Escolar Documentos
Profissional Documentos
Cultura Documentos
PAOLO SANTI
Istituto di Informatica e Telematica
Topology Control (TC) is one of the most important techniques used in wireless ad hoc
and sensor networks to reduce energy consumption (which is essential to extend the
network operational time) and radio interference (with a positive effect on the network
traffic carrying capacity). The goal of this technique is to control the topology of the
graph representing the communication links between network nodes with the purpose
of maintaining some global graph property (e.g., connectivity), while reducing energy
consumption and/or interference that are strictly related to the nodes’ transmitting
range. In this article, we state several problems related to topology control in wireless
ad hoc and sensor networks, and we survey state-of-the-art solutions which have been
proposed to tackle them. We also outline several directions for further research which
we hope will motivate researchers to undertake additional studies in this field.
Author’s address: Istituto di Informatica e Telematica del CNR, Area della Ricerca, Via G. Moruzzi 1, 56124
Pisa, Italy; email: paolo.santi@iit.cnr.it.
Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted
without fee provided that copies are not made or distributed for profit or direct commercial advantage and
that copies show this notice on the first page or initial screen of a display along with the full citation.
Copyrights for components of this work owned by others than ACM must be honored. Abstracting with
credit is permitted. To copy otherwise, to republish, to post on servers, to redistribute to lists, or to use any
component of this work in other works requires prior specific permission and/or a fee. Permissions may be
requested from Publications Dept., ACM, Inc., 1515 Broadway, New York, NY 10036 USA, fax: +1 (212)
869-0481, or permissions@acm.org.
2005
c ACM 0360-0300/05/0600-0164 $5.00
ACM Computing Surveys, Vol. 37, No. 2, June 2005, pp. 164–194.
Topology Control in Wireless Ad Hoc and Sensor Networks 165
events (expos, concerts, etc.), or in hostile et al. [2002], Pottie and Kaiser [2000],
environments. Sadler et al. [2004], Schwiebert et al.
Wireless sensor networks (WSNs) are [2001], Srivastava et al. [2001], Steere
a special class of ad hoc networks. In et al. [2000], and Szewczyk et al. [2004].
a WSN, the interconnected units are The following aspects that have to be
battery-operated microsensors, each of carefully taken into account in the de-
which is integrated in a single package sign stage are peculiar to wireless ad hoc
with low-power signal processing, compu- networks.
tation, and a wireless transceiver. Sensor
nodes collect the data of interest (e.g., — Energy conservation. Contrary to the
temperature, pressure, soil makeup, etc.), case of wired networks, units in ad hoc
and transmit them, possibly compressed networks are typically equipped with
and/or aggregated with those of neigh- limited energy supplies. Hence, one of
boring nodes, to the other nodes. In this the primary goals of the design is to
way, every node in the network acquires use this limited energy as efficiently
a global view of the monitored area that as possible. Energy efficiency is espe-
can be accessed by the external user con- cially important in WSNs where re-
nected to the WSN through one or more placing/refilling sensor batteries is, in
gateway nodes (see Figure 1). Potential general, infeasible. If energy conserva-
applications of sensor networks abound; tion techniques are used at different
they can be used to monitor remote and/or levels of the wireless architecture, the
hostile geographical regions, to trace ani- functional lifetime of both individual
mals movement, to improve weather fore- units and the network can be extended
cast, and so on. Examples of scenarios considerably.
where WSN can be used are described in — Limited bandwidth. Typically wire-
Estrin et al. [1999], Heinzelman et al. less multihop networks are character-
[1999], Khan et al. [2000], Mainwaring ized by a limited bandwidth available
to the nodes. Although the theoreti- nication quality, sensed data must be
cal bandwidth in industrial standards compressed and/or aggregated with data
such as IEEE 802.11 can be as high of neighboring sensors before sending
as 54Mb/sec [IEEE 1999], the situa- them to the gateway node(s).
tion is far worse in practical situations —Scalability. Depending on the scenario
mainly because of the radio interfer- considered, WSNs might be composed of
ence caused by simultaneous commu- several thousands of sensors. Thus, the
nications. Thus, a major problem in the scalability of the proposed protocols is
design of ad hoc networks is to keep the an important issue.
network traffic carrying capacity at a
reasonable level even in the presence Several solutions have been proposed
of dense node deployments. in the literature that address at least
— Unstructured and time-varying net- some of the issues raised above. In par-
work topology. Nodes in the network ticular, great efforts have been devoted
may, in principle, be arbitrarily placed to the design of energy-efficient and
in the deployment region; hence, the mobility-resilient routing, broadcast, and
graph representing the communication multicast protocols [Basagni et al. 1999;
links between the nodes is usually un- Gerla and Tsai 1995; Ko and Vaidya
structured. Furthermore, due to node 1998; Michail and Ephremides 2003;
mobility and/or failure, the network Murthy and Garcia-Luna-Aceves 1996;
topology may vary with time. As a con- Papadimitriou and Georgiadis 2004;
sequence, determining the appropri- Rajaraman 2002; Seada et al. 2004].
ate value of fundamental network pa- Routing and broadcast protocols are
rameters (e.g., the critical transmitting usually concerned with energy-efficient
range for connectivity, see Section 5.1) message delivery on a given communica-
is a difficult task. tion graph which is considered as an in-
put to the protocol. However, contrary to
— Low-quality communications. Commu-
the case of wired networks, the network
nication on wireless channels is, in gen-
topology in wireless networks is not fixed
eral, much less reliable than in wired
and can be changed by varying the nodes’
channels. Furthermore, the quality of
transmitting range. So, further energy can
communication is strongly influenced
be saved if the network topology used
by environmental factors which can be
to route/broadcast messages is energy-
time-varying. Considering that ad hoc
efficient itself. The goal of topology control
networks, and especially WSNs, are
is to dynamically change the nodes’ trans-
likely to be deployed in hostile environ-
mitting range in order to maintain some
ments, low communication quality is to
property of the communication graph (e.g.,
be expected in general, with nonnegli-
connectivity), while reducing the energy
gible off-service time intervals.
consumed by node transceivers (which is
In the case of WSNs, the following as- strictly related to the transmitting range).
pects must also be considered. Since transceivers are one of the pri-
mary sources of energy consumption in the
—Operation in hostile environments. In wireless unit, especially in WSNs, topol-
many scenarios, WSNs are expected to ogy control mechanisms are fundamen-
operate in hostile environments so sen- tal to achieving a good network energy
sors must be explicitly designed to work efficiency.
under extreme conditions which may Besides reducing energy consumption,
make individual unit failure a likely topology control has the positive effect of
event. Hence, resilience to sensor faults reducing contention when accessing the
must be explicitly addressed at different wireless channel. In general, when the
network layers. nodes’ transmitting ranges are relatively
—Data processing. Given the energy con- short, many nodes can transmit simul-
straints and the expected poor commu- taneously without interfering with each
other, and the network capacity is in- approaches to the topology control prob-
creased. Ideally, the nodes’ transmitting lem that has appeared in the literature. In
range should be set to the minimum value Section 4, we review the probabilistic the-
such that the graph that represents the ories that have been used in the derivation
communication links between units is con- of theoretical results concerning topology
nected. How to compute this value under control. In Section 5, we introduce several
different hypotheses on the initial node problems related to topology control in sta-
distribution, presence, and type of mobil- tionary networks, and we survey state-
ity, and so on, is the subject of this survey. of-the-art solutions which have been pro-
Before proceeding, some observations posed to tackle them. In Section 6, we
regarding terminology are in order. The will discuss how node mobility affects the
term topology control has been used with picture drawn in Section 5. Finally, in
at least two different meanings in the ad Section 7, we outline several directions for
hoc and sensor networks literature. For further research.
instance, several authors consider as
topology control techniques aimed at su-
2. A WIRELESS AD HOC NETWORK MODEL
perimposing a hierarchy on an otherwise
flat network organization in order to re- In this section, we introduce a simplified
duce, typically, energy consumption. This but widely accepted model of a wireless ad
is the case, for instance, with clustering hoc network which will be used in the def-
algorithms which select some of the nodes inition of the various problems related to
in the network as clusterheads whose pur- topology control considered in the litera-
pose is to optimize energy and communica- ture.
tion efficiency in their cluster. Although, in The node configuration of a
a sense, clustering algorithms can be seen d -dimensional mobile wireless ad hoc
as a way of controlling the network topol- network with d = 1, 2, 3, is represented
ogy, they cannot be classified as topology by a pair M d = (N , P ), where N is
control mechanisms according to the infor- the set of nodes, with |N | = n, and
mal definition previously presented since P : N × T → [0, l ]d , for some l > 0, is the
the transmit power of the nodes is usually placement function. The placement func-
not modified by a clustering algorithm. tion assigns to every element of N and to
Also, the terms power control and topol- any time t ∈ T a set of coordinates in the
ogy control are often confused with each d -dimensional cube of side l , representing
other in the current literature. In our view, the node’s physical position at time t. The
we classify as power control those tech- choice of limiting the admissible physical
niques that, by acting on the transmit placement of nodes to a bounded region
power level of the nodes, aim at optimizing of Rd of the form [0, l ]d , for some l > 0,
a single wireless transmission. Although is realistic and eases the treatment of
this transmission might, in general, be some of the problems considered in the
multihop, the focus of power control is on following.
