You are on page 1of 11

Cross-Layer Design of Wireless

Multihop Backhaul Networks with


Multiantenna Beamforming
Seung-Jun Kim, Member, IEEE, Xiaodong Wang, Senior Member, IEEE, and
Mohammad Madihian, Fellow, IEEE
AbstractA cross-layer design approach is considered for joint routing and resource allocation for the physical (PHY) and the medium
access control (MAC) layers in multihop wireless backhaul networks. The access points (APs) are assumed to be equipped with
multiple antennas capable of both transmit and receive beamforming. A nonlinear optimization problem is formulated, which maximizes
the fair throughput of the APs in the network under the routing and the PHY/MAC constraints. Dual decomposition is employed to
decouple the original problem into smaller subproblems in different layers, which are coordinated by the dual prices. The network layer
subproblem can be solved in a distributed manner and the PHY layer subproblem in a semidistributed manner. To solve the PHY layer
subproblem, an iterative minimum mean square error (IMMSE) algorithm is used with the target link signal-to-interference-and-noise-
ratio (SINR) set dynamically based on the price generated from the upper layers. A scheduling heuristic is also developed, which
improves the choice of the transmission sets over time. Simulation results illustrate the efficacy of the proposed cross-layer design.
Index TermsWireless backhaul networks, cross-layer design, beamforming, dual decomposition, scheduling.

1 INTRODUCTION
W
ITH the advent of the high-speed access technologies
such as Third-Generation (3G) 802.11 a/g (WiFi) and
802.16 (WiMAX), the wireless backhaul network is emer-
ging as a cost-effective solution to transport the aggregate
traffic from the access points (APs) to the core network [1],
[2], [3]. The deployment cost of the backhaul network can be
significantly reduced if the wireless technology is adopted
compared to the existing solutions of leasing bundled
copper wires or installing a fiber-optic network.
A multihop wireless network is particularly promising
for the backhaul application, as it can provide increased
reliability and flexibility in the adverse wireless environ-
ment [4], [5]. In addition, advanced physical (PHY) layer
technologies such as power control, rate control, beamform-
ing, and multiple-input/multiple-output (MIMO) techni-
ques can be employed to meet the increasing demand for
bandwidth, quality of service (QoS), and scalability [6], [7],
[8]. An important issue is how we can efficiently coordinate
the available resources under the constraints from different
layers to maximize the performance.
In order to further reduce operation costs and increase
scalability, it is desirable to develop a distributed algorithm.
There are many tools that can facilitate the developing of
distributed algorithms [9]. If the problem has a separable
structure, then dual decomposition can be applied to yield a
distributed solution [10]. In particular, if the problem can be
formulated as a convex optimization problem, then dual
decomposition can find the primal and the dual optimal
solutions in a distributed manner. More generally, when the
problem does not admit a convex structure and a duality
gap exists, the dual methods are still useful in finding a
bound for the optimal value [10].
Recently, the dual decomposition technique has drawn
attention in the cross-layer design of wireless networks, as
it can divide a complex system problem into smaller
subproblems corresponding to different layers in the
protocol stack [11], [12]. The subproblems are naturally
coordinated by dual variables, which have an intuitive
interpretation of market price that balances the supply and
demand [12], [13]. In this work, we propose a cross-layer
design approach that jointly coordinates multipath routing
and PHY/medium access control (MAC) resource alloca-
tion for the beamforming multihop wireless backhaul
networks. The nodes in the network are assumed to be
equipped with multiple antenna arrays, which are capable
of both transmit and receive beamforming.
The receive beamforming technique is a rather classical
topic in the area of array signal processing [14]. The
technique steers the high-gain beams toward the direction
of the intended signal while forming nulls toward the
interferers to suppress them. On the other hand, transmit
beamforming employs a similar technique at the transmit-
ter side and steers the beam toward the intended receiver
while placing nulls toward the unintended ones to avoid
causing interference to them. Transmit beamforming has
received attention relatively recently [15], [16], [17], [18].
The transmit beamforming problem is usually formulated
as minimizing the total transmit power subject to the
signal-to-interference-and-noise-ratio (SINR) constraints at
IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 6, NO. 11, NOVEMBER 2007 1259
. S.-J. Kim and M. Madihian are with NEC Laboratories America, Inc.,
4 Independence Way, Suite 200, Princeton, NJ 08540.
E-mail: {sjkim, madihian}@nec-labs.com.
. X. Wang is with the Electrical Engineering Department, Columbia
University, New York, NY 10027. E-mail: wangx@ee.columbia.edu.
Manuscript received 26 Jan. 2006; revised 13 Nov. 2006; accepted 13 Feb.
2007; published online 1 Mar. 2007.
For information on obtaining reprints of this article, please send e-mail to:
tmc@computer.org, and reference IEEECS Log Number TMC-0030-0106.
Digital Object Identifier no. 10.1109/TMC.2007.1052.
1536-1233/07/$25.00 2007 IEEE Published by the IEEE CS, CASS, ComSoc, IES, & SPS
the receivers. In [18], a complete solution to this problem
is provided, including the solution to the associated
feasibility problem. The key observation that leads to the
solution is the concept of duality. Using duality, one can
transform the difficult transmit beamforming problem to a
receive beamforming problem, which is easier to solve. In
fact, duality holds even for the joint receive and transmit
beamforming problem, where both the receiver and
transmitter employ adaptive antenna arrays. Moreover,
under the channel reciprocity condition, a distributed
computation of optimal beamformers is also possible in
the context of ad hoc wireless networks [19], [20], [21].
This is the most relevant case to our problem setting. The
channel reciprocity condition is typically applicable to
time-division-duplex (TDD) systems. Other works that
consider the optimization of a MIMO wireless network
include [22] and [23]. In [22], a game theoretic greedy
algorithm is considered, where each user selfishly opti-
mizes the MIMO sum capacity of the network in a
distributed manner. Since the algorithm does not take into
account the interference that the unilateral optimization
incurs, the resulting Nash equilibrium is not as efficient as
that obtained from a centralized algorithm. In this work,
the PHY layer employs the iterative minimum mean
square error (IMMSE) algorithm proposed in [21] as the
core, but the target SINRs are adjusted dynamically based
on the prices generated by the upper layers.
An important issue in the multihop wireless backhaul
network is spatial fairness. Since different APs are different
hop distances away from the core network, the throughput
experienced by the users can vary significantly, depending
on which APs they are associated with. In [2], a reference
model for spatial fairness was proposed, and the perfor-
mance benefit of imposing the fairness constraints was
shown. Here, we impose the fairness constraint to the
aggregate throughput of the APs. For simplicity, all partici-
pating APs are assumed to have equal traffic demands to
the core network. The problem is then formulated to
maximize the fair throughput of the network, given the
constraints of the network and the PHY/MAC layers.
Related work on the resource allocation problem in
wireless backhaul networks includes [5], [24], and [25]. In
[5], the total average transmission power expended in the
network is minimized, subject to the given minimum link
rate constraints. However, it is assumed that the supported
data rate is proportional to the SINR, which holds only in
the low-signal-to-noise-ratio (SNR) regime. The joint rout-
ing, scheduling, and power control problem is considered
in [24], where a 3-approximation polynomial time solution
is proposed. The scheduling objective is to avoid primary
conflicts, and interference is not considered by assuming
orthogonal channelization in the PHY layer. A cross-layer
optimization algorithm that takes the effect of interference
into account is proposed in [25]. Adopting a high-SNR
approximation for the link capacity, the work transforms
the joint routing and PHY/MAC problem to a convex
program, which can be solved efficiently by a centralized
processor.
This paper is organized as follows: In Section 2, the fair
throughput maximization problem is formulated as a
nonlinear optimization problem. In Section 3, the problem
is decoupled using dual decomposition into the network
layer and the PHY/MAC layer subproblems, and a
distributed solution for the network layer subproblem is
derived. The PHY layer subproblem is addressed in
Section 4, where a semidistributed solution is proposed.
The MAC layer scheduling is discussed in Section 5. The
performance of the proposed algorithm is verified by
simulations in Section 6. Section 7 contains concluding
remarks.
2 PROBLEM FORMULATION
Consider a wireless network of a set of nodes (either APs or
gateways) denoted by N. The set of links between the nodes
is denoted by L. Denote the set of APs by S and the set of
gateways by D. Also, denote the set of links that are
connected to the gateways by L
1
and the rest L
o
LnL
1
.
From each AP i 2 S, traffic is injected into the network
at the rate of :
i
. Each link 2 L is unidirectional. The
bidirectional links can be represented by a pair of unidirec-
tional links in opposite directions.
It is assumed that the transceivers are subject to the
half-duplex constraint; that is, the radio cannot transmit
and receive at the same time. Such constraint influences
the choice of the set of links that can be activated
simultaneously. This set is called the transmission set in
this paper. The transmission set in time slot t is denoted by
T
t
for t 1. 2. . T. The sequence of transmission sets
T
1
. . T
T
, where

