Resource Allocation for QoS Provisioning

in Wireless Ad Hoc Networks

Mung Chiang1 , Daniel ONeill, David Julian2 and Stephen Boyd
Electrical Engineering Department
Stanford University, CA 94305, USA

Abstract- For wireless ad hoc networks with mul- P 1 As a special case of resource allocation, the power con-
tihop transmissions and Rayleigh fading, this paper trol of user nodes are optimized to maximize the overall
maximizes the overall system throughput subject to system throughput.
QoS constraints on power, probability of outage, and P 2 Turning to the general cases, feasibility of service level
data rates. Formulations are also given which mini- agreement (SLA) terms are determined under network
mize delay and optimize network resources in a wire- resource constraints.
less ad hoc network, where each link is shared by mul-
tiple streams of traffic from different QoS classes, and P 3 Taking delay into consideration, the total delay for the
each traffic traverses many links. Although these opti- most time sensitive class of traffic is minimized by opti-
mal resource allocation problems are non-linear, they mizing over powers, capacities, and SLA terms.
can be posed as geometric programs, which are trans- P 4 Optimizing over powers, capacities, and SLA terms; the
formed into convex optimizations, and can be solved unused capacity of the network is maximized.
globally and efficiently through interior-point meth-
ods. Because the mobile radio channel is fast varying and the
number of user nodes is large, a fast and robust decision mak-
ing algorithm is needed that accommodates a large number of
I. Introduction variables for dynamic resource allocation to be feasible. Sev-
eral ad hoc heuristics have been proposed to tackle the above
Quality of service (QoS) has become an important issue in problem, but they cannot meet all of the following criteria:
various kinds of data networks as some users are no longer sat- optimality, speed, and the ability to accommodate a variety
isfied with resource allocation based on service provisioning. of constraints and a large number of variables. Solutions for
Three major considerations of QoS support are bandwidth, wireless cellular networks have been proposed in [1]. This pa-
delay and delivery guarantee. Voice, data, image, and video per tackles the more difficult problems of resource allocation
have different bandwidth requirements. Some classes of traf- for wireless ad hoc networks.
fic, such as voice, are also much more sensitive to delay than
background classes, such as data. QoS provisioning in a wire- A global solution to non-linear problems P 1 to P 4 is found
by transforming the problems into convex optimization prob-
less network is a particularly difficult issue because physical
layer problems; such as path loss, fading, and multipath; can lems. Solution methods for these problems not only produce
make the communication links unreliable. This makes delivery globally optimal solutions as efficiently as for linear program-
s, but also unambiguously determine feasibility. This second
guarantee a necessary feature in wireless ad hoc network QoS
provisioning. property is used to determine the feasibility and pricing scheme
of admitting a new user with a defined QoS requirement.
The challenge is to first prescribe a feasible QoS scheme for
different classes of traffic, and then to optimize the use of net-
work resources, mainly link capacities and transmitter powers,
II. Convex optimization and geometric
to satisfy QoS requirements for all classes while maximizing
programming
either the total network performance, or the QoS for the pre- An efficient algorithm is needed in order to find the opti-
mier class. Within the wireless arena, ad hoc wireless networks mal solution to the above nonlinear problems in a high speed
pose additional technical challenges for QoS support. Unlike dynamic network with a large number of links and nodes. For-
cellular wireless networks, ad hoc networks have no fixed in- mulations for P 1 to P 4 are provided that can be turned into
frastructure, and long range communications require multihop convex optimizations, which have fast algorithms, such as the
transmissions where a packet is routed through the network interior point method and the primal dual method, that make
by other transceivers that act as relay nodes. them as easy to solve as linear programs.
In sections 3 and 4, the following resource allocation prob- Convex optimization refers to minimizing a convex objec-
lems for QoS provisioning in wireless ad hoc networks are tive function over convex constraint sets. The particular type
solved: of convex optimization used in this paper is in the form of geo-
metric programming [2]. First consider the following definition
1 This work was supported by Hertz Foundation Fellowship and
Stanford Graduate Fellowship.
2 This work was supported by NSF Grant CCR-9973134 and the Definition 1 A monomial is a function h : Rn R, where
Stanford Network Research Center. the domain contains all real vectors with positive components:

0-7803-7206-9/01/$17.00 2001 IEEE

2911
 The Rayleigh fading between each transmitter j and receiv-
h(x) = cxa1 1 xa2 2 xann , c 0 and ai R (1) er i is given by Fij . The Fij s are assumed to be independent
P and have unit mean. The Gij s are appropriately scaled to re-
A sum of monomials f (x) = c xa1k xa2 2k xannk is
k k 1 flect variations from this assumption. The distribution of the
called a posynomial. Geometric programming is an optimiza- received power between any pair of transmitter j and receiver
tion problem with the following form: i is exponential with mean value,

minimize f0 (x) E [Gij Fij Pj ] = Gij Pj (5)

subject to fi (x) 1 (2)
hj (x) = 1 The signal to interference ratio (SIR) for user i determines
the quality of the received signal and is defined as
where f0 and fi are posynomials and hj are monomials.
Geometric programming in the above form is not a convex Pi Gii
optimization problem. However, with a change of variables: SIRi = PN (6)
yi = log xi and bik = log cik , the geometric programming form j6=i
Pj Gij + ni
is put into convex form: Unlike SNR, SIR cannot be increased by simply increasing
all users transmitting powers since that would raise both the
P signal level and the interference level. This introduces a bit
minimize p0 (y) = log P k exp(aT0k y + b0k ) error floor and a QoS bottleneck.
subject to pi (y) = log k exp(aTik y + bik ) 0
qj (y) = aTj y + bj = 0
B. Outage probability and system throughput
(3)
It can be verified that the log of sum of exponentials is a convex An outage is declared when the received SIR falls below a
function [2]. Therefore pi are convex functions and qj are affine given threshold defined as SIRth , often computed from a BER
functions, and the problem is a convex optimization problem. requirement. The outage probability associated with the ith
Note that if all posynomials are in fact monomials, geometric hop is given by
programming becomes linear programming.
Oi = P r(SIRi SIRth ) P
Convex optimization problems can be solved globally and (7)
= P r(Gii Fii Pi SIRth k6=i Gik Fik Pk )
efficiently through interior point and primal dual methods [3],
with running times that usually scale to the square root of the The outage probability can be expressed as [4]
problem size. These methods also offer duality interpretations, Q 1
stability analyses and accommodate a variety of constraints. Oi = 1 k6=i 1+ SIRth Gik Pk (8)
G P ii i
This paper shows how geometric programming can solve many
versions of QoS provisioning problems in wireless ad hoc net- Outage probability over a hop induces an outage probability
works. over a path S

OpathS = Prob(outage
Q along the path S)
III. Power control for throughput
= 1 QsS (1
Q Oi ) (9)
optimization 1
= 1 sS k6=s SIRth Gik Pk .
(1+ Gii Pi
)
First a special case of resource allocation optimization is
formulated and solved in this section. The variables are us- The constellation size M used by a hop can be closely ap-
er node powers and the objective is to maximize the overall proximated for MQAM modulation as follows
system throughput in bps. The method in this paper explic-
itly takes into account the statistical variation of the received 1.5
M =1+ SIR (10)
signal and the interference power. ln(5BER)
1.5
where BER is the bit error rate. Defining K = ln(5BER)
leads
A. Multi-hop network model and Rayleigh fading th
to a monotonic expression for the data rate of the i hop as
Consider a wireless ad hoc network with n transmit- a function of the received SIR:
ter/receiver pairs, labeled 1, . . . , n, which transmit at powers
Ri = (1/T ) log2 (1 + KSIRi ) (11)
P1 . . . , Pn . The power received from transmitter j, at receiver
i is given by
The aggregate data rate for the system can then be written
simply as the sum of terms of this form.
Gij Fij Pj (4) P Q
Rsystem = i
Ri = (1/T ) log 2 i
(1 + KSIRi ) (12)
The nonnegative number Gij represents the path gain in
the absence of fading from the j th transmitter to the ith re- Overall system throughput is defined as the maximum aggre-
ceiver. Gij can encompass path loss, shadowing, antenna gain, gate data rate supportable by the system given a set of users
coding gain, and other factors. with defined QoS.

