Korea Advanced Institute of Science and Technology
Network Systems Lab.
Cross-layer Control of Wireless Networks:
From Theory to Practice
Professor Song ChongNetwork Systems Laboratory
EECS, [email protected]
Korea Advanced Institute of Science and Technology
Network Systems Lab.
Multi-user Opportunistic Communication
Multi-user diversity
In a large system with users fading independently, there is likely to be a user with a very good channel at any time. Long-term total throughput can be maximized by always serving the user with the strongest channel.
Korea Advanced Institute of Science and Technology
Network Systems Lab.
Capacity Region: A Realization of Channel Consider a single channel realizationCDMA downlink with two users
θ: orthogonality factor in [0,1]
Capacity region [Kum03]
1 1 2 21 2 2 2
1 2 1 2 1 2
log 1 , log 1g p g p
r rg p n g p n
1 2p p P
0 0.5 0.5 1
User 1 User 2
1 2g p1 1g p
2 1g p2 2g p
1r
2rConvex
1r
2rNonconvex
Korea Advanced Institute of Science and Technology
Network Systems Lab.
Long-term Capacity RegionTime-varying achievable rate region
Long-term rate region
: long-term rate region
(can be shown to be convex)
F
1R
2R
F
1r
2r
1r
2r
1r
2r
t
: long-term rate of user iR i
Korea Advanced Institute of Science and Technology
Network Systems Lab.
Convexity Proof [Stol05]
Case of finite channel states and scheduling policiesNotation
S: finite set of channel statesSequence of channel states s(t)∈S, t=0,1,... forms an irreducible Markov chain with stationary distribution
K(s): set of all possible scheduling decisions for given channel state s∈Sri
s(k)≥0: rate allocated to user i for channel state s∈S and scheduling decision k∈K(s)rs(k): rate vector, i.e., rs(k)=[ri
s(k),∀i]
For each channel state s, a probability distribution φs=[φsk,∀k∈K(s)] is fixed, i.e.,
, , , , 1s s ss S
s S s S
( )
0, ( ), , 1sk skk K s
k K s s S
Korea Advanced Institute of Science and Technology
Network Systems Lab.
Convexity Proof [Stol05]
Rate vector for a set of distributions φ=[φs,∀s∈S]
If we interpret φsk as the long-term average fraction of time slots when the channel state is s and the rate allocation is k, then R(φ) is the corresponding vector of long-term average service rates
The long-term rate region FF is defined as the set of all average service rate vectors R(φ) corresponding to all possible φThe convexity of FF immediately follows as it is a convex hull of all possible instantaneous rates
Consider which is a convex combination of all possible rate vectors rs(k), ∀k∈K(s), ∀s∈S
( )
( )ss sk
s S k K s
R r k
( )
( )ss sk
s S k K s
R r k
Korea Advanced Institute of Science and Technology
Network Systems Lab.
Long-term NUM
Utility function [Mo00]
Network Utility Maximization (NUM)
max ( )
where = long-term rate of user i
= long-term capacity region
iR F
i
i
U R
R
F
11
, if 0 & 11log , if 1
ii i
i
RU R
R
α→0: throughput maximizationα=1: proportional fairness (PF)α→∞: max-min fairness
Korea Advanced Institute of Science and Technology
Network Systems Lab.
Maximization of sum of weighted rates
Both problems yield an unique and identical solution if we set , where is the optimal solution of the long-term NUM problem.
Sum of Weighted Rates (SWR)
max i iR F
i
R
*'( )i iU R *iR
2R
1R
F
*2R
*1R
1 2( ) ( )U R U R J
0
1 1 2 2R R K
Korea Advanced Institute of Science and Technology
Network Systems Lab.
Gradient-based SchedulingAssuming stationarity and ergodicity, we have
The long-term NUM problem can be solved if we solve with at each state sThe resource allocation problem during slot t
where Ri(t) is the average rate of user i up to time t and is the replacement of Ri
* which is unknown a prioriConvergence of Ri(t) to Ri
* can be proved by stochastic approximation theory [Kush04] or fluid limit technique [Stol05].