the efficiency of a single (possibly multi- Node i ∈ N is said to be stationary if
hop) wireless channel. Again, this feature its physical placement does not vary with
of power control does not fulfill our infor- time. If all the nodes are stationary, the
mal definition of topology control in which network is said to be stationary, and func-
nodes adjust their transmitting range in tion P can be represented simply as P :
order to achieve a certain network-wide N → [0, l ]d .
target goal (e.g., network connectivity). A range assignment for a d -dimensional
The rest of this article is organized as node configuration M d = (N , P ) is a func-
follows. In Section 2, we introduce a sim- tion R A : N → (0, rmax ] that assigns to
plified but widely accepted model of a wire- every element of N a value in (0, rmax ],
less ad hoc network which will be used representing its transmitting range.
in the rest of the article. In Section 3, we Parameter rmax is called the maximum
propose a taxonomy to classify the many transmitting range of the nodes in the
network and depends on the features G t = (N , E(t)), where the directed edge
of the radio transceivers equipping the [i, j ] exists if and only if RA(i) ≥ δ P (i,t), P ( j,t) .
mobile nodes. A common assumption In other words, the directed edge [i, j ] ex-
is that all the nodes are equipped with ists if and only if nodes i and j are at
transceivers having the same features; a distance of at most RA(i) at time t. In
hence, we have a single value of rmax for this case, node j is said to be a neigh-
all the nodes in the network. bor of i. A range assignment RA is said
It is known [Rappaport 2002] that the to be connecting at time t if the resulting
power pi required by node i to correctly communication graph at time t is strongly
transmit data to node j must satisfy in- connected.2 If the network is stationary,
equality we simply say that the range assignment
pi RA is connecting. A range assignment in
≥β, (1) which all the nodes have the same trans-
δi,α j
mitting range r, for some 0 < r ≤ rmax , is
where α ≥ 2 is the distance-power gradi- called r-homogeneous range assignment.3
ent, β ≥ 1 is the transmission quality pa- Observe that the communication graph
rameter, and δi, j is the Euclidean distance generated by a homogeneous range as-
between the nodes. While the value of β signment can be considered as undirected
is usually set to 1, the value of α depends since [i, j ] ∈ E(t) ⇔ [ j, i] ∈ E(t).
on environmental conditions. In the ideal In general, the range assignment may
case, we have α = 2; however, α is typi- vary with time in order to ensure tar-
cally 4 in realistic situations. A value of α get properties (e.g., strong connectivity, a
in the interval [2, 6] is commonly accepted. given network diameter h < n) of the com-
Given the previous formula, we can define munication graph. Hence, a sequence of
the energy cost
of a range assignment RA range assignments RAt1 , RAt2 , . . . can be
as c(RA) = i∈N (RA(i))α . defined, where RAti is the range assign-
Formula (1) holds for free-space envi- ment at time ti , and the transition between
ronments with nonobstructed line of sight, range assignments is determined by the
and it does not consider the possible occur- topology control mechanism.
rence of reflections, scattering, and diffrac- The communication graph as defined
tion caused by buildings, terrain, and so here is essentially the point graph model
on. Although more complicated formulas introduced in Sen and Huson [1996], but
of the radio signal attenuation with dis- it is more often called the unit disk graph
tance are known, such as that recently de- model in the topology control (TC) litera-
rived in Bruck et al. [2002], Inequality (1) ture. If node positions are chosen accord-
is widely accepted in the ad hoc network ing to some probability distribution, the
community. point graph model coincides with the con-
Note that Inequality (1) accounts for cept of Random Geometric Graph (RGG)
only the power consumed by the sender which is a generalization of the notion of
node (transmit power). In practice, in Random Graph introduced in the applied
a radio communication, a nonnegligible probability community (see Section 4 for
amount of energy is also consumed at details).
the receiver node to receive and decode The main weakness of the point graph
the transmitted signal. Most current lit- model is the assumption that the radio
erature does not account for the receiver coverage area is a perfect circle. This as-
energy, and the design of topology control sumption is quite realistic in open-air flat
protocols based on more realistic energy
models is one of the main open issues in
2 A directed graph G = (N , E) is strongly connected
the field (see Section 7).
if and only if, for any two nodes u, v ∈ N , there exists
Given a node configuration M d = (N , P ) a directed path from u to v in G.
and a range assignment RA, the commu- 3 When the value of r is not relevant, the r-
nication graph induced by RA on M d at homogeneous range assignment is simply called the
time t is defined as the directed graph homogeneous range assignment.
infinite-order component with high prob- conditions all the cells are filled with at
ability. However, the existence of an least one node (ball). This technique has
infinite-order component is not sufficient been used in Santi and Blough [2003,
to ensure the connectivity of the network. 2002].
In fact, there could exist (infinitely many)
nodes which do not belong to the giant
5. STATIONARY NETWORKS
component, thus leading to a disconnected
communication graph. Hence, the quality In this Section, we will consider several
of connectivity is related to the fraction θ problems related to topology control in sta-
of nodes belonging to the giant component tionary ad hoc networks. The generaliza-
[Janson et al. 1993] which, in turn, depend tion of some of these problems to the more
on the percolation probability. The perco- complicated scenario of mobile networks
lation probability is the probability that is presented in Section 6.
an arbitrary node belongs to a connected
component of infinite order. The main re- 5.1. Homogeneous Topology Control
sult of the theory of continuum percola-
tion is that there exists a finite, positive First, we consider the following prob-
value λc of λ, called critical density under lem concerning homogeneous range
which the percolation probability is zero assignments:
and above which it is nonzero. However,
Definition 5.1 CTR (Critical Transmit-
no explicit expression of the percolation
ting Range). Suppose n nodes are placed
probability is known to date. The theory
in R = [0, l ]d , with d = 1, 2, 3. What is
of continuum percolation have been used
the minimum value of r such that the r-
in Dousse et al. [2002], and Gupta and
homogeneous range assignment for this
Kumar [1998] to analyze the connectivity
placement is connecting?
of ad hoc networks.
In the occupancy theory [Kolchin et al. The minimum value of r such that the
1978], it is assumed that n balls are r-homogeneous range assignment is con-
thrown independently, at random, into C necting is known as the critical trans-
cells. The allocation of balls into cells can mitting range for connectivity in the
be characterized by means of random vari- literature.
ables describing some property of the cells. The motivation for studying CTR stems
The occupancy theory is aimed at deter- from the fact that, in many situations
mining the probability distribution of such the dynamically-adjusting node transmit-
variables as n and C grow to infinity (i.e., ting range is not feasible. In fact, in-
the limit distribution). The most studied expensive radio transceivers might not
random variable is the number of empty allow the transmission range to be ad-
cells after all the balls have been thrown, justed [Ramanathan and Rosales-Hain
which we denote µ(n, C). Of course, the 2000]. In this scenario, setting the same
limit distribution of µ(n, C) depends on the transmitting range r for all the units
relative magnitude of n and C, that is, on is a reasonable choice, and the only op-
the asymptotic behavior of ρ = n/C. De- tion to reduce power consumption and
pending on the asymptotic behavior of ρ, increase network capacity is to set r to
five domains such that n, C → ∞ for which the minimum possible value that ensures
the limit distribution of ρ(n, C) is differ- connectivity.
ent have been determined. Depending on Characterizing the critical transmitting
the domain, the limit distribution can be range helps the system designer to an-
either Poisson or Normal with different swer fundamental questions such as given
parameters. The occupancy theory can be a number of nodes n to be deployed in a
used to analyze connectivity in ad hoc region R, what is the minimum value of
networks by subdividing the deployment the transmitting range that ensures net-
region R into equal subregions (cells) of work connectivity? Conversely, for a given
size ≈ r d and by determining under which transmitter technology, how many nodes
Although interesting, the theory of GRG The critical transmitting range for
can be used only to derive results con- connectivity in sparse ad hoc networks
cerning dense ad hoc networks. In fact, have been analyzed in Santi et al. [2001]
a standard assumption in this theory is and Santi and Blough [2003, 2002] using
that the deployment region R is fixed, and the occupancy theory. It has been proven
the asymptotic behavior of r as n grows to that, under the assumption that n nodes
infinity is investigated, that is, the node are distributed uniformly at random in
density is assumed to grow to infinity. A R = [0, l ]d , the r-homogeneous range
similar limitation applies to the model of
assignment is connecting with high prob-
Gupta and Kumar [1998]. In their case, R ability if r = l d c logn l for some constant
is the disk of unit area, and the authors c > 0. The authors also prove that, if
show that if the units’ transmitting range
r ∈ O(l d n1 ), then the r-homogeneous
is set to r = log n+c(n)
πn
, then the resulting
range assignment is not connected with
network is connected with high probabil-
high probability.
ity if and only if c(n) → ∞. This result is
obtained making use of the theory of
continuum percolation Meester and Roy
[1996] which is also used in Dousse et al. 5.1.3. More Practical Characterizations of
[2002] to investigate the connectivity of the CTR. Besides analytical characteriza-
hybrid ad hoc networks in which base sta- tion, the critical transmitting range has
tions can be used to improve connectivity. been investigated from a more practi-
cal viewpoint. In Narayanaswamy et al.
