wavelet and time -domain modeling of multi -layer vbr...

1

Wavelet and TimeWavelet and Time--Domain Domain Modeling of MultiModeling of Multi--Layer Layer

VBR Video TrafficVBR Video Traffic

Min Dai, Dmitri LoguinovTexas A&M University

2

Agenda

• Background– Importance of traffic modeling

– Goals of traffic modeling

– Preliminary knowledge of video traffic

• Challenges & Current Status

• Our Work– Modeling single-layer video traffic

– Modeling multi-layer video traffic

• Conclusion

3

Importance of Traffic Modeling

• Properly allocate network resources

• Evaluate protocols and effectively design networks

• Use as traffic descriptor to achieve certain Quality of Service (QoS) requirements

• Analyze and characterize a queue or a network

4

Goals of Traffic Modeling

• Capture the characteristics of video frame size sequences

– The marginal probability density function (PDF) of frame sizes

– The autocorrelation function (ACF) of video traffic

• Accurately predict network performance

– Buffer overflow probabilities

– Temporal burstiness

5

Preliminary

I0 B1 B2 P3 B4 B5 P6 I7

Intra-GOP CorrelationInter-GOP Correlation

Group of Pictures (GOP)

6

Agenda

• Background– Importance of traffic modeling

– Goals of traffic modeling

– Preliminary knowledge

• Challenges & Current Status

• Our Work– Modeling single-layer video traffic

– Modeling multi-layer video traffic

• Conclusion

7

Challenges

• The PDF is different among I, P, and B-frame sizes

• VBR video traffic exhibits both long range dependency (LRD) and short range dependency (SRD).

• Single-layer and base layer video traffic:

– Coexistence of inter- and intra-GOP correlation

• Multi-layer video traffic:

– Strong cross-layer correlation

8

Current Status

• It is hard to capture both LRD and SRD properties of the autocorrelation function of video traffic

• Little work has considered the intra-GOP correlation

• Most existing models only apply to single-layer VBR video traffic

• Current multi-layer traffic models do not capture the cross-layer correlation

9

Agenda

• Background– Importance of traffic modeling

– Goals of traffic modeling

– Preliminary knowledge

• Challenges & Current Status

• Our Work– Modeling single-layer video traffic

– Modeling multi-layer video traffic

• Conclusion

10

Our Work

Intra-GOP

CorrelationI-frame

Size ModelingP and B-frame Size Modeling

Inter-GOP Correlation

Wavelet Domain Time Domain

Enhancement Layer Frame Size Modeling

Cross-layer Correlation

Single-layer & Base layer

11

Wavelet Decomposition

• Wavelet function generates the detailed coefficients {Dj} and scaling function generates the approximation coefficients {Aj}, where j is the decomposition level.

A typical wavelet decomposition.

DA

D

A

DA

wavelet functionscaling

function

12

Wavelet Decomposition (cont.)

The ACF structures of {D3} and {A3} (left). The PDF of I-frame sizes and that of {A3} (right).

0

0.01

0.02

0.03

0.04

0.05

0 2000 4000 6000 8000 10000

bytesp

rob

abil

ity

I-frame size

approx. coefficients

-0.2

0

0.2

0.4

0.6

0.8

1

0 10 20 30 40 50

lag

auto

corr

elat

ion

ACF approx.

ACF detailed

13

Modeling I-Frame Sizes

• Estimate the coarsest approximation coefficients {AJ} :– Prior work — independent random Gaussian

or Beta variable

– Our model — dependent random variables with marginal Gamma distribution

• Estimate detailed coefficients {Dj} at each level:– Prior work — i.i.d. Gaussian random variables

– Our model — i.i.d. mixture Laplacian random variables

14

Estimate Detailed Coefficients

1

10

100

1000

10000

-700 -500 -300 -100 100 300 500 700

coefficients (bytes)

1

10

100

1000

10000

-700 -500 -300 -100 100 300 500 700

coefficients (bytes)

1

10

100

1000

10000

-700 -500 -300 -100 100 300 500 700

coefficients (bytes)

Mixture-Laplacian

1

10

100

1000

10000

-700 -500 -300 -100 100 300 500 700

coefficients (bytes)

GGD

GaussianActual

Mixture – Laplacian

15

Performance Comparison

-0.2

0

0.2

0.4

0.6

0.8

1

0 100 200 300

lag

auto

corr

elat

ion

actual

our model

nested AR

The ACF structure of the actual and that of the synthetic traffic in a long range (left) and short range (right).

