IEEE INFOCOM 2002

QoS and Fairness Constrained Convex

Optimization of Resource Allocation for Wireless
Cellular and Ad Hoc Networks
David Julian, Mung Chiang, Daniel ONeill and Stephen Boyd.

Abstract
For wireless cellular and ad hoc networks with QoS constraints,
we propose a suite of problem formulations that allocate network
resources to optimize SIR, maximize throughput and minimize delay. The distinguishing characteristics of these resource allocation
formulations is that, by using convex optimization, they accommodate a variety of realistic QoS and fairness constraints. Their globally optimal solutions can be computed efficiently through polynomial time interior point methods, even though they use nonlinear
objectives and constraints.
Through power control in wireless cellular networks, we optimize SIR and delay for a particular QoS class, subject to QoS
constraints for all other QoS classes. For wireless ad hoc networks with multihop transmissions and Rayleigh fading, we optimize various objectives, such as the overall system throughput,
subject to constraints on power, probability of outage, and data
rates. These formulations can also be used for admission control
and relative pricing. Both proportional and minmax fairness can
be implemented under the convex optimization framework, where
fairness parameters can be jointly optimized with QoS criteria.
Simple heuristics are also shown and tested using the convex optimization tools.
Index Terms Ad hoc Networks, Cellular Networks, Convex
Optimization, QoS Constrained Resource Allocation, Fairness

I. I NTRODUCTION

S users of communication networks become less satisfied

with best efforts transmission, Quality of Service (QoS)
has become an important research and commercial issue. QoS
covers a wide array of network attributes, including bandwidth,
delay, and packet delivery guarantee. Voice, data, image, and
video have different bandwidth requirements. Some classes of
traffic, such as voice, are also much more sensitive to delays
than other classes, such as data.
QoS provisioning in a wireless network is a particularly difficult problem due to the time varying and unreliable physical
channel. We present a new framework of convex optimization
as a computationally efficient tool for resource allocation, including power control and admission control, under QoS and
fairness constraints.
In wireless cellular networks, power control can be used to
control interference, and in doing so, indirectly control the QoS
seen by users on the network. In the downlink, a mobile user

Information Systems Laboratory,Electrical Engineering Department, Stanford University, CA 94305, USA. E-mail:{djulian, chiangm, dconeill, boyd}
@Stanford.edu
This work was supported by NSF Grant CCR-9973134, Stanford Network
Research Center, Hertz Foundation Fellowship, and Stanford Graduate Fellowship.

can receive interfering transmissions form base stations in adjacent cells resulting in adjacent channel interference. In the
uplink, a base station experiences adjacent channel interference
from users in adjacent cells, and also co-channel interference
from mobile users in the same cell interfering with one another.
The Signal to Interference Ratio (SIR) is often used to capture
the effect of both co-channel and adjacent channel interference,
and is routinely used in this paper to characterize the QoS parameter of throughput of a particular link. Extending our work
in [10], sections IV and V formulate the following problems P 1
to P 3 for wireless cellular networks and show that they can be
solved using convex optimization techniques.
P1

Determining feasibility of a set of SIR requirements.

P2

Maximizing SIR for a particular class of users with

lower bounds on the QoS of all other users.

P3

Satisfying queuing delay requirements for users in

various QoS classes.

Ad hoc wireless networks pose additional technical challenges for QoS support. Unlike cellular wireless networks, ad
hoc networks have no cells or base stations, but are composed
of a set of nodes that transmit, receive and relay information
among each other. Packets traverse the network by multihop
transmissions from node to node until arriving at the destination. Consequently user QoS requirements are transformed into
a set of QoS link requirements for the hops taken from source
to destination.
Extending the results for cellular networks to ad hoc networks and our work in [7], the following problems P 4 to P 7
are solved for multi-hop networks in sections VI and VII:
P4

Finding the optimum power control to maximize overall system throughput consistent with QoS guarantees
in a fading environment.

P5

Determining feasibility of a set of service level agreements (SLA) under network resource constraints.

P6

Solving for the minimum total transmission delay of

the most time sensitive class of traffic by optimizing
over powers, capacities, and SLA terms.

P7

Maximizing the unused capacity of the network.

Apart from performance optimization, fairness is another important issue in QoS provisioning. We show that both propor-

0-7803-7476-2/02/$17.00 (c) 2002 IEEE.