[2002], the authors present a distributed
5.1.2. Sparse Networks. Given the pre- protocol, called COMPOW, that attempts
ceding discussion, the applicability of the- to determine the minimum common trans-
oretical results concerning connectivity mitting range needed to ensure network
in ad hoc networks to realistic scenarios connectivity. They show that setting the
could be impaired. In fact, it is known transmitting range to this value has
that real wireless networks cannot be too the beneficial effects of maximizing net-
dense, due to the problem of spatial reuse: work capacity, reducing the contention to
when a node is transmitting, it interferes access the wireless channel, and mini-
with all the nodes within its interference mizing energy consumption. Bettstetter
range which is typically larger than the [2002a] analyzes network connectivity un-
transmitting range. If the node density is der the assumption that some of the nodes
very high, the level of interference is very have transmitting range r1 , and the re-
high as well, and the overall network ca- maining ones have transmitting range
pacity is compromised [Gupta and Kumar r2
= r1 . Santi and Blough [2003] in-
2000]. vestigate through simulation the trade-
In order to circumvent this problem, off between the transmitting range and
other authors have characterized the criti- the size of the largest connected compo-
cal transmitting range in the more general nent in the communication graph. The
model in which the side l of the deploy- experimental results presented in Santi
ment region is a further parameter, and and Blough [2003] show that, in sparse
n and r can be arbitrary functions of l . In two and three-dimensional networks, the
this case, the critical transmitting range is transmitting range can be reduced signif-
analyzed asymptotically as l → ∞. Note icantly if weaker requirements on connec-
that, using this model, the node density tivity are acceptable: halving the critical
might either converge to 0 or to a con- transmitting range, the largest connected
stant c > 0, or diverge as the size of the component has an average size of ap-
deployment region grows to infinity. Thus, proximately 0.9n. This means that a con-
results based on this framework can be siderable amount of energy is spent to
applied to dense as well as sparse ad hoc connect relatively few nodes. This be-
networks. havior is not displayed in the case of
Network coverage is defined as follows: ev- time (more specifically, in time O(n4 )) in
ery node covers a circular area of radius rc , the one-dimensional case (i.e., nodes in a
and the monitored area R is covered if ev- line), while it is shown to be NP-hard in
ery point of R is at a distance of at most the case of three-dimensional networks. In
rc from at least one node. The goal is to a later paper, Clementi et al. [1999] have
find the critical value of rc that ensures shown that RA is NP-hard also in the two-
coverage with high probability. This prob- dimensional case. Thus, computing the
lem has been investigated in Philips et al. optimal range assignment in two and
[1989] for the case of nodes distributed in three-dimensional networks is a virtually
a square with a side of length l according impossible task. However, the optimal so-
to a Poisson process of fixed density. The lution can be approximated within a factor
critical transmitting and coverage range of 2 using the range assignment generated
for Poisson distributed points on a line of as follows [Kirousis et al. 2000]. Let T be
length l is derived in Piret [1991]. the MST built on N , where the weight of
edge (ui , u j ) is the power δuαi ,u j needed to
5.2. Nonhomogeneous Topology Control transmit a message between ui and u j ; for
every node ui ∈ N , define RA(ui ) as the
In the previous Section, we have analyzed maximum of distances δui ,u j , for all nodes
the problem of determining a minimum u j which are neighbors of ui in T . In the
common value of the transmitting range following, we will denote this range as-
that generates a connected communica- signment with RAMST .
tion graph under the hypothesis that only Several variants of RA have been consid-
probabilistic information about node po- ered in the literature. In Clementi et al.
sitions is available. In this Section, we [1999, 2000a, 2000b] and Kirousis et al.
survey the considerable body of results ob- [2000], the focus is on a constrained ver-
tained for the more general problem in sion of RA in which the additional require-
which nodes are allowed to have differ- ment of having a communication graph
ent transmitting ranges. As in the case with diameter at most h, for some constant
of homogeneous topology control, in this h < n, is imposed on the communication
Section, we only report results concern- graph. However, we believe this version
ing the stationary case. Nonhomogeneous of the problem is less interesting from a
topology control techniques for mobile net- practical point of view. In fact, imposing
works will be discussed in Section 6. a topology which is too connected would
often cause communication interference
5.2.1. The Range Assignment Problem. to occur even between nodes that are far
The problem of assigning a transmit- apart, thus decreasing the network capac-
ting range to nodes in such a way that ity. This phenomenon is confirmed by the-
the resulting communication graph is oretical as well as experimental results
strongly connected and the energy cost is [Grossglauser and Tse 2001; Gupta and
minimum is called the range assignment Kumar 2000; Li et al. 2001] which show
problem (RA), and it was first studied in that the communication graph in wireless
Kirousis et al. [2000]. More formally, the ad hoc networks should be as sparse as
problem is defined as follows. possible, while preserving connectivity.
Definition 5.2 RA. Let N = {u1 , . . . , un } Two important variants of RA which
be a set of points in the d -dimensional have been recently studied are based
space (d = 1, 2, 3), denoting the positions on the concept of symmetry of the
of the network nodes. Determine a con- communication graph. In general, the
necting range assignment RA such that communication graph generated by a
range assignment is not symmetric,
c(RA) = ui ∈N (RA(ui ))α is minimum.
that is, it might contain unidirectional
The computational complexity of RA has links. Although implementing wireless
been analyzed in Kirousis et al. [2000]. unidirectional links is technically feasible
The problem is solvable in polynomial (see Bao and Garcia-Luna-Aceves [2001],
Calinescu et al. [2002] introduce two puting topologies which have energy-
polynomial approximation algorithms for efficient paths between potential source-
WSRA which improve on the approxima- destination pairs. More specifically, the
tion ratio of 2, previously known.5 The first following problem has been considered
algorithm has an approximation ratio of (see Li et al. [2002] and Rajaraman
1 + ln 2 ≈ 1.69, while the second, which is [2002]).
more computationally efficient, has an ap- Let G be the communication graph ob-
proximation ratio of 15 8
. These ratios have tained when all the nodes transmit at
been recently have been recently improved maximum power (the maxpower graph),
for any positive constant to 53 + , and to and assume G is connected. Every edge
11 (ui , u j ) in G is weighted with the power
, respectively [Althaus et al. 2003]. Fur-
6 δuαi ,u j needed to transmit a message be-
ther, the authors of Althaus et al. [2003]
present an exact branch and cut algorithm tween ui and u j . Given any path P =
for solving WSRA based on a new integer u1 , u2 , . . . , uk in G, the power cost of P
linear program formulation of the prob- is defined as the sum of the power costs
lem. Experimental results show that the of the single edges, that is, pc(P ) =
k−1 α
branch and cut algorithm solves instances i=1 δui ,ui+1 . Let pcG (u, v) denote the min-
with up to 35-40 nodes (with randomly imum of pc(P ) over all paths P that
generated positions) in 1 hour. Most im- connect nodes u and v in G. A path in
portantly, the experimental results show G connecting u and v and consuming
that the average improvement of the ex- the minimum power pcG (u, v) is called a
act solution over RAMST , which can be eas- minimum-power path between u and v.
ily calculated, is in the range 4–6%. This Let G be an arbitrary subgraph of G. The
means that the average case approxima- power stretch factor of G with respect to
tion ratio of RAMST is much smaller than G is the maximum over all possible node
the worst case ratio of 2. pairs of the ratio between the cost of the
The problem of ensuring k-connectivity minimum-power path in G and in G. For-
(i.e., fault-tolerance) of the communication mally, ρG = maxu,v∈N pc G (u,v)
pcG (u,v)
.
graph has also been considered in the lit- The power stretch factor is a generaliza-
erature. It was first studied in Loyd et al. tion of the concept of distance stretch factor
[2002] in the weakly symmetric version which is well known in computational ge-
when k = 2, and further analyzed in ometry. Another similar concept is the hop
Calinescu and Wan [2003]. In particular, stretch factor, which measures the ratio of
Calinescu and Wan prove that the weakly the hopcounts rather than that of power
symmetric version of the problem, with or distance.
