Opti l All i fimal Resource Allocation for Video Streaming...
Transcript of Opti l All i fimal Resource Allocation for Video Streaming...
i l All i fOptimal Resource Allocation for Video Streaming over DistributedVideo Streaming over Distributed Communication Networks
Ling GuanRyerson Multimedia Research Laboratory &Ryerson Multimedia Research Laboratory &Centre for Interactive Multimedia Information MiningDepartment of Electrical and Computer Engineering,R U i it T t C d
12/7/2009 1
Ryerson University, Toronto, [email protected], http://www.rml.ryerson.ca/
Acknowledgmentg
The presenter would like to thank Dr. Yifeng He p gfor his persistent effort in making this research a true success
The presenter also would like to thank Dr. Ivan Lee for his continuous contributions to this workLee for his continuous contributions to this work
This research is supported by The Canada Research Chair (CRC) Program The Canada Research Chair (CRC) Program,
Canada Foundation for Innovations (CFI), Th O t i I ti T t (OIT) d The Ontario Innovation Trust (OIT), and
Ryerson University
12/7/2009 2
Major Publicationsj Y. He, I. Lee and L. Guan, “"Distributed throughput optimization in
P2P VoD systems ” to appear in IEEE Transactions on MultimediaP2P VoD systems, to appear in IEEE Transactions on Multimedia. Y. He, I. Lee and L. Guan, “Distributed algorithms for network
lifetime maximization in wireless visual sensor networks,” IEEE Transactions on Circuits and Systems for Video TechnologyTransactions on Circuits and Systems for Video Technology(subject to minor revision).
Y. He, I. Lee and L. Guan, “Optimized video multicast in wireless ad hoc network using network coding ” to appear in IEEE Transactionshoc network using network coding, to appear in IEEE Transactions on Circuits and Systems for Video Technology.
Y. He, G. Shen, Y. Xiong and L. Guan, “Optimal prefetching scheme in P2P VoD applications with guided seeks ” accepted byscheme in P2P VoD applications with guided seeks, accepted by IEEE Transactions on Multimedia
Y. He and L. Guan, “Optimal resource allocation in distributed visual communications ” in Intelligent Multimedia Communication:visual communications, in Intelligent Multimedia Communication: Techniques and Applications, C.W. Chen, Z. Li and S. Lian, eds, Springer-Verlag, to be published in 2009.
12/7/2009 3
Outline
Motivation and contributions Principles of convex optimization Resource allocation in distributed video Resource allocation in distributed video
communication systemsThro ghp t ma imi ation in P2P VoD applications Throughput maximization in P2P VoD applications
Network lifetime maximization in wireless visual sensor networksnetworks
Optimization for video streaming over wireless ad hoc networksnetworks
Conclusions
12/7/2009 4
Motivation
Many multimedia applications involve real-a y u ed a app ca o s o e eatime video transmissions over distributed networks Some examples:networks. Some examples:
P2P VoD applications Video streaming over wireless ad hoc networks Video streaming over wireless ad hoc networks Wireless visual sensor networks
Distributed algorithms to optimize resource Distributed algorithms to optimize resource allocations in distributed networks
Each node has local knowledge Each node has local knowledge No centralized controller Scalability
12/7/2009 5
Scalability
Challengesg P2P VoD applicationspp
Limited bandwidth Unreliable and dynamic peers Random seeks
Video streaming over wireless ad hoc networks High transmission error rate Source rate allocation Optimized routing scheme Optimized routing scheme
Wireless visual sensor networks Video compression consuming a large amount of power Video compression, consuming a large amount of power Trade-off between network life time and video quality
12/7/2009 6
Contributions
Formulate resource allocation problems in o u a e esou ce a oca o p ob e sdistributed networks based on convex optimization andoptimization, and
Solve them using distributed algorithmsTh h t i i ti i P2P V D li ti Throughput maximization in P2P VoD applications
Optimized video streaming over wireless ad hoc t knetworks
Network lifetime maximization in wireless visual t ksensor networks
12/7/2009 7
Outline
Motivation and contributions Principles of convex optimization Resource allocation in distributed video Resource allocation in distributed video
communication systemsThro ghp t ma imi ation in P2P VoD applications Throughput maximization in P2P VoD applications
Network lifetime maximization in wireless visual sensor networksnetworks
Optimization for video streaming over wireless ad hoc networksnetworks
Conclusions
12/7/2009 8
Convex optimization problemp p
Convexity is often viewed as the “watershed” between easy and hard y yoptimization problems.
Convex optimization: the primal problem [Boyd2004]
miww
i ,...,1 ,0 :Subject to :minimize 0
xx
piqi ,...,1 ,0 x
functions,convex are , e,on variabloptimizati is where 0 xxRx in ww
function. affine is xiq
Solution:Distributed algorithm: Lagrange duality properties construct a dual problem
dual decomposition solve dual problem with subgradient method obtain primal optimization variable from dual variables
12/7/2009 9
Convex function and affine function
Convex function:
xqWh t i ffi f ti ? xiqWhat is affine function?