IEEE INFOCOM 2002

tional fairness and minmax fairness can be formulated within

the convex optimization framework. A joint optimization of
the fairness parameters and the QoS criteria can also be implemented under the proposed framework.
This paper unifies and extends our work in cellular networks
[10] and ad hoc networks [7]. In addition to unifying results
into one framework, the following new extensions are made.
Admission control and pricing in the ad hoc network is described in detail with a simulation of the concepts, and delay
constraints in the geometric programming framework are addressed. Also, the fairness weights are extended to the noninteger case, and it is shown that the weights can be used as optimization variables. Finally, two simple heuristics are described,
and illustrate that the convex optimization formulations can be
used as a benchmark to compare the performance of heuristics
against.
II. R ELATED W ORK
As an important special case of resource allocation in wireless networks, power control in cellular networks has been studied extensively, and various schemes have been proposed or
adopted in 2G and 3G networks. For example, the classical
Qualcomm power control is used to solve the near far problem
in CDMA networks. We show in section IV that this particular
power control scheme is a special case of the proposed general
convex optimization framework.
Various iterative methods have been proposed to optimally
maximize the minimum SIR, to minimize total or individual
power, or to maximize throughput in [1], [3], [4], [6], [8], [9],
[13], [15], [18], [19], but these methods are not general enough
to allow a diverse set of QoS constraints and other objective
functions. We present a new framework of resource allocation
based on the computational efficiency and versatility of convex optimization, which strikes a balance between efficiently
achieving global optimality and flexibly allowing different constraints, objectives and variables.
A similar idea has been used to minimize outage probability
in cellular networks under Rayleigh fading without QoS constraints in [11]. We show that the convex optimization framework can be used to incorporate a variety of QoS constraints
and objectives, not just for cellular networks, but for ad hoc
networks as well.
III. C ONVEX O PTIMIZATION AND G EOMETRIC
P ROGRAMMING
We need efficient algorithms to find the optimal solution to
nonlinear problems P 1 to P 7. Fortunately, these problems can
be turned into convex optimization formulations, which have
efficient polynomial time algorithms such as the primal dual
interior point method.
Convex optimization refers to minimizing a convex objective function over convex constraint sets. The particular type
of convex optimization we use is in the form of geometric program. [5]. Geometric programming focuses on monomial and
posynomial functions.
Definition 1: A monomial is a function f : Rn R, where
the domain contains all real vectors with non-negative components:

h(x) = cxa1 1 xa2 2 xann , c 0 and ai R

(1)

2: A posynomial is a sum of monomials f (x) =

Definition
a1k a2k
ank
k ck x1 x2 xn .
Geometric program is an optimization problem with the following form:
minimize
subject to

f0 (x)
fi (x)
hj (x)

1
= 1

(2)

where f0 and fi are posynomials and hj are monomials. Geometric programming in the above form is not a convex optimization problem. However, with a change of variables:
yi = log xi and bik = log cik , we can put it into convex form:
minimize
subject to

p0 (y) = log
k exp(aT0k y + b0k )
pi (y) = log k exp(aTik y + bik ) 0
= 0
qj (y) = aTj y + bj

(3)
It can be verified that the logarithm of a sum of exponentials is a convex function. Therefore pi are convex functions
and qj are affine functions, and we have a convex optimization
problem. Note that if all posynomials are in fact monomials,
geometric programming becomes linear programming.
Convex optimization problems can be solved globally and
efficiently through the interior point primal dual method
[14],

with polynomial running times that are often O( N ) where

N is the size of the problem. Apart from computational efficiency, convex optimization also offer duality interpretations,
stability analysis and accommodate a variety of constraints. Solution algorithms also unambiguously and efficiently determine
feasibility. This paper shows how geometric programming can
solve many versions of QoS provisioning and resource allocation problems in wireless cellular and ad hoc networks.
IV. T HROUGHPUT O PTIMIZATION FOR W IRELESS
C ELLULAR N ETWORKS
A. Problem formulations
We first consider power control in a wireless cellular network
in this section. For notational simplicity, this section considers
a single base station and N links. Extensions to multiple base
stations and the associated links are straight forward. Each link
is a unidirectional path from the transmitter to the receiver. The
propagation model used in this section is as follows:
 
d0
(4)
Pr = P K
d
where Pr is the received power, P is the transmitted power, d
is the propagation path length, and d0 is a reference distance
for the antenna far-field, usually taken so that the normalization
constant K equals 1. The path loss exponent is usually between 2 and 6 for most indoor and outdoor environments. The
interfering users powers are decreased by the inverse of Ks ,
which can be the spreading gain for a CDMA system or the
power falloff with frequency for an FDMA system. Accordingly, SIRi for the ith link is defined as

0-7803-7476-2/02/$17.00 (c) 2002 IEEE.

IEEE INFOCOM 2002

SIRi =
N

j=i

i
Pi d
i
i

Pj Ks1 dj

j + ni

(5)

where the factors j are introduced to accommodate normalization constants and other factors, such as the effects of beamforming in multiantenna systems. SIR is well justified to be
used as a throughput QoS parameter. For example, channel capacity scales with log SIR for high SIR, and the probability
of symbol decoding error
for coherent MPSK and MQAM is
approximately M Q( M SIR), where M and M depend
on the modulation type and the Q function is the complementary Gaussian CDF.
The problem of SIR maximization can be formulated as a geometric program. In the following basic formulation, the SIR
is maximized for a particular mobile i. At the same time QoS
for the other mobiles should also satisfy certain requirements
or constraints. The following four kinds of constraints are reflected in Formulation 1 below.
1) Interference due to users, including base stations and mobiles, in index set I1,k must be smaller than a positive
constant ck because their assigned QoS values are relatively low.
2) Interference due to users in index set I2,k has to be
smaller than the received signal power for some mobile
k so as to achieve a required SIR, k .
3) The received signal power for some mobile k needs to be
exactly equal to a positive constant Ck .
4) As in the special case of the classical power control
scheme to solve the near-far problem in CDMA, the received signal power for one mobile k1 needs to be equal
to that of another mobile k2 .
With the objective and constraints thus formulated and upper
bounds Pi,U B on all transmitted powers Pi included, we obtain
the following non-linear optimization formulation:
Formulation 1: (SIR constrained optimization of power control) The following nonlinear problem of optimizing node powers to maximize SIR for a particular user under QoS constraints for other users in a cellular network is a convex optimization problem.

