Dude, Where is My Packet?

1
Dude, Where is My Packet?
2
Overview

• Characteristics of the Internet
• General techniques
• Error recovery for audio
• Effect of loss on MPEG
• Error recovery for MPEG

3
Loss Characteristics of The Internet
4
Characteristics of Internet

• 60-70 of paths do not show any loss
• Those with loss have an average of 4.5 6
  packet loss
• Paxson97 End-to-end Internet packet dynamics

5
Packet Loss Pattern
1000
30
6
Characteristic of Internet

• Bursts of loss are typically short (2-3
  consecutively loss packets)
• Long burst do occur
• Burst may occur periodically

7
Wireless Link

• Loss rate measured in my office 10 20
• up to 50 reported!

8
General Error Recovery Techniques

• Perkins, Hodson and Hardman
• IEEE Network Magazine 1998

9
Retransmission
1
2
3
4
3
1
2
4
3
10
Redundant Data
1
2
3
4
2
3
4
5
1
2
4
2
3
5
3
11
Error Concealment
1
2
3
4
2
4
1
12
Retransmission

• for audio

13
Retransmission is Bad

• Need scalable feedback
• NACK consumes bandwidth
• May need to retransmit frequently
• Retransmission maybe useless

14
Retransmission is Good

• Only retransmit when needed
• Exact recovery

15
Scalable Retransmission

• On packet loss
• T random(0, RTT)
• wait for T
• multicast NACK
• On receiving NACK from others
• suppress own NACK

16
Retransmit when

• group size is small
• loss rate is low
• large latency acceptable

17
Redundant Data

• for audio

18
Parity FEC
1011001
XOR
1000010
0010010
0001001
19
Parity FEC
1011001
1000010
0001001
0010010
20
Parity FEC Ordering
21
Parity FEC Ordering
22
Parity FEC Ordering
23
Parity FEC Ordering
24
Parity FEC Ordering
25
Reed-Solomon Code

• RS(n,k)

RS






26
Reed-Solomon Code








27
Media Specific FEC
1
2
3
4
2
3
4
5
1
2
4
2
3
5
3
28
Interleaving
29
Pro and Cons

• Media Independent FEC
• Overhead for redundant data
• Exact recovery
• Could be computationally costly
• Media Specific FEC
• Approximate recovery only

30
Error Concealment
31
Effect of Loss on Audio

• Speech
• Human ears can interpolate
• Loss up to length of phoneme can still be
  tolerable

32
Insertion-based Repair

• Splice
• Silence Substitution
• Noise Substitution
• Repetition

33
Other Repair Methods

• Interpolation
• Regeneration

34
Pro and Cons

• Could be computationally costly
• Approximate recovery only
• Does not work well for long burst of packet loss

35
Colins Recommendations
36
Non-Interactive Apps

• Interleaving
• FEC
• Retransmission for unicast only

37
Interactive Applications

• Media Specific FEC

38
Error Concealment

• Repeat

39
Effects of Packet Loss on MPEG

• Jill Boyce and Robert Gaglianello
• ACM Multimedia 1998

40
Measurement-based Study

• Need to understand the problem before proposing
  solution

41
Data Gathering Method

• From NYC13 , Austin21, London18
• To Holmdel, NJ

42
Data Gathering Method

• Video
• Two 5-mins MPEG
• 30 fps
• 384 kbps and 1 Mbps
• IBBPBBPBBPBBPBB
• QSIF 176x112 and SIF 352x240
• One row per slice

43
Average Packet Loss
44
Packet Loss vs Time
45
How to Fill Packets

• B 1 frame 1 packet
• P 1 slice 1 packet
• I 1 slice 1 packet

46
Packet Size Distributions
47
Packet Size vs Loss Rate
48
Lesson

• 1 slice per packet even if larger than MTU
• If smaller packets means higher loss rate fill
  packet until MTU
• Else one slice per MTU
• Alt Change size of slice

49
Frame Effected by Errors
50
Correlation of Packet Loss
51
Overview of Error Recovery for Video
52
Methods

• Retransmission
• Error Concealment
• FEC
• Limiting Error Propagation
• Reference Frame Selection

53
Reference Frame Selection
I P B B P

I
X
54
Methods

• Retransmission
• Error Concealment
• FEC
• Limiting Error Propagation
• Reference Frame Selection
• Changing Temporal Pattern

55
Changing Temporal Patterns

• Injong Rhee
• SIGCOMM 98

56
Basic Idea

• Better Late Than Never!, or
• Late packet is still useful

57
MPEG Frame Pattern
I
B
B
P
B
B
P
I
P
B
B
P
B
B
58
H.261 Error Propagation
I
P
P
P
P
P
P
X
X
X
X
X
X
loss
X
X
X
X
retransmission
59
H.261 Frame Pattern
I
P
P
P
P
P
P
X
loss
60
H.261 Frame Pattern
X
X
X
retransmission
loss
61
PTDD

• Periodic Temporal Dependency Distance
• Large PTDD
• Later Deadlines
• Longer Error Propagations
• Less Temporal Dependencies

62
QAL
I
P
P
P
Enhancement Layer
I
P
P
P
Base Layer
63
QAL
I
P
P
P
X
Enhancement Layer
I
P
P
P
Base Layer FEC
64
QAL PTDD
I
P
P
P
Enhancement Layer
I
P
P
P
Base Layer
65
SR-RTP

• Feamster and Balakrishnan
• PV02

66
Overview

• An analytical model for packet loss
• SR-RTP
• Post Processing

67
Analytic Model

• p prob. packet loss
• f observed frame rate
• ? frame drop rate
• Si mean num. of packets in frames of
  type i
• P(i) prob. frame type is i
• P(F) prob. frame is useless

68
p prob. packet loss f observed frame
rate ? frame drop rate Si mean num. of
packets in frames of type i P(i)
prob. frame type is i P(F) prob. frame is
useless
69
Is it accurate?
70
SR-RTP

• Extensions to RTP for selective reliability

71
SR-RTP
0
Length
ADU Sequence Number
ADU Length
ADU Offset
Priority
Layer Number
72
Feamster Balakrishnans ADU

• 1 ADU 1 Frame

73
Loss Detection
LEN 50 ADU 0 ADU Size 100 ADU Off 0
LEN 20 ADU 0 ADU Size 100 ADU Off 80
LEN 50 ADU 1 ADU Size 150 ADU Off 40
LEN 50 ADU 3 ADU Size 50 ADU Off 0
74
Retransmission Request

• Based on priority
• Example
• I-Frame highest priority
• P-Frame higher priority if closer to prev I
• B-Frame no retransmition

75
Error Concealment for Video

• Repeat pixels from previous frame
• Interpolate pixels from neighbouring region
• Interpolate motion vectors from previous frame

76
Inferred MV for P Frame
I
P
77
Inferred MV for I Frame
P
P
I
78
Evaluation

• Peak Signal to Noise Ratio (PSNR)
• 2552 over average square of differences in pixel
  values
• Calculated in dB (10 log 10)

79
Performance of SR-RTP
80
Todays Summary
81
How to recover packet loss

• Retransmission
• FEC
• Error Concealment

82
How to limit the damage of error

• Interleaving
• Key frame selection
• Change reference frame