-0.2

0

0.2

0.4

0.6

0.8

1

0 5 10 15 20 25 30 35

lagau

toco

rrel

atio

n

actual

our model

GBAR

Gamma_A

16

Our Work

Intra-GOP

Correlation

I-frame Size Modeling

P and B-frame Size Modeling

Inter-GOP Correlation

Wavelet Domain Time Domain

Enhancement Layer Frame Size Modeling

Cross-layer Correlation

Single-layer & Base layer

17

Modeling P/B-Frame Sizes

• Assume that GOP structure is fixed, e.g., IBBPBBPBBPBB

• Definition: In the n-th GOP, – — the I-frame size

– — the size of the i-th P-frame

– — the size of the i-th B-frame

• For example, represents the size of the third P-frame in the 10-th GOP

18

Intra-GOP Correlation

• Most previous work does not consider intra-GOP correlation and estimates P and B-frame sizes as i.i.d. random variables

• However, intra-GOP correlation is important and has similar structures between and , with respect to different i.

19

Intra-GOP Correlation (cont.)

-0.1

0

0.1

0.2

0.3

0 50 100 150 200 250 300 350

lag

co

rrel

ati

on

cov(P1,I)

cov(P2,I)

cov(P3,I)

The correlation between different B-frame sequences and the I-frame sequence (left). That between different P-frame sequences and the I-frame sequence (right).

-0.1

0

0.1

0.2

0.3

0 50 100 150 200 250 300 350

lag

co

rrela

tio

n

cov(B1,I)

cov(B2,I)

cov(B7,I)

20

Modeling P-Frame Sizes

• The size of the i-th P-frame in the n-th GOP:

– Process and is independent of

– Parameters σP and σI are the standard deviation of and , respectively.

– Parameter is the lag-0 correlation coefficient.

21

Performance Comparison

-0.1

0.1

0.3

0.5

0 70 140 210 280 350

lag

corr

elat

ion

actual

our model

i.i.d methods

The correlation between and in Star Wars.

22

Agenda

• Background– Importance of traffic modeling

– Goals of traffic modeling

– Preliminary knowledge

• Challenges & Current Status

• Our Work– Modeling single-layer video traffic

– Modeling multi-layer video traffic

• Conclusion

23

Our Work

Intra-GOP

CorrelationI-frame

Size ModelingP and B-frame Size Modeling

Inter-GOP Correlation

Wavelet Domain Time Domain

Enhancement Layer Frame Size Modeling

Cross-layer Correlation

Single-layer & Base layer

24

Modeling Enhancement Layer

• We estimate I-frame sizes in wavelet domain

• We estimate P and B-frame sizes using the cross-layer correlation:

– where is the size of the i-th P-frame, and is the size of the i-th B-frame

– Parameter is the lag-0 cross correlation coefficient, are the standard deviation of the enhancement layer sequence and its corresponding base layer sequence.

25

Performance

The cross correlation in the original and synthetic The silence of the Lambs.

-0.3

-0.1

0.1

0.3

0.5

0.7

0.9

0 100 200 300 400

lag

cro

ss c

orr

elat

ion

actual

our model

-0.3

-0.1

0.1

0.3

0.5

0.7

0.9

0 100 200 300 400

lagcr

oss

co

rrel

atio

n

actual

Zhao et al.

26

Model Accuracy Study

• QQ plots— Verify the distribution similarity between the original traffic and the synthetic one

0

1000

2000

3000

4000

5000

0 1000 2000 3000 4000 5000

original frame size

syn

thet

ic f

ram

e si

ze

0

3000

6000

9000

12000

0 3000 6000 9000 12000

original frame size

syn

thet

ic f

ram

e si

ze

QQ plots for the synthetic single-layer Star Wars (left) and the synthetic enhancement layer The Silence of the Lambs (right).

27

Model Accuracy Study (cont.)

• Leaky-bucket simulation– Examine how well the traffic model preserves

the temporal information of the original traffic

– Implementation: Pass the original and synthetic traffic through a generic buffer with capacity cand drain rate d

– Evaluation metric:

• where p is the actual packet loss ratio and pmodelis the synthetic one.

28

Model Accuracy Study (cont.)

0

0.2

0.4

0.6

0.8

1

0 1 2 3 4 5

drain rate

bit

lo

ss r

atio

actual

synthetic

0%

1%

2%

3%

4%

10 20 30 40

buffer capacity (ms)re

lati

ve e

rro

r

our modelNested ARGamma_AGamma_B

Comparison of several models in H.26L coded Starship Troopers

The loss ratio p of the original and synthetic The Silence of the Lambs enhancement layer

29

Conclusion

• This paper proposes a traffic model applicable to both single-layer and multi-layer VBR video traffic.

• The presented traffic modeling framework captures both LRD and SRD properties of video traffic.

• This framework accurately describes the intra-GOP correlation and the cross-layer correlation.