Definition: 0 and ,0Let 21 xqxq ii [Boyd2004]
,0haveweif
,1 21
yqxxy
i R
affinealwaysisfunctionLinearfunction. affine is Then ,
xqyq
i
i
12/7/2009 10
affine. alwaysisfunction Linear
Dual solutionConstruct a dual problem [Boyd2004]
The Lagrangian:
,, 0 xxxvλx p
ii
m
ii qvwwL
. variablesdual are ,wher 11
vλ
ii
The dual function:
p
ii
m
ii qvwwLg 0minmin ,,, xxxvλxvλ The dual function:
iii
iii qvwwLg
110minmin ,,, xxxvλxvλ
XX
The dual function is always concave
The minimum of several linear functions is always concave
The dual problem: 0:Subject to, :maximize
λvλg
The dual function is always concave.
D l bj ti f ti0 :Subject to λ Dual objective function
xxxxxxxvλX
011
011
0min, wqvwwqvwwgp
iii
m
iii
p
iii
m
iii
12/7/2009 11
Dual objective value is no larger than primal objective value.
Dual solutionDuality gap: 0 -*
0 *gwd** gg(u,v)g value,objective dual maximal theis
*0
*0 ) w( value,objective primal minimal theis ww x
Strong duality: under Slater's condition,*gwd *
0 :is that ,0gapduality
00
g0,g pySlater’s condition: There exists a x that satisfies:
miw 10 x piq 10xand miwi ,...,1 ,0 x piqi ,...,1 ,0 xand
In other word, there exist a strictly feasible point
Weak duality: d > 0
12/7/2009 12
Optimalityp y
Objective value of primal problem
Objective value of dual problem Duality gap=0,Obtain minimal primal objective and maximal dual objective at the
Optimal primal variables
same time
12/7/2009 13
Optimal dual variables
Subgradientg• A subgradient of function f at point x is any vector g, g p y g
that satisfies the inequality: f(z)>=f(x)+gT(z-x) for all z, f is a convex function
[B d2004]
12/7/2009 14
[Boyd2004]
SubgradientgSubgradient method to update dual variables [Bertsekas2003]
p,...ifvv
m,...i hk
ikk
ik
i
ki
kki
ki
,1
,1 ,0max1
1
p,fiii , lyrespective ,at oft subgradien theis , where k
ik
ik
ik
i v-gfh miwh kkk 1* vλx p,...iqf
m,...iwhkk
ik
i
ii
,1
,1 *
v,λx
v,λx
Non-summable diminishing Step size for convergence:Non summable diminishing Step size for convergence:
1
,0lim ,0k
kk
k
k
•Find optimal primal variables from dual variables: vλ,xvλ,x , inf arg* L
12/7/2009 15
Example: convex optimizationp p 04S bj
2056.0 :minimize2
20 xxxw
5]. 0[ ,04 :Subject to 2
1
xxxxw
16
18
20
16
18
20
10
12
14
l obj
ectiv
e va
lue
10
12
14
obje
ctiv
e va
lue
11.13 11.13
2
4
6
8
Prim
al
2
4
6
8
Dua
l
Optimal primal variable=2.56Optimal dual variable=0.6
0 0.5 1 1.5 2 2.50
Primal variable0 0.5 1 1.5 2 2.5 3
0
Dual variable
Primal Objective Value Dual Objective Value
12/7/2009 16
Primal Objective Value Dual Objective Value
Regularization term makes dual function gdifferentiable
Non-differentiable Differentiable
.-xx
01 :Subject to2 :minimize
2
2.5=0=0.05=0.1=0.15
Non differentiable Differentiable
jLagrange dual function:
-xxg 1 2 )( 1.5
2
l fun
ctio
n va
lue
2i i i 2
1
Lagr
ange
dua
l
.-xxx01 :Subject to
2 :minimize 2
Lagrange dual function:0 0.5 1 1.5 2 2.5 3 3.5 4
0
0.5
Lagrange dual function:
-xxg 1x 2 )( 2
12/7/2009 17
Dual decompositionpWhy solve the primal problem by first solving the dual problem?
