École nationale supérieure des télécommunications 1 Z. Li, 2007 Distributed Coordination and Cross-Layer Optimization in Multi-Access Wireless Video Streaming System Zhu Li, PhD Principal Staff Research Engineer Multimedia Research Lab Motorola Labs, USA
58
Embed
Distributed Coordination and Cross-Layer Optimization in ...
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
École nationale supérieure des télécommunications 1 Z. Li, 2007
Distributed Coordination and Cross-Layer Optimization in
Multi-Access Wireless Video Streaming System
Zhu Li, PhD
Principal Staff Research Engineer
Multimedia Research Lab
Motorola Labs, USA
École nationale supérieure des télécommunications 2 Z. Li, 2007
Outline
• Overview of my Multimedia Research at Motorola Labs
• Motivation
• Problem Formulation– Interference limited multiple access channel,
– Typically operate at VLBR for mutli-media traffic,
– Multi-user diversity in channel states and content rate-distortion characteristics
– How to achieve a optimal received quality among users that also best utilizes radio
resource ?
• Solution– Problem parallelization and co-ordination via dual decomposition
– Cross-Layer optimization of video adaptation with resource pricing
• Simulation Results
• Conclusion & Future Work
École nationale supérieure des télécommunications 3 Z. Li, 2007
An overview of my
Multimedia Computing & Communication (MC2)
research at Motorola Labs
École nationale supérieure des télécommunications 4 Z. Li, 2007
2007
École nationale supérieure des télécommunications 5 Z. Li, 2007
Devices
• Explosive growth of devices:
– Billions of cell phones/PDAs
– Billions of computers
– Billions of TVs
– Billions of Media Players
• Different Multimedia
Capabilities in:
– display,
– capture,
– storage,
– computing,
– communication
École nationale supérieure des télécommunications 6 Z. Li, 2007
Networks
• Better technology from
equipment makers
– Better wireless spectrum efficiency,
WiMAX/LTE
– High speed DLS/Cable
– Fiber optical solutions
• More capacity from service
providers
– More bandwidth, better coverage,
– Convergence of data, voice and
media service from service
providers
– Vertical integration of application
and services
École nationale supérieure des télécommunications 7 Z. Li, 2007
Content
• Explosive growth of digital
media
– Web, Email, Audio, Video, Game
– News, Music, Movie, Talk show,
Game, 2nd Life.
• Rapid changes in the way
contents are produced and
consumed
– Personal vs Commercial
– Passive (TV) vs Interactive (Blog,
Game)
– Centralized vs P2P
École nationale supérieure des télécommunications 8 Z. Li, 2007
People and Technology
• People’s need:– Good Access, be able to get what you want, a storage and communication
problem
– Mobility across devices and access points: anywhere, on any device, not tied to TV only, get what they want, with good media quality (coding) and availability (communication/networking).
– Intelligence and Personalization: be able to find what they are interested in and locate what they want, browsing with (implicit and explicit) personal preference.
– Self-expression, Interaction and Social Networking, P2P video,video blog, live events streaming, social group based video sharing. Immersive video interaction.
• Technology Gap ?– Distribution: multimedia coding, streaming and networking
– Search & Mining, multimedia analysis, indexing and retrieval, search and mining.
– Interaction, visual/audio/motion sensor data processing, pattern recognition, tracking.
École nationale supérieure des télécommunications 9 Z. Li, 2007
It is a good time for MC2 research….
• Networked multimedia experience is still in its infancy, like
web browsing before broadband access and search
engines, there are,
• Real challenges and exciting research opportunities for
MC2 applications
– multimedia distribution (coding/communication) and,
– multimedia searching (computing) problems,
– multimedia based interaction (computing) problems
• Opportunities to advance the state-of-art in MC2 techniques:
– Systems: novel multimedia computing & communication systems
– Algorithms: visual signal processing, analysis, computer vision and
pattern recognition
– Tools: optimization, statistics, and machine learning
École nationale supérieure des télécommunications 10 Z. Li, 2007
MC2 problems under investigation
• Multimedia Computing Problems:
– Video Search: LUminance Field Trajectory (LUFT) Based Video Indexing
and Retrieval (With A. Katsaggelos at IVPL/Northwestern)
– Large Subject Set Visual Pattern Recognition: Localized subspace
learning for large label set (head pose and motion, suspect face database)
visual pattern recognition problems, (with Y. Fu and T. Huang at
IFP/UIUC)
– Spatio-Temporal Visual Pattern Recognition: Human behavior
recognition, accelerometer sensorial data based human motion/behavior
recognition, spatio-temporal volume tensorial modeling, (with Y. Fu, S.
