characteristics of its end clients, how should the server allocate the bandwidth and the corresponding packet-and-byte-level FEC of each video layer in order to maximize the overall video quality? Given the heterogeneous error and bandwidth characteristics of its end clients, how should the server allocate the bandwidth and the corresponding packet-and-byte-level FEC of each video layer in order to maximize the overall video quality? Are there any differences in performance between a simple “nontranscoding” gateway and the more complicated “transcoding” gateway? Are there any differences in performance between a simple “nontranscoding” gateway and the more complicated “transcoding” gateway?

6. Measure of video quality PSNR (peak signal-to-noise ratio) PSNR (peak signal-to-noise ratio) Packet-loss rate after error correction has to be below a certain (low) value, e.g., ≤1% for base layer and ≤2% for enhancement layers. Packet-loss rate after error correction has to be below a certain (low) value, e.g., ≤1% for base layer and ≤2% for enhancement layers. The PSNR is proportional to the video goodput defined as the useful data bits per second (after error correction) received by a client[15]. The PSNR is proportional to the video goodput defined as the useful data bits per second (after error correction) received by a client[15]. Maximizing the overall goodput. Maximizing the overall goodput.

7. Key issues in this paper How FEC can be inroduced and applied in a video system vs. How much FECs is required, and other important issues shch as the optimal bandwidth of each layer and the value of a transcoding gateway. How FEC can be inroduced and applied in a video system vs. How much FECs is required, and other important issues shch as the optimal bandwidth of each layer and the value of a transcoding gateway. Packet-level and byte-level FECs should be combined for optimal system operation. Packet-level and byte-level FECs should be combined for optimal system operation. A receiver-driven multicast system with allocation at the sender. A receiver-driven multicast system with allocation at the sender. Examined of layered multicast over mixed media with joint bandwidth and FEC allocation, and advantages of using a transcoding wireless gateway[32]. Examined of layered multicast over mixed media with joint bandwidth and FEC allocation, and advantages of using a transcoding wireless gateway[32].

8. Base-layer transmission and its optimization Allocated as the minimum end-to-end bit rate. Allocated as the minimum end-to-end bit rate. How much error control should be applied? How much error control should be applied? Quality is measured by the aggregate goodput or average goodput of the clients. Quality is measured by the aggregate goodput or average goodput of the clients.

9. The FEC Schemes Nontranscoding gateway Nontranscoding gateway Packet-level and byte-level FEC encoding are done at the video server, and error correction are only done at the end clients. Packet-level and byte-level FEC encoding are done at the video server, and error correction are only done at the end clients.

10. The FEC generation scheme Byte-Level: Each symbol consists of m bits ( m =8 in general) Packet size: n b bytes k b (≤1)bytes of source data packed with n b - k b parity bytes, where k b = n b, n b -2, …. This is the so-called RS ( n b, k b ), which is able to correct up to t b symbol errors in a packet, where t b = ceiling[ ( n b – k b )/2 ]. The packet size n b is limited by 2 m -1 symbols; therefore, for m =8, n b ≤255.Packet-Level: 把 k p 個 byte-encoded video packets 中每個 packet 的每一個 byte 分別抽出來形成 n b 個大小為 n p 的 packets. 比照 Byte-level 的做法，可以產生 n p - k p 個 parity packets. 因為 packet 是連續的，所以 t p = n p - k p 個 packet losses 可以獲得修正。 The delay of the system will increase with n p.

11. Formulation To find out the optimal allocation between the video data rate, the packet-level FEC rate, and the byte-level FEC rate. To find out the optimal allocation between the video data rate, the packet-level FEC rate, and the byte-level FEC rate.

12. Transcoding gateway Transcodes video packets from packet- level FEC to byte-level FEC before forwarding the packets to the wireless clients. Transcodes video packets from packet- level FEC to byte-level FEC before forwarding the packets to the wireless clients.

13. The FEC generation scheme

14. Quality Optimization

15. Optimization for Nontranscoding Gateway Bit error rate Bit error rate Symbol error rate Symbol error rate The probability that a random packet cannot be recovered by byte-level FEC is given by The probability that a random packet cannot be recovered by byte-level FEC is given by For the wired clients, bit error rate is 0. For the wired clients, bit error rate is 0. The end-to-end packet drop rate The end-to-end packet drop rate

16. Optimization for Nontranscoding Gateway cont. The probability that a random packet is permanently “lost” is given by[6][7] The probability that a random packet is permanently “lost” is given by[6][7] The goodput of the client g is hence given by The goodput of the client g is hence given by

17. Search on k p and k b O( Gn p n b ) => O( G(n p + n b ) ) O( Gn p n b ) => O( G(n p + n b ) ) For k p, which the residual loss rate over the wired netwoek is no more than ε o. For k p, which the residual loss rate over the wired netwoek is no more than ε o. α g =0 for all clients. Let if ( ) { STOP and goto next step. } else { for all the clients with search for the largest } search for the largest }

18. Search on k p and k b (cont.) For k b, For k b, Reintroduce α g for all wireless clients. Given k p * such that ε g for all the wireless clients are no more than ε o.

19. Optimization for Transcoding Gateway The probability that a random packet is permanently lost over the wired network is given by The probability that a random packet is permanently lost over the wired network is given by

20. Optimization for Transcoding Gateway (cont.) The end-to-end packet-loss rate after error correction is given by The end-to-end packet-loss rate after error correction is given by and hence, the goodput of the clients is and hence, the goodput of the clients is

21. Compare the performance of transcoding and notranscoding gateways G=10, half of them being wireless clients. G=10, half of them being wireless clients.