T
t1
T
t
& L, are repeated periodically to
yield a transmission schedule of period T. All transmissions
are assumed to occupy the same frequency band. There-
fore, the links activated at the same time may experience
interference from each other. Obviously, the network
performance will be dependent on the choice of the
transmission sets. This scheduling issue is relegated to
Section 5. For the time being, it is assumed that a suitable
transmission schedule has been given.
For each 2 T
t
, define the link flow variable r
t

that
represents the average traffic flow through link in slot t.
Then, the flow conservation law at each node i 2 S can be
written as
1
T

T
t1

2T
t
\Oi
r
t

2T
t
\Ii
r
t

_
_
_
_
:
i
. i 2 S. 1
where Oi is the set of links emanating from node i and
Ii is the set of links entering node i.
The link flow variables must also satisfy the link capacity
constraints given by r
t

c
t

for all 2 T
t
and t 1. 2. . T,
where c
t

is the PHY layer capacity of link in time slot t


determined by the PHY layer resources such as power,
bandwidth, and so forth, as well as interference. If the
interference can be eliminated by orthogonal multiple
access schemes such as frequency division multiple access
(FDMA) and time division multiple access (TDMA), then
the PHY layer constraints can be typically described by
convex constraints [12]. In this work, multiantenna beam-
forming is employed to mitigate the interference. Thus, the
1260 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 6, NO. 11, NOVEMBER 2007
interference cannot be completely eliminated. The link
capacity is given by
c
t

\ log
2
1 i
t

. 2
where the SINR
t

is given by


G
t

1
t

0
6

0
2T
t
G
t

0 1
t

0 1
. 3
In (2), \ is the transmission bandwidth and i is a factor
called the SNR gap that models the loss in the rate when
using realistic modulation schemes [7]. Without loss of
generality, we set i and \ to 1 in the subsequent
discussion. In (3), 1
t

is the transmit power at the transmitter


of link in time slot t. Note that the variance of the additive
white noise has been normalized to unity. The total transmit
power of the network is assumed to be not more than 1
:ni
in each time slot. Let us denote by 1 and T the
receiving node and the transmitting node of link ,
respectively. Then, G
t

0 is the channel gain from T


0
to
1 in time slot t as a result of beamforming, as well as
path loss and fading, which are assumed to be quasi-static.
Let ` be the number of antenna elements in the nodes,
which is assumed to be common throughout the network. It
is straightforward to extend the formulation to the case
where the number of antennas varies depending on the
nodes. Define w
t

2 CC
`
and g
t

2 CC
`
as the receive and the
transmit beamformer weights at 1 and T, respectively,
in time slot t. Assume that kw
t

k kg
t

k 1. Also, denote
the matrix channel from T
0
to 1 as H
1.T
0

. Then, G
t

0
in (3) is given by
G
t

0 jw
t

H
H
1.T
0

g
t

0 j
2
. 4
Fairness is ensured in terms of the per-AP aggregate
throughput and ensuring fairness among the users within
each AP is beyond the scope of this work. Without loss of
generality, it is assumed that the aggregate throughput
demands from the APs are equal. The extension to an
asymmetric throughput demand vector is straightforward.
Then, define o as the minimum per-AP throughput in the
network. Our objective is to maximize the minimum per-AP
throughput given the constraints from the network and the
PHY/MAC layers:
max o. 5
subject to