( )max maxi i s i iR F r F s
i i
R E r
where rate of user at state
( ) capacity region at state ir i s
F s s
( )max i ir F s
i
r
*'( )i iU R
( )
max '( ( )) ( )
subject to ( ) ( )
i ir t
i
U R t r t
r t F s t
Korea Advanced Institute of Science and Technology
Network Systems Lab.
Gradient-based Scheduling
This coincides with the optimality condition given by directional derivative
Consider
The optimality condition is given by
The optimal solution to the following problem is R*
Thus we set
* arg max iR F
i
R U R
* *
* * *
0,
,
i i ii
i i i ii i
U R R R R F
U R R U R R R F
*max i iR F
i
U R R
*'( )i iU R
Korea Advanced Institute of Science and Technology
Network Systems Lab.
HDR PF Scheduler
PF scheduler is a special case of gradient-based scheduler
Logarithmic utility functionFeasible region (TDMA)
PF scheduler serves user i* such that
( ), 1*
max ' ( ) ( )arg max
( ) subject to ( ) ( )
i ir ti t i
ii
U R t r tri
R t r t F s t
logi iU R R
1, 1
2, 1
, 1
00 0
0 0( ) , , , ...,
0
0 0 0
t
t
M t
r
rF s t
r
, 1where available data rate of user during slot i tr i t
Korea Advanced Institute of Science and Technology
Network Systems Lab.
Opportunistic Communication in OFDMA Downlink
Exploit multi-user diversity in time and frequencyIn a large system with users fading independently, there is likely to be a user with a very good channel at some carrier frequency for each time. Long-term total throughput can be maximized by always serving the user with the strongest channel.Challenge is to share the benefit among the users in a fair way.
User M
MobileUser 1
frequency
Channel
gain
frequency
Channel
gain
Fading channel
Korea Advanced Institute of Science and Technology
Network Systems Lab.
Frequency Selectivity in Channel
Frequency response in multipath environment
Delay spread
Coherence bandwidth Bc
Frequency separation at which the attenuation of two frequency-domain samples becomes decorrelated
For given delay spread,Frequency-selective channel if B>>Bc
Frequency-flat channel if B<<Bc
1
2cd
BT
2 ( )( ; ) ( ) ij f ti
i
H f t a t e ( ) : attenuation on path ( ) : propagation delay on path i
i
a t it i
,max ( ) ( )d i ji j
T t t
( ), ( )i ia t t
Bc
B
freq.
gain
B
Korea Advanced Institute of Science and Technology
Network Systems Lab.
Long-term NUM Problem in OFDMA Downlink
max ii
U R
User 1
User 2
User 3
User 3
User 4
User 3
Frequency (subcarrier)
Time slot
P
Power allocation
Subcarrier allocation
(user selection)Dynamic subcarrier and power allocation achieving
Korea Advanced Institute of Science and Technology
Network Systems Lab.
Joint OptimizationConsider M mobile users and N subcarriersJoint optimization of subcarrier and power allocation at each time slot t
Mixed integer nonlinear programming
, 1 , 1
, 1 , 1, 1 2
, 1
, 1
, 1
, 1 , 1
maximize ' ( )
subject to log 1
1,
0, 0,1 , ,
j ji i t i t
i M j N
ji t j tj
i t ji t
j tj
ji t
i
jj t i t
U R t r x
g pBr
N N
p P
x j
p x i j
( )F s t
, 1 where ( )i tr r F s t
Korea Advanced Institute of Science and Technology
Network Systems Lab.
Suboptimal Allocation [Lee08]
Iteratively solve two subproblemsFor fixed p, subcarrier allocation problem
Opportunistic scheduling over each subcarrier
For fixed x, power allocation problemConvex optimization (water filling)
Each subproblem is easy to solve
Frequency-selective power allocation (FPA)
Equal power allocationEqual power allocation
Subcarrier allocation for givenpower allocation
Subcarrier allocation for givenpower allocation
Power allocation for givensubcarrier allocation
Power allocation for givensubcarrier allocation
While subcarrier allocationis changing
InitializationEqual power allocation
Korea Advanced Institute of Science and Technology
Network Systems Lab.