1. Dual problem is always convex, can be efficiently solved.2. Dual problem can be decomposed easily, leading to distributed algorithm
D l d iti ( t bt i di t ib t d l ith )Dual decomposition: ( to obtain distributed algorithm)
Primal problem:
xfxf :minimize
D l bl
cx xxfxf
21
21
:Subject to :minimize
Dual problem:
infinf
inf :maximize 2121
cxxfxxfcxxxfxf
Dual variables update:
0 :Subject to
infinf 2211
λcxxfxxf
211 ,0max xxckkk
12/7/2009 18
Primal decompositionp
Primal problem:p
cx x
xfxf
21
2211
:Subject to :minimize
*2
*1 :max ffimize
α
Master problem
c x
xf
2
22
:Subject to :minimize
1
11
:Subject to :minimize
xxf
Subproblem 1 Subproblem 2
12/7/2009 19
Distributed solution
Original optimization problemProblems for FormulationOriginal optimization problem
(primal problem)video streaming over distributed networks
Dual problemUpdate dual variables
Subproblem 1 Subproblem N
Compute primal variable x1at node 1
Compute primal variable xNat node N
12/7/2009 20
Outline
Motivation and contributions Principles of convex optimization Resource allocation in distributed video Resource allocation in distributed video
communication systemsThro ghp t ma imi ation in P2P VoD applications Throughput maximization in P2P VoD applications
Network lifetime maximization in wireless visual sensor networksnetworks
Optimization for video streaming over wireless ad hoc networksnetworks
Conclusions
12/7/2009 21
Outline
Motivation and contributions Principles of convex optimization Resource allocation in distributed video Resource allocation in distributed video
communication systemsThro ghp t ma imi ation in P2P VoD applications Throughput maximization in P2P VoD applications
Network lifetime maximization in wireless visual sensor networksnetworks
Optimization for video streaming over wireless ad hoc networksnetworks
Conclusions
12/7/2009 22
Client/server VoD vs. P2P VoD/ Client/server VoD Server
Server bandwidth bottleneck
P2P li ti P2P applications P2P VoD, P2P live TV
Peer
Client/server
Single-layer coded P2P VoD Each user receives the same quality
Source node
q y
Scalable P2P VoD Lower bandwidth lower quality High bandwidth higher quality
Peer
12/7/2009 23
P2P
Related work P2P VoD applications
Existing P2P architectures Existing P2P architectures Buffer-forwarding architectures:
Tree-based: P2VoD [Do2004]Mesh based [Li2006] Mesh-based [Li2006]
Storage-forwarding architecture: VMesh [Yiu2007] Hybrid-forwarding architecture [[YHe-ICASSP08,YHe-TMM08a]
E i i i i i i P2P Existing optimizations in P2P Minimum-Delay for constant-bit-rate (CBR) P2P media session
[Wu2005] Distributed auction algorithm for rate allocation [Li2006] Maximization of throughput in scalable P2P VoD systems taking
into account packet loss due to excessive delay at each link [YHe-CICME07a]
Prefetching in P2P [YHe-TMM08b]
12/7/2009 24
Scalable VoD: prioritized coding schemep g
Prioritized coding scheme g Originally proposed in [Chou2003], useful for video
broadcast Layered coding + prioritized packetization + network
coding (at source and intermediate nodes) Advantages:
Scalable Resilient to packet loss Duplicate-free
A l th h t t i l d t hi h lit A larger throughput at a receiver leads to a higher quality
12/7/2009 25
Buffer-forwarding overlay for P2P VoDg y
Existing approaches to construct Existing approaches to construct buffer-forwarding overlay Tree-based Mesh-based
Our work is on the module of throughput maximizationthroughput maximization
Peer i
Module of buffer-forwarding overlay construction
Module of throughput maximization
Allocate rate at each outgoing link
Peer i
overlay construction maximization each outgoing link
12/7/2009 26
Graph modelp A network can be modeled as a directed graph
G=(N,L) N is the set of nodes
L i th t f li k L is the set of links Matrix to represent the node-link relationship
00011A: the relationship between the node and its connected links
A+ th l ti hi b t th
11 2
110001011001101000
A
00011
Node 2
A+ : the relationship between the node and its outgoing links
A th l ti hi b t th
23
4 5
3
00000100000110000011
A
A- : the relationship between the node and its incoming links
A= A+ A-
45
11000001100000100000
AA network
12/7/2009 27
A= A+ - A
Graph modelp
Network graphNetwork graphA P2P is modeled by a directed graph G = (N, L)
otherwise.,0, node intolink incomingan is link if,1, node fromlink outgoingan is link if,1
ilil
ailMatrix A:
.otherwise,0, node fromlink outgoingan is link if,1 il
ailMatrix A+:
.otherwise,0, node intolink incomingan is link if,1 il
ailMatrix A-:
12/7/2009 28
Throughput maximization in buffer-
P bl f l ti
g pforwarding systemsProblem formulation:
1 maximize pxa l
llil Aggregate throughputAggregate throughput Ll
lNi Ll
llil xpxa 21 minimize
, subject to
NiI
Ni, sxa rLl
lil
Ni Ll
Source rate constraint
D l d b d idth t i t
gg g g p LlNi Ll
, ,
,
NiOxa
Ni, Ixa
ilil
iLl
lil
Download bandwidth constraint
Upload bandwidth constraint
0
, ,
Ll
Llxcx lLm
mlml
Ll
Link-forwarding constraint
. ,0 Llxl
Original optimization problem:Linear Programming (LP) problem
Converted to:Strictly convex optimization problem distributed algorithm to solve it
12/7/2009 29
Linear Programming (LP) problemStrictly convex optimization problem, distributed algorithm to solve it
Throughput maximization in buffer-g pforwarding systems (cont.)Simulation results:Simulation results:
0.5
Centralized LPDistributed optimizationProportional allocation0.5
Centralized LPDistributed optimizationProportional allocation
Achievable throughput in buffer-forwarding architecture
0.4
Proportional allocationEqual allocation
Mbp
s] 0.4
Proportional allocationEqual allocation
Mbp
s]
Any better architecture to improve the achievable th h t?
02
0.3
thro
ughp
ut [M
02
0.3
thro
ughp
ut [M throughput?