T
t1

2T
t
\Oi
r
t

2T
t
\Ii
r
t

_
_
_
_
T:
i
. i 2 S. 6
o :
i
. i 2 S. 7
r
t

c
t

log
2
1
t

. 2 T
t
. t 1. 2. . T. 8

2T
t
1
t

1
:ni
. t 1. 2. . T. 9
over o ! 0. :
i
! 0. r
t

! 0. 10
1
t

! 0. kg
t

k kw
t

k 1.
i 2 S. 2 T
t
. t 1. . T.
11
If the optimal value of o is not smaller than the given
throughput demand, then the optimized network and
PHY/MAC parameters can fully support this demand. If
the optimal value of o is smaller than the given demand,
then the optimized parameters can support only a fraction
of thdemanded throughput, but the fraction is the same for
all APs, which is fair. Note also that our framework can
easily accommodate the utility-based fairness concepts [26]
such as proportional fairness by substituting (5) with an
appropriate utility function of f:
i
g and removing the
constraints in (7).
The above optimization problem is not a convex problem
because of (8) and (11), representing the PHY capacity
constraints. However, when the transmission sets T
t
are
constructed such that strongly interfering links are not
scheduled in the same time slot and/or the number of
antenna elements ` is large so that the interference is
effectively rejected by beamforming, these constraints
become approximately convex given the beamformers w
t

and g
t

. Thus, to exploit the inherent separable structure of


the problem, we take the approach of dual decomposition,
which is usually applied to convex problems [10], [12].
1
Intuitively, one can expect the duality gap to be small if the
nonconvexity is mild. Hence, the primal variables recovered
from the dual optimal variables will be close to the optimal
point. The simulation results in Section 6 support this claim
by demonstrating a significant throughput increase ob-
tained by the joint design compared to the case when the
routing and the PHY layer are separately optimized.
3 DISTRIBUTED COORDINATION USING DUAL
DECOMPOSITION
To facilitate algorithm development, the problem formula-
tion in (5), (6), (7), (8), (9), (10), and (11) needs to be
modified slightly. First, it is noted that the fair throughput
o must be shared among all the nodes in the network if
dual decomposition is applied to the present formulation.
To eliminate this overhead, the objective function o is
replaced by

i2S
:
i
and (7) is replaced by the explicit
equalities given by :
1
:
T
, 2 L
o
. Second, since the
objective function is not strictly concave, the primal
variables cannot be immediately recovered from the dual
variables. To make the objective function strictly concave
with respect to :
i
and r
t

, the variables :
i
in the objective
are again replaced by log :
i
, which does not change the
optimal solution. Moreover, a small regularization term
c

2T t
r
t

_ _
2
is added to the objective. These changes
facilitate the recovery of the primal optimal variables from
the dual optimal solutions [12].
KIM ET AL.: CROSS-LAYER DESIGN OF WIRELESS MULTIHOP BACKHAUL NETWORKS WITH MULTIANTENNA BEAMFORMING 1261
1. Exactly accounting for nonconvexity makes the problem much more
difficult to solve. For example, in [27], the PHY part, with an exact model for
interference, is formulated as an integer program. Our pragmatic approach
is to ease the burden by exploiting the approximate convex structure of the
problem when the MAC scheduling and PHY beamforming eliminate the
interference reasonably well. Although the problem is not exactly convex,
the dual decomposition approach applied to our problem provides an
excellent performance in practice, as demonstrated in the simulation results.
The resulting reformulation is given by
min

i2S
log :
i
c

T
t1

2T t
r
t

_ _
2
. 12
subject to

T
t1

2T
t
\Oi
r
t

2T
t
\Ii
r
t

_
_
_
_
T:
i
. i 2 S. 13
:
T
:
1
. 2 L
o
. 14
r
t

c
t

log
2
1
t

. 2 T
t
. t 1. 2. . T. 15

2T
t
1
t

1
:ni
. t 1. 2. . T. 16
over :
i
! 0. r
t

! 0. 1
t

! 0. kg
t

k kw
t

k 1.
i 2 S. 2 T
t
. t 1. . T.
17
The partial Lagrangian for the above problem can be
written as
1f:
i
g. fr
t

g. f1
t

g. fg
t

g. fw
t

g. fi
i
g. f`

g. fj
t

i2S
log :
i
c

2T
t
r
t

_ _
2

i2S
i
i

2T t\Oi
r
t

2T t\Ii
r
t

_
_
_
_
T:
i
_
_
_
_

2L
o
`

:
T
:
1
_ _

2T
t
j
t

r
t

c
t

_ _
.
18
The Lagrange dual function is then given by
1fi
i
g. f`

g. fj
t

i2S
inf
:
i
!0
log :
i
Ti
i

2Oi
`

2Ii
`

_
_
_
_
:
i
_
_
_
_
_
_

2T
t
inf
r
t

!0
c r
t

_ _
2
i
T
i
1
j
t

_ _
r
t

_ _

t
inf
1
t

!0. kg
t

kkw
t

k1. 2T
t

2T
t
1
t

1:ni

2T t
j
t

c
t

_ _
.
19
where i
i
s for i 2 D are defined to be zero. Thus, given the
dual variables i
/
i
, `
/

, and j
t

/
at the /th iteration, the
optimal primal variables for the first and the second infima
in (19), :

i
/
and r
t

/
, respectively, can be recovered in
closed form:
:

i
/

1
Ti
/
i


2Oi
`
/



2Ii
`
/

_
_

. i 2 S. 20
r
t

/

1
2c
i
/
T
i
/
1
j
t

/
_ _
_ _

. 2 T
t
. t 1. . T.
21
where

maxf0. g, and the denominator in (20) is


assumed to be positive (when it is not positive, :

i
/
can be
assigned an upper bound, namely, :
ior
). Assuming that the
solution c
t