maximize

SIRi

i
i

Pj Ks1 dj j j +ni
j=i

N

Pi di

subject to


j
j
jI1,k Pj dj

k jI2,k Pj Ks1 dj j j + k nk
k
Pk d
k
k
k1
Pk1 dk1 k1
Pj
Pj

<
<
=
=

ck
k
Pk d
k
k
Ck
k2
Pk2 d
k2 k2
Pj,U B
j
0
j

(6)
where the first four constraints are for all k in the appropriate
index sets. While the objective function is not a posynomial,
1
is, and minimizing ISR over the same
its inverse ISR = SIR

constraints is equivalent to the original problem. The inequality constraints above are posynomials, since posynomials when
divided by monomials are necessarily posynomials in the parameters Pi , di and i . The equality constraints are monomials in
the same parameters. The variables are the transmitted powers
Pi . Therefore, this is indeed a geometric program, and therefore a convex optimization problem with efficient algorithms
that obtain the global optimality.
This general formulation can be applied to different power
control situations. For example, if there is no objective function, the above formulation reduces to a SIR requirement feasibility problem.

Also, the objective function can be replaced
by minimize i Pi as in the following formulation, and then
the minimum power vector under the QoS constraints can be
determined.
Formulation 2: (SIR constrained optimization for minimum
power)
The following nonlinear problem of minimum power allocation in a cellular network
is a convex optimization problem.


P
minimize
i i
(7)
subject to Same constraints as in Formulation 1.
Additionally, a weighted sum or powers, or the maximum
user power can be minimized. The di can also be treated as optimization variables for optimization of antenna sectoring, which
is a popular technique for interference mitigation.
B. Interpretations of the QoS Constrained Power Control
The log-sum-exp function can be interpreted as a smooth approximation of the maximum function [5]:
max(xi ) log

n


exi log(n) + max(xi )

(8)

Therefore, the above convex optimization of power control is

minimizing a smooth approximation of the maximum of
log(

d i
Pj
) + log( i j ).
Pi
dj

(9)

When i = j , this is a weighted sum of the difference in powers (measured in dB) and the difference in distance (also measured in dB) for users j and i.
The dual problem is a generalized entropy maximization [5].
By duality analysis, it can be shown that solving the QoS constrained power control problem is equivalent to finding the linear combiners of the Lagrangian function (the augmented objective function) with the maximum weighted sum of entropy,
where the weights are induced by the constraints of the dual
problem.
C. Proportional and Minmax Fairness Extensions
Fairness is another major issue in a QoS policy. Both proportional fairness and minmax fairness can be accommodated
in the framework of geometric programming.
Formulation 3: (SIR constrained optimization with proportional fairness)

0-7803-7476-2/02/$17.00 (c) 2002 IEEE.

IEEE INFOCOM 2002

The following nonlinear problem of weighted fair power allocation in a cellular

network is a convex optimization problem.
maximize
i wi log SIRi
(10)
subject to Same constraints as in Formulation 1.
This extended version of power control for general wi R is
still a convex optimization

problem in geometric program form
because maximizing i wi log SIRi is equivalent to maximizing log i SIRiwi , which is in turn equivalent to minimizing
i ISRiwi . Since a product of posynomials is also a posynomial, both the objective function and constraints are posynomials if the weights wi are integers.
While in general posynomials to noninteger powers can not
be handled in the geometric programming framework, the structure of this problem allows for noninteger weights. Auxiliary
be introduced, the objective function changed
variables
 ti can
i
,
and
the constraints ISRi ti added to the exto min i tw
i
isting set of constraints for all i. The optimization variables are
now ti and Pi , and the objective function and constraints are
posynomials in the optimization variables. Further, the value of
the objective function and the optimizing powers are the same
as in the original formulation.
The minimum SIRi can also be maximized subject to QoS
constraints for other users through convex optimization. This
minimax algorithm is useful in situations where the worst case
is of concern.
Formulation 4: (SIR constrained optimization with minmax
fairness)
The following nonlinear problem of minmax fair power allocation in a cellular network is a convex optimization problem.
maximize mini SIRi
subject to Same constraints as in Formulation 1.

(11)

This
programming problem because minimiz
 is a geometric
1
is
equivalent
to minimizing over an auxiliary
ing maxi SIR
i
scalar variable t such that maxi ISRi t, which is in turn
equivalent to minimizing t such that ISRi t i. So the
auxiliary variable t acts as an upper bound on all ISRs. When
minimized over all feasible P , the value of t is reduced until it
achieves the minimax value.
D. SIR optimization simulation
A simple system comprised of five users is used for a simulation of Formulation 1. The setup is as follows. First, the five
users are spaced at distances d of 1, 5, 10, 15, and 20 units from
the base station. The power drop off factor = 4, and = 1.
Each user has a maximum power constraint of PU B = 0.5W .
The noise power is 0.5W for all users. CDMA is used with
a spreading gain of Ks = 10. The SIR of all users, other than
the user we are optimizing for, must be greater than a common
threshold SIR level . is varied to observe the effect on the
optimized users SIR. This is done independently for the near
user at d = 1, a medium distance user at d = 15, and the far
user at d = 20. The results are plotted in figure 1.
Several interesting effects are illustrated by this simulation.
First, when the required threshold SIR for the non-optimized