Yan at IFP/UIUC)
École nationale supérieure des télécommunications 11 Z. Li, 2007
MC2 problems under investigation
• Multimedia Communication Problems:
– Video Coding and Adaptation (Motorola Lab):
» Video Summarization and Coding for VLBR (12~48kbps) Streaming
» H.265 research: joint scalability and error-resilience coding, motion field
scalable coding, new visual signal decomposition schemes.
» Multi-View Video Coding and Networking,
– Video Networking (with J. Huang and M. Chiang at Princeton):
» Video over Wireless Multi-Access Channel: Adaptation and Resource
Pricing for Multiple Access Wireless Video Communication
» Video over P2P networks: self-organizing multicasting, distributed resource
pricing.
» Video over Wireless Broadcast Channel: Joint source-channel coding for
wireless video broadcasting (mobile TV) with limited feedback, relay with
network coding, optimization.
École nationale supérieure des télécommunications 12 Z. Li, 2007
In this talk
• Purpose:
– To lay a landscape of my current MC2 research and collaborations
– To show some in-depth techniques and results in Video over Mutli-Access
Network problems,
– To share some of my views and opinions on MC2 research and
applications,
– To have potential collaborations in the future with interested faculties
École nationale supérieure des télécommunications 13 Z. Li, 2007
Mixed Voice/Video over CDMA Up Link
Radio tower
• Mixed QoS requirements for
Video/Voice traffics
• Limited resource, video has to
operate at VLBR
• Shared radio resource, and
interference limited capacity,
Ri=f(Pi;P-i),
• Diversity of channel gains and
source rate-distortion
characteristics among users.
• How to optimize video adaptation
and transmission to achieve better
QoS and radio resource efficiency
?
École nationale supérieure des télécommunications 14 Z. Li, 2007
A General Formulation
• Total utility maximization subject to a shared resource
constraint,
– Where utility function Ui() is a concave differentiable function reflecting the
quality-bit rate/resource trade-offs. (true for most video source’s PSNR-R
function)
– Difficult to solve the primal problem by allocating {xi} directly, because of
coupling of {xi} in constraint.
– Transform the problem for a distributed solution, utilizing computing
capability at mobiles
max,,,
..,)(max21
xxtsxUi
i
i
iixxx n
École nationale supérieure des télécommunications 15 Z. Li, 2007
Distributed Solution of the Dual Problem
•Lagrangian relaxation:
•The dual problem:
•Decomposed into n separable video adaptation problems at
mobiles :
•And a base station resource pricing problem:
]}))(([max{min max,,,0 21
xxxUi
iiixxx n
))((maxarg*
))((max
))((max
,,,
max,,,
21
21
iiix
i
i
iiixxx
i
iiixxx
xxUx
xxU
xxxU
i
n
n
i
ig )(min0
)()()( max
** xxxUg iiii
)()(),,,,( max21 xxxUxxxLi
i
i
iin
École nationale supérieure des télécommunications 16 Z. Li, 2007
BTS Mobile i
Announce resource price in iteration kk
Mobile optimization:
Protocol for Distributed Optimization
))((maxarg* i
k
iix
i xxUxi
Report back resource used xi* in iteration k
Increase price, if
Otherwise decrease price i
i xx max
k
kk 1 Mobile optimization:
))((maxarg* 1
i
k
iix
i xxUxi
École nationale supérieure des télécommunications 17 Z. Li, 2007
Distributed Optimization for Multiple Access Video Network
• Geometrical
Interpretation on price:
– From the Karush-Kuhn-
Tucker (KKT) condition:
– Allocations {xi*} will have
the same marginal utility
(slope) as -price.
– Optimal price must also be
tight on all available
resource.