/
for the third infimum in (19) has been found,
the update for the dual variables are given by
i
/1
i

i
/
i
c
/

2T t\Oi
r
t

2T t\Ii
r
t

/
_
_
_
_
T:

i
/
_
_
_
_
. i 2 S.
22
`
/1

`
/

c
/
:

T
/
:

1
/
_ _
. 2 L
o
. 23
j
t

/1
j
t

/
c
/
r
t

/
c
t

/
_ _ _ _

. 2 T
t
. t 1. 2. . T.
24
These update equations are an instance of the subgradient
method since the quantities in the brackets in (22), (23), and
(24) are the subgradients of 1 with respect to i
i
, `

, and
j
t

, respectively [10]. The update can be initialized to any


starting point. Various convergence results have been
established for the subgradient method [28], [10]. For
example, the step size can be chosen to satisfy c
/
! 0 and

1
/0
c
/
1.
It should be noted that the dual-decomposition approach
has decoupled the overall problem into smaller subpro-
blems in the network layer and the PHY layer. Coordinating
the different layer problems are the set of dual variables
j
t

_ _
, which have the natural interpretation of price. The
network layer solves the first and the second minimization
problems in (19), which includes the cost of using the PHY
layer resource in the form of

2T
t
j
t

r
t

, whereas the
PHY layer maximizes the revenue generated by supplying
the resource in the third minimization in (19). The price
update equation (24) essentially represents the market-
based balancing process that corrects the discrepancy in the
supply and demand by finding the best price for goods. For
example, if the quantity r
t

c
t

_ _
is negative, then it means
that the link capacity is underutilized (more supply than
demand). Thus, the price j
t

for the PHY resource is


decreased. This, in turn, would lead to an increase of the
flow rate of the link according to (20). On the other hand,
when the quantity r
t

c
t

_ _
is positive, the price is
increased since the demand for the PHY resource exceeds
what can be supported. This results in a decrease in the link
flow. Over time, the dual prices converge to the values that
can best match the supply and demand.
The level of decomposition of the network layer problem
actually goes down to the individual node/link level such
that the developed algorithm in (20), (21), (22), (23), and (24)
would be fully distributed, provided that the PHY layer
problem could also be solved in a distributed manner.
In the next section, a semidistributed algorithm is
developed, which can find appropriate values of c
t

/
by
providing the power allocation and the beamformer
vectors. This makes the whole algorithm efficiently im-
plementable in a semidistributed manner.
1262 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 6, NO. 11, NOVEMBER 2007
4 A SEMIDISTRIBUTED BEAMFORMING ALGORITHM
In this section, the PHY layer subproblem (the third
infimum in (19)) identified in Section 3 is addressed. The
focus is to find a semidistributed algorithm to minimize
the global message passing overhead. Since the PHY layer
subproblem needs to be solved for each time slot t
1. 2. . T separately, we drop the time slot index t from
the variables. Then, the PHY layer subproblem can be
written as
max

2T
j

log
2
1

. 25
subject to

2T
1

1
:ni
. 26
over 1

! 0. kw

k kg

k 1. 2 T . 27
This is a weighted sum capacity maximization problem
under the condition of using linear receive and transmit
beamforming schemes and treating other link signals as
interference.
4.1 The IMMSE Algorithm
The IMMSE algorithm developed in [21] computes a set of
beampatterns and transmit powers by solving an optimiza-
tion problem in a distributed manner. Explicit feedback of
channel-state information is not necessary, as the algorithm
relies on channel reciprocity. Channel reciprocity holds
when the matrix channels in both directions of a MIMO link
are identical. This property is often assumed in TDD
systems. The algorithm iterates MMSE beamformer com-
putation and power control for each of the (scheduled) links
in the network. To perform MMSE beamforming and power
control, the PHY overheads of transmitting training
sequences and power control commands are incurred.
Since the IMMSE algorithm performs the MMSE beamform-
ing procedure iteratively, the PHY overheads may be
incurred multiple times, depending on the number of
iterations. However, it should be emphasized that these
overheads are generated only locally, since the training and
the power control are done only within the individual links
in a distributed manner. That is, given the number of links
1 jT j in the network, the overhead is on the order of
O1, which is very minor compared to a centralized
algorithm where the interfering links are explicitly con-
sidered, giving a complexity on the order of O11 1.
The IMMSE algorithm minimizes the sum transmit
power subject to the link SINR constraints under the
assumption of channel reciprocity represented by
H
1.T
0

H
T
T
0
.1
. for all .
0
. 28
The algorithm converges to a local optimum of the
following optimization problem:
min
kw

kkg

k1. 1

!0. 2T

2T
1

subject to

. 2 T .
29
where

is the target SINR for link and

is the SINR of
link given by


jw
H

H
1.T
g

j
2
1

0
6

0
2T
jw
H

H
1.T
0

0 j
2
1

0 1
. 30
The inputs to the IMMSE algorithm are the transmission
set T and the target SINRs

, 2 T . How we can determine


the target SINR values to achieve the optimal solution to the
original problem in (25), (26), and (27) is explained in the
next section. In the following, the operation of the IMMSE
algorithm is described.
The beamformers can be initialized to the omnidirec-
tional beampatterns and the transmit powers to suitable
nonzero values. In each iteration , the following sub-
iteration is performed for each link 2 T sequentially.
First, node T and all the other transmitters T
0
,
0
6 ,
transmit packets that contain training sequences. At
node 1, an MMSE receive beamformer is computed
and normalized, which is given by
w


R
1

H
1.T
g

kR
1

H
1.T
g

k
. 31
where the interference-plus-noise covariance matrix R

is
given by
R

0
6

0
2T
1

0 H
1.T
0

0 g
H

0 H
H
1.T
0

I. 32
In the next time slot, transmissions occur in the reverse
directions. Node 1 transmits by using the beamformer
~ g

, which is the complex conjugate of the receive beamfor-


mer just computed, that is, ~ g

. The transmit power of


this transmission is denoted by
~
1

. At the same time, all the


other nodes 1
0
,
0
6 transmit by using their (un-
changed) transmit beamformers ~ g