users is high there are no feasible power control solutions. At

moderate threshold SIR, as is decreased, the optimized SIR
initially increases rapidly. This is because it is allowed to increase its own power by the sum of the power decrease in
the four other users, and the noise is relatively insignificant.
At low threshold SIR, the noise becomes more significant and
the power trade-off from the other users less significant, so
the curve starts to bend over. Eventually, the optimized user
reaches its upper bound on power and cannot utilize the excess power allowed by the lower threshold SIR. Therefore, during that stage, the only gain in the optimized SIR is the lower
power transmitted by the other users. This is exhibited by the
sharp bend in the curve to a much shallower sloped curve. We
also note that the most distant user in the constraint set dictates
feasibility.
E. Admission Control and Pricing
As shown by the interpretation of the above simulation, convex optimization can also be used for admission control in wireless communication networks. A new user is only allowed admission into the network when a feasible solution of this geometric program exists.
The effect of adding a new user to the system could be used
to establish pricing for that user. The concept is to charge more
to users who consume more of the total system user capacity.
Unlike fixed QoS systems, the effect of admitting a new user
depends heavily on the QoS requirements of the new user. A
user with a high QoS requirement will reduce system user capacity more than a user with easily supported QoS requirements
and should be charged commensurately more. For example, a
user seeking high data rates close to a group of other users will
more adversely affect user capacity, through interference and
power limitations, than a user seeking a lower data rate in a low
interference region.
The effect on capacity can be modeled using geometric programming by determining the number of standardized users that
could be added to the system both before and after the new user
is admitted. The difference between these two numbers can be
taken as the reduction in the system (standard) user capacity
that results from admitting the new user to the system. The
price could then be set as an linear function of this difference.
Under this approach, a user could experience different spot
pricing at different times depending on the existing load on the
system when the user sought to access the network. Spot pricing is different from the current rate based pricing approaches,
and more realistically models the effect a user has on the network from a revenue potential point of view.
V. Q UEUING D ELAY O PTIMIZATION FOR W IRELESS
C ELLULAR N ETWORKS
Delay can be an important part of QoS for a wireless cellular
network. There are three main component of the overall delay: propagation delay, transmission delay and queuing delay.
Queuing delay is particularly important for bursty digital data,
where the short term data rate of the information to be transmitted may exceed the data rate supported by the wireless link.
Buffering is used to address this short term imbalance. This

0-7803-7476-2/02/$17.00 (c) 2002 IEEE.

IEEE INFOCOM 2002

Optimized SIR vs. Threshold SIR

Gij can encompass path loss, shadowing, antenna gain, coding

gain, and other factors.
The Rayleigh fading between each transmitter j and receiver
i is given by Fij . The Fij s are assumed to be independent
and have unit mean. The Gij s are appropriately scaled to reflect variations from this assumption. The distribution of the
received power between any pair of transmitters j and receivers
i is exponential with mean value,

20
near
medium
far

15

Optimized SIR (dB)

10

E [Gij Fij Pj ] = Gij Pj

The signal to interference ratio (SIR) for user i now becomes

10

SIRi =
N

15

20
5

(14)

j=i

Fig. 1. Simulation results for constrained optimization of power control in a

cellular network, with implications for admission control and pricing scheme.

queuing delay can dominate the propagation delay for reasonable link data rates.
By assuming that packets arrive according to a Poisson distribution and that packets are of variable length, the system can
be modeled as an M/M/1 queue. The average queueing delay,
D, can then be expressed as
1
(P )

(12)

where (P ) is link transmission rate, or service rate, and is

the arrival rate.
If a QoS agreement specifies an average delay bound Db and
an average maximum arrival rate, then this bound can be met by
constraining the SIR on this link to exceed a minimum threshold, so that link transmission rate, as determined by the modulation type and the SIR, is larger than D1b + .
VI. P OWER CONTROL FOR THROUGHPUT OPTIMIZATION IN
W IRELESS A D H OC N ETWORKS
In this section, we turn to power control in wireless ad hoc
networks with multihop transmission. We will further expand
the suite of formulations to more general resource allocation
settings in the next section.
The formulation used in this section explicitly takes into account the statistical variation of the received signal and the interference power over a multi-hop network.
A. Multi-hop network model and Rayleigh fading
Consider a wireless ad hoc network with n transmitter/receiver pairs, labelled 1, . . . , n, which transmit at powers
P1 . . . , Pn . The power received from transmitter j, at receiver
i is given by
Gij Fij Pj

(15)

Pj Gij Fij + ni

10

Threshold SIR (dB)

D=

Pi Gii Fii

(13)

The nonnegative number Gij represents the path gain in the

absence of fading from the j th transmitter to the ith receiver.