Subproblem I: Opportunistic Scheduling
Find x for a fixed power vector p0
Separable w.r.t. subcarriersFor each subcarrier j, select user ij* such that
0max log 1
subject to 1,
0,1 , ,
ij i ij jx
i M j N
iji M
ij
x w g p
x j
x i j
* 0arg max log 1j i ij ji M
i w g p
Korea Advanced Institute of Science and Technology
Network Systems Lab.
Subproblem II: Water FillingFind p for a fixed subcarrier allocation x0
Convex optimizationWater filling is optimal
λ is a nonnegative value satisfying
( ) ( )max log 1
subject to
0,
m j m j j jp
j N
jj N
j
w g p
p P
p j
( )*
( )
1m jj
m j j
wp
g
*j
j N
p P
subcarrier
( )m jw
*jp
( )
1
m j jg
Korea Advanced Institute of Science and Technology
Network Systems Lab.
FPA vs. EPA
FPA gives significant throughput gain (up to 40%) in OFDMA downlink when
Sharing policy becomes more fairness-oriented Delay spread (frequency selectivity) increasesSystem bandwidth becomes wider
00.5
12
510
50100
12
34
56
78
1
1.1
1.2
1.3
1.4
s
No
rma
lize
d T
ota
l Th
rou
gh
pu
t (T F/T
E)
0 0.51
25
10 50100
12
34
56
78
1
1.1
1.2
1.3
1.4
s
No
rma
lize
d T
ota
l Th
rou
gh
pu
t (T F/T
E)B=5MHz B=20MHz
MT
MM
MT
MM
Korea Advanced Institute of Science and Technology
Network Systems Lab.
Impact of α: InterpretationEfficiency-oriented policy (α=0)
Only best user for each subcarrier
Fairness-oriented policy (α→∞)Bad-channel users are also selected
( ) 1, m j jg j N *j
Pp
N
FPA ≈ EPA
subcarrier
*jp
subcarrier
*jp
High, medium, low gm(j)j’s
*jp
P
N
Korea Advanced Institute of Science and Technology
Network Systems Lab.
Impact of System Bandwidth: Interpretation
Narrowband (less frequency-selective)
Wideband (more frequency-selective)
frequency
Channelgain
B
frequency
Channelgain
Bsubcarrier
*jp
Extreme case(frequency flat)
*j
Pp
N
subcarrier
*jp
*jp
P
N
Korea Advanced Institute of Science and Technology
Network Systems Lab.
Impact of SNR Distribution
EPA is comparable to FPA only when all the mobiles are located in high SNR regime
0 0.5
1 2
5 10
50 100
020
4060
80100
11.11.21.31.41.51.6
Low SNR Percentage (%)
No
rma
lize
d T
ota
l Th
rou
gh
pu
t (T F/T
E)
B=20MHzs=6
MT
MM
Low SNR: gij<5dB
Korea Advanced Institute of Science and Technology
Network Systems Lab.
Impact of SNR Distribution: Interpretation
High SNR
Mix of high and low SNR
Low SNR
*jp
subcarrier
*jp
p
( )log(1 )m j j jg p
p
( )log(1 )m j j jg p
Sensitive to power variation
Insensitive to power variation
Subcarriers with low SNR usersare more sensitive to power thanhigh SNR users
*jp
subcarrier
subcarrier
Korea Advanced Institute of Science and Technology
Network Systems Lab.
Throughput-optimal Scheduling and Flow Control
Joint scheduling and flow control— Stabilize the system whenever the long-term
input (demand) rate vector lies within the capacity region
— Stabilize the system while achieving throughput optimality even if the long-term input (demand) rate vector lies outside of the capacity region
Long-term NUM for arbitrary input rates [Nee05]
max
subject to (stability constraint)
0 (demand constraint)
iR
i
U R
R F
R D
where = long-term capacity region
= long-term demand vector
F
D
Korea Advanced Institute of Science and Technology
Network Systems Lab.
