Impact of FEC Overhead on Scalable Video Streaming
1. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Impact of FEC Overhead on
Scalable Video Streaming
Seong-ryong Kang and Dmitri Loguinov
Department of Computer Science
Texas A&M University
College Station, TX 77843
June 14, 2005
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 1/30
2. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Outline
Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 2/30
3. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Outline
Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 3/30
4. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Motivation of this work
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 4/30
5. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Motivation of this work
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 4/30
6. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Motivation of this work
Internet streaming is an important part of the Internet
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 4/30
7. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Motivation of this work
Internet streaming is an important part of the Internet
Streaming applications usually require special mechanisms
that can overcome packet loss without utilizing retransmission
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 4/30
8. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Motivation of this work
Internet streaming is an important part of the Internet
Streaming applications usually require special mechanisms
that can overcome packet loss without utilizing retransmission
FEC is often considered for recovering lost data segments
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 4/30
9. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Motivation of this work
Internet streaming is an important part of the Internet
Streaming applications usually require special mechanisms
that can overcome packet loss without utilizing retransmission
FEC is often considered for recovering lost data segments
However, studies reported conflicting results on the benefits of
FEC
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 4/30
10. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Motivation of this work
Internet streaming is an important part of the Internet
Streaming applications usually require special mechanisms
that can overcome packet loss without utilizing retransmission
FEC is often considered for recovering lost data segments
However, studies reported conflicting results on the benefits of
FEC
Our work aims to:
address this uncertainty and
provide additional insight into understanding how FEC
overhead affects the performance of scalable video streaming
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 4/30
11. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Outline
Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 5/30
12. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Background
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 6/30
13. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Background
FEC
FEC schemes require application servers to send extra
information along with the original data
Media independent FEC
Based on (N, k) block codes (such as parity or Reed-Solomon
codes), where N is the size of an FEC block and k is the
number of FEC packets in the block
All data packets are recovered if the number of lost packets in
a block is no more than k
If more than k packets are lost, none of them can be
recovered by the receiver
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 6/30
14. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Background
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 7/30
15. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Background
FGS
Streaming profile of the ISO/IEC MPEG-4 standard
Method of compressing residual video signal into a single
enhancement layer
Allows application servers to scale the enhancement layer to
match variable network capacity during streaming
The enhancement layer is typically coded at some fixed bitrate
and can be rescaled to any desired bitrate
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 7/30
16. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Background
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 8/30
17. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Background
Because of dependency in
the enhancement layer,
higher sections of FGS
cannot be used in decoding
the frame without the
presence of lower sections
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 8/30
18. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Background
Because of dependency in loss
the enhancement layer,
higher sections of FGS
cannot be used in decoding loss
the frame without the useful
presence of lower sections packets
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 8/30
19. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Outline
Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 9/30
20. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Analysis of video streaming
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 10/30
21. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Analysis of video streaming
We investigate the performance of FEC-based streaming
considering Markov and renewal patterns of packet loss
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 10/30
22. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Analysis of video streaming
We investigate the performance of FEC-based streaming
considering Markov and renewal patterns of packet loss
We use MPEG-4 FGS as an example and only examine the
enhancement layer
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 10/30
23. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Analysis of video streaming
We investigate the performance of FEC-based streaming
considering Markov and renewal patterns of packet loss
We use MPEG-4 FGS as an example and only examine the
enhancement layer
Note that many studies show that Internet packet loss can be
captured by Markov models
Alternating ON/OFF renewal process can model more general
distribution of packet loss and allows heavy-tailed burst lengths
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 10/30
24. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Analysis of video streaming
We investigate the performance of FEC-based streaming
considering Markov and renewal patterns of packet loss
We use MPEG-4 FGS as an example and only examine the
enhancement layer
Note that many studies show that Internet packet loss can be
captured by Markov models
Alternating ON/OFF renewal process can model more general
distribution of packet loss and allows heavy-tailed burst lengths
In what follows, we derive the expected amount of useful data
E[Zj ] recovered from each frame