k = 2 is NP-hard, and they provide In general, we would like to identify a
approximation algorithms for both the subgraph G (also called a routing graph
weakly symmetric and asymmetric ver- in the following) of the maxpower graph G
sion of the problem. which has a low-power stretch factor and
5.2.2. Minimum Energy Unicast and which is sparser than the original graph.
Broadcast The routing graph can be used to compute
routes between nodes with the guaran-
5.2.2.1. Unicast. In the previous Section, tee that the power needed to communicate
the emphasis was on finding a range along these routes is almost minimal. The
assignment that generates a connected advantage of using G instead of G is that
topology of minimum energy cost. An- computing the optimal routes in G is eas-
other branch of research focused on com- ier than in G and generates little message
overhead, and that a sparse communica-
5 Itcan be easily observed that the RAMST range as- tion graph requires little maintenance in
signment used in Kirousis et al. [2000] to approx-
imate RA within a factor of 2 is weakly symmet-
the presence of node mobility.
ric. This observation has been used in Blough et al. Given the maxpower graph G, the prob-
[2002] to prove that cWS − cRA ∈ O(1). lem of computing a subgraph G with
Fig. 4. Edges in the relative neighborhood graph (left) and in the gabriel graph (right).
low-power stretch factor has been widely Several routing graphs that satisfy
studied in the literature. Ideally, the some of the previous requirements have
routing graph should have the following been proposed in the literature. Most of
features: them are based on subgraphs of G which
have been shown to be good distance span-
(a) constant power stretch factor, that is, ners. In fact, it can be easily seen that,
ρG ∈ O(1). Using the terminology of if a subgraph G is a distance spanner of
geometric graphs, G should be a power graph G, then it is also a power spanner
spanner of G; of G (note that the reverse implication, in
(b) linear number of edges, in other words, general, is not true). Thus, the consider-
G should be sparse; able body of research devoted to distance
(c) bounded node degree, and spanners in computational geometry can
(d) easily computable in a distributed be used to design good routing graphs.
and localized fashion. By localized, we The following geometric graphs have
mean that every node should be able to been considered in the literature.
compute the set of its neighbors in G Definition 5.6 Let N be a set of points
using only information provided by its in the Euclidean two-dimensional space.
neighbor nodes in G.
Property (a) ensures that the routes cal- The Relative Neighborhood Graph
culated on G are at most a constant fac- (RNG) of N has an edge between two
tor away from the energy-optimal routes. nodes ui and u j if there is no node uk
Property (b) eases the task of finding such that max{δui ,uk , δu j ,uk } ≤ δui ,u j (see
routes in G and of maintaining the rout- Figure 4(a)).
ing graph in the presence of node mobil- The Gabriel Graph (GG) of N has an
ity, and it reduces the routing overhead. edge between two nodes ui and u j if there
The requirement of bounded node degree is no node uk such that δu2i ,uk +δu2 j ,uk ≤ δu2i ,u j ;
is motivated by the fact that nodes with a in other words, (ui , u j ) ∈ GG(N ) if and only
high degree are likely to be bottlenecks in if the disk obtained using ui u j as its diam-
the communication graph. Finally, prop- eter does not contain any node from N (see
erty (d) is fundamental for a fast and Figure 4(b)).
effective computation of the routing graph The Delaunay Graph (DG) of N is
in a real wireless ad hoc network. the unique triangulation such that the
circumcircle of every triangle contains no Table I. Distance Stretch Factor, Power Stretch
points of N in its interior; Factor, and Maximum Node Degree of Different
Proximity Graphs
The Yao Graph (YG) of N of parameter c Distance Power Degree
for any integer c ≥ 6 is denoted YGc , and is RNG n−1 n−1 n−1
√
defined as follows. At each node ui ∈ N , GG n−1 1 n−1
√ √ α
any c equally separated rays originated 1+ 5 1+ 5
RDG π π
(n)
at ui define c equal cones. In each cone, 2 2
topology control), and the problem is one of with high probability) composed of
finding a range assignment (and, thus, a bidirectional links7 ;
network topology) which is optimal with — rely on low quality information.
respect to a certain measure. Hence in
these approaches, the emphasis is on the 5.2.3.1. Location-Based TC Protocols. In
quality of the topology produced rather Rodoplu and Meng [1999], the authors
than on the process of building the topol- presented a distributed topology control
ogy itself. Another branch of research fo- algorithm that leverages on location in-
cused on more practical approaches to formation (provided by low-power GPS re-
the TC problem, trying to design sim- ceivers) to build a topology that is proven
ple, fully distributed protocols that build to minimize the energy required to com-
and maintain a reasonably good topology. municate with a given master node. In Li
We call these protocols topology control and Wan [2001], the authors described a
protocols. more efficient implementation of the pro-
Ideally, a topology control protocol tocol which, however, computes only an ap-
should be fully distributed, asynchronous, proximation of the minimum energy topol-
and localized. As discussed previously, ogy.
these requirements are vital for an effec- In Ramanathan and Rosales-Hain
tive implementation of the protocol, espe- [2000], the authors considered the prob-
cially in the presence of node mobility. An- lem of minimizing the maximum of node
other aspect to be considered is the quality transmitting ranges while achieving
of the information needed by the topology connectedness. They also considered the
control protocol. In general, there is a stronger requirement of 2-connectivity of
trade-off between information quality and the communication graph. They present
energy consumption and/or interference centralized topology control algorithms
reduction: the more accurate the informa- that provide the optimal solution for both
tion required (e.g., exact node positions), versions of the problem. The range as-
the more energy savings/interference signment returned by the algorithm has
reductions can be achieved. However, the the additional property of being per-node
price to be paid (in terms of additional minimal, that is, no transmitting range
hardware on the nodes or of additional can be reduced further without impairing
messages to be exchanged) to obtain high connectivity (or 2-connectivity).
quality information must be carefully con- In Li et al. [2003], the authors intro-
sidered. For example, suppose protocol P1 duced LMST, a fully distributed and local-
is based on location information, and pro- ized protocol aimed at building an MST-
tocol P2 is based on distance estimation. like topology. The authors show that (1)
Clearly, the cost of implementing P2 in a the protocol generates a strongly con-
real network is lower than that required nected communication graph; (2) the node
by P1 since the hardware needed to esti- degree of any node in the generated topol-
mate distance between nodes is cheaper ogy is at most 6; and (3) the topology can be
than that required to estimate node po- made symmetric by removing asymmetric
sitions. So, if the energy savings provided links without impairing connectivity. Fur-
by protocol P1 are not considerably higher thermore, the authors show through simu-
than those achieved by P2 , a solution lation that LMST outperforms CBTC (see
based on protocol P2 may be preferable in the following) and the protocol of Rodoplu
practice. and Meng [1999] in terms of both aver-
Summarizing, a topology control proto- age node degree and average node trans-
col should: mitting range. A drawback of LMST is
that it requires location information that
can be provided only with a considerable
— be fully distributed and asynchronous;
— rely on local information only; 7 The motivation for using bidirectional links is given
— generate a connected topology (at least in Section 5.2.1.
when k ∈
(log n) is connected with high execution of a topology control protocol.
probability. From a practical viewpoint, On the other hand, the critical transmit-
Blough et al. [2003] show through simula- ting range for connectivity considered in
tion that setting k = 9 is sufficient to ob- Section 5.1 is representative of the sce-
tain connected networks with high proba- nario in which only a straightforward type
bility for networks with n ranging from 50 of topology control is feasible.
to 500. Furthermore, the authors analyze The following theorem is a consequence
the time and message complexity of the of the results presented in Santi and
protocol and present simulation results Blough [2003].
that show that the topology generated by
k-NEIGH is, on average, 20% more energy THEOREM 5.7 Let l be a positive real
efficient than that generated by CBTC. number sufficiently large, and let N be a
A protocol that shares many similar- set of n nodes positioned uniformly and in-
ities with k-NEIGH is the XTC protocol dependently at random in R = [0, l ]d , with
presented in Wattenhofer and Zollinger d = 1, 2, 3. Assume the distance-power gra-
[2004]: the neighbors of a node u are or- dient α is 2, and denote by cmin (N ) the cost
dered according to some metric (e.g., dis- of the r-homogeneous range assignment
tance or link quality), and u decides which such that r is minimum, and the resulting
nodes are kept as immediate neighbors communication graph is connected. Then,
in the final network topology based on a with high probability:
simple rule. Contrary to k-NEIGH, which l 2 log2 l
achieves connectivity with high probabil-
O( n ) for d = 1
ity, XTC builds a topology which is con- cmin (N ) = O(l log l )
2
for d = 2
nected whenever the maxpower communi-
2/3
cation graph is connected. To achieve this, O(l n log l ) for d = 3.