0.1
0.2
Ave
rage
t
0.1
0.2
Ave
rage
t
C i f th h t
100 200 300 400 5000
Number of peers
100 200 300 400 5000
Number of peers
12/7/2009 30
Comparison of average throughput
Impact of regularization factorp g
3000
3500
4000
erge
nce
0.4
0.405
0.41
ps]
2000
2500
3000
ratio
ns fo
r con
ve
0.39
0.395
0.4
thro
ughp
ut [M
bp
500
1000
1500
Num
ber o
f ite
r
0.375
0.38
0.385
Ave
rage
t
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.80
Regulation factor0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
0.37
Regulation factor
C l i Sub optimalityComplexity Sub-optimality
Throughput maximization in buffer-forwarding P2P VoD
12/7/2009 31
Impact of step size to convergence speedp p g p
Non summable diminishing1
=1.0Non-summable diminishing step size sequence:
0.7
0.8
0.9 1.0=2.0
k 0.5
0.6
Link
rate
kk
0.2
0.3
0.4
0 10 20 30 40 50 60 70 80 90 1000
0.1
Iteration No.
Iteration of a link rate in a 100-peer P2P VoD buffer-forwarding system
12/7/2009 32
Dynamics handling in P2P VoDy g
Convergence speed during transition3
76
-75
Convergence speed during transition
1
2
X2=2X3=3
1
4
X4=0
u2
78
-77
-76
ion
valu
e
1
5X1=5
u1
80
-79
-78
Dua
l fun
ct5
Steady state (flow conservation): x1=x2+x3Dual variable: u1=0.4, u2=0.5, u3=0.3
980 990 1000 1010 1020 1030 1040 1050 1060 1070 1080-81
-80
No. of iteration
Peer 3 left, peer 4 joined
Update dual variables: u1=0 6 u2=0 3 u4=0 1Dynamic 200-peer and 4-parent/peer scenario, 10 peers left and 20 new peers joined in the previous time slot
Update dual variables: u1 0.6, u2 0.3, u4 0.1
Update primal variables: x1=4, x2=3, x4=1Reach new steady state: x1=x2+x4
12/7/2009 33
Communication overhead for throughput maximization bl i b ff f diproblem in buffer-forwarding systems
ii vu ,3 4
2Link 1 Link 2
The dual variables at node i:The dual variables at link l: l
Th i l i bl (li k t ) t li k l x2
1Link 3
The primal variable (link rate) at link l: lxEach node is responsible for updating the primal and dual variables at this node
d t it t i li kand at its outgoing links
Node 2 computes the link rate at link 3: 2132133 1 vupx 2132133 1 vupxCommunication overhead: node 2 needs to request u from node 1 requestCommunication overhead: node 2 needs to request u1 from node 1, request λ1 from node 3, and λ2 from node 4
The update of dual variables requires the link rates of connected links, no communication overhead
ui: incoming link ratesvi: outgoing link rates
12/7/2009 34
λl: incoming link rates
A Hybrid-forwarding Architecture for y gP2P VoD
Server 1
130
116
104
43
22
3
7
130
134
34
115
27
12Video-1 buffer-forwarding 12
75
Buffer-forwarding link
Idle peers
Video-2 buffer-forwarding overlay
Video 1 buffer forwarding overlay
Hybrid-forwarding architectureShare both buffer and storage, to improve throughput
Storage-forwarding link
g p g pPeers contribute their stored segments to other peersImplement service differentiation among videosStored segments are stable robust to peer dynamics
12/7/2009 35
Stored segments are stable, robust to peer dynamics
Buffer-forwarding and storage-forwarding g g goverlay construction
1 Source node 1 Source node
St d t 22
3
Buffer2
Stored segment 2Requesting segment 4
33
4Peer 3
S
3
Stored segment 4Requesting segment 5
Playback progress10 15 20
Stored segment 1Requesting segment 2
Buffer-forwarding overlay Storage-forwarding overlay
12/7/2009 36
Throughput maximization in hybrid-
P bl f l ti
g p yforwarding architectureProblem formulation:
Aggregate weighted throughput 1 maximize pxa Ni Ll
llili
Aggregate throughput
Ll
lNi Ll
llili xpxa 21 minimize
Source rate constraint
Download bandwidth constraint,
, subject to
N i, Ixa
Ni, sxa
ilil
rLl
lil
Upload bandwidth constraint
Buffer-forwarding constraint, ,
,
Llxcx
Ni, Oxhxd
Blmlml
iLl
lilLl
lil
Ll
Buffer forwarding constraint
0
, ,
,,
Llx
LlFhx Sl
Niiill
lLm
mlml
Storage-forwarding constraint
Original optimization problem:Linear Programming (LP) problem
Converted to:Strictly convex optimization problem distributed algorithm to solve it
. ,0 Llxl
12/7/2009 37
Linear Programming (LP) problemStrictly convex optimization problem, distributed algorithm to solve it
Throughput maximization in hybrid-
Simulation results:
g p yforwarding architecture (cont.)Simulation results:
0 6
0.7Hybrid-forwarding centralizedHybrid-forwarding distributedBuffer-forwarding distributed
0.5
0.6ut
[Mbp
s]g
Storage-forwarding distributed
0.3
0.4
age
thro
ughp
u
0.1
0.2
Ave
ra
100 200 300 400 5000
Number of peers
Comparison of average throughput
12/7/2009 38
Comparison of average throughput
Other observation 1 – Number of Peers
Simulation results (Poisson distribution):Simulation results (Poisson distribution):
Buffer forwarding Hybrid forwarding
12/7/2009 39
Other observation 2 – Scalability
Simulation results:
y
Simulation results:
Comparison of the cost introduced by the proposed distributed algorithm in buffer-forwarding P2P VoD systems with different network sizes: (a)in buffer forwarding P2P VoD systems with different network sizes: (a) the number of the overlay links, (b) the average number of iterations per link, and (c) the average communication overhead per node
12/7/2009 40
Outline
Motivation and contributions Principles of convex optimization Resource allocation in distributed video Resource allocation in distributed video
communication systemsThro ghp t ma imi ation in P2P VoD applications Throughput maximization in P2P VoD applications
Network lifetime maximization in wireless visual sensor networksnetworks
Optimization for video streaming over wireless ad hoc networksnetworks
Conclusions
12/7/2009 41
Wireless visual sensor networks Applications of WVSN Video sensor
Video surveillance, environmental tracking
The proposed distributed The proposed distributed algorithm: Maximize the network lifetime by
jointl optimi ing so rce ratesjointly optimizing source rates, encoding powers, and routing scheme
Total power consumption at a Sinkp pnode
fEncoding power + transmission power + reception power
A wireless visual sensor network
Network lifetime: defined as minimum node lifetime
12/7/2009 42
Related work Network Lifetime maximization for Wireless visual sensor
networks (WVSNs)networks (WVSNs) Conventional wireless sensor networks
Collect data (e.g., temperature), negligible power consumption on ( g ) g gsignal processing at sensor node
Existing distributed optimizations for conventional wireless sensor networkwireless sensor network Tradeoff between the source rate allocation and the network
lifetime [Nama2006] Distributed algorithm to maximize lifetime [Madan2006] These methods cannot be applied directly to WVSNs, since they
omit the processing power consumption at the sensor nodes Maximization of network lifetime for WVSN by jointly optimizing source
rates, encoding powers and routing scheme. [YHe-ICME07b, YHe-TCSVT2008a]
12/7/2009 43
Markov model and encoding powerg p
Channel error model01lq01lqChannel error model
Two-state Markov model0 1
lq011 lq 101 lq
0 1
lq011 lq 101 lq
A bit b bilit10lb qp
10lq10lq
Average bit error probability: ,0110ll
ll qq
p
packet loss rate (PLR): .11Gb
lpl pp p ( ) ll pp
Power consumption modelEncoding power consumptionEncoding power consumption
Power-rate-distortion model [He2006]3/22 Ps .2 shh Ps
sh ed
Under the same encoding power, increasing rate reducing distortion
12/7/2009 44
Under the same rate, increasing encoding power reducing distortion
Graph modelp
Network graphNetwork graphA WVSN is modeled by a directed graph G = (N, L)
otherwise.,0, node intolink incomingan is link if,1, node fromlink outgoingan is link if,1
ilil
ailMatrix A:
.otherwise,0, node fromlink outgoingan is link if,1 il
ailMatrix A+:
.otherwise,0, node intolink incomingan is link if,1 il
ailMatrix A-:
Each session follows the law of flow conservation:
, , , NiVhxa hihlil
12/7/2009 45
hiLl
hlil
Achievable maximum network lifetime Without loss Problem formulation:
iii i BTT i
Network lifetime
, , , :subject to
minmin maximize
NiVh xa
yacycaPTT
hiLl
hlil
Lllil
r
Lll
slilsi
iinet
Flow conservation
Network lifetime
, ,
, ,
3/2..2 VhDe
Llyx
hPs
Vhlhl
Ll
shh
Aggregate flow rate
Video quality requirement
. ,0 , ,0 , , ,0
VhPVhsLlVhx
shh
hl
Original optimization problem
12/7/2009 46
Achievable maximum network lifetime (cont.)
P bl i Problem conversion: minimize q
Flow conservation
Aggregate flow rate, ,
, , , :subject to
Llyx
NiVh xa
lhl
hiLl
hlil
Video quality requirement , ,
, ,//log 3/22
NiqByacycaP
VhsPD
ililr
lslilsi
hshh
Vh
Power constraint
. ,0 , ,0 , , ,0
VhPVhsLlVhx
shh
hl
LlLl
First conversion: change the variable: q=1/Tnet
12/7/2009 47
g q net
Problem conversion in wireless visual sensor networks (cont.)
P2 i i iq q1
P2:
.sconstraint :subject to
minimize q1
2
1
20q
23
23
q3
q2
(q q q )P3:
bj t t
N minimize 2
NiVh
q
(q1=q2=q3)
P4::subject to
minimize 2
NiVhxa
qNi
i
Vh
hVh Ll
hl sx 22
//l
, ,
, , , :subject to
3/22 VhPD
Llyx
NiVh xa
Vhlhl
hiLl
hlil
, ,//log
, ,
, , , :subject to
3/22 VhsPD
Llyx
NiVhxa
hshh
Vhlhl
hiLl
hlil
, , ,0
, ,
, ,//log 3/2
LlVhx
NiqByacycaP
VhsPD
hl
iLl
lilr
Lll
slilsi
hshh
, , ,0
, ,0
, ,
LlVhx
Llqa
NiBqyacycaP
hl
Niiil
iiLl
lilr
Lll
slilsi
12/7/2009 48
. ,0 , ,0 VhPVhs shh . ,0 , ,0 ,,,
VhPVhsV
shh
hl
Achievable maximum network lifetime (cont.) Problem conversion:
minimize 222 sxqVh
hVh Ll
hlNi
i
Flow conservation
Video quality requirement , ,//log
, , , :subject to
3/22 VhsPD
NiVh xa
hshh
hiLl
hlil
VVN
,,0
, ,
Llqa
NiBqxacxcaP
iil
iiLl Vh
hlilr
Ll Vhhl
slilsi
Power constraint
Auxiliary variables are equal
. ,0 , ,0 , 0 , , ,0
, ,0
VhPVhsNi,qLlVhx
Llqa
shh
ihl
Niiil
Second conversion: introduce auxiliary variables: qi
12/7/2009 49
y qi
Achievable maximum network lifetime (cont.)