0 and the transmit powers


~
1

0 . At node T, an MMSE receive beamformer is


computed and normalized:
~ w


~
R
1

H
T.1
~ g

k
~
R
1

H
T.1
~ g

k
. 33
where the matrix
~
R

is defined similar to (32) by


~
R

0
6

0
2T
~
1

0 H
T.1
0

~ g

0 ~ g
H

0 H
H
T.1
0

I. 34
This procedure is repeated until the beamformers for link
converge. After that, the transmit powers 1

and
~
1

are
updated to meet the SINR target:
1

and
~
1


~
1

. 35
where
~

is the SINR of link in the reverse transmission,


defined similarly as in (30). The summary of the IMMSE
algorithm is given in Table 1 in pseudocode.
Note that the receive beamformer computation can be
performed by various adaptive algorithms such as the
least mean squares (LMS) or the recursive least squares
(RLS) algorithms by using the training sequences present
KIM ET AL.: CROSS-LAYER DESIGN OF WIRELESS MULTIHOP BACKHAUL NETWORKS WITH MULTIANTENNA BEAMFORMING 1263
in the packets. The Gauss-Seidel version of the IMMSE
algorithm is shown in Table 1 and explained in the
preceding paragraphs. A Jacobi iteration can be imple-
mented similarly to take advantage of the parallel update
of the links to speed up the algorithm, but the Gauss-
Seidel iteration has a better convergence property [9], [21].
4.2 DETERMINING THE TARGET SINR VALUES
In order to apply this technique to our PHY layer sub-
problem, appropriate target SINR values must be set [20].
Here, for simplicity, the use of a water-filling-type
algorithm is considered. The justification is that, if the
interference is mitigated sufficiently well by beamforming
(that is, G

0 % 0 for 6
0
), then the links can be closely
modeled by parallel channels [29]. Thus, the power
allocation problem can be solved by water filling, except
that in our case, the weighting has to be taken into account.
Given the channel gains G

, 2 T , the target SINR

for
link 2 T is given by

ij

. 36
where i ! 0 is the water level adjusted so as to satisfy the
condition

2T
1

1
:ni
. To determine the value of i, a
central processor collects the transmit powers of all the
active nodes in the network and performs the following
adaptation:
i
/1
i
/
c
0
/
1

2T
1

1
:ni
_ _
. 37
Then, the new value of i is broadcast to all the active nodes
in the network.
Since the update of the beamformers affects the channel
gains G

, the convergence of the update (36) cannot be


guaranteed. A suboptimal heuristic that does not depend on
the values of G

would be to set the target SINR


proportional to j

ij

. 38
For comparison, the case of setting all the link capacities
equal is also considered:

i. 39
Since the target SINR is independent of the dual price, (39)
corresponds to the case where there is no cross-layer
coordination.
The overall algorithm that solves the cross-layer problem
including the PHY layer is summarized in Table 2.
5 HEURISTIC SCHEDULING ALGORITHM
Up to now, we have assumed that the transmission
schedule represented by T
t
, t 1. . T, was given. In this
section, a heuristic algorithm that can provide a good choice
of transmission sets is developed. A similar approach has
been reported in [25]. The idea is to shut down the links that
have minimal contribution in terms of link flow and
schedule those links that have high demands in additional
time slots.
The algorithm starts with an initial schedule T
t
,
t 1. 2. . T, that has all the links scheduled at least once.
For this given schedule, the optimization is performed to
find the optimal route and PHY layer parameters by using
the algorithm in Table 2. Then, the links with the optimized
link capacity c
t

smaller than a threshold c


c
are removed
1264 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 6, NO. 11, NOVEMBER 2007
TABLE 1
The IMMSE Algorithm for Joint Beamforming and Power Control
TABLE 2
Algorithm for Cross-Layer Optimization with a
Given Transmission Schedule T
t
, t 1. . T
from the transmission sets. After that, the algorithm picks
the link with the largest dual price j
t

over all and t and


sees if this link is admissible in the other time slots under
the half-duplex transceiver constraints. If there is no such
time slot, then the algorithm tries the link with the next
largest dual price. Note that j
t

represents the sensitivity of


the optimal objective value at c
t

[30]. Therefore, it is
natural to try increasing the capacity of the link with the
largest dual price to achieve the most increase in the fair
throughput. If there are multiple time slots that can
accommodate the chosen link, then the time slot in which
the least interference is experienced by this link is selected.
For the purpose of estimating the amount of interference
experienced by the candidate link, the beamformers of the
link are set to the optimal beamformers under the zero-
interference condition.
By using the candidate schedule, the cross-layer optimi-
zation is performed again to see if the throughput is
increased. If the throughput has been increased, then this
schedule is accepted and the scheduling algorithm is
iterated from the beginning. Otherwise, the previous
schedule is reinstated and the link with the next highest
j
t

is tried. This procedure terminates when there are no


more links left to try. The flowchart of the described
heuristic scheduling algorithm is given in Fig. 1.
Since the proposed scheduling heuristic is a centralized
algorithm, the scalability of the algorithm is inherently
limited in terms of the computational complexity and the
overhead associated with the scheduled broadcast. How-
ever, in practice, the scheduling update can occur only
infrequently, whereas the network operates with an inter-
mediate suboptimal schedule. In this case, the proposed
algorithm will evolve the schedule to the optimized one in a
longer time scale.
6 SIMULATION RESULTS
6.1 Cross-Layer Optimization with a Given
Schedule
6.1.1 Simulation Setup
The performance of the proposed algorithm is verified by
computer simulations. First, consider the case with fixed
scheduling. A network model, as shown in Fig. 2a, is used.
In this particular network topology, there are six APs
(denoted by the circles), two gateway nodes (denoted by
diamonds), and 19 unidirectional links. The transmission
sets are given by
T
1
f1. 4. 5. 7. 13. 14. 15. 16. 19g.
T
2
f2. 3. 6. 8. 10. 12. 13. 14. 17. 18g. and
T
3
f2. 3. 4. 8. 9. 11. 15. 16. 17g.
where the link numbers are given in Fig. 2a. Each node in
the network is assumed to have a multiple antenna array
KIM ET AL.: CROSS-LAYER DESIGN OF WIRELESS MULTIHOP BACKHAUL NETWORKS WITH MULTIANTENNA BEAMFORMING 1265
Fig. 1. The flowchart of the scheduling algorithm.
Fig. 2. Network topologies used in the simulations.
with ` 4 elements. The matrix channel H
1.T
0