2 1/3
2
Proposition 5.8 can be restated as
( ln ). It — Increased message overhead. The im-
is not difficult to show that equally spacing plementation of any distributed topol-
nodes is the most energy-efficient place- ogy control protocol causes a certain
ment. It follows that the energy cost of any message overhead which is due to the
instance (including a random one9 ) of RA fact that nodes need to exchange mes-
2 sages in order to set the transmitting
is ( ln ). Comparing this bound with the range to the appropriate value. In the
upper bound reported in Theorem 5.7 for case of stationary networks, the topol-
d = 1, we have that the asymptotic gap be- ogy control protocol is, in general, ex-
tween the energy cost of the optimal range ecuted once at the beginning of the
assignment and that of the optimal ho- network operational time, and then pe-
mogeneous range assignment is at most riodically to account for node join/leave.
log2 l . Hence, the asymptotic benefit of the Thus, the efficiency of the protocol (ex-
adoption of a topology control mechanism pressed here in terms of message over-
in one-dimensional networks is at most a head) has relatively little importance,
factor of log2 l . and the emphasis is more on the quality
Bounds on the energy cost of the solu- of the produced topology. In the pres-
tion of the random instance of RA in two ence of mobility, the topology control
and three dimensions have been obtained protocol must be executed frequently in
in Blough et al. [2002], and are
(l 2 ) for order to account for the new positions
d = 2, and
(l 2 n1/3 ) for d = 3. By Theorem of the nodes. Thus, reducing message
5.7, we can conclude that the asymptotic overhead is fundamental when imple-
benefit of the adoption of a topology control menting topology control mechanisms
mechanism is at most a factor of log l in in mobile networks (especially in the
two-dimensional networks, and at most a case of high mobility scenarios) even
factor of l o g 2/3l in three-dimensional net- if reducing message overhead comes at
works. the cost of a lower quality of the con-
The comparison of the bounds on the en- structed topology.
ergy cost of the optimal solution of RA and
— Nonuniform node spatial distribution.
CTR in one, two, and three-dimensional
As it will be discussed in detail later,
networks indicates that the benefit, ex-
some mobility patterns cause a nonuni-
pressed in terms of energy cost, of the
form node spatial distribution. This
adoption of a topology control mechanism
fact should be carefully taken into ac-
increases with the length l of the side
count in setting important network pa-
of the deployment region but becomes
rameters (e.g., the critical transmitting
less significant for networks of higher
range) at the design stage.
dimension.
From this discussion, it is clear that
6. MOBILE NETWORKS the impact of mobility on the effective-
ness of topology control techniques heav-
In Section 5, we have analyzed several ily depends on the mobility pattern. For
problems related to energy-efficient com- this reason, we first present the mobility
munication in stationary wireless ad hoc models which have been considered in the
networks. In this section, we will discuss literature.
how node mobility affects topology control
in general.
6.1. Mobility Models
The impact of mobility on topology con-
trol is twofold: The most widely used mobility model in
the ad hoc network community is the ran-
9 Here,
dom waypoint model [Johnson and Maltz
with random instance we mean an instance
of the problem in which node positions are chosen
1996]. In this model, every node chooses
uniformly at random in the deployment region R = uniformly at random a destination in
[0, l ]d . [0, l ]d (the waypoint) and moves towards
it along a straight line with a velocity cho- node moves far away from its position in
sen uniformly at random in the interval the previous step.
[vmin , vmax ]. When it reaches the destina- Observe that, in the case of random
tion, it remains stationary for a predefined direction or Brownian-like motion, nodes
pause time tpause , and then it starts moving may, in principle, move out of the deploy-
again according to the same rule. ment region. Since a standard approach in
A similar model is the random direc- simulations is to keep the number of net-
tion model [Bettstetter 2001; Royer et al. work nodes constant, we need a so-called
2001] in which nodes move with the di- border rule [Bettstetter 2001]/ that defines
rection chosen uniformly in the interval what to do with nodes that are about to
[0, 2π] and the velocity chosen uniformly leave the deployment region. In this situ-
at random in the interval [vmin , vmax ]. Af- ation, a node can be:
ter a randomly chosen time taken usually
from an exponential distribution, the node (1) bounced back according to some rule;
chooses a new direction. A similar proce- (2) positioned at the point of intersection
dure is used to change velocity, using an of the boundary with the line connect-
independent stochastic process. ing the current and the desired next
Contrary to the case of the random position;
waypoint and the random direction model (3) wrapped around to the other side of the
which resemble (at least to some extent) region which is considered as a torus;
intentional motion, the class of Brownian- (4) deleted, and a new node initialized ac-
like mobility models resembles noninten- cording to the initial distribution;
tional movement. For example, in the
model used in Blough et al. [2002], mo- (5) forced to choose another position until
bility is modeled using parameters pstat , the chosen position is inside the bound-
pmove , and m. Parameter pstat represents aries of the deployment region.
the probability that a node remains sta- Depending on the choice of the border
tionary during the entire simulation time. rule, nonuniformity in the node spatial
Hence, only (1 − pstat )n nodes (on the av- distribution can be produced. For example,
erage) will move. Introducing pstat into, the second rule described places nodes ex-
the model accounts for those situations in actly on the boundary of the region with
which some nodes are not able to move. For higher probability than at other points. In
example, this could be the case when sen- fact, the only two rules that do not appear
sors are spread from a moving vehicle, and to favor one part of the region over another
some of them remain entangled, for exam- are the torus rule (3) and rule (5) one in
ple, in a bush or tree. This can also model which a node is eliminated when it would
a situation where two types of nodes are cross the boundary and a new node is cre-
used, one type that is stationary and an- ated in its place. However, these rules ap-
other type that is mobile. Parameter pmove pear quite unrealistic and are used mainly
is the probability that a node moves at a to artificially generate a more uniform
given step. This parameter accounts for node spatial distribution.
heterogeneous mobility patterns in which For a more exhaustive survey of mo-
nodes may move at different times. Intu- bility models in wireless networks, the
itively, the smaller the value of pmove , the reader is referred to Bettstetter [2001] and
more heterogeneous the mobility pattern Camp et al. [2002].
is. However, values of pmove close to 0 re-
sult in an almost stationary network. If a
6.2. Homogeneous Topology Control
node is moving at step i, its position in step
i + 1 is chosen uniformly at random in the If deriving analytical results for station-
square of side 2m centered at the current ary networks is difficult, deriving theo-
node location. Parameter m models, to a retical results regarding mobile ad hoc
certain extent, the velocity of the nodes: networks is even more challenging, even
the larger m is, the more likely it is that a in the simpler case of topology control,
that is, in case of homogeneous range Unfortunately, the significance of the find-
assignment. ings of Sanchez et al. [1999] is partly im-
When the range assignment is homoge- paired by the fact that the toroidal border
neous, the message overhead is not an is- rule is used in simulations and that the
sue since the nodes’ transmitting range values of the mobility parameters used in
is set at the design stage, and it can- the experiments (such as t pause in the ran-
not be changed dynamically. However, the dom waypoint model) are not reported.
node spatial distribution generated by the Santi and Blough [2003, 2002] investi-
mobility model could be an issue. For gate the relationship between the critical
instance, it is known [Bettstetter 2001; transmitting range in stationary and in
Bettstetter and Krause 2001; Bettstetter mobile networks through extensive sim-
et al. 2003; Blough et al. 2002] that the ulation. They consider random waypoint
random waypoint model generates a node and Brownian-like motion and analyze
spatial distribution which is independent different critical values for the node trans-
of the initial node positions and in which mitting range that are representative of
nodes are concentrated in the center of different requirements on network con-
the deployment region. This phenomenon, nectivity (for instance, connectivity dur-
which is known as the border effect, is ing 100% and 90% of the simulation time).
due to the fact that, in the random way- The simulation results show that a rela-
point model, a node chooses a uniformly tively modest increase of the transmitting
distributed destination point rather than range with respect to the critical value
a uniformly distributed angle. Therefore, in the stationary case is sufficient to en-
nodes located at the border of the region sure network connectivity during 100% of
are very likely to cross the center of the the simulation time. The increase is about
region on their way to the next waypoint. 21% in the random waypoint and about
The intensity of the border effect mainly 25% in the Brownian-like model. Further-
depends on the pause time tpause . In fact, more, the simulation results show that the
a longer pause time tends to increase the transmitting range can be considerably re-
percentage of nodes that are resting at any duced (in the order of 35–40%) if the re-
given time. Since the starting and desti- quirement for connectivity is only on 90%
nation points of a movement are chosen of the simulation time.