B. Outage probability and system throughput

Due to multihop transmission over unreliable links, outage
probability is an important QoS parameter in wireless ad hoc
networks. An outage is declared when the received SIR falls below a given threshold defined as SIRth , often computed from
a BER requirement. Neglecting the noise in the high power interference limited case, the outage probability associated with
the ith hop is given by
Oi

= P r(SIRi SIRth )

= P r(Gii Fii Pi SIRth k=i Gik Fik Pk )

The outage probability can be expressed as [11]


1
Oi = 1 k=i SIRth
Gik Pk
1+

(16)

(17)

Gii Pi

Outage probability over a hop induces an outage probability

over a path S
OpathS

= Prob(outage
along the path S)

= 1 sS (1
 Oi )
1
= 1 sS k=s
SIRth Gik Pk .
(1+

Gii Pi

(18)

The constellation size M used by a hop can be closely approximated for MQAM modulation as follows
M =1+

1.5
SIR
ln(5BER)

(19)

1.5
where BER is the bit error rate. Defining K = ln(5BER)
leads
th
to a monotonic expression for the data rate of the i hop as a
function of the received SIR:

Ri = (1/T ) log2 (1 + KSIRi )

(20)

The aggregate data rate for the system can then be written as
a sum of terms of this form.



Rsystem = i Ri = (1/T ) log2 i (1 + KSIRi )
(21)
So throughput maximization is equivalent to maximizing the
product of SIR. This was also observed by Qiu and Chawla in
[6], [15] where they used it for optimizing throughput in cellular networks. Overall system throughput is now defined as the
maximum aggregate data rate supportable by the system given
a set of users with defined QoS.

0-7803-7476-2/02/$17.00 (c) 2002 IEEE.

IEEE INFOCOM 2002

C. Throughput optimization

Formulation 5: (Optimize power for throughput maximization) The following nonlinear problem of optimizing user node
powers to maximize total network throughput is a convex optimization problem.
maximize Rsystem
subject to
Ri 
1
1 k=i SIRth
Gik Pk
1+
Gii Pi


1
1 sS k=s
SIRth Gik Pk
(1+

Pi

Gii Pi

Ri,LB , i
P routi

P rout

Pi,U B

20m

path s

(22)
The objective function is the overall system throughput.
In
 the actual optimization the posynomial objective function
i ISRi is minimized; which, as shown previously, is equivalent to maximizing the system throughput. The objective function is now optimized over the set of all feasible powers Pi .
The first constraint is the data rates demanded by existing
system users. The second constraint is the outage probability
limitations demanded by users using single hops. The third
constraint is the outage probability limitations for users using
a multi-hop path. Lastly, the forth constraint is regulatory or
system limitations on transmitted powers.
D. Throughput maximization simulation
A simple four node multi-hop network is considered in the
following simulation. As shown in figure 2, the network consists of 4 nodes A, B, C, and D, and 4 links 1, 2, 3, and 4. On
link 1 node A is the transmitter and node B is the receiver; similarly, the transmitter and receiver nodes for each link are shown
in the figure. Note that node A is the transmitter on both links
1 and 3, illustrating that a node can be a transmitter and/or receiver on many links. Nodes A and D as well as B and C are
separated by a distanceof 20m. By geometry the distance of
each transmit path is 10 2m.
For our simulation each link has a maximum transmit power
of 1W. Alternatively, we could also have placed the power constraint on each node instead of each link by adding a constraint
that P1 + P3 1W. All nodes are using MQAM modulation.
The baseband bandwidth for each link is 10kHz, the minimum
data rate for each link is 100bps for maintenance data, and the
target BER is 103 . For the Rayleigh fading we require a probability of outage of Pout = 0.1 for an SIR threshold of 10dB.
 1 4
1
The gains for each link are computed as Gij = 200
for
d
 1 4
i = j, and Gii = d , with the exception of G12 and G34
which we set equal to 0 since we assume that a node does not
1
can
transmit and receive at the same time. The factor of 200
be viewed as the spreading gain in a CDMA system, or power
falloff with frequency in a FDMA system. This gives the following gain matrix:

0.2500 0.0003 0.0012 0.0003

0
0.2500 0.0003 0.0012

G = 104
0.0012 0.0003 0.2500 0.0003 (23)
0.0003 0.0012 0
0.2500

C
20m
Fig. 2. Network Topology for Simulation

Using the geometric programming optimization method we

find the maximum aggregate data rate is R = 216.8kbps, with
M = 42.8QAM modulation for each link, Ri = 54.2kbps for
each link, and P1 = P3 = 0.709W and P2 = P4 = 1W link
transmit powers. The resulting SIR = 21.7dB on each link.
The symmetry in modulation levels and SIR is due to the symmetries in the network topology, and not due to any explicit
optimization constraint.

E. Admission control and pricing

In this subsection admission control and a possible approach
to pricing are considered. As discussed in section IV E, a new
user is admissible if his QoS requirements can be supported by
the system without disturbing current users. In this model a
user is admissible if a feasible solution of the problem in formulation 5 exists after the new users QoS constraints have been
added.
In the pricing discussion of section IV, a new user is charged
according to the equivalent number of standard users the new
user costs the system in lost future user capacity. In economics
this approach is termed the opportunity cost associated with
serving a new user.
The concept used for multi-hop networks is similar, but a
different measure of opportunity cost is used. In a multi-hop
network, the route taken by a stream of packets, not just its
QoS, jointly determine its effect on the network. Therefore, the
opportunity cost is taken as the data transport capacity lost by
the entire network in supporting a new user. When a new user
is added to the system the ability of the system to support additional data traffic is reduced. The opportunity cost can be estimated for the multi-hop network by subtracting the maximum
data transport capacity of the system, in bits per second, after
the user is added to the system, from the capacity before the
user is added to the system. The value of the objective function
used in the multi-hop formulation is precisely the maximum
aggregate data rate for the network.