2
CC
``
is assumed to be of rank 1, corresponding to the
case where strong line-of-sight paths exist between the
nodes. A path loss exponent of 3.5 is assumed and the
path loss reference distance is set to the minimum of all
the link distances. Note that the proposed algorithm also
works with the channel with a higher rank that models the
multipath channel, but the rank 1 channel is used here
to make the beampattern plots in Fig. 4 intuitive. The
number of subiterations Q in the IMMSE algorithm is set
to 1, and the do-while loop in Table 1 is also executed only
once per link (that is, c 1) to minimize the computa-
tional overhead.
6.1.2 Simulation Results
Table 3 summarizes the maximized per-AP throughput for
sum transmit powers 1
:ni
10
2
. 10
3
. and 10
4
under dif-
ferent rules of computing the target SINRs. It can be seen
that the water-filling-based rule gives the best performance,
but the simple rule of setting the target SINR in proportion
to the dual price j
t

also gives a quite good performance.


However, it can be clearly seen that just provisioning the
same capacity for all the scheduled links in each time slot
results in much performance degradation. This highlights
the benefit of our cross-layer optimization.
Fig. 3 shows the evolution of the optimized variables as
the algorithm is iterated under the sum power constraint,
with 1
:ni
10
4
. The water-filling-based SINR selection rule
in (36) was applied. The value of c 0.01 was chosen, and
the step size c
/
was fixed to 0.001 for all /. The traffic
variables :
i
and r
t

were normalized by a factor of 10 when


running the algorithm to improve the convergence [13]. In
the top panel of Fig. 3a, the equalized per-AP traffic rate :

i
is shown to converge to a value of approximately 7.2.
Shown in the bottom panel of Fig. 3a are the traffic rates
over the individual links in the transmission set T
1
. Note
that some of the scheduled links carry zero traffic after the
optimization. This is because two unidirectional links have
been scheduled for each bidirectional link (except the ones
1266 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 6, NO. 11, NOVEMBER 2007
TABLE 3
Maximized Throughput for the Network in Fig. 2a
Fig. 3. (a) Evolution of equalized throughput :

i
(top) and link flow r
1

for transmission set T


1
(bottom). (b) Dual variables `

(top) and i
i
(bottom).
(c) Dual variable j
1

for transmission set T


1
(top) and the corresponding target SINR
1

(bottom). (d) Evolution of transmit powers 1


1

for transmission
set T
1
(top) and the total transmit power

2T 1
1
1

(bottom).
that connect to the gateways), since one would not know in
general in which direction the traffic would flow until the
optimization algorithm is actually run. Furthermore, there
may be links that are not used at all in either direction.
These observations motivate the adaptation of the transmis-
sion sets themselves.
The plot in Fig. 3b depicts the evolution of the dual
variables `

and i
i
. The network layer and the PHY layer
subproblems are coordinated by the dual variables j
t

.
Fig. 3c shows the evolution of j
1

for the transmission set T


1
and the corresponding target SINR values
1

determined by
the water-filling-type rule in (36). Fig. 3d shows the
corresponding transmit power allocation for each link and
the total transmit power for transmission set T
1
. Recall from
the theory of Lagrange multipliers that whenever the dual
price j
t

is nonzero, the corresponding inequality constraint


in (8) must be satisfied with equality at the stationary
points. Therefore, those links with nonzero j
t

are provided
with the link capacity just enough to meet the actual traffic
rate. According to the SINR setting rules (36) and (38), the
zero values of j
t

lead to turning off the corresponding links


in the PHY layer. This prevents wasting transmit power
resource by overprovisioning the link capacity.
The beampatterns obtained are shown in Fig. 4 for each
of the transmission sets T
t
, where t 1. 2. and 3. When the
flow variable r
t

is not 0, the value of the flow rate is


approximately depicted by the thickness of the link. It can
be seen that the transmit and receive beampatterns are
formed such that high-gain beams are constructed toward
the desired signal, whereas nulls are placed at the directions
of the interferers/unintended receivers. Also, it can be seen
that more resources are allocated to the links connecting to
the gateways, as those links must carry more traffic.
6.2 Cross-Layer Optimization with Schedule Update
To verify the efficacy of the scheduling heuristic and to
show that the proposed algorithm maintains good perfor-
mance in a larger network, the network topology, as shown
in Fig. 2b, is considered, which has 20 nodes and 77 links.
The matrix channel H
1.T
0