*
**,,0
i
i
i x
U
x
L
U1(x1)
U2(x2)
U3(x3)
x1* x2
* x3*
U1(x1)
U2(x2)
U3(x3)
x1* x2
* x3*
École nationale supérieure des télécommunications 18 Z. Li, 2007
Video Over Multiple Access Channel
• In solving real world problems with this distributed pricing
scheme:
– Source coding: scalability, adaptability issues
– Diversity in Channel state
– Diversity in content
– Collaboration in resource allocation, scheduling
– Uplink problem: interference limited
– Downlink problem: power limited.
– Computational complexity
École nationale supérieure des télécommunications 19 Z. Li, 2007
CDMA Uplink with Mixed Voice/Video Traffic
• Consider a single cell CDMA uplink:
– Pvoice – received power for a voice user
– M – total voice users
– Pvideo – total received power for all video users
École nationale supérieure des télécommunications 41 Z. Li, 2007
Related Publications
– Z. Li, J. Huang, and A. K. Katsaggelos, “Pricing Based Collaborative Mutli-User Video Streaming Over Power Constrained Wireless Down Link”, oral paper, IEEE Int’l Conference on Acoustics, Speech and Signal Processing (ICASSP), Toulouse, France, 2006.
– Z. Li, J. Huang, M. Chiang, and A. K. Katsaggelos, “Intelligent Wireless Video Communication: Source Adaptation and Multi-User Collaboration”, invited paper, special issue on Multimedia Communication, Ed. Changwen Chen, China Journal of Communication, December, 2006.
– Z. Li, J. Huang, and A. K. Katsaggelos, “Utility Driven Video Segment Scheduling for Peer-to-Peer Live Video Streaming System”, 45th Allerton Conference on Communication, Control and Computing, Monticello, IL, USA, 2007.
– J. Huang, Z. Li, M. Chiang, and A. K. Katsaggelos, “Pricing Based Efficient Multi-User Wireless Video Communication over a CDMA Downlink”, accepted to IEEE Trans. on Circuits & System for Video Tech.
– Y. Yang, Z. Li, W. Shi, Y. Chen, and H. Xu, “Network-Aware Mobile Gaming Traffic Shaping and Scheduling”, submitted to IEEE Trans. on Multimedia.
École nationale supérieure des télécommunications 42 Z. Li, 2007
Questions
?
…Thanks!
École nationale supérieure des télécommunications 43 Z. Li, 2007
Backup: Network Device Icons
Radio tower
École nationale supérieure des télécommunications 44 Z. Li, 2007
Power Constrained CDMA Down Link Video
• Code Division
• Total Transmitting Power Constrained
• Content R-D diversity
• Maximize total video quality
Video source 1
Video source 2
Video source n
Radio tower
École nationale supérieure des télécommunications 45 Z. Li, 2007
Power Constrained CDMA Down Link Video
• Problem Formulation:
– Considering a segment of video of duration T for all users
– Allocate a power function for each user, subject to a total power constraint
• Similar solution to the uplink problem:
– Goal: achieve max total quality among users
– Two stage solution:
» Power Level Allocation
» Joint Packet Scheduling
» Base station iterates on power price, until total utility converges.
0)(,)(..),)((max1
max
01)(,),(),( 21
tPPtPtsdttPU j
n
j
j
T
t
j
n
j
jtPtPtP n
École nationale supérieure des télécommunications 46 Z. Li, 2007
Down Link Problem Dual Decomposition
• Dual Decomposition:
– Lagrangian:
– Dual problem
– Where source problem becomes separable, for given price :
– Base station problem: pricing control:
)()(max)(1
max1,,, 21
n
jjj
n
jj
PPPPPPUJ
n
)}()(max{min1
max
1,,,0 21
n
j
jj
n
j
jPPP
PPPUn
j
i
jjP
j PPUPj
)(max*
i
)}(,0max{ max
*1
j
j
ii PP
École nationale supérieure des télécommunications 47 Z. Li, 2007
Down Link Source Problem
• Source Problem:
– Maximize a surplus function:
– For each user, find an optimal power level Pji that maximizes surplus
– Power price is given by base station,
– Utility depends on video coding and available adaptation scheme
ji
jjP
ij PPUP
j
)(max
École nationale supérieure des télécommunications 48 Z. Li, 2007
Down Link Source Problem
• Source Problem:
– Video Summarization for VLBR case, very similar to the uplink case:
D(S) is the summarization distortion, P(S, W, h) is the power level needed to
transmit all summary frames with bandwidth W and channel gain h.