0-7803-7476-2/02/$17.00 (c) 2002 IEEE.

IEEE INFOCOM 2002

F. Pricing Simulation
Consider admission control and pricing for the simulation in
Section VI-D above. Initially the system has no users with QoS
constraints beyond the basic setup given previously. So current
user data is admitted and priced based on a best effort transmission. Three new users U1 , U2 , and U3 are going to arrive
to the system in order. U1 and U2 require 30kbps sent along
the upper path A B D, while U3 requires 10kbps sent
from A B. All three users require the outage probability to
be less than 0.1. When U1 arrives to the system the optimization with his QoS demands has the same solution as without the
demands, so his price is the baseline price. Next, U2 arrives,
and his QoS demands decrease the maximum system throughput from 216.82 kbps to 216.63 kbps, so his price is the baseline price plus an amount proportional to the change in system
throughput. Finally, U3 arrives, and his QoS demands have no
feasible solution, so he is not admitted to the system.
Note that the prices charged are a function of system demands when the user arrives. If U2 had arrived before U1 , U2
would have paid less and U1 more. Similarly, U3 would have
been admitted to the system for the baseline price if he had arrived before U2 , where as his price was effectively infinite when
he arrived after U2 .
VII. R ESOURCE ALLOCATION FOR DELAY AND
EFFICIENCY OPTIMIZATION IN WIRELESS AD HOC
NETWORKS

In this section multi-hop networks are treated from a general

perspective of resource allocation. Resources include power,
the number of flows in each category of service, bandwidth and
capacity of each link. These resources are allocated according
to the optimization criteria of transmission delay, unused capacity and overall system throughput.
A. Problem formulations
Consider a network with J links with capacity of Cj packets
per second for each link j. There are K classes of traffic with
different QoS requirements to be transported over the network.
For each QoS class k, the bandwidth required is bk Hz, and the
delay guarantee in the service level agreement (SLA) is dk,U B
seconds. End to end total delay consists of propagation delay,
transmission delay and queuing delay. Complementary to the
discussion on queuing delay in section V, in this section we
assume that transmission delay is the dominant term for ad hoc
networks. The minimum acceptable probability of delivering
the packet across the unreliable network in the SLA is denoted
by pk,LB .
Similar to the last section, each stream of traffic from source
s to destination d will traverse certain specific links as dictated
by the particular routing protocol used for the network. Denote
by Kj the set of traffic using link j and by Jk the set of links
traversed by QoS class k. Denote by nk the number of packets
dynamically admitted in the k th class of traffic.
In an ad hoc network each link may fail due to a node leaving
the network or due to an outage. pj is defined as the probability
that this link will be maintained during the transmission. Note

that by increasing transmitter power over a link j while keeping other parameters of the network constant, the SIR of link j
can be increased. Consequently, the outage probability of link
j will decrease and pj will increase. Therefore, power control
is reflected through the optimization variable pj . The sixth formulation is the following:
Formulation 6: (SLA feasibility under network constraints)
The following nonlinear problem of testing SLA feasibility is a
convex optimization problem.

minimize
subject to

No

Objective Function
bk nk
kKj 



ni


iKj


jJk
jJk

Cj

pj

bk nk
bk nk
nk
Cj

pj
bk , Cj , pj , dk,U B , pk,LB

Cj , j
dk,U B , k

=
=

pk,LB , k
Rk , k
Cj
dk,j
pj,U B
0
(24)

No objective function is necessary to test feasibility of the

SLA terms pj , dk,U B and pk,LB . Note that the first constraint
is the link capacity constraint, the second one is the delay guarantee constraint and the third one the delivery probability constraint. The fourth constraint delivers a guaranteed data rate to
each class of traffic. The fifth constraint makes room for SLA
terms that give a class of traffic the sole right to traverse a link
j . This could be for bandwidth requirements or for security
reasons as in virtual private networks. The sixth constraint allows for SLA terms that specify not just an end to end total
delay guarantee, but also an exact delay requirement for a particular traffic class k on a link j . The other constraints are
positivity constraints on the variables, and upper bound constraints on pj .
The following parameters can all become variables in the
optimization: bk , nk , pj , Cj , dk,U B and pk,LB . Variables
bk , dk,U B and pk,LB are terms in the SLA. The link capacities Cj and probability of maintaining a link pj are network
resources to be optimized over. Admission control is reflected
in nk .
In the seventh formulation, the unused capacity of a particular link j0 is maximized. This link could be a bottleneck link
or the most often traveled link in the network where capacity is
considered scarce or of great value.
Formulation 7: (Unused capacity maximization) The following nonlinear problem of maximizing the unused capacity
under SLA and network constraints is a convex optimization
problem.

0-7803-7476-2/02/$17.00 (c) 2002 IEEE.