2 CC
``
used in the simula-
tions in this section has a rank of 3: one line-of-sight path
and two multipaths that are 45 degrees off the line-of-
sight path with half the power of the line-of-sight path. The
path loss exponent is set to 3.5. The initial transmission
schedule, with T 5 time slots, is given in Table 4. Then,
the scheduling heuristic described in Section 5 was
KIM ET AL.: CROSS-LAYER DESIGN OF WIRELESS MULTIHOP BACKHAUL NETWORKS WITH MULTIANTENNA BEAMFORMING 1267
Fig. 4. Optimized beampatterns and link traffic assignments. (a) t 1. (b) t 2. (c) t 3.
performed with (36), (37), (38), and (39) for determining the
target SINR.
The result is depicted in Fig. 5, where it can be seen that
the algorithm quickly improves the initial schedule. In the
case of using the water-filling-based target SINR rule, the
algorithm terminates in 46 iterations, but the best schedule
is reached in 43 iterations. The resulting transmission
schedule is tabulated in Table 4, which achieves about an
80 percent increase in the fair throughput. This is accom-
plished by dropping the minimally used links from the
transmission sets and assigning more time slots to the links
that are heavily used. Naturally, the links that are near the
gateways, such as links 12, 13, 14, 71, 72, and 77, are given
more chances to be activated. Comparing the curves with
the cross-layer coordination (labeled Water filling and
Proportional) with the one without the cross-layer
coordination (labeled Equal), one can again clearly see
the benefit of our design.
7 CONCLUSIONS
A joint multipath routing and PHY/MAC resource alloca-
tion algorithm has been proposed for wireless backhaul
networks. A MIMO beamforming PHY layer has been
employed for higher spectral efficiency. The spatial unfair-
ness prone to appear in multihop networks has been
addressed by explicitly imposing fairness constraints in
terms of the per-AP aggregate throughput. A nonlinear
optimization problem that captures the cross-layer con-
straints has been formulated. Although the resulting
formulation is nonconvex and coupled due to the effect of
interference, it was observed that the nonconvexity of the
problem is mild, provided that the beamforming can
separate the signals effectively. The dual decomposition
method is applied to decouple the problem into the network
layer and the PHY layer subproblems, which are coordi-
nated by the dual variables. The network layer subproblem
admits a distributed solution and the beamforming PHY
layer subproblem is solved by a semidistributed algorithm.
A heuristic scheduling update algorithm has also been
developed. The performance of the proposed algorithm has
been verified by simulations, which demonstrates the
efficacy of the proposed cross-layer design paradigm.
REFERENCES
[1] R. Karrer, A. Sabharwal, and E. Knightly, Enabling Large-Scale
Wireless Broadband: The Case for TAPs, Evaluation of a Wireless
Enterprise Backbone Network Architecture, vol. 34, no. 1, pp. 27-32,
Jan. 2004.
[2] V. Gambiroza, B. Sadeghi, and E.W. Knightly, End-to-End
Performance and Fairness in Multihop Wireless Backhaul Net-
works, Proc. ACM MobiCom, pp. 287-301, Sept.-Oct. 2004.
[3] C. Eklund, R.B. Marks, K.L. Stanwood, and S. Wang, IEEE
Standard 802.16: A Technical Overview of WirelessMAN Air
Interface for Broadband Wireless Access, IEEE Comm. Magazine,
vol. 40, no. 6, pp. 98-107, June 2002.
[4] R. Bruno, M. Conti, and E. Gregori, Mesh Networks: Commodity
Multihop Ad Hoc Networks, IEEE Comm. Magazine, vol. 43, no. 3,
pp. 123-131, Mar. 2005.
[5] R.L. Cruz and A.V. Santhanam, Optimal Routing, Link Schedul-
ing and Power Control in Multi-Hop Wireless Networks, Proc.
IEEE INFOCOM, vol. 1, pp. 702-711, Mar.-Apr. 2003.
[6] S.A. Grandhi, R. Vijayan, D.J. Goodman, and J. Zander,
Centralized Power Control in Cellular Radio Systems, IEEE
Trans. Vehicular Technology, vol. 42, no. 4, pp. 466-468, Nov. 1993.
[7] A.J. Goldsmith and S. Chua, Variable-Rate Variable-Power
MQAM for Fading Channels, IEEE Trans. Comm., vol. 45,
pp. 1218-1230, Oct. 1997.
[8] G.J. Foschini, Layered Space-Time Architecture for Wireless
Communication in a Fading Environment When Using Multi-
Element Antennas, Bell Labs Technical J., pp. 41-59, Autumn 1996.
[9] D.P. Bertsekas and J.N. Tsitsiklis, Parallel and Distributed Computa-
tion: Numerical Methods. Prentice Hall, 1989.
[10] D.P. Bertsekas, Nonlinear Programming, second ed. Athena
Scientific, 1999.
[11] M. Chiang, To Layer or Not to Layer: Balancing Transport and
Physical Layers in Wireless Multihop Networks, Proc. IEEE
INFOCOM, vol. 4, pp. 2525-2536, Mar. 2004.
[12] L. Xiao, M. Johansson, and S.P. Boyd, Simultaneous Routing and
Resource Allocation via Dual Decomposition, IEEE Trans. Comm.,
vol. 52, no. 7, pp. 1136-1144, July 2004.
[13] R. Madan and S. Lall, Distributed Algorithms for Maximum
Lifetime Routing in Wireless Sensor Networks, Proc. 47th Ann.
IEEE Global Telecomm. Conf. (GLOBECOM 04), pp. 748-753, 2004.
[14] R.A. Monzingo and T.W. Miller, Introduction to Adaptive Arrays.
John Wiley & Sons, 1980.
[15] F. Rashid-Farrokhi, K. Liu, and L. Tassiulas, Transmit Beam-
forming and Power Control for Cellular Wireless Systems, IEEE
J. Selected Areas in Comm., vol. 16, no. 8, pp. 1437-1450, Oct. 1998.
1268 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 6, NO. 11, NOVEMBER 2007
TABLE 4
Initial Schedule for Optimizing the Network Shown in Fig. 2b
and Final Schedule Obtained by the Proposed Scheduling
Update Algorithm with the Water-Filling Heuristic in (36)
Fig. 5. Iterative scheduling update.
[16] E. Visotsky and U. Madhow, Optimum Beamforming Using
Transmit Antenna Arrays, Proc. 49th IEEE Vehicular Technology
Conf. (VTC 99), pp. 851-856, May 1999.
[17] M. Bengtsson and B. Ottersten, Optimal Downlink Beamforming