2912
 C. Throughput optimization B

Theorem 1 (Optimize power for throughput maximization) 1 2

The following problem of optimizing user node powers to max-

imize total network throughput is a convex optimization prob- A 20m
D
lem.

maximize Rsystem 3 4

C
subject to
20m
Ri Q Ri,LB , i
1
1 k6=i SIRth Gik Pk P routi i Fig. 1: Network Topology for Simulation
1+
Q Q Gii Pi
1
1 sS k6=s (1+ SIRth Gik Pk )
P rout path s S
G P ii i
Pi Pmax The maximum aggregate data rate, found using the geo-
(13) metric programming optimization method, is R = 216.8kbps,
The objective function is the overall system throughput. It with M = 42.8QAM modulation for each link, Ri = 54.2kbps
is optimized over the set of all feasible powers Pi . The first set for each link, and P1 = P3 = 0.709W and P2 = P4 = 1W
of constraints are the data rates demanded by existing system link transmit powers. The resulting SIR = 21.7dB on each
users. The second set of constraints are the outage probability link. The symmetry in modulation levels and SIR is due to
limitations demanded by users using single hops. The third set the symmetries in the network topology, and not due to any
of constraints are the outage probability limitations for users explicit optimization constraint.
using a multi-hop path. Lastly, the forth set of constraints are
regulatory or system limitations on transmitter powers. E. Admission control and pricing

The methodology used in this section can also be extended

D. Simulation
for other resource management purposes in a wireless ad hoc
A simple four node multi-hop network is considered in the network. For example, a new user is admissible if his QoS re-
following simulation. As shown in figure 1, the network con- quirements can be supported by the system without disturbing
sists of 4 nodes A, B, C, and D, and 4 links 1, 2, 3, and 4. On the existing QoS requirements of current users. In this mod-
link 1 node A is the transmitter and node B is the receiver; el a user is admissible if a feasible solution of the problem in
similarly, the transmitter and receiver nodes for each link are Theorem 1 exists after the new users QoS constraints have
shown in the figure. Note that node A is also the transmitter been added. An infeasible solution is a definitive statement
on link 3, illustrating that a node can be a transmitter and/or that this new user may not be added to the system without a
receiver on many links. Nodes A and D as well as B and C change in required QoS.
are separated by a distanceof 20m. By geometry the distance Our model also supports several service-pricing approach-
of each transmit path is 10 2m. es. The decrease in the overall system throughput associated
For this simulation each link has a maximum transmit pow- with a new user can be used to estimate the incremental cost
er of 1W. Alternatively, the power constraint could be placed of supporting that new user, and should be proportional to the
on each node instead of each link by adding a constraint that price that might be charged. A second approach is to calcu-
P1 + P3 1W. All nodes are using MQAM modulation. The late the effect of different levels of QoS for a given data rate.
baseband bandwidth for each link is 10kHz, the minimum da- The effect on overall system throughput for each of the chosen
ta rate for each link is 100bps for maintenance data, and the different levels of QoS can then be used to set relative pricing
target BER is 103 . For the Rayleigh fading a probability of for these levels. One example is to set pricing as a function of
outage of Pout = 0.1 is required for an SIR threshold of 10dB. probability of outage for a given data rate.
1
 1 4
The gains for each link are computed as Gij = 200 for
 4 d
IV. Resource allocation for delay and
i 6= j, and Gii = d1 , with the exception of G12 and G34 efficiency optimization
which are set equal to 0 since it is assumed that a node does
1
not transmit and receive at the same time. The factor of 200 The more general cases of resource allocation are formulat-
can be viewed as the spreading gain in a CDMA system, or ed and solved in this section. Optimization variables include
the power falloff with frequency in a FDMA system. This gives powers, the number of packets in each traffic, bandwidth, de-
the following gain matrix: lay and delivery guarantee required for each QoS class, and
capacity for each link. Potential objective criteria include de-
0.2500 0.0003 0.0012 0.0003 lay, unused capacity and SLA feasibility, in addition to the
4 0 0.2500 0.0003 0.0012 overall system throughput in the last section.
G = 10
0.0003
(14)
0.0012 0.0003 0.2500
0.0003 0.0012 0 0.2500 A. Problem formulations