Zj is the number of consecutively received packets in a frame j
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 10/30
25. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Analysis of video streaming
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 11/30
26. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Analysis of video streaming
Assume that long-term network packet loss is given by p
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 11/30
27. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Analysis of video streaming
Assume that long-term network packet loss is given by p
Loss process can be modeled by a two-state Markov chain:
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 11/30
28. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Analysis of video streaming
Assume that long-term network packet loss is given by p
Loss process can be modeled by a two-state Markov chain:
α
1—α 0 1 1—β
β
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 11/30
29. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Analysis of video streaming
Assume that long-term network packet loss is given by p
Loss process can be modeled by a two-state Markov chain:
α
1—α 0 1 1—β
β
α and β are transition probabilities
In the stationary state, probabilities π0 and π1 to find the
process in each of its two states are given by:
β α
π0 = , π1 = p = . (1)
α+β α+β
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 11/30
30. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Analysis of video streaming
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 12/30
31. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Analysis of video streaming
To derive E[Zj ], we define:
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 12/30
32. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Analysis of video streaming
To derive E[Zj ], we define:
L to be the number of packets lost in an FEC block
¯
Q = E[Zj |L > k] to be the expected number of useful video
packets recovered from the front of an FEC block when L > k
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 12/30
33. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Analysis of video streaming
To derive E[Zj ], we define:
L to be the number of packets lost in an FEC block
¯
Q = E[Zj |L > k] to be the expected number of useful video
packets recovered from the front of an FEC block when L > k
Then, we have the following result:
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 12/30
34. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Analysis of video streaming
To derive E[Zj ], we define:
L to be the number of packets lost in an FEC block
¯
Q = E[Zj |L > k] to be the expected number of useful video
packets recovered from the front of an FEC block when L > k
Then, we have the following result:
Lemma 1
Assuming a two-state Markov packet loss and L > k, the expected
number of useful video packets recovered per frame is:
¯ 1−p
Q = E[Zj |L > k] = 1 − (1 − α)H , (2)
α
where H = N − k is the number of video packets in an FEC block
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 12/30
35. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Analysis of video streaming
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 13/30
36. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Analysis of video streaming
Theorem 1
Assuming two-state Markov packet loss with average loss
probability p, the expected number of useful packets recovered per
FEC block of size N is:
k
E[Zj ] = P (N, i)H (3)
i=0
N
1−p
+ P (N, i) 1 − (1 − α)H ,
α
i=k+1
where P (N, i) is the probability of losing exactly i packets out of
N transmitted packets.
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 13/30
37. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Simulation results
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 14/30
38. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Simulation results
Define ψ to be the fraction of FEC packets in a block
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 14/30
39. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Simulation results
Define ψ to be the fraction of FEC packets in a block
To verify the model, we simulate Markov loss process with
two different values of ψ
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 14/30
40. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Simulation results
Define ψ to be the fraction of FEC packets in a block
To verify the model, we simulate Markov loss process with
two different values of ψ
2500 350
2000 300
Decoding rate (kb/s)
Decoding rate (kb/s)
250
1500
200
1000
150
Model Model
500 Simulation Simulation
100
500 1000 1500 2000 2500 3000 500 1000 1500 2000 2500 3000
Sending rate S (kb/s) Sending rate S (kb/s)
(a) ψ = 1.1p (b) ψ = 0.9p
˜
Figure: Expected decoding rate R (p = 0.1, α = 0.08, and β = 0.72)
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 14/30
41. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Simulation results
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 15/30
42. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Simulation results
Note that the behavior of expected decoding rate changes for
different ψ
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 15/30
43. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Simulation results
Note that the behavior of expected decoding rate changes for
different ψ
The amount of overhead in FEC-based streaming plays a
significant role in determining video quality
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 15/30
44. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Simulation results
Note that the behavior of expected decoding rate changes for
different ψ
The amount of overhead in FEC-based streaming plays a
significant role in determining video quality
Next, we derive the utility of received video and examine how
FEC overhead affects the quality of video
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 15/30
45. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Simulation results
Note that the behavior of expected decoding rate changes for
different ψ
The amount of overhead in FEC-based streaming plays a
significant role in determining video quality
Next, we derive the utility of received video and examine how
FEC overhead affects the quality of video
Define the utility U as the fraction of received data that is
useful for decoding
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 15/30
46. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Analysis of video streaming
Theorem 2
Assuming Bernoulli packet loss in an FEC block of size N , average
loss probability p, and FEC overhead rate ψ = ηp, (0 < ψ < 1),
the utility of received video for each FEC block converges to the
following as H → ∞:
0
0<η<1
lim U = 0.5 η=1 , (4)
H→∞ 1−ψ
1 < η < 1/p
1−p
where η is constant.