uniformly in [0, l ]d , this implies that a rel- Further insights into the relationship
atively long pause time generates a more between the stationary and mobile criti-
uniform node spatial distribution. cal transmitting range can be derived from
An immediate consequence of the fact the statistical analysis of the node spatial
that the node spatial distribution in the distribution of mobile networks reported
presence of mobility is, in general, nonuni- in Blough et al. [2002]. Again, the authors
form is that results concerning the critical consider random waypoint and Brownian-
transmitting range in stationary networks like mobility and perform several statisti-
(which are based on the uniformity as- cal tests on the node spatial distribution
sumption) cannot be directly used. For this generated by these models. The results
reason, the relationship between the crit- of these tests show that the distribution
ical transmitting range with and without generated by Brownian-like motion is vir-
mobility must be carefully investigated. tually indistinguishable from the uniform
Sanchez et al. [1999] analyze the distribution and confirm the occurrence of
probability distribution of the critical the border effect in random waypoint mo-
transmitting range in the presence of dif- tion, whose intensity heavily depends on
ferent mobility patterns (random way- the value of tpause . In the extreme case of
point, random direction, and Brownian- tpause = 0, the random waypoint model gen-
like) through simulation. The simulation erates a node spatial distribution which is
results seem to indicate that the mobility considerably different from uniform. Over-
pattern has little influence on the distri- all, the analysis of Blough et al. [2002] in-
bution of the critical transmitting range. dicate that Brownian-like mobility should
have little influence on the value of the ogy is executed. In turn, this depends on
critical transmitting range, while the ef- several factors such as the mobility pat-
fect of random waypoint mobility on the tern and the properties of the topology
critical transmitting range should heav- generated by the protocol. To clarify this
ily depend on the settings of the mobility point, let us consider two topology con-
parameters. trol protocols P1 and P2 . Protocol P1 builds
The quality of the observation above is the MST in a distributed fashion and sets
confirmed by the probabilistic analysis re- the nodes’ transmitting range accordingly,
ported in Santi [2005] which is, to the while protocol P2 attempts to keep the
best of our knowledge, the only theoret- number of neighbors of each node below
ical result concerning the critical trans- a certain value k as in the k-NEIGH proto-
mitting range in the presence of mobility col of Blough et al. [2003]. Protocol P1 is
reported in the literature so far. Denot- based on global and very precise informa-
ing with r and rmp the critical transmitting tion, since the MST can be built only if the
range in the case of uniformly distributed exact position of every node in the network
nodes and of random waypoint mobile net- is known. In principle, P1 should be re-
works with t pause = p, respectively, and configured every time the relative position
with v = vmin = vmax the node velocity, the of any two nodes in the network changes
author shows that since this change could cause edge inser-
tion/removal in the MST. On the other
rmp p + 0.521405 hand, P2 can be easily computed in a lo-
= v
>1 calized fashion and can be implemented
r p using relatively inaccurate information
p
such as distance estimation. In this case,
if p > 0, and that rrm → ∞ otherwise the protocol should be reexecuted only
(asymptotically, as n → ∞). The author when the relative neighborhood relation
validates this result through simulations, of some node changes. It is quite intuitive
whose results show an interesting thresh- that this occurs less frequently than edge
old phenomenon: for small values of n (n≤ insertion/removal in the MST. It should
50), rmp is less than r, while for larger value also be observed that having a topology
of n the situation is reversed. This phe- that is not up-to-date is much more crit-
nomenon is caused by the border effect in- ical in the case of the MST than in case
duced by random waypoint mobility which of the k-neighbors graph. In fact, a sin-
tends to concentrate nodes in the center of gle edge removal in the MST is sufficient
the deployment region. When n is small, to disconnect the network, while several
the probability of finding at least one node edges can, in general, be removed from the
close to the border is very low, and the crit- k-neighbors graph without impairing con-
ical transmitting range is smaller than in nectivity. Overall, we can reasonably state
the stationary case. However, when n is that P1 should be reexecuted much more
large enough, some of the nodes actually frequently than P2 . Further, we observe
lie close to the border of the deployment that the reconfiguration procedure needed
region, forcing a higher value of rmp . to maintain the MST is more complicated
than that required by the k-neighbors
graph since it relies on global information.
6.3. Nonhomogeneous Topology Control
So, we can conclude that protocol P1 is not
In the case of nonhomogeneous topology suitable to be implemented in a mobile sce-
control, the more relevant effect of mobil- nario; in other words, it is not resilient to
ity is the message overhead generated to mobility.
update the nodes’ transmitting range in From the previous discussion, it is clear
response to node mobility. The amount of that a mobility resilient topology con-
this overhead depends on the frequency trol protocol should be based on a topol-
with which the reconfiguration protocol ogy which can be computed locally and
used to restore the desired network topol- which requires little maintenance in the
There are several ways in which the the packet loss probability is below 1 but
point graph model can be modified in order greater than 0. This fact, which has been
to be more realistic. For instance, we could observed in Seada et al. [2004], should be
define the occurrence of links between accounted for in the design of topology con-
nodes in probabilistic rather than deter- trol mechanisms.
ministic terms. A possible model could be Although some research on the charac-
the following. Given nodes u and v at dis- terization of fundamental network prop-
tance δu,v , we have a link between u and v erties with a more realistic link model has
with probability 1 if δu,v ≤ δ, where δ is an been recently done, further investigation
arbitrary constant, and with probability in this direction is needed.
p(δu,v ) < 1 otherwise, where p(δu,v ) is an ar-
bitrary decreasing function of the distance More Realistic Node Distribution. A sim-
with values in [0, 1]. This characterization plifying assumption commonly used in
of the occurrence of a wireless link is far the analysis of ad hoc networks is that
more realistic than the 1/0 characteriza- nodes are uniformly distributed in the de-
tion used in the point graph model. For ex- ployment region. Although this assump-
ample, there could exist nodes u, v, w with tion seems reasonable in some settings,
δu,v = δu,w > δ such that link (u, v) exists it is quite unrealistic in many scenarios.
and link (u, w) does not. Thus, the radio For instance, as discussed earlier, this as-
coverage area is, in general, not regular sumption does not hold when the nodes
as is the case in real wireless networks. move according to the random waypoint
Radio link models similar to the one de- model. Further, when nodes are dispersed
scribed previously have been introduced from a moving vehicle, the assumption of
in Faragó [2002] and Booth et al. [2003]. uniform distribution is only a rough ap-
In particular, Booth et al. study network proximation of the actual node distribu-
connectivity under this more realistic link tion. Thus, the analysis of network prop-
model and argue that the characteriza- erties in the presence of nonuniform node
tion of the critical range for connectivity spatial distributions is another step for-
based on the assumption of circular cover- ward in the direction of a more realistic
age area can be seen as a worst-case anal- characterization of ad hoc networks.
ysis provided the (possibly irregular) area
covered by the radio signal remains the More Accurate Analysis of Mobile Networks.
same. More work needs to be done to investigate
Another possibility to make the net- the effect of mobility on topology control.
work model more realistic is to take into In particular, the following issues need to
account interferences between nodes. For be addressed.
example, in Dousse et al. [2003] a bidirec-
tional link between nodes u and v exists — Is mobility beneficial or detrimental?
if the signal to noise ratio at the receiver On the one hand, we have seen that
is larger than some threshold where the mobility causes an increased message
noise is the sum of the contribution of overhead to restore the desired topol-
interferences from all other nodes and of ogy. On the other hand, mobility has
a background noise. The authors analyze the positive effect of balancing the node
the impact of such a wireless link model energy consumption. In stationary net-
on network connectivity. works, if a node u has twice the trans-
Note that there is another major driver mitting range of node v, it is likely to de-
for more realistic network models, namely plete its battery much faster than node
the usage of link-layer retransmission pro- v. In the presence of mobility, nodes
tocols. In fact, it turns out that it usually change the transmitting range dynam-
pays off in term of minimal overall en- ically and a more balanced energy
ergy consumption in the presence of re- consumption is likely to occur. Since
transmissions to use connections at the one of the ultimate goals of topology
boundary of the radio coverage area where control is to extend network lifetime,
the overall effect of mobility on the nodes allow the transmitted power to be
network lifetime should be carefully dynamically adjusted.
investigated.
— Determination of the optimal frequency ACKNOWLEDGMENTS
for reconfiguration. As outlined in The author wishes to thank the anonymous review-
Section 6, there is a trade-off between ers for the many suggestions that helped improving
the message overhead caused by a the presentation quality of the article.
topology control protocol and the qual-
ity of the topology generated. In gen- REFERENCES
eral, to have a high quality topology
ALDOUS, D. AND STEELE, J. 1992. Asymptotics for
(e.g., a connected topology), we should euclidean minimal spanning trees on random
execute the reconfiguration protocol points. Probab. Theo. Relat. Fields 92, 247–258.
frequently. On the other hand, each ex- ALTHAUS, E., CALINESCU, G., MANDOIU, I., PRASAD,
ecution of the reconfiguration protocol S., TCHERVENSKI, N., AND ZELIKOVSKY, A. 2003.
causes a significant message overhead. Power efficient range assignment in ad hoc
The careful investigation of this trade wireless networks. In Proceedings of the IEEE
Wireless Communications and Networking Con-
off would help in answering the previ- ference (WCNC)’03.
ous issue. BAHRAMGIRI, M., HAJIAGHAYI, M., AND MIRROKNI, V.