2913
 Consider a network with J links with capacity of Cj pack- constraint, the second one is the delay guarantee constraint
ets per second for each link j. There are K classes of traffic and the third one the delivery probability constraint. The
with different QoS requirements to be transported over the fourth constraint delivers a guaranteed data rate to each class
network. For each QoS class k, the bandwidth required is bk of traffic. The fifth constraint makes room for SLA terms that
Hz, and the delay guarantee in the service level agreement give a class of traffic the sole right to traverse a link j . This
(SLA) is dk,U B seconds. Also, a minimum probability of de- could be for bandwidth requirements or for security reasons.
livering the packet across the unreliable network is required in The sixth constraint allows for SLA terms that specify not just
the SLA, denoted by pk,LB . In this problem formulation the an end to end total delay guarantee, but also an exact delay
delay is the delay due to transmission time; propagation delay requirement for a particular traffic class k on a link j . The
is ignored because it is constant for the optimization parame- other constraints are positivity constraints on the variables,
ters. The more refined model with queueing delay is treated and upper bound constraints on pj .
in the extension. The following parameters are all potential optimization
Similar to the last section, each stream of traffic from variables: bk , nk , pj , Cj , dk,U B and pk,LB . Variables bk , dk,U B
source s to destination d will traverse certain specific links as and pk,LB are terms in the SLA. The link capacities Cj and
dictated by the particular routing protocol used for the net- probability of maintaining a link pj are network resources to
work. Denote by Kj the set of traffic using link j and by Jk be optimized over. Admission control is reflected in nk .
the set of links traversed by QoS class k. Denote by nk the The formulation is a non-linear optimization problem be-
number of packets dynamically admitted in the kth class of cause optimization variables are multiplied together, such as
traffic. bk nk in the first constraint or the product of pj in the third
In an ad hoc network each link may fail due to either pow- constraint, and appear in the denominator, such as Cj in the
er shut down of a user or deep fading that causes an outage. second and sixth constraints. However, all the inequality con-
Therefore pj , a real number between 0 and 1, is attached to straints are in posynomial form and all the equality constraints
each link as the probability that this link will be maintained are in monomial form.
during the transmission. By increasing transmitter power over In the second formulation, the unused capacity of a partic-
a link j while keeping other parameters of the network con- ular link j0 is maximized. The link may be a bottleneck link,
stant, SIR of link j and therefore pj can be increased. or the most often traversed link in the network. [5].
As will be shown in the following problem formulations,
with the above constraints on link capacity, bandwidth require-
ment, delay, and delivery probability guarantees, the problem Theorem 3 (Unused capacity maximization) The following
is not a linear programming problem. However, these non- problem of maximizing the unused capacity under SLA and
linear optimizations can be turned into geometric program- network constraints is a convex optimization problem.
ming problems and solved as efficiently as linear programs.
The first formulation is the following.

Theorem 2 (SLA feasibility under network constraints) The P

maximize Cj0 bk nk
following problem of testing SLA feasibility is a convex opti- P kKj0
subject to bk nk Cj , j
mization problem. kKj
P !
P iKj
ni

jJk Cj
dk,U B , k
Q
minimize No pj pk,LB , k
PObjective Function jJk
bk nk Rk , k
subject to bk nk Cj , j
kK j P ! bk nk = Cj
P iKj
ni nk
= dk,j
jJk Cj
dk,U B , k Cj

Q pj pj,U B
jJk
pj pk,LB , k bk , Cj , pj , dk,U B , pk,LB 0
bk nk Rk , k (16)
bk nk = Cj The objective function is to maximize unused capacity of a
dk,j
nk
Cj
= link j0 by keeping the used capacity to the minimum under all
pj pj,U B network and QoS constraints. The constraints are the same as
bk , Cj , pj , dk,U B , pk,LB 0 in Theorem 2.
(15) In the third formulation, the total delay for a particular
No objective function is necessary since in this formulation class of traffic is minimized.
only the feasibility of the SLA terms pj , dk,U B and pk,LB is
being tested. Alternatively, a cost function as the objective
function could be used for relative pricing during admission Theorem 4 (Delay Minimization) Delay minimization under
control. Note that the first constraint is the link capacity SLA and network constraints is a convex optimization problem.

