An Analytical Model for Progressive Mesh Streaming. WEI TSANG OOI National University of Singapore

Transcription

1 An Analytical Model for Progressive Mesh Streaming WEI TSANG OOI National University of Singapore 1

2 joint work with Cheng Wei National University of Singapore Sebastian Mondet Romulus Grigoras Geraldine Morin IRIT, University of Toulouse 2

3 3

4 4

5 5

6 6 10 MB

7 7 2 GB

8 Hoppe s Progressive Mesh Edge Collapse Vertex Split 8

9 At the sender = vk... v4 v3 v2 v1 + base model 9

10 Transmission TCP base model v1 v2 v3 v4... UDP vk 10

11 At the receiver base model v1 v2 v3 v4... vk

12 What happen if some data is lost? base model X v1 v2 v3 v ? 12

13 Vertex Split v v1 v2 v3 v 13

14 Dependency Graph 14

15 Error Propagation X 15

16 Retransmission upon detecting loss 16

17 Retransmission takes precedence over new vertex splits 17

18 Normally send multiple vertex splits per packet X 18

19 How serious is error propagation? What is the effect of dependencies? 19

20 Decoded Mesh Quality 20

21 Quality versus Time Quality Time 21

22 Importance of a vertex Quality Importance Time v1 22

23 Case 1: complete dependency Quality lost retransmitted Time 23

24 Case 2: no dependency Quality lost retransmitted Time 24

25 higher quality earlier is better Quality Time 25

26 Evaluation metric: area under the graph Larger area = better Quality 26 Time

27 Given a progressive mesh, what affects the area? 27

28 Dependency Pattern Given a progressive mesh, the dependencies among the vertex splits are fixed, but packetization can affect dependencies among the data packets. 28

29 Sending Order Given a set of packets, which one to send first? 29

30 Loss Rate Different loss pattern gives different area. We are interested in the expected area given a loss rate. 30

31 Available Bandwidth Faster sending rate means the quality increases quickly. 31

32 Round Trip Time Larger round trip time means longer time till realizing that a packet is lost and retransmit. Quality lost retransmitted Time 32

33 Network Properties Mesh Properties some formula Expected Area 33

34 Quality t Dv wv 34 Time

35 Quality Dv random variable t wv 35 Time

36 The Analytical Model 36

37 1 unit time X 37

38 Clock at sender starts when sending first packet Clock at receiver starts RTT/2 later. t = 0 X t = 0 38

39 Packet i is sent at time i if there is no retransmission X 39

40 Packet i is sent at time i+k if there are k retransmissions before i X 40

41 ???????????????? losses out of time slot when packet i is sent loss probability 41

42 Packet i is received at time X : X 42

43 approximate using E[Si] 43

44 A packet p is a parent packet of a vertex v if a vertex that v depends on belongs to p A B C D E F 44

45 P(v) = {A, B, C, E} A B C D E F v 45

46 A vertex v is decoded at time t, if 1. a parent packet of v is received at time t, and 2. all other parent packets are received before t. 46

47 47

48 Simulation with HORSE model with 10% Losses 48

49 Quality Dv random variable t wv 49 Time

50 50

51 50

52 50

53 Do dependencies matter? 51

54 best case worst case 52

55 best case worst case 53

56 best case worst case 54

57 55

58 Gap between the two extreme cases at t = Td - 1 RTT = 250 ms, Packet Size = 1500 bytes, Sending rate = 1.5 Mbps Td = 30, p = 5% 100 vertex splits per packet Gap = 1500 vertex splits 56

59 A Better Packetization Algorithm 57

60 FIFO strategy: send the most important vertex split first [Gu05] s strategy: minimize the dependencies among the vertex splits 58

61 Need to consider both importance and dependencies 59

62 put into curr packet, or next packet??? curr packet next packet 60

63 (only consider nodes whose parents are packed) maintain a max heap of all nodes using δi as key while heap is not empty and packet is not full pop a node i from heap and packed i for each child k of i insert k into heap 61

64 60 (b)happy Buddha, p = Time Slot FIFO greedy bf bsub 62

65 40 (e)happy Buddha, p = Confident Quality Time Slot FIFO greedy 63

66 Summary 64

67 Network Properties Mesh Properties some formula Expected Area 65

68 Dependencies matter only for a short time initially 66

69 67