Si l ti lt Simulation results
12x 106
Source-and-Routing Optimized (Proposed) 0.7 Encoding power0.7 Encoding power
10
11
ork
lifet
ime
[s]
g p ( p )Source-Optimized Scheme (SOS)Routing-Optimized Scheme (ROS)
0.5
0.6
n [W
]
g pTransmission and reception power
0.5
0.6
n [W
]
g pTransmission and reception power
A B C
8
9
max
imum
net
wo
0.3
0.4
wer
con
sum
ptio
n
0.3
0.4
wer
con
sum
ptio
n
7
8
Ach
ieva
ble
m
0.1
0.2Pow
e
0.1
0.2Pow
e0 50 100 150 200 250 300
6
Encoding distortion requirement
Network lifetime comparison Power consumption at each node
0 5 10 15 20 25 30 350
Sensor node No.1 2 3 4 5 6 7 8 90 5 10 15 20 25 30 35
0
Sensor node No.1 2 3 4 5 6 7 8 9
12/7/2009 50
p Power consumption at each node
Network lifetime for large-delay g yapplications Large-delay applications:
Example: visual data 11
12x 106
Average PLR=0Average PLR=0.072Average PLR=0.134Average PLR=0.264
collection Retransmissions to recover
th t d k t9
10
twor
k lif
etim
e [s
]
the corrupted packets Retransmission consumes
extra power reduce 6
7
8
Max
imum
net
w
extra power, reduce network lifetime 0 50 100 150 200 250 300
5
6
Encoding distortion requirement
N k lif i i h i iNetwork lifetime with retransmissions
12/7/2009 51
Network lifetime for small-delay yapplications Small-delay applications:
Example: real-time traffic 10
11
12x 106
[s]
Average PLR=0Average PLR=0.072Average PLR=0.134Average PLR=0.264
monitoring FEC to recover the
t d k t8
9
10
netw
ork
lifet
ime
corrupted packets Introduce extra encoding
and decoding power 5
6
7
Max
imum
n
and decoding power consumption, reduce network lifetime
0 50 100 150 200 250 3004
Encoding distortion requirement
N k lif i i h FECnetwork lifetime
Network lifetime with FEC
12/7/2009 52
Outline
Motivation and contributions Principles of convex optimization Resource allocation in distributed video Resource allocation in distributed video
communication systemsThro ghp t ma imi ation in P2P VoD applications Throughput maximization in P2P VoD applications
Network lifetime maximization in wireless visual sensor networksnetworks
Optimization for video streaming over wireless ad hoc networksnetworks
Conclusions
12/7/2009 53
Optimized video unicasting over wireless p gad hoc networks Video unicasting Video unicasting
From a source to a receiver Example: A group of visitors in a park,
Source Example: A group of visitors in a park,
person A receives video streaming from person B via relays.
Relay node The proposed optimized
video unicasting scheme
y
Prioritized coding + network coding Minimize the distortion by jointly
optimizing the source rate allocation
Receiver
Video unicastingoptimizing the source rate allocationand the routing scheme
Video unicasting
12/7/2009 54
Related work Video streaming over wireless ad hoc networks
Distributed optimization for data communications over Distributed optimization for data communications over wireless ad hoc networks [Xiao2004, Chen2006]
The optimizations for data communications cannot be The optimizations for data communications cannot be applied directly to real-time streaming applications
Existing optimizations for video streaming over wireless ad hoc networks Maximize the expected video quality using genetic
algorithm [Mao2004] (centralized thus high complexity)algorithm [Mao2004] (centralized thus high complexity) Optimization of resource competition among multiple
unicast video sessions [Zhu2006] (distributed and low l it t h d )complexity at each node)
Optimal resource allocation for both video unicast streaming [and video multicast streaming [YHe-ISCAS06, YHe-PCM07,
12/7/2009 55
[ g [ , ,YHe-TCSVT08b] over wireless ad hoc networks.
Video distortion model
180
200Akyio QCIF sequence
Fitting distortion modelExperimental data
Video distortion modelDistortion vs. received throughput
120
140
160
on
Modeled as:
0 Dd
60
80
100
Dis
torti
o d=2.97+ 474.97/(R+0.13),0
0
RDd
Convex function
0 50 100 150 200 250 3000
20
40Convex functionData fitting technique to find Parameters: 000 ,, D
Throughput [Kbps]
0D : Parameter related to encoding distortion
P t f t i i di t ti
12/7/2009 56
: Parameters for transmission distortion00 ,
Optimized video unicasting over wireless p gad hoc networks (cont.)