Single-carrier Downlink Problem
AssumptionInfinite demandsInfinite backlog at every transport layer queue
Cross-layer controlJoint optimization of flow control and scheduling
1( )d t
2 ( )d t
( )Kd t
1( )x t
2 ( )x t
( )Kx t
1( )q t
2 ( )q t
( )Kq t
demands
Flow Control at Source
Base Station
Scheduling
fading channel
1( )r t
2 ( )r t
( )Kr tfeedback: achievable rates
1( )L t
2 ( )L t
( )KL t
Korea Advanced Institute of Science and Technology
Network Systems Lab.
Cross-layer Control Scheduling at BS
Flow control at source i
Algorithm Performance
*
( ) ( ( ))max ( ) ( ) arg max ( ) ( )i i i i
r t F s t ii
q t r t i q t r t
0 : control parameterV 1 ( )( ) max ,0i
i
q tx t U
V
1
0
1limsup ( )
t
it i
q O Vt
1
*
0
1 1liminf ( )
t
i it
i i
U E x U R Ot V
Stability
Optimality
Korea Advanced Institute of Science and Technology
Network Systems Lab.
Derivation of Cross-layer Control Primal problem
Dual problem
Dual decomposition
,max ( )
subject to x ,
1
, 0
iX
i
i i i
ii
U x
r i
X
,min max ( ) ( )
subject to 1
, 0
i i i i iQ X
i i
ii
U x q x r
X
1 ,i ix U q i
1max
ii
i i ii
q r
,i i iiq x r i
Korea Advanced Institute of Science and Technology
Network Systems Lab.
Multi-hop Wireless Networks:Cross-layer Control
Korea Advanced Institute of Science and Technology
Network Systems Lab.
Multi-hop Wireline Network
Network Utility Maximization
―Link capacity is given and constant―Rate allocation problem
( )
max
s.t. ,
0
sr
s
s ls S l
U r
r c l
r
1r
2r 3r
1c 2c
Korea Advanced Institute of Science and Technology
Network Systems Lab.
Functional DecompositionLagrangian function
Dual problem
Dual decomposition― Flow control at source
― Congestion price at link
TCP is an approximation of this dual decomposition
( )
, = s l s ls l s S l
L r U r r c
min max ,r
L r
1
( ) ( )
max s s l s lr
s l L s l L s
U r r r U
( ) ( )
min ll s l s ll s S l s S l
r c r c
Korea Advanced Institute of Science and Technology
Network Systems Lab.
Multi-hop Wireless Network
Long-term Network Utility Maximization
― Link capacity is time-varying and a function of resource control
― Joint rate, power allocation and link scheduling
, , max
s.t. ,
sR P I
s
U R
R F P I
1R
2R 3R
1 , ,C P I h 2 , ,C P I h : power allocation: link scheduling: channel state
PIh
Korea Advanced Institute of Science and Technology
Network Systems Lab.
Functional DecompositionFor a realization of channelsLagrangian function
Dual problem
Dual decomposition― Flow control at source
― Scheduling/power control at link
― Congestion price at link
Joint MAC and transport problemDistributed scheduling/power control is a challenge
( )
, , , = , ,s l s ls l s S l
L r P I U r r C P I h
, ,
min max , , ,r P I
L r P I
1
( ) ( )
max s s l s lr
s l L s l L s
U r r r U
( ) ( )
min , , , ,ll s l s ll s S l s S l
r C P I h r C P I h
,
max , ,l lP I
l
C P I h
Korea Advanced Institute of Science and Technology
Network Systems Lab.
Per-link Queueing Case
User 0
User 2
)( 11 xU
)( 00 xU
)( 22 xU
cA=1 cB=1
a is the fraction oftime link A is used
0 1 2, ,
0 1
1 2
max ( )
subject to
1
, 0
i ix x x
i
a
b
a b
U x
x x
x x
x
Korea Advanced Institute of Science and Technology
Network Systems Lab.