Using Semidefinite Optimization, Proc. 37th Ann. Allerton Conf.
Comm., Control and Computing, pp. 987-996, Sept. 1999.
[18] M. Schubert and H. Boche, Solution of the Multiuser Downlink
Beamforming Problem with Individual SINR Constraints, IEEE
Trans. Vehicular Technology, vol. 53, no. 1, pp. 18-28, Jan. 2004.
[19] M.C. Bromberg and B.G. Agee, Optimization of Spatially
Adaptive Reciprocal Multipoint Communication Networks, IEEE
Trans. Comm., vol. 51, no. 10, pp. 2352-2362, Oct. 2003.
[20] M.C. Bromberg, Optimizing MIMO Multipoint Wireless Net-
works Assuming Gaussian Other-User Interference, IEEE Trans.
Information Theory, vol. 49, no. 10, pp. 2352-2362, Oct. 2003.
[21] R.A. Iltis, S.-J. Kim, and D. Hoang, Noncooperative Iterative
MMSE Beamforming Algorithms for Ad Hoc Networks, IEEE
Trans. Comm., vol. 54, no. 4, pp. 748-759, Apr. 2006.
[22] S. Ye and R.S. Blum, Optimized Signaling for MIMO Interference
Systems with Feedback, IEEE Trans. Signal Processing, vol. 51,
no. 11, pp. 2839-2848, Nov. 2003.
[23] S. Serbetli and A. Yener, Transceiver Optimization for Multiuser
MIMO Systems, IEEE Trans. Signal Processing, vol. 52, no. 1,
pp. 214-226, Jan. 2004.
[24] R. Bhatia and M. Kodialam, On Power-Efficient Communication
over Multi-Hop Wireless Networks: Joint Routing, Scheduling
and Power Control, Proc. IEEE INFOCOM, vol. 2, pp. 1457-1466,
Mar. 2004.
[25] R. Madan, S. Cui, S. Lall, and A. Goldsmith, Cross-Layer Design
for Lifetime Maximization in Interference-Limited Wireless Sensor
Networks, Proc. IEEE INFOCOM, pp. 1964-1975, 2005.
[26] J. Mo and J. Walrand, Fair End-to-End Window-Based Conges-
tion Control, IEEE/ACM Trans. Networking, vol. 8, no. 5, pp. 556-
567, Oct. 2000.
[27] M. Johansson and L. Xiao, Cross-Layer Optimization of Wireless
Networks Using Nonlinear Column Generation, IEEE Trans.
Wireless Comm., vol. 5, no. 2, pp. 435-445, Feb. 2006.
[28] N.Z. Shor, Minimization Methods for Non-Differentiable Functions.
Springer-Verlag, 1985.
[29] T. Cover and J. Thomas, Elements of Information Theory. John Wiley
& Sons, 1991.
[30] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge
Univ. Press, 2004.
Seung-Jun Kim received the BS and MS
degrees from Seoul National University, Korea,
in 1996 and 1998, respectively, and the PhD
degree from the University of California, Santa
Barbara in 2005, all in electrical engineering.
From 1998 to 2000, he served as a Korea
Overseas Volunteer at Chiangrai Teachers
College in Chiangrai, Thailand. Since 2005, he
has been with NEC Laboratories America in
Princeton, New Jersey, where he is currently a
research staff member. His general research interests are in the area of
physical and MAC layer design for wireless communication systems
including space-time processing, cross-layer optimization, and wireless
backbone networks. He is a member of the IEEE.
Xiaodong Wang received the PhD degree in
electrical engineering from Princeton University.
He is now on the faculty of the Department of
Electrical Engineering, Columbia University.
Dr. Wangs research interests fall in the general
areas of computing, signal processing, and
communications, and he has published exten-
sively in these areas. Among his publications is
a book titled Wireless Communication Systems:
Advanced Techniques for Signal Reception
(Prentice Hall, 2003). His current research interests include wireless
communications, statistical signal processing, and genomic signal
processing. Dr. Wang received the 1999 US National Science
Foundation CAREER Award and the 2001 IEEE Communications
Society and Information Theory Society Joint Paper Award. He has
served as an associate editor for the IEEE Transactions on Commu-
nications, the IEEE Transactions on Wireless Communications, the
IEEE Transactions on Signal Processing, and the IEEE Transactions on
Information Theory. He is a senior member of the IEEE.
Mohammad Madihian received the PhD degree
in electronic engineering from Shizuoka Univer-
sity, Japan, in 1983. He joined NEC Central
Research Laboratories, Kawasaki, Japan,
where he worked on research and development
of Si and GaAs device-based digital as well as
microwave and millimeter-wave monolithic ICs.
In 1999, he moved to NEC Laboratories Amer-
ica, Inc., Princeton, New Jersey, and is presently
the Department Head and Chief Patent Officer.
He conducts PHY/MAC layer signal processing activities for high-speed
wireless networks and personal communications applications. He has
authored or coauthored more than 130 scientific publications, including
20 invited talks, and holds 35 Japanese/US patents. Dr. Madihian
received the IEEE MTT-S Best Paper Microwave Prize in 1988 and
became an IEEE fellow in 1998. He holds eight NEC Distinguished R&D
Achievement Awards. He has served as guest editor of the IEEE Journal
of Solid-State Circuits, Japan IEICE Transactions on Electronics, and
IEEE Transactions on Microwave Theory and Techniques. He is
presently serving on the IEEE Speakers Bureau, the IEEE Compound
Semiconductor IC Symposium (CSICS) executive committee, the IEEE
Radio and Wireless Conference steering committee, the IEEE Interna-
tional Microwave Symposium (IMS) technical program committee, the
IEEE MTT-6 subcommittee, the IEEE MTT editorial board, and the
technical program committee of the International Conference on Solid
State Devices and Materials (SSDM). Dr. Madihian is an adjunct
professor in the Electrical and Computer Engineering Department,
Drexel University, Philadelphia, Pennsylvania.
> For more information on this or any other computing topic,
please visit our Digital Library at www.computer.org/publications/dlib.
KIM ET AL.: CROSS-LAYER DESIGN OF WIRELESS MULTIHOP BACKHAUL NETWORKS WITH MULTIANTENNA BEAMFORMING 1269

You might also like