2002. Fault-tolerant ad 3-dimensional dis-
Group Mobility. In most of the mobil- tributed topology control algorithms in wireless
ity models considered in the literature multihop networks. In Proceedings of the IEEE
(such as the random waypoint, random di- International Conference on Computer Commu-
rection, and Brownian-like model), nodes nications and Networks. 392–397.
move independently one of each other. BAO, L. AND GARCIA-LUNA-ACEVES, J. 2001. Channel
access scheduling in ad hoc networks with unidi-
However, in many realistic scenarios, net- rectional links. In Proceedings of Discrete Algo-
work nodes move in groups. This could be rithms and Methods for Mobile Computing and
the case, for instance, of sensors dispersed Communications (DIALM). 9–18.
in the ocean to monitor water temperature BASAGNI, S., BRUSCHI, D., AND CHLAMTAC, I. 1999.
which are moved by ocean flows, or the A mobility-transparent deterministic broadcast
mechanism for ad hoc networks. IEEE Trans.
case of cars on a freeway which exchange Netw. 7, 6, 799–807.
messages with the purpose of rapidly BETTSTETTER, C. 2001. Smooth is better than
propagating information about traffic con- sharp: A random mobility model for simulation
ditions. Thus, the impact of group mobility of wireless networks. In Proceedings of the ACM
on topology control should be carefully in- Workshop on Modeling, Analysis and Simulation
vestigated. of Wireless and Mobile Systems (MSWiM). 19–
27.
Implementation of TC. Despite the consid- BETTSTETTER, C. 2002a. On the connectivity of
wireless multihop networks with homogeneous
erable body of research devoted to topology and inhomogeneous range assignment. In Pro-
control presented in this article, and the ceedings of the 56th IEEE Vehicular Technology
many theoretical and simulation-based Conference (VTC). 1706–1710.
evidences of the effectiveness of topology BETTSTETTER, C. 2002b. On the minimum node
control techniques in reducing energy con- degree and connectivity of a wireless multihop
sumption and/or increasing network ca- network. In Proceedings of the ACM MobiHoc 02.
80–91.
pacity, to date there is little experimental
BETTSTETTER, C. AND KRAUSE, O. 2001. On border
evidence that topology control can actually effects in modeling and simulation of wireless
be used to these purposes. This is perhaps ad hoc networks. In Proceedings of the IEEE In-
the main open issue in the field. ternational Conference on Mobile and Wireless
Note that the almost complete lack of Communication Network (MWCN).
experimental results about topology con- BETTSTETTER, C., RESTA, G., AND SANTI, P. 2003. The
node distribution of the random waypoint mo-
trol techniques is not due to technologi- bility model for wireless ad hoc networks. IEEE
cal problems as current wireless network Trans. Mobile Comput. 2, 3, 257–269.
cards (see, e.g., the CISCO Aironet 802.11 BLOUGH, D., LEONCINI, M., RESTA, G., AND SANTI,
cards [Cisco 2004]) and wireless sensor P. 2002. On the symmetric range assignment
problem in wireless ad hoc networks. In Proceed- CLEMENTI, A., PENNA, P., AND SILVESTRI, R. 1999.
ings of the IFIP Conference on Theoretical Com- Hardness results for the power range assign-
puter Science. 71–82. ment problem in packet radio networks. In
BLOUGH, D., LEONCINI, M., RESTA, G., AND SANTI, P. Proceedings of the 2nd International Workshop
2003. The k-neighbors protocol for symmetric on Approximation Algorithms for Combinatorial
topology control in ad hoc networks. In Proceed- Optimization Problems (RANDOM/APPROX).
ings of the ACM MobiHoc 03. 141–152. 197–208.
BLOUGH, D., RESTA, G., AND SANTI, P. 2002. A sta- CLEMENTI, A., PENNA, P., AND SILVESTRI, R. 2000b.
tistical analysis of the long-run node spatial dis- The power range assignment problem in radio
tribution in mobile ad hoc networks. In Proceed- networks on the plane. In Proceedings of the 17th
ings of the ACM Workshop on Modeling, Analysis, Symposium on Theoretical Aspects of Computer
and Simulation of Wireless and Mobile Systems Science (STACS). 651–660.
(MSWiM). 30–37. DETTE, H. AND HENZE, N. 1989. The limit distribu-
BOLLOBÁS, B. 1985. Random Graphs. Academic tion of the largest nearest-neighbor link in the
Press, London, UK. unit d -cube. J. Appl. Probab. 26, 67–80.
BOOTH, L., BRUCK, J., COOK, M., AND FRANCESCHETTI, DIAZ, J., PENROSE, M., PETITA, J., AND SERNA, M. 2000.
M. 2003. Ad hoc wireless networks with noisy Convergence theorems for some layout measures
links. In Proceedings of the IEEE International on random lattice and random geometric graphs.
Symposium on Information Theory (ISIT). Combin., Probab., Comput. 6, 489–511.
BORBASH, S. AND JENNINGS, E. 2002. Distributed DOUSSE, O., BACCELLI, F., AND THIRAN, P. 2003. Im-
topology control algorithm for multihop wire- pact of interferences on connectivity in ad hoc
less networks. In Proceedings of the IEEE Inter- networks. In Proceedings of IEEE Infocom.
national Joint Conference on Neural Networks. 1724–1733.
355–360. DOUSSE, O., THIRAN, P., AND HASLER, M. 2002. Con-
nectivity in ad hoc and hybrid networks. In
BRUCK, J., FRANCESCHETTI, M., AND SCHULMAN, L.
Proceedings of IEEE Infocom. 1079–1088.
2002. Microcellular systems, random walks,
and wave propagation. In Proceedings of the ESTRIN, D., GOVINDAN, R., HEIDEMANN, J., AND KUMAR, S.
IEEE Symposium on Antennas and Propaga- 1999. Next century challenges: Scalable coor-
tion. 220–223. dination in sensor networks. In Proceedings of
BURKHART, M., RICKENBACH, P. V., WATTENHOFER, R., AND the ACM Mobicom. 263–270.
ZOLLINGER, A. 2004. Does topology control re- FARAGÓ, A. 2002. Scalable analysis and design of
duce interference? In Proceedings of ACM Mobi- ad hoc networks via random graph theory. In
Hoc 04. 9–19. Proceedings of ACM Discrete Algorithms and
CAGALI, M., HUBAUX, J., AND ENZ, C. 2002. Methods for Mobile Computing and Communi-
cation (DIALM). 43–50.
Minimum-energy broadcast in all-wireless
networks: Np-completeness and distribution GAO, J., GUIBAS, L., HERSHBERGER, J., ZHANG, L., AND
issues. In Proceedings of the ACM Mobicom 02. ZHU, A. 2001. Geometric spanners for routing
172–182. in mobile networks. In Proceedings of the ACM
MobiHoc. 45–55.
CALINESCU, G., MANDOIU, I., AND ZELIKOVSKY, A. 2002.
Symmetric connectivity with minimum power GERLA, M. AND TSAI, J. T.-C. 1995. Multicluster, mo-
consumption in radio networks. In Proceedings bile, multimedia radio networks. ACM/Baltzer
of the IFIP Conference on Theoretical Computer Wirel. Netw. 1, 255–265.
Science. 119–130. GOODMAN, J. AND O’ROURKE, J. 1997. Handbook
CALINESCU, G. AND WAN, P. 2003. Range assignment of Discrete and Computational Geometry. CRC
for high connectivity in wireless ad hoc net- Press, New York, NY.
works. In Proceedings of the Ad Hoc Networks GROSSGLAUSER, M. AND TSE, D. 2001. Mobility in-
and Wireless. 235–246. creases the capacity of ad hoc wireless networks.
CAMP, T., BOLENG, J., AND DAVIES, V. 2002. A survey In Proceedings of IEEE Infocom. 1360–1369.
of mobility models for ad hoc network research. GUPTA, P. AND KUMAR, P. 1998. Critical power for
Wirel. Comm. Mobile Comput. 2, 5, 483–502. asymptotic connectivity in wireless networks.
CHLAMTAC, I. AND FARAGÓ, A. 1999. A new approach Stochastic Analysis, Control, Optimization and
to the design and analysis of peer-to-peer mobile Applications. Birkpauser, Boston, MA. 547–566.
networks. ACM/Baltzer Wirel. Netw. 5, 149–156. GUPTA, P. AND KUMAR, P. 2000. The capacity of wire-
CISCO. 2004. Aironet data sheets. Available at less networks. IEEE Trans. Inf. Theo. 46, 2, 388–
http://www.cisco.com/en/US/products/hw/ 404.
/wireless. HEINZELMAN, W., KULIK, J., AND BALAKRISHNAN, H.
CLEMENTI, A., FERREIRA, A., PENNA, P., PERENNES, S., 1999. Adaptive protocols for information
AND SILVESTRI, R. 2000a. The minimum range dissemination in wireless sensor networks. In
assignment problem on linear radio networks. In Proceedings of ACM Mobicom. 174–185.
Proceedings of the 8th European Symposium on HOLST, L. 1980. On multiple covering of a circle
Algorithms (ESA). 143–154. with random arcs. J. Appl. Probab. 16, 284–290.