22. Compare the performance of transcoding and notranscoding gateways (cont.)

23. e s,g =0.18

24. Compare the performance of transcoding and notranscoding gateways (cont.)

27. Joint bandwidth and FEC optimization for the enhancement layers The base-layer focuses mainly on FEC allocation, the optimization of the enhancement layers has two dimensions: both FEC and bandwidth allocations. The base-layer focuses mainly on FEC allocation, the optimization of the enhancement layers has two dimensions: both FEC and bandwidth allocations. The video is encoded into L enhancement layers. ( L+1 layers included the base layer) The video is encoded into L enhancement layers. ( L+1 layers included the base layer) A client cannot decode the layer i without receiving all of its preceding i-1 layers. A client cannot decode the layer i without receiving all of its preceding i-1 layers.

28. Joint bandwidth and FEC optimization for the enhancement layers Each of the layers is carried by a multicast group. Each of the layers is carried by a multicast group. Assume that the video quality is enhanced due to the enhancement layers is linearly dependent on the aggregate goodput of the layers received. Assume that the video quality is enhanced due to the enhancement layers is linearly dependent on the aggregate goodput of the layers received.

29. Optimization of the enhancement layers What are the bandwidth and FEC of each of the enhancement layers in order to maximize the sum of video quality enhanced in terms of the goodput of each client. What are the bandwidth and FEC of each of the enhancement layers in order to maximize the sum of video quality enhanced in terms of the goodput of each client.

30. Dynamic Program Optimization 1. Ordering the end-to-end bandwidth in increasing order.

31. Dynamic Program Optimization Cumulative transmission rate of the enhancement layers up to and including layer l Cumulative transmission rate of the enhancement layers up to and including layer l We only need to consider We only need to consider

32. Dynamic Program Optimization Let S l be the set of clients who join the l- th enhancement layer, the sum of the good put for all the clients joining enhancement layer l is given by Let S l be the set of clients who join the l- th enhancement layer, the sum of the good put for all the clients joining enhancement layer l is given by

33. Dynamic Program Optimization The total good put of the system due to the L enhancement layers The total good put of the system due to the L enhancement layers Denote the maximum goodput Denote the maximum goodput l enhancement layers l enhancement layers maximum end-to-end bandwidth is x

34. Dynamic Program Optimization

35. Efficient Approximation on Layer Bandwidths In each of the recursive steps in the dynamic program above, there are O(G) possibilities of R (l). The search space of the above bit rate and FEC allocation probem is O(GL(n p +n b )). In each of the recursive steps in the dynamic program above, there are O(G) possibilities of R (l). The search space of the above bit rate and FEC allocation probem is O(GL(n p +n b )). Reduce the search space to O(L(n p +n b )). Reduce the search space to O(L(n p +n b )).

36. The Approximation Consider a large number of clients (i.e. G→∞ ) with their end-to-end bandwidth distributed according to some probability density function(pdf) f(x) which ranges from B min to B max. Consider a large number of clients (i.e. G→∞ ) with their end-to-end bandwidth distributed according to some probability density function(pdf) f(x) which ranges from B min to B max. A total of clients are with enhancement bit rate of A total of clients are with enhancement bit rate of ^^

37. The Approximation (cont.) The aggregate goodput of all the clients for the enhancement layers The aggregate goodput of all the clients for the enhancement layers

38. The Approximation (cont.) Consider that the end-to-end bandwidth of the clients is uniformly distributed between B min and B max ( with mean(B max -B min )/2 ), i.e., f(x) = 1/(B max -B min ). Thus Consider that the end-to-end bandwidth of the clients is uniformly distributed between B min and B max ( with mean(B max -B min )/2 ), i.e., f(x) = 1/(B max -B min ). Thus ^^^^ ^ ^

39. The Approximation (cont.) The approximated layered bit rate is

40. End-to-end loss rate for End-to-end loss rate for Enhancement layers are Enhancement layers are Base-layer is Base-layer is The client end-to-end bandwidths are uniformly distributed between The client end-to-end bandwidths are uniformly distributed between the standard deviation is the standard deviation is Results of the joint bandwidth and FEC optimization

41. Results of the joint bandwidth and FEC optimization (cont.) The and of each client are independent distributed with mean The and of each client are independent distributed with mean

42. Transmission rate of enhancement layers R l versus the standard deviation of the end-to-end bandwidth of the clients

43. Average goodput versus the standard deviation of client bandwidth

44. Average goodput versus the standard deviation of client bandwidth for transcoding and nontranscoding gateways

45. versus the standard deviation of client bandwidth

46. Average goodput versus the standard deviation of client bandwidth for different allocation strategies

48. Main contributions A study of a video multicast system over wired and wireless networks with joint bandwidth and FEC allocation for each layer in order to maximize the over all video quality. A study of a video multicast system over wired and wireless networks with joint bandwidth and FEC allocation for each layer in order to maximize the over all video quality. Present of an analytic model of the system, and efficient algorithm on optimal FEC allocation for the base layer, and a dynamic program formulation with a fast and accurate approximation on the optimal allocation of the enhancement layers Present of an analytic model of the system, and efficient algorithm on optimal FEC allocation for the base layer, and a dynamic program formulation with a fast and accurate approximation on the optimal allocation of the enhancement layers

49. Main contributions Investigated of the advantages of using a gateway which transcodes from packet- level FECs to byte-level FECs for the wireless link. Investigated of the advantages of using a gateway which transcodes from packet- level FECs to byte-level FECs for the wireless link.