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 16/30
47. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Simulation results
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 17/30
48. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Simulation results
Simulation results under Bernoulli loss
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 17/30
49. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Simulation results
Simulation results under Bernoulli loss
1
0.8 η=1.1
0.6 η=1.0
Utility
0.4
Model
η=0.9 Simulation
0.2
0
0 1000 2000 3000 4000 5000
Frame size H
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 17/30
50. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Simulation results
Simulation results under Bernoulli loss
1
0.8 η=1.1
0.6 η=1.0
Utility
0.4
Model
η=0.9 Simulation
0.2
0
0 1000 2000 3000 4000 5000
Frame size H
Note that U indeed converges to 0, 0.5 or (1 − ψ)/(1 − p)
depending on the value of ψ as the streaming rate becomes large
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 17/30
51. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Simulation results
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 18/30
52. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Simulation results
Note that the asymptotic behavior of U in Theorem 2 holds
for Markov and renewal patterns of packet loss
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 18/30
53. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Simulation results
Note that the asymptotic behavior of U in Theorem 2 holds
for Markov and renewal patterns of packet loss
1 1
h=1.2
0.8 0.8
0.6 h=1.0 0.6
Utility
Utility
0.4 0.4 h = 1.2
h=0.8
Model h = 1.0
Simulation h = 0.8
0.2 0.2
0 0
500 1000 1500 500 1000 1500
Frame size H Frame size H
(a) Markov loss (b) Renewal loss
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 18/30
54. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Simulation results
Note that the asymptotic behavior of U in Theorem 2 holds
for Markov and renewal patterns of packet loss
1 1
h=1.2
0.8 0.8
0.6 h=1.0 0.6
Utility
Utility
0.4 0.4 h = 1.2
h=0.8
Model h = 1.0
Simulation h = 0.8
0.2 0.2
0 0
500 1000 1500 500 1000 1500
Frame size H Frame size H
(a) Markov loss (b) Renewal loss
Again, U exhibits percolation and converges to 0, 0.5 or
(1 − ψ)/(1 − p) depending on ψ
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 18/30
55. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Discussion
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 19/30
56. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Discussion
Effectiveness of FEC
depends on how the server
uses redundant packets
based on packet-loss
dynamics
Using fixed amount of
overhead may cause
significant quality
degradation when packet
loss fluctuates
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 19/30
57. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Discussion
Simulation results of U for
Effectiveness of FEC
different packet loss p under
depends on how the server
improper amount of FEC overhead
uses redundant packets
based on packet-loss
1
dynamics
0.8
Using fixed amount of 0.6
Utility
overhead may cause 0.4
ideal amount
improper amount
significant quality
0.2
degradation when packet
0
loss fluctuates 0 0.1 0.2 0.3
Packet loss
0.4 0.5
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 19/30
58. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Outline
Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 20/30
59. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Adaptive FEC control
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 21/30
60. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Adaptive FEC control
Needs for adaptive control
In a practical network environment, packet loss changes
dynamically, depending on
cross-traffic
link quality
routing updates, etc
The amount of FEC needs to be adjusted according to
changing packet loss to maintain high end-user utility
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 21/30
61. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Adaptive FEC control
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 22/30
62. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Adaptive FEC control
For adaptive FEC rate control, we use a simple proportional
controller:
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 22/30
63. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Adaptive FEC control
For adaptive FEC rate control, we use a simple proportional
controller:
ψi (n) = ψi (n − Di ) + τ (ηpi (n − Di ) − ψi (n − Di )) , (5)
where index i represents flow number, pi (n) is the measured
average packet loss in the FGS layer for flow i during interval
n, τ is the controller’s gain parameter, Di is the round-trip
delay for flow i.
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 22/30
64. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Adaptive FEC control
For adaptive FEC rate control, we use a simple proportional
controller:
ψi (n) = ψi (n − Di ) + τ (ηpi (n − Di ) − ψi (n − Di )) , (5)
where index i represents flow number, pi (n) is the measured
average packet loss in the FGS layer for flow i during interval
n, τ is the controller’s gain parameter, Di is the round-trip
delay for flow i.
Lemma 2
Controller (5) is stable if and only if 0 < τ < 2.