– Scalable video stream extraction for medium rate range:
),,()(minarg)(*jj
ij
S
ij hWSPSDS
j
),,()(maxarg)(*
jj
i
jL
i
j hWLPLPSNRLj
École nationale supérieure des télécommunications 49 Z. Li, 2007
Source Problem
• Solutions for both Summarization and Bit Extraction are
similar in structure
– A “Convex Hull” solution similar to the bit constrained summarization.
– FGS scalable stream is quantized into packets, an optimal extraction for
given price on resource is a path thru the all possible extraction routes.
– Has the following recursive relation:
– Which gives us a polynomial complexity Viterbi algorithm like solution.
}{max
}][max{max
max
*
1
1
1,,,
1,,,
*
121
21
nnn
nnkknn
kkn
llnl
ll
n
k
llllll
n
k
lllll
n
RUV
RURU
RUV
École nationale supérieure des télécommunications 50 Z. Li, 2007
FGS Packets extraction
• Example of DP trellis and optimal path
• GoP size n=8, data from “foreman” sequence
• Optimal extraction paths for prices 0.002 and 0.008
École nationale supérieure des télécommunications 51 Z. Li, 2007
Joint Packet Scheduling
• Now each video source come up with a set of packets
(frames) with different size and deliver deadline, is it
actually schedulable ?
• For scalable case, no need to schedule, because of receiver
buffering .
• For video summary case, need to guarantee each frame’s
data arrive on time.
• Solution: A greedy water filling scheduling algorithm.
École nationale supérieure des télécommunications 52 Z. Li, 2007
Greedy Water Filling Solution
• Packets are identified by triplets: {Bki, tk
j, Tkj} sort by their
deliver deadlines
• For video summary case, need to guarantee each frame’s
data arrive on time.
• Solution: A greedy water filling scheduling algorithm.
– For a given packet and its deadline, find the min power level that will be
able to send it on time.
École nationale supérieure des télécommunications 53 Z. Li, 2007
Greedy Water Filling Solution
• Power function for user j in transmitting packet k
else
TtttPLLtP
kkjk
,0
],[),();(
• Determine the level L
thru water-filling
• B(L*) = Bkj
dtWN
tPh
WN
LtPhWLB
j
jj
T
t
j
j
kjk
k
))(
1log(
));(
1log()(
0
0
École nationale supérieure des télécommunications 54 Z. Li, 2007
Simulation
• For Pmax=2.4, resulting optimal video summaries for 4 users:
– Optimal price =101.45
– Average bit rates: 20.1 43.3 8.1 9.4 kbps
– Channel: H=[0.75 1.0 0.8 0.65]
*
Pmax=2.4/4=0.6
École nationale supérieure des télécommunications 55 Z. Li, 2007
Simulation
• Joint packet scheduling vs. single user greedy scheduling
– Pmax = 2.4
– Left: Joint Scheduling, Right: Single user based, not schedulable for
Pmax=2.4
École nationale supérieure des télécommunications 56 Z. Li, 2007
Summary for Down Link work
• Solution based on Dual Decomposition
• Coordination thru pricing on power
• Collaboration thru joint packet scheduling
• Computational Complexity can be distributed
• Works for a variety of adaptation scheme like
summarization, scalable stream extraction.
École nationale supérieure des télécommunications 57 Z. Li, 2007
Summary for pricing scheme
• Dual decomposition is a powerful framework in distributed
optimization
– Source granularity and utility modeling is essential
– Similarities to a set of economics problems, methodologies like pricing
and auction can be applied
• Future work
– Investigate pricing scheme for P2P, convergence issues, stability issues
– Source-Channel coding and optimization scheme for video broadcasting
– Auction schemes for multi-user video over wireless mesh network.
• Thanks to my research collaborators
– Prof. Mung Chiang, Princeton
– Prof. Jianwei Huang, CUHK
– Prof. Aggelos Katsaggelos, Northwestern
– Ms. Ying Li, PhD Student, Princeton
École nationale supérieure des télécommunications 58 Z. Li, 2007