IEEE INFOCOM 2002

Cj0

kKj0 bk nk
b
nk
kKj k


ni


iKj

jJk
jJk

Cj

pj

bk nk
bk nk
nk
Cj

pj
bk , Cj , pj , dk,U B , pk,LB

Cj , j

dk,U B , k

=
=

Formulation 8: (Weighted Joint Capacity and Delay Minimization) The following nonlinear problem of minimizing
transmission delay under SLA and network constraints is a convex optimization problem.

minimize
subject to




ni


iKj
+

C
j
j
Cj

jJk0
bk nk
Cj , j
kKj 



ni


iKj
dk,U B , k
jJk
Cj

pk,LB , k
jJk pj
Rk , k
bk nk
= Cj
bk nk
nk
= dk,j
Cj
pj
pj,U B
bk , Cj , pj , dk,U B , pk,LB 0

(26)
where is the marginal tradeoff of capacity for delay. By increasing capacities available on each link at the relative cost
through dynamic bandwidth allocation or bandwidth leasing,
delay of the most time sensitive QoS class can be decreased.
B. Joint capacity and delay minimization simulation
This simulation investigates the tradeoff between delay and
cost of capacity in Formulation 8. For the network in Fig. 3
there are three classes of traffic. The first class is data traffic
sent along path ABCD requiring a rate of 50 packets/second
and a maximum delay of 0.2 seconds. The second class is also
data traffic sent along path DFEA with the same rate and delay
requirements. The third class of traffic is voice sent along path
ABFD with a rate requirement of 250 packets/second. We want
to minimize both the delay of the voice traffic and the cost of
capacity that we must provision or lease. We accomplish this
by minimizing a weighted sum of the voice traffic delay and the
total capacity used, subject to meeting the rate constraints on all
traffic classes and the delay constraints on the data traffic. For
each value of , the marginal tradeoff parameter between delay and capacity, we find the minimum delay achievable for the

3
8

pk,LB , k
Rk , k
Cj
dk,j
pj,U B
0

(25)
The objective function is to maximize unused capacity of a
link j0 by keeping the used capacity to the minimum under all
network and QoS constraints. The constraints are the same as
in formulation 6.
In the eighth formulation, the total delay for a particular class
of traffic is minimized.

4
E

Fig. 3. Network Topology for Simulation

Minimized Delay for Voice Traffic

maximize
subject to

1.4

1.2

0.8

0.6

0.4

0.2

0
6
10

10

10

10

10

10

,seconds of delay per packet of capacity

10

Fig. 4. Trade off between voice traffic delay and capacity cost

voice traffic given in Fig. 4 with log scale for the x-axis. The
tradeoff curve shows that the minimum delay increases rapidly
with increasing cost of capacity until it reaches the delay associated with the minimum capacity required to support the voice
signal; from that point onwards the tradeoff curve is flat.
C. Extensions
A number of extensions can be made to Formulations 6 to 8.
One extension is minimizing the maximum transmission delay
for users. A second extension is accounting for queueing delay
at the nodes. The formulations, proofs and interpretations are
similar to those in section IV C and section V, respectively.
Another extension is weighted fairness formulations which
can also be solved by geometric programming. The weight
parameters can become variables, therefore fairness and QoS
criteria can be jointly optimized using geometric programming.
This is based on the observation that the posynomial form of the
objective and constraint functions is maintained when weights
of proportional fairness are variables; thus preserving the geometric programming framework. Solving these problems
would give the globally optimal tradeoff between the weights
attached to each user and the resources allocated among the
users.

0-7803-7476-2/02/$17.00 (c) 2002 IEEE.

IEEE INFOCOM 2002

VIII. S IMPLE H EURISTICS

IX. S UMMARY

Although geometric programming has highly efficient interior point algorithms that find the global optimal solution in
polynomial time, in some practical systems, suboptimal simple
heuristics with even lower computational load are desired. In
this section, we briefly outline two heuristics for some of the
formulations. Geometric programming now becomes an efficient tool for heuristic validation and testing.

Various QoS provisioning problems are considered for wireless cellular and ad-hoc networks from a resource allocation
point of view. Such formulations are nonlinear problems, but
can be efficiently solved by convex optimization. The geometric programming framework makes possible formulations that
include both throughput and delay as objective functions and
allow for a variety of general network models. Proportional
and minmax fairness algorithms are also possible. Simulations
show that these algorithms can be used for admission control
and pricing schemes that are based on the relative disturbances
on the use of network resources by the new users. Additionally, two simple heuristics for cellular and ad hoc networks are
outlined.