Problem formulation: Problem formulation:Video distortionFlow conservation
45
Flow conservation
Link capacity constraint
35
40
Original optimization problemConverted to:strictly convex optimization problem
20
25
30
PS
NR
[dB
]
strictly convex optimization problem, develop distributed algorithm to solve it
5
10
15
Proposed routingCongestion-minimized routingDouble-disjoint-path routing
Simulation results: Comparison of frame PSNR
12/7/2009 57
0 10 20 30 40 50 60 70 80 90 1005
Frame No.
Comparison of frame PSNR
Optimized video multicasting over p gwireless ad hoc networks Video multicasting:Video multicasting:
Streaming from a source node to H receivers simultaneously With network coding, a multicast flow = H conceptual
unicast sessions [Ahlswede2000]unicast sessions [Ahlswede2000] The proposed optimized multicasting scheme
Minimize the distortion by jointly optimizing the source rate allocation, the routing scheme and the powerrouting scheme and the power
11 1
Source node
11 1
Source node
11
Source node
1
2
4
3
110
0
2
4
3
001
1
2
4
3
111
1
56 711 0
Receiver 1 56 710 1
Receiver 2 56 711 1
Receiver 2Receiver 1
C C f
12/7/2009 58
(a) (b) (c)Conceptual unicast session 1 Conceptual unicast session 2 Multicast flow
Optimized video multicasting over p gwireless ad hoc networks Packet loss rate (PLR) in wireless ad hoc networks
Loss due to transmission error (pi2)
Loss due to congestion (p )
(pi2)
Loss due to congestion (pi1)
•PLR at link i: 21 111 iii ppp
where
TL
RCTTdelayobpi expexpPr1
ik
kikthi
PGPGSIRp
1
112
12/7/2009 59
iii PG
Optimized video multicasting over p gwireless ad hoc networks (cont.) Problem formulation: Problem formulation:
Total distortionAggregate throughputFlow conservationMulticast flow rate yl
Capacity constraint
gg g g p
p yLink capacity under CDMA
Transmit power constraint
O i i l i i i blOriginal optimization problemNetwork coding eliminates duplicate packets, change the objective function to maximize the aggregate throughput
12/7/2009 60
Optimized video multicasting over p gwireless ad hoc networks (cont.) Distributed solution using hierarchical dual decompositions Distributed solution using hierarchical dual decompositions
Optimization variables (s,x,y,P)Fi t l d l d iti
Network flow variables (s,x,y) Power variable (P)
First-layer dual decomposition
Second-layer dual decompositionSecond-layer decompositionusing game theory
Compute link rate and transmit power
Compute link rate and transmit power
for each outgoing link
Node 1 Node N
for each outgoing link
12/7/2009 61
Optimized video multicasting over p gwireless ad hoc networks (cont.) Simulation results:
035
0.4Single-treeDouble-treeOptimized035
0.4Uniform-powerOptimized
0.25
0.3
0.35
Mbp
s]
p
0.25
0.3
0.35
Mbp
s]
0.15
0.2
Thro
ughp
ut [M
0.15
0.2
Thro
ughp
ut [M
1 2 3 40
0.05
0.1
1 2 3 40
0.05
0.1
1 2 3 4
Receiver ID
1 2 3 4Receiver ID
Compare to uniform-power scheme Compare to tree-based schemes
12/7/2009 62
Optimized video multicasting over wireless ad hoc k
4
4.5
5
1.6
1.8
2
networks
1.5
2
2.5
3
3.5
Lam
da
0.6
0.8
1
1.2
1.4
Pow
er [W
]
0 7 0 3
0 50 100 150 200 250 300 350 400 450 5000
0.5
1
Iteration No.0 50 100 150 200 250 300 350 400 450 500
0
0.2
0.4
Iteration No.
0 4
0.5
0.6
0.7
e ra
te [M
bps]
0.2
0.25
0.3
te [M
bps]
0.1
0.2
0.3
0.4
Con
cept
ual s
ourc
e
0.05
0.1
0.15
Mul
ticas
t lin
k ra
t0 50 100 150 200 250 300 350 400 450 500
0
0.1
Iteration No.
0 50 100 150 200 250 300 350 400 450 500
0
Iteration No.
Optimization results for a 50-node wireless ad hoc network
12/7/2009 63
A multicast flow for a source to 8 receivers
References [Bertsekas2003] D. P. Bertsekas, A. Nedic, and A. E. Ozdaglar, Convex Analysis and Optimization,
Athena Scientific, 2003. [Boyd2004] S. Boyd and L. Vandenberghe, Convex Optimization, Cambridge University Press, 2004. [Do2004] T. T. Do, K. A. Hua, and M. A. Tantaoui, “P2VoD: Providing Fault Tolerant Video-on-Demand [ ] , , , g
Streaming in Peer-to-Peer Environment,” in Proc. of IEEE ICC, vol. 25, no. 1, pp. 119-130, Jan. 2004. [Li2006] Z. Li and A. Mahanti, “A Progressive Flow Auction Approach for Low-Cost On-Demand P2P
Media Streaming,” in Proc. of ACM QShine, Aug. 2006. [Yiu2007] W. P. Yiu, X. Jin and S. H. Chan, “VMesh: Distributed segment storage for peer-to-peer
interactive video streaming,” IEEE Journal on Selected Areas in Communications, vol. 25, no. 9, pp.interactive video streaming, IEEE Journal on Selected Areas in Communications, vol. 25, no. 9, pp. 1717-1731, Dec. 2007.