Lagrange Multipliers
0 1,
0 2
max ( ) ( )
( )
1
, 0
i i A ax
i
B b
a b
U x p x x
p x x
x
Korea Advanced Institute of Science and Technology
Network Systems Lab.
Functional DecompositionCongestion control (sources and nodes)
MAC or scheduling (network)
0 10
0 1
max ( ) ( )
( )
i i Ax
i
B
U x p x x
p x x
1maxA B
A A B Bp p
Korea Advanced Institute of Science and Technology
Network Systems Lab.
Per-flow Queueing Case
User 0
User 1
User 2
)( 11 xU
)( 00 xU
)( 22 xU
cA=1 cB=1
a0 is the fraction oftime link A is used foruser 0
, 0
0 0
1 1
0 0
2 2
max ( )
subject to
1
i ix
i
a
a
a b
b
ij
U x
x
x
x
Korea Advanced Institute of Science and Technology
Network Systems Lab.
Functional DecompositionCongestion control (sources)
MAC or scheduling (network)
x0 μa0 μb0
x1 μa1 x2 μb2
pa0 pb0
pa1 pb2
0 0 1 1 2 20
max ( )i i a a ax
i
U x p x p x p x
0 0 0 0 0 1 1 2 21
max ( )i
i
a a b b b a a b bp p p p p
Korea Advanced Institute of Science and Technology
Network Systems Lab.
Interference Model
2r
1 2 3 4
5
node
link
5 2
1 3
4
1r
Network connectivity graph G
Conflict graph CG
- Links in G = nodes in CG- CG-Edge if links in G interfere with each other
Korea Advanced Institute of Science and Technology
Network Systems Lab.
Interference Model Maximal independent set model
- Only one maximal independent set can be active at a time
- - NUM problem
{1, 4}, {2}, {3}, {4, 5}
5 2
1 3
4
CGMaximal independent sets
( )I t I where {1, 4}, {2}, {3}, {4, 5}I
2
1
max ( )ii
U r
1 1
1 2 2
1 2 3 41 2 3
41 2 4
52
4
1
subject to independent set constraints
00 0
00 0
00 0
0 0
0 0 0
1ii
r c
r r c
a a a ar r c
cr r c
cr
a
Korea Advanced Institute of Science and Technology
Network Systems Lab.
Jointly Optimal Power and Congestion Control
NUM at particular state s [Chiang05]
is a nonconcave function of p Assuming high SINR regime, i.e,
can be converted into a concave function of p through a log transformation (geometric programming) Joint optimization of congestion control and power control
,
( )
max ( )
subject to
( , ) ,
where
ip r
i
i li I l
U r
r C p s l
( , )lC p slog(1 ( , )) log( ( , ))l lSINR p s SINR p s
( , )lC p s
0
( )( , ) log(1 ( , )) log 1
( )l ll
l lk kl
k Lk l
p sC p s SINR p s
N p s
Korea Advanced Institute of Science and Technology
Network Systems Lab.
Jointly Optimal Power and Congestion Control
Flow control at source
Power control at link
Congestion price at link
Interpretation
1
( )i l
l L i
r U
( ) ( ),( )
( ) ( )k k
kk kk
t SINR p t sm t
p t s
( )
( 1) ( ) ( ) ( )( )
ll l kl k
k Llk l
tp t p t s m t
p t
( )
( 1) ( ) ( ) ( ),l l i li I l
t t r t C p t s
Physical layer
r C
Transport layer
Source NodeFlow Control
LinkPower Control
Congestion
Price
r
Korea Advanced Institute of Science and Technology
Network Systems Lab.
Routing and Network Layer Queueing1 2
3
45
transport
layer
network
layer
(3)1A
(4)1A
(3)5A
(1)4A (1)
4R (1)4U
= set of commodities in the network = the amount of new commodity c data that exogenously arrives to node i during slot t = the amount of commodity c data allowed to enter the network layer from the transport layer at node i during slot t = the backlog of commodity c data stored in the network layer queue at node i during slot t
K( ) ( )ciA t
( ) ( )ciR t
( ) ( )ciU t
( and )i N c K
Korea Advanced Institute of Science and Technology
Network Systems Lab.