A. A heuristic for cellular networks

Using the geometric programming formulations in section IV
as an efficient tool of validating heuristics, we find that the following simple heuristic performs well with a small suboptimality gap for Formulation 1. Denote by Dij the total path loss
from the transmitter on link j to the receiver on link i that takes
into account all factors other than power:
ij

Dij =

dij

ij Ks1
ni

(27)

where dij , ij , and ij are the distance, path loss exponent,

and normalization constant from the transmitter on link i to the
receiver on link j, respectively. Denote by Di the mean of Dij
for j = i. The constraint set can then be rewritten as a system
of inequalities in terms of Pi and Di . Let the SIR of user i
be the objective function to be maximized under this constraint
set. For the heuristic, set Pi to its maximum Pi ,U B , and make
all constraints active by turning the system of inequalities into a
system of equalities. Feasibility of the SIR requirements can be
analytically determined, and if feasible, the system of equalities
can easily be solved for Pi , i = i . Substituting the resulting
set of powers into SIRi gives the maximized SIR for user i
under this heuristic. Note that the computational load for this
heuristic is very small.
From empirical results in simulations, the difference in SIR,
for both user i and all the other users, between the convex optimization algorithm and this heuristic is smaller than 5 percent
for networks with sizes larger than 15.
B. A distributed heuristic for ad hoc networks
Extending similar ideas in [11], the following is a simple
iterative heuristic for throughput maximization in ad hoc networks under outage probability constraints as in formulation 5.
Briefly, the heuristic makes the constraints in formulation 5
active, and reexpress
For example, the con each constraint.

1
P rout path s bestraint 1 sS k=s
SIRth Gik Pk
(1+
)
Gii Pi


Pk
) = 1P rout1 path s . Next,
comes sS k=s (1 + SIRGthiiGPik
i
taking the log of both sides and multiplying both sides by Pi ,
the constraints are rewritten as a system of linear equations,
which can be put into matrix form Q(P )P = P . Thus, an initial power vector P (0) can be randomly selected, and the next
iterate of the power vector P (1) can be computed as the Perron Frobenius eigenvector of Q(P (0) ). This heuristic becomes
a sequence of Perron Frobenius eigenvector problems, which is
computationally easy to solve.

ACKNOWLEDGMENTS
We would like to thank Nicholas Bambos and Andrea Goldsmith at Stanford University, and Lizhong Zheng at UC Berkeley for helpful discussions.
R EFERENCES
[1] M. Andersin, Z. Rosberg, and J. Zander, Gradual removals in cellular PCS
with constrained power control and noise. Wireless Networks, 2(1):27-43,
1996.
[2] M. Andersin, Z. Rosberg, and J. Zander, Gradual removals in celllular
PCS with constrained power control and noise. IEEE/ACM Trans. in Networking, 5(2):255-265, 1997.
[3] N. Bambos, Toward power-sensitive network architectures in wireless
communications: Concepts, issues, and design aspects. IEEE Personal
Communications Magazine, 5(3):50-59, 1998.
[4] N. Bambos, S. Chen, and G. Pottie, Radio link admission algorithms for
wireless networks with power control and active link quality protections.
Proc. of the IEEE Infocom, 1995.
[5] S. Boyd and L. Vandenberghe, Convex Optimization and Its Engineering
Applications, Stanford University EE 364 Course Reader 2001.
[6] K. Chawla and X. Qiu, Throughput performance of adaptive modulation
in cellular systems. IEEE Proc. on Universal Personal Communications,
2:945-950, 1998.
[7] M. Chiang, D. ONeill, D. Julian, and S. Boyd, Resource allocation for
QoS provisioning in wireless ad hoc networks. Proc. IEEE Globecom,
San Antonio, TX, Nov. 2001.
[8] G. Foschini and Z. Miljanic, A simple distributed autonomour power control algorithm and its convergence. IEEE Trans. Vehicular Technology,
42(4), 1993.
[9] G. Foschini and Z. Miljanic, Distributed autonomous wireless channel assignment with power control. IEEE Trans. Vehicular Technology,
44(3):420-429, 1995.
[10] D. Julian, M. Chiang, and D. ONeill, Robust and QoS constrained optimization of power control in wireless cellular networks. Proc. IEEE VTC
Fall, Atlantic City, NJ, Oct. 2001.
[11] S. Kandukuri and S. Boyd, Optimal power control in interference limited
fading wireless channels with outage probability specifications. To appear
IEEE JSAC, 2002.
[12] S. Lal and E. Sousa, Distributed resource allocation for DS-CDMA
based multimedia ad hoc wireless LANs. IEEE J. Sel. Areas in Communications, 17(5):947-967, May 1999.
[13] D. Mitra, An asynchronous distributed algorithm for power control in
cellular radio systems. Proceedings of 4th WINLAB Workshop Rutgers
University, NJ, 1993.
[14] Yu. Nesterov and A. Nemirovsky, Interior Point Polynomial Methods in
Convex Programming, SIAM 1994.
[15] X. Qiu and K. Chawla, On the performance of adaptive modulation in
cellular systems. IEEE Transaction on Communication, pp. 884-895, June
1999.
[16] R. Ramanathan and R. Rosales-Hain, Topology control of multihop
wireless networks using transmit power adjustment. Proc. IEEE Infocom,
March 2000.
[17] M.J. Shah and P.G. Flikkema, Power-based leader selection in ad-hoc
wireless networks. IEEE Int. Performance, Computing and Communications Conf., pp. 134-139, 1999.

0-7803-7476-2/02/$17.00 (c) 2002 IEEE.

IEEE INFOCOM 2002

10

[18] S. Ulukus and R. Yates, Adaptive power control and MMSE interference
suppression. ACM/Baltzer Wireless Networks, 999(4), 1998.
[19] R. Yates, A framework for uplink power control in cellular radio systems. IEEE J. Sel. Areas in Comm., 13(7):1341-1347, 1995.

0-7803-7476-2/02/$17.00 (c) 2002 IEEE.