[YHe-ICASSP08] Y. He, I. Lee, and L. Guan, “Distributed throughput maximization in hybrid-forwarding P2P VoD applications”, in Proc. of IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), pp. 2165─2168, Las Vegas, USA, April 2008.
[YHe-TMM2008a] Y He I Lee and L Guan “"Distributed throughput optimization in P2P VoD [YHe TMM2008a] Y. He, I. Lee and L. Guan, Distributed throughput optimization in P2P VoD systems,” IEEE Transactions on Multimedia, vol. 11, no. 3, pp. 509-522, April 2009..
[YHe-TMM2008b] Y. He, G. Shen, Y. Xiong and L. Guan, “Optimal prefetching scheme in P2P VoD applications with guided seeks,” IEEE Transactions on Multimedia, vol. 11, no. 1, pp. 138-151, Jan. 2009.
[Wu2005] C Wu and B Li “Optimal Peer Selection for Minimum-Delay Peer-to-Peer Streaming with [Wu2005] C. Wu and B. Li, Optimal Peer Selection for Minimum-Delay Peer-to-Peer Streaming with Rateless Codes,” in Proc. of ACM MM, pp. 69-78, Nov. 2005.
[YHe-ICME07a] Y. He, I. Lee, and L. Guan, “Distributed rate allocation in p2p streaming”, in Proc. of IEEE International Conference on Multimedia & Expo (ICME), Special Session on P2P Multimedia Content Access and Distribution, pp. 388─391, Beijing, China, July 2007.
[Xiao2004] L Xiao M Johansson and S Boyd “Simultaneous routing and resource allocation via dual [Xiao2004] L. Xiao, M. Johansson and S. Boyd, Simultaneous routing and resource allocation via dual decomposition,” IEEETransactions on Communications, vol. 52, no. 7, pp. 1136-1144, Jul. 2004.
[Chen2006] L. Chen, S. H. Low, M. Chiang, and J. C. Doyle, “Cross-layer congestion control, routing and scheduling design in ad hoc wireless networks,” in Proc. of IEEE INFOCOM, pp. 1-13, Apr. 2006.
[Mao2004] S. Mao, X. Cheng, Y. T. Hou, and H. D. Sherali, “Multiple description video multicast in wireless ad hoc networks ” in Proc of IEEE BROADNETS pp 671 680 Oct 2004
12/7/2009 64
wireless ad hoc networks, in Proc. of IEEE BROADNETS, pp. 671-680, Oct. 2004.
References (cont.)( ) [Zhu2006] X. Zhu, J. P. Singh, and B. Girod, “Joint routing and rate allocation for multiple video streams
in ad hoc wireless networks,” Journal of Zhejiang University, Science A, vol. 7, no. 5, pp. 727-736, May 2006.[YH ISCAS06] Y H I L d L G “O ti i d lti th ti i d l d iti f [YHe-ISCAS06] Y. He, I. Lee, and L. Guan, “Optimized multi-path routing using dual decomposition for wireless video streaming”, in Proc. of IEEE International Symposium on Circuits and Systems (ISCAS), pp. 977─980, New Orleans, USA, May 2007.
[YHe-PCM07] Y. He, I. Lee, and L. Guan, “Video multicast over wireless ad hoc networks using distributed optimization”, in Proc. of Pacific-Rim Conference on Multimedia (PCM), pp. 296─305, Hongkong Dec 2007Hongkong, Dec. 2007.
[Nama2006] H. Nama, M. Chiang, and N. Mandayam, “Utility-lifetime trade-off in self-regulating wireless sensor networks: A cross-layer design approach,” in Proc. of IEEE ICC, vol. 8, pp. 3511-3516, Jun. 2006.
[Madan2006] R. Madan, S. Lall, “Distributed algorithms for maximum lifetime routing in wireless sensor networks,” IEEE Transactions on Wireless Communications, vol. 5, no. 8, pp. 2185-2193, Aug. 2006.
[YHe-ICME07b] Y. He, I. Lee, and L. Guan, “Network lifetime maximization in wireless visual sensor networks using a distributed algorithm”, in Proc. of IEEE International Conference on Multimedia & Expo (ICME), pp. 2174─2177, Beijing, China, July 2007.
[He2006] Z. He and D. Wu, “Resource allocation and performance analysis of wireless video sensors,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 16, no. 5, pp. 590-599, May 2006.
[Ahlswede2000] R. Ahlswede, N. Cai, S.-Y. R. Li and R. W. Yeung, “Network information flow,” IEEE Transactions on Information Theory, vol. 46, pp. 1204-1216, Jul. 2000.
[YHe-TCSVT2008a]Y. He, I. Lee and L. Guan, “Distributed algorithms for network lifetime maximization in wireless visual sensor networks,” IEEE Transactions on Circuits and Systems for Video Technology, vol.19, no. 5, pp. 704-718, May 2009. pp y
[YHe-TCSVT2008b] Y. He, I. Lee and L. Guan, “Optimized video multicast in wireless ad hoc network using network coding,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 19, no. 6, pp. 796-807, June 2009..
12/7/2009 65