Dynamic Control for Network Stability
The stabilizing dynamic backpressure algorithm [Tassiulas92]
- An algorithm for resource allocation and routing which stabilizes the network whenever the vector of arrival rates lies within the capacity region of the network
Resource allocation- For each link , determine optimal commodity and optimal weight by
- Find optimal resource allocation action by solving
)(* tCab
)(* tab
]0,[max)(
)]()([maxarg)(
))(())((*
)()(
}),|({
*
** tCb
tCaab
cb
ca
Lbacab
abab
c
UUt
tUtUtC
)(
*
)(
)(.
))(),(()(max
ts
ababab
tI
ItIts
tstICt
),( ba
)(* tI
Korea Advanced Institute of Science and Technology
Network Systems Lab.
Dynamic Control for Network Stability Routing
- For each link such that , offer a transmission rate of to data of commodity .
The algorithm requires in general knowledge of the whole network state. However, there are important special cases where the algorithm can run in a distributed fashion with each node requiring knowledge only of the local state information on each of its outgoing links.
Interpretation- The resulting algorithm assigns larger transmission rates to links with larger differential backlog, and zero transmission rates to links with negative differential backlog.
),( ba 0)(* tab))(,)(()( ** tstICt abab )(* tCab
Korea Advanced Institute of Science and Technology
Network Systems Lab.
Dynamic Control for Infinite DemandsAssumption
Infinite backlog at every transport layer queueCross-layer control
Flow control at node iEach time t, set Ri
(c)(t) to
Routing and resource allocationSame as previous
Performance
Tradeoff between utility and delay
0 : control parameterV ( )
( ) 1( ) max ,0c
c ii
UR t U
V
1
( )
0
1limsup ( )
tc
it i
U O Vt
1
( ) *
, 0 ,
1 1liminf ( )
tc
i it
i c i c
U E R U R Ot V
Korea Advanced Institute of Science and Technology
Network Systems Lab.
References[Kum03] K. Kumaran and L. Qian, “Uplink Scheduling in CDMA Packet-Data Systems,” IEEE INFOCOM 2003.[Mo00] J. Mo and J. Walrand, “Fair End-to-End Window-Based Congestion Control,” IEEE/ACM Trans. Networking, Vol. 8, No. 5, pp. 556-567, Oct. 2000.[Kush04] H. J. Kushner and P. A. Whiting, “Convergence of Proportional-Fair Sharing Algorithms Under General Conditions,” IEEE Trans. Wireless Comm., vol. , no., 2004.[Stol05] A. L. Stolyar, “On the Asymptotic Optimality of the Gradient Scheduling Algorithm for Multiuser Throughput Allocation,” Operations Research, vol. 53, no. 1, pp. 12-25, Jan. 2005.[Lee08] H. W. Lee and S. Chong, "Downlink Resource Allocation in Multi-carrier Systems: Frequency-selective vs. Equal Power Allocation", IEEE Trans. on Wireless Communications, Vol. 7, No. 10, Oct. 2008, pp. 3738-3747.[Nee05] M. J. Neely et al., “Fairness and Optimal Stochastic Control for Heterogeneous Networks,” IEEE INFOCOM 2005. [Chiang05] M. Chiang, “Balancing Transport and Physical Layers in Wireless Multihop Networks: Jointly Optimal Congestion Control and Power Control,” IEEE J. Sel. Areas Comm., vol. 23, no. 1, pp. 104-116, Jan. 2005.[Tassiulas92] L. Tassiulas and A. Ephremides, “Stability Properties of Constrained Queueing Systems and Scheduling Policies for Maximum Throughput in Multihop Radio Networks,” IEEE Trans. Automatic Control, vol. 37, no. 12, Dec. 1992.