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 22/30
65. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Outline
Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 23/30
66. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Packet loss pattern
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 24/30
67. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Packet loss pattern
We simulate a streaming session with a hypothetical packet
loss pattern
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 24/30
68. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Packet loss pattern
We simulate a streaming session with a hypothetical packet
loss pattern
0.5 0.8
0.4
Transition probability
0.6
Packet loss
0.3 β
0.4 α
0.2
0.2
0.1
0 0
0 10 20 30 5 10 15 20 25 30
Time (seconds) Time (seconds)
(a) Packet loss pattern (b) Transition probability α,β
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 24/30
69. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Achieved utility
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 25/30
70. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Achieved utility
We investigate adaptive FEC overhead controller (5) with the
behavior of achieved utility
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 25/30
71. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Achieved utility
We investigate adaptive FEC overhead controller (5) with the
behavior of achieved utility
To illustrate the adaptivity of the controller, we use target
utility UT = 0.8
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 25/30
72. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Achieved utility
We investigate adaptive FEC overhead controller (5) with the
behavior of achieved utility
To illustrate the adaptivity of the controller, we use target
utility UT = 0.8
For comparison, we apply two different scenarios that use
fixed amount of overhead
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 25/30
73. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Achieved utility
We investigate adaptive FEC overhead controller (5) with the
behavior of achieved utility
To illustrate the adaptivity of the controller, we use target
utility UT = 0.8
For comparison, we apply two different scenarios that use
fixed amount of overhead
The fixed-overhead amount is driven by the lower and upper
bounds on packet loss p
˜
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 25/30
74. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
The evolution of achieved utility U
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 26/30
75. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
The evolution of achieved utility U
adaptive adaptive
1 fixed 1 fixed
0.8 0.8
Utility
Utility
0.6 0.6
0.4 0.4
0.2 0.2
0 0
0 10 20 30 0 10 20 30
Time (seconds) Time (seconds)
(a) p = 0.1
˜ (b) p = 0.4
˜
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 26/30
76. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
The evolution of achieved utility U
adaptive adaptive
1 fixed 1 fixed
0.8 0.8
Utility
Utility
0.6 0.6
0.4 0.4
0.2 0.2
0 0
0 10 20 30 0 10 20 30
Time (seconds) Time (seconds)
(a) p = 0.1
˜ (b) p = 0.4
˜
Our adaptive controller maintains the target utility UT very
well along the entire streaming session
However, fixed-overhead schemes cannot maintain high utility
as packet loss varies
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 26/30
77. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
PSNR quality
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 27/30
78. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
PSNR quality
PSNR of CIF Foreman reconstructed with different FEC
overhead control
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 27/30
79. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
PSNR quality
PSNR of CIF Foreman reconstructed with different FEC
overhead control
M1 and M2 use fixed amount of overhead driven by p = 0.1
˜
and p = 0.4, respectively
˜
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 27/30
80. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
PSNR quality
PSNR of CIF Foreman reconstructed with different FEC
overhead control
M1 and M2 use fixed amount of overhead driven by p = 0.1
˜
and p = 0.4, respectively
˜
45
40
PSNR (dB)
35
adaptive
30 M1
M2
25
0 20 40 60 80 100
Frame number
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 27/30
81. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
PSNR quality
PSNR of CIF Foreman reconstructed with different FEC
overhead control
M1 and M2 use fixed amount of overhead driven by p = 0.1
˜
and p = 0.4, respectively
˜
The adaptive method offers
45 as much as 2.5 dB higher
40
PSNR than M2 for the first
PSNR (dB)
60 frames
35
adaptive
Also, outperforms M1 by
30 M1
M2 almost 10 dB for the last 40
25
0 20 40 60 80 100 frames
Frame number
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 27/30
82. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Outline
Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 28/30
83. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Conclusion
FEC has conflicting effects on video quality depending on the
amount of overhead used
Adaptive FEC overhead control can provide a high quality of
video to end-users
Proper control of FEC overhead can significantly improve the
utility of received video over lossy channels
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 29/30
84. Motivation
Background
Impact of FEC on Scalable Video
Adaptive FEC Control
Evaluation
Conclusion
Thank you!
Any questions?
S. Kang and D. Loguinov Impact of FEC Overhead on Scalable Video Streaming 30/30