2914
 2
B C
P
P ni P  1 3
iKj
8
minimize + Cj 7
P jJk0 Cj j
subject to bk nk Cj , j A D
kKj
P !
P iKj
ni
6
jJk Cj
dk,U B , k 4

Q E F
jJk
pj pk,LB , k 5

bk nk Rk , k
bk nk = Cj Fig. 2: Network Topology for Simulation
dk,j
nk
Cj
=
pj pj,U B 1.4

Minimized Delay for Video Traffic

bk , Cj , pj , dk,U B , pk,LB 0 1.2

(17)
1

Where is the marginal tradeoff of capacity for delay. By

0.8

increasing capacities available on each link at the relative cost

through bandwidth allocation or bandwidth leasing, delay of 0.6

the most time sensitive QoS class can be decreased. Therefore, 0.4

a weighted sum, parametrized by , of the premier QoS classs 0.2

delay and the cost of capacity provisioning is maximized.

0
6 5 4 3 2 1 0
10 10 10 10 10 10 10

, seconds of delay per packet of capacity

B. Simulation
Fig. 3: Trade off between video traffic delay and capacity cost
The following is a simulation for Theorem 4, which investigates
the tradeoff between delay and cost of capacity.
For the network in Fig. 2 there are three classes of traffic. Important performance metrics are optimized, such as
The first class is audio data traffic sent along path ABCD re- throughput and delay, subject to a variety of realistic con-
quiring a rate of 50 packets/second and a maximum delay of straints on both SLA terms and network resources. This is
0.2 seconds. The second class is also audio data traffic sent done for wireless ad hoc networks with multihop transmission-
along path DFEA with the same rate and delay requirements. s, mutual interference, and intrinsic unreliability of links and
The third class of traffic is video data sent along path ABFD nodes. Although these resource allocation optimizations are
with a rate requirement of 250 packets/second. The goal is non-linear problems, they can be solved efficiently using fast
to minimize both the delay of the video data and the cost of convex optimization algorithms.
capacity that must be provisioned or leased. This is accom-
plished by minimizing a weighted sum of the video data delay
Acknowledgments
and the total capacity used subject to meeting the rate con-
straints on all traffic classes, and the delay constraints on the We would like to thank Nicholas Bambos and Andrea Gold-
audio data traffic. For each , the marginal tradeoff value smith for discussions of wireless network problems.
between delay and capacity, Fig. 3 shows the minimum delay
achievable for the video traffic given. The x-axis uses a log
References
scale. The tradeoff curve shows that the minimum delay in-
creases rapidly with increasing cost of capacity until it reaches [1] D. Julian, M. Chiang and D. ONeill, Constrained power con-
the delay associated with the minimum capacity required to trol optimization for wireless cellular networks, to appear in
IEEE Vehicular Technology Conference, Oct. 2001.
support the video signal; from that point onwards the tradeoff
curve is flat. [2] S. Boyd and L. Vandenberghe, Convex Optimization and Its
Engineering Applications, Stanford University EE 364 Course
Reader 1999.
C. Extensions [3] Yu. Nesterov and A. Nemirovsky, Interior Point Polynomial
Methods in Convex Programming, SIAM 1994.
Some extensions to Theorems 2 to 4 include minimizing
[4] S. Kandukuri and S. Boyd, Optimal power control in inter-
the maximum delay and accounting for queueing delay at the ference limited fading wireless channels with outage probability
nodes. Minimizing the maximum delay can be accomplished specifications. To appear IEEE JSAC, 2001.
by minimizing a dummy variable subject to all delays less than
[5] D. ONeill, D. Julian, M. Chiang and S. Boyd, Resource alloca-
that dummy variable. Queueing delay in a store and forward tion optimization for QoS provisioning in wireless cellular and
network can also be accounted for in the geometric program- ad hoc networks, Submitted to IEEE Transaction on Commu-
ming construction [5]. nication, 2001.

V. Summary

2915