HUANG, Z., SHEN, C., SRISATHAPORNPHAT, C., AND LIANG, W. 2002. Constructing minimum-energy
JAIKAEO, C. 2002. Topology control for ad hoc broadcast trees in wireless ad hoc networks. In
networks with directional antennas. In Proceed- Proceedings of ACM Mobihoc. 112–122.
ings of the IEEE International Conference on LIU, J. AND LI, B. 2002. Mobilegrid: Capacity-aware
Computer Communications and Networks. 16– topology control in mobile ad hoc networks. In
21. Proceedings of the IEEE International Confer-
IEEE. 1999. Wireless lan medium access control ence on Computer Communications and Net-
and physical layer specifications. In IEEE 802.11 works. 570–574.
Standard (IEEE Computer Society LAN MAN LOYD, E., LIU, R., MARATHE, M., RAMANATHAN, R., AND
Standards Committee). RAVI, S. 2002. Algorithmic aspects of topology
JANSON, S., KNUTH, D., LUCZAK, T., AND PITTEL, B. control problems for ad hoc networks. In Pro-
1993. The birth of the giant component. Rand. ceedings of ACM Mobihoc. 123–134.
Struct. Algor. 4, 3, 233–359. MAINWARING, A., POLASTRE, J., SZEWCZYK, R., CULLER,
JOHNSON, D. AND MALTZ, D. 1996. Dynamic source D., AND ANDERSON, J. 2002. Wireless sensor
routing in ad hoc wireless networks. In Mobile networks for habitat monitoring. In Proceedings
Computing. Kluwer Academic Publishers. 153– of ACM Wireless Sensor Networks and Applica-
181. tions (WSNA). 88–97.
KHAN, J., KATZ, R., AND PISTER, K. 2000. Emerging MARINA, M. AND DAS, S. 2002. Routing performance
challenges: Mobile networking for smart dust. J. in the presence of unidirectional links in multi-
Comm. Netw. 2, 3, 186–196. hop wireless networks. In Proceedings of ACM
KIM, D., TOH, C., AND CHOI, Y. 2001. On supporting Mobihoc. 12–23.
link asymmetry in mobile ad hoc networks. In MEESTER, R. AND ROY, R. 1996. Continuum Perco-
Proceedings of IEEE Globecom. 2798–2803. lation. Cambridge University Press, Cambridge,
KIROUSIS, L., KRANAKIS, E., KRIZANC, D., AND PELC, A. U.K.
2000. Power consumption in packet radio net- MICHAIL, A. AND EPHREMIDES, A. 2003. Energy-
works. Theoret. Comput. Sci. 243, 289–305. efficient routing for connection-oriented traf-
KO, Y. AND VAIDYA, N. 1998. Location-aided routing fic in wireless ad hoc networks. Mobile Netw.
(lar) in mobile ad hoc networks. In Proceedings Appl. 8, 5, 517–533.
of ACM Mobicom. 66–75. MOAVENI-NEJAD, K. AND LI, X. 2005. Low-
KOLCHIN, V., SEVAST’YANOV, B., AND CHISTYAKOV, V. interference topology control for wireless
1978. Random Allocations. V.H. Winston and ad hoc networks. Ad Hoc Sensor Netw.: Int. J.
Sons, Washington D.C. To appear.
LI, J., BLAKE, C., COUTO, D. D., LEE, H. I., AND MURTHY, S. AND GARCIA-LUNA-ACEVES, J. 1996. An
MORRIS, R. 2001. Capacity of ad hoc wireless efficient routing protocol for wireless networks.
networks. In Proceedings of ACM Mobicom. 61– Mobile Netw. Appls. 1, 2, 183–197.
69. NARAYANASWAMY, S., KAWADIA, V., SREENIVAS, R., AND
LI, L., HALPERN, J., BAHL, P., WANG, Y., AND KUMAR, P. 2002. Power control in ad hoc
WATTENHOFER, R. 2001. Analysis of a cone- networks: Theory, architecture, algorithm and
based distributed topology control algorithm implementation of the compow protocol. In Pro-
for wireless multi-hop networks. In Proceed- ceedings of European Wireless. 156–162.
ings of ACM Principles of Distributed Company PALMER, E. 1985. Graphical Evolution. John Wiley
(PODC). 264–273. and Sons, New York, NY.
LI, N. AND HOU, J. 2004. Flss: a fault-tolerant topol- PANCHAPAKESAN, P. AND MANJUNATH, D. 2001. On the
ogy control algorithm for wireless networks. In transmission range in dense ad hoc radio net-
Proceedings of ACM Mobicom. 275–286. works. In Proceedings of IEEE Signal Processing
LI, N., HOU, J., AND SHA, L. 2003. Design and anal- Communication (SPCOM).
ysis of an mst-based topology control algorithm. PAPADIMITRIOU, I. AND GEORGIADIS, L. 2004. Energy-
In Proceedings of the IEEE Infocom. 1702– aware broadcast trees in wireless networks. Mo-
1712. bile Netw. Appls. 9, 6, 567–581.
LI, X. AND WAN, P. 2001. Constructing minimum PEARLMAN, M., HAAS, Z., AND MANVELL, B. 2000. Us-
energy mobile wireless networks. In Proceedings ing multi-hop acknowledgements to discover and
of ACM Mobihoc. 283–286. reliably communicate over unidirectional links
LI, X., WAN, P., WANG, Y., AND FRIEDER, O. 2002. in ad hoc networks. In Proceedings of Wire-
Sparse power efficient topology for wireless less Communications and Networking Confer-
networks. In Proceedings of the IEEE Hawaii ence (WCNC). 532–537.
International Conference on System Sciences PENROSE, M. 1997. The longest edge of the ran-
(HICSS). dom minimal spanning tree. Annals Appl.
LI, X., WANG, Y., AND SONG, W. 2004. Applications of Probab. 7, 2, 340–361.
k-local mst for topology control and broadcasting PENROSE, M. 1998. Extremes for the minimal
in wireless ad hoc networks. IEEE Trans. Paral. spanning tree on normally distributed points.
Distr. Syst. 15, 12, 1057–1069. Advances Appl. Probab. 30, 628–639.
PENROSE, M. 1999a. On k-connectivity for a ge- less ad hoc networks. IEEE Trans. Mobile
ometric random graph. Rand. Struct. Al- Comput. 2, 1, 25–39.
gori. 15, 2, 145–164. SANTI, P., BLOUGH, D., AND VAINSTEIN, F. 2001. A
PENROSE, M. 1999b. A strong law for the largest probabilistic analysis for the range assignment
nearest-neighbour link between random points. problem in ad hoc networks. In Proceedings of
J. London Math. Soci. 60, 2, 951–960. ACM Mobihoc. 212–220.
PENROSE, M. 1999c. A strong law for the longest SCHWIEBERT, L., GUPTA, S., AND WEINMANN, J. 2001.
edge of the minimal spanning tree. The Annals Research challenges in wireless networks of
Probab. 27, 1, 246–260. biomedical sensors. In Proceedings of ACM Mo-
PHILIPS, T., PANWAR, S., AND TANTAWI, A. 1989. Con- bicom. 151–165.
nectivity properties of a packet radio network SEADA, K., ZUNIGA, M., HELMY, A., AND KRISHNAMACHARI,
model. IEEE Trans. Inform. Theo. 35, 5, 1044– B. 2004. Energy-efficient forwarding strate-
1047. gies for geographic routing in lossy wireless sen-
PIRET, P. 1991. On the connectivity of radio net- sor networks. In Proceedings of ACM SenSys.
works. IEEE Trans. Inform. Theo. 37, 5, 1490– 108–121.
1492. SEN, A. AND HUSON, M. 1996. A new model for
POTTIE, G. AND KAISER, W. 2000. Wireless inte- scheduling packet radio networks. In Proceed-
grated network sensors. Comm. ACM 43, 5, 51– ings of IEEE Infocom. 1116–1124.
58. SONG, W., WANG, Y., LI, X., AND FRIEDER, O. 2004.
PRAKASH, R. 2001. A routing algorithm for wire- Localized algorithms for energy efficient topol-
less ad hoc networks with unidirectional links. ogy in wireless ad hoc networks. In Proceedings
ACM/Kluwer Wirel. Netw. 7, 6, 617–625. of ACM MobiHoc. 98–108.
RAJARAMAN, R. 2002. Topology control and rout- SRIVASTAVA, M., MUNTZ, R., AND POTKONJAK, M. 2001.
ing in ad hoc networks: A survey. SIGACT Smart kindergarten: Sensor-based wireless net-
News 33, 2, 60–73. works for smart developmental problem-solving
RAMANATHAN, R. AND ROSALES-HAIN, R. 2000. Topol- environments. In Proceedings of ACM Mobicom.
ogy control of multihop wireless networks using 132–138.
transmit power adjustment. In Proceedings of STEELE, J. 1988. Growth rates of euclidean mini-
IEEE Infocom 00. 404–413. mal spanning trees with power weighted edges.
RAMASUBRAMANIAN, V., CHANDRA, R., AND MOSSE, D. Annals Probab. 16, 1767–1787.
2002. Providing a bidirectional abstraction for STEERE, D., BAPTISTA, A., MCNAMEE, D., PU, C., AND
unidirectional ad hoc networks. In Proceedings WALPOLE, J. 2000. