Transcript DiffServ Aware Video Streaming over 3G Wireless Networks

DiffServ Aware Video
Streaming over 3G Wireless
Networks
Julio Orozco, David Ros
Novembre Project
Sophia Antipolis, 26/11/2004
Agenda




Context
The DiffServ-aware streaming approach
Quality Assessment
Performance Evaluation
Agenda

Context






Overview
Technical challenges
Requirements
The DiffServ-aware streaming approach
Quality Assessment
Performance Evaluation
Example Scenario






I’m at the airport and have a two-hour wait
ahead …
Real Madrid faces Milan in the Champions
League final …
After a simple procedure, I start watching the
match …
Transmitted from an Internet server …
On my mobile terminal …
With decent quality and an affordable price.
Example Scenario
Internet
UMTS
Video Server
Video Client
(mobile terminal)
Technical Challenges

UMTS radio link



Heterogeneous architecture


High and variable delay/jitter
Variable and limited bandwidth
Internet + UMTS
Business/billing models
Requirements

Video compression



Network architecture


Highly efficient
Error resilient
Affordable QoS
Integration


Video information – Network
Internet - UMTS
Agenda


Context
The DiffServ-aware streaming approach






Concept
H.264
DiffServ architecture
Semantic mapping
Quality Assessment
Performance Evaluation
Our Approach



DiffServ-aware streaming of H.264 video
Pseudo-subjective quality assessment
Goals


Reduce visual distortion caused by network losses
(induced by variable bandwidth and delay)
Validate performance in terms of visual quality
DiffServ Aware Streaming of
H.264 Video

H.264




DiffServ IP network


State-of-the-art standard codec
High efficiency
Improved network adaptation and error resilience
Simple, scalable QoS at the IP level
Mapping

Video semantics <-> DiffServ packet priorities
H.264 Video Codec


State of the art (May 2003)
High compression efficiency



Designed for network applications


50% rate gain against MPEG-2
30% against MPEG-4
Network adaptation layer (NAL)
Novel error resilience features
DiffServ Architecture


AF prioritized discard
Three packet priorities per class




Green
Yellow
Red
Under congestion:

Red packets get discarded first then yellow
packets, and finally green packets.
AF Prioritized Discard

RIO algorithm
Semantic Mapping

Original idea (MPEG-2)



Video is transported in a single AF class
AF packet priorities <-> MPEG frame types
Reduces visual distortion caused by losses
DiffServ Mapping of H.264

General strategy


Map coarse syntax elements in a single AF
class
Take advantage of H.264 advanced
network adaptation and error resilience
features



Slices
Flexible Macroblock Ordering
Data Partitioning
Agenda



Context
The DiffServ-aware streaming approach
Quality Assessment




Motivation
Classical methods
Pseudo-subjective assessment
Performance Evaluation
Motivation

Streaming in Internet/UMTS


Distorsion = f(compression, network
losses)
Network losses = f (congestion, rate,
delay, jitter)
Motivation

We need to measure the quality of the
streamed signals


Does DiffServ-Aware Streaming yield a
better quality?
Which mapping strategy is better? (from a
perceived-quality point of view)
Classical Quality Assessment

Subjective




Reflects human
perception
Difficult
Expensive
Not feasible in real
time

Objective





Automated
Repeateble
Does not necessarily
reflects human
perception
Requires acces to the
original signal
Can be computerintensive
Pseudo-Subjective Assessment



Novel Approach
Based on Neural Networks
Link network and coding parameters to
human perception

MOS = f (network, coding)
Pseudo-Subjective Assesment

Methodology




Identification of the quality-affecting
parameters
Generation of distorted samples and
recording of parameter values
Subjective assessment of distorted samples
NN training and validation
Performance Evaluation
Specification
Preliminary evaluation
• NS-2 simulation
• Wired scenario
• Objective quality assessment
Pseudo-subjective quality
assessment 1
• Markov chain loss simulator
• Subjective assessment
• Neural network training
Pseudo-subjective
quality assessment 2
• Prediction with trained
neural network
Developement of UMTS
simulation models/tools
UMTS scenario
• NS-2 simulation
Preliminary Evaluation
Specification
Preliminary evaluation
• NS-2 simulation
• Wired scenario
• Objective quality assessment
Pseudo-subjective quality
assessment 1
• Markov chain loss simulator
• Subjective assessment
• Neural network training
Pseudo-subjective
quality assessment 2
• Prediction with trained
neural network
Developement of UMTS
simulation models/tools
UMTS scenario
• NS-2 simulation
Preliminary Evaluation

Goal


NS-2 simulation


Verify that our proposal effectively yields a
better visual quality than plain best-effort
Wireline scenario
Objective quality assessment

ITS impairment metric (ANSI standard)
Preliminary Evaluation
Results: visual impairment
ITS Video Quality Metric
VQM score (impairment)

1
0,8
BE
AF-30
AF-60
AF-90
AF-120
0,6
0,4
0,2
0
110
120
130
140
Network Load (% of bottleneck capacity)
Development of UMTS
Simulation Models & Tools
Specification
Preliminary evaluation
• NS-2 simulation
• Wired scenario
• Objective quality assessment
Pseudo-subjective quality
assessment 1
• Markov chain loss simulator
• Subjective assessment
• Neural network training
Pseudo-subjective
quality assessment 2
• Prediction with trained
neural network
Developement of UMTS
simulation models/tools
UMTS scenario
• NS-2 simulation
Development of UMTS
Simulation Models & Tools

Goal


NS link object with variable bandwidth and
delay
Tradeoff between simplicity and realism
Target Abstraction
GGSN
Video
server
Internet
SGSN
Mobiles
RAN
UMTS backbone
video source
DiffServ queue
background sources
rtr0
mobile terminal
UMTS link
• Low multiplexing (1-5 flows)
• Variable bandwidth
• Variable delay
Bandwidth Oscillation

A single mobile in HS-DSCH
BW (bit/s)
Time (s)
Approach

Markov-chain model


One state per bandwidth level
Transitions possible between all states
P(0,n)
P(0,1)
P(0,0)
0
P(1,n)
1
n
P(1,1)
P(1,0)
P(n,1)
P(n,0)
P(n,n)
Model Definition

We need to define:



Number and values of bandwidth levels
Transition period
Transition probability matrix
BW (bit/s)
To state
From state
Time (s)
0
1
2
n
0
1
2
P0,0 P0,1 P0,2
P1,0 P1,1 P1,2
P2,0 P2,1 P2,2
P0,3
P1,n
n
Pn,0 Pn,1 Pn,2
Pn,n
P2,n
Model Solution

Trace-based measurement
1.
2.

Run simulations with the Eurane UMTS
extensions
Measure transitions
Three « variability » scenarios



Low, medium, high
Combination of number of users, speed and
distance
One transition matrix per scenario
First Model


Transition period: 20 ms
12 bandwitdh levels (kbit/s)







0
208
318
400
680
920






1,272
1,848
2,136
2,632
3,040
3,600
Mean bandwidth: 290 kbit/s
Model Implementation

Main issue in NS-2:


Packet scheduling when bandwidth goes to
zero
Solved!
Pseudo-Subjective Quality
Assessment 1
Specification
Preliminary evaluation
• NS-2 simulation
• Wired scenario
• Objective quality assessment
Pseudo-subjective quality
assessment 1
• Markov chain loss simulator
• Subjective assessment
• Neural network training
Pseudo-subjective
quality assessment 2
• Prediction with trained
neural network
Developement of UMTS
simulation models/tools
UMTS scenario
• NS-2 simulation
Quality Affecting Parameters

Per-color packet loss rate





Green
Yellow
Red packet
Mean green loss burst size
Coding/mapping strategy
Example Generation



100 distorted clips
Each clip is associated to a combination
of parameter values
Markov-chain loss simulator
Subjective Assessment

20 assessors rated the 100 clips
4
2
Learning: Training the Neural
Network
Examples
Random Neural
Network
Subjective
grade
Network and source
parameters
Training
algorithm
UMTS Scenario
Specification
Preliminary evaluation
• NS-2 simulation
• Wired scenario
• Objective quality assessment
Pseudo-subjective quality
assessment 1
• Markov chain loss simulator
• Subjective assessment
• Neural network training
Pseudo-subjective
quality assessment 2
• Prediction with trained
neural network
Developement of UMTS
simulation models/tools
UMTS scenario
• NS-2 simulation
Topology
Clip:
318 kbit/s
video source
10 s/CIF/15 fps
60 byte packets
DiffServ RIO queue
10 Mbit/s 5ms
mobile
terminal
10 Mbit/s 10ms
10 Mbit/s 15ms
Mean rate: 1 Mbit/s
background sources
Pareto/TCP
UMTS link
Downlink delay: 200 ms
Uplink delay: 200 ms
Varying bandwith
Scenarios


Best Effort
AF

Two threshold models



Overlapped (G-RIO)
Scattered (RIO)
Three values of Wq



0.0017 « Normal »
0.5
1 (maximum reactiveness)
Pseudo-Subjective Quality
Assessment 2
Specification
Preliminary evaluation
• NS-2 simulation
• Wired scenario
• Objective quality assessment
Pseudo-subjective quality
assessment 1
• Markov chain loss simulator
• Subjective assessment
• Neural network training
Pseudo-subjective
quality assessment 2
• Prediction with trained
neural network
Developement of UMTS
simulation models/tools
UMTS scenario
• NS-2 simulation
Prediction: Using the Neural
Network
New network
(simulation output)
and source data
Trained Neural
Network
New subjective
score
Results
Predicted Visual Quality
Visual Quality
2,5
2
1,5
wq 0,0017
wq 0,5
1
wq 1,0
0,5
0
BE
RIO
GRIO
G
R
wq
wq
Case
IO
wq
1
1
0,
5
0,
5
wq
wq
wq
0,
00
17
0,
00
17
RI
O
IO
G
R
G
R
RI
O
IO
RI
O
BE
Predicted Visual Quality
2,5
2
0,4
1,5
0,3
1
0,2
0,5
0,1
0
0
Total Loss Rate
Results
Visual Quality and Total Loss Rate
0,6
0,5
Quality
Loss Rate
Conclusions



DiffServ-aware streaming can help
reduce visual distortion under drastic
bandwidth variations in UMTS for H.264
video
RIO thresholds affect visual quality
RIO must be highly reactive

Increase Wq -> react more to
instantaneous than to average queue size
Perspectives



Introduce delay oscillations in the UMTS
link model
Detailed study of RIO parameters
Introduce losses due to excessive delay
Questions

Thank you!
Quality assessment

There is a need to measure the quality
of the streamed signals


Does DiffServ-Aware Streaming yield a
better quality?
Which mapping strategy is better? (from a
perceived-quality point of view)
Pseudo-subjective assessment


Neural Networks approach
Link network and coding parameters to
human perception

MOS = f (network, coding)
Pseudo-subjective assesment

Methodology




Identification of the quality-affecting
parameters
Generation of distorted samples and
recording of parameter values
Subjective assessment of distorted samples
NN training and validation
DiffServ Mapping of H.264

Slice mapping
Slices
I
P
P
P
P
DiffServ Mapping of H.264

FMO mapping
FMO checker board
Slice group 0
Slice group 1
I
P
P
P
P
DiffServ Mapping of H.264

FMO mapping
FMO interleaved slices
Slice group 0
Slice group 1
I
P
P
P
P
Preliminary Evaluation
Specification
Preliminary evaluation
• NS-2 simulation
• Wired scenario
• Objective quality assessment
Pseudo-subjective quality
assessment 1
• Markov chain loss simulator
• Subjective assessment
• Neural network training
Pseudo-subjective
quality assessment 2
• Prediction with trained
neural network
Developement of UMTS
simulation models/tools
UMTS scenario
• NS-2 simulation
Preliminary Evaluation

Goal


NS-2 simulation


Verify that our proposal effectively yields a
better visual quality than plain best-effort
Wireline scenario
Objective quality assessment

ITS impairment metric (ANSI standard)
Preliminary Evaluation

Methodology
Preliminary Evaluation

Topology
video source
video sink
10 Mbit/s - 40 ms
TrTCM
agg0
agg1
agg9
background
aggregates
10 Mbit/s - 5 ms
RIO queue
rtr0
rtr1
30 Mbit/s - 5 ms
10 Mbit/s - 20 + 5k ms
100 Mbit/s - 5 ms
background sink
Preliminary Evaluation
Results: visual impairment
ITS Video Quality Metric
VQM score (impairment)

1
0,8
BE
AF-30
AF-60
AF-90
AF-120
0,6
0,4
0,2
0
110
120
130
140
Network Load (% of bottleneck capacity)
Preliminary Evaluation

Results: example clips

No DiffServ



Total loss rate: 11%
Impairment: 0.76
DiffServ


Total loss rate: 19%
Impairment: 0.66
Preliminary Evaluation

Paper accepted in Packet Video 2004


To be held next December in Irvine, Ca
« Novembre » acknowledged
Delay Variation



Exponentially distributed.
Max: 1 s.
Taken from:

Gurtov and Floyd. « Modelling Wireless
links for transport Protocols ». ACM
computer communications review. Dec.
2003.
Status

Current


Short term


Implementation of the first matrix in NS-2
Application
Mid term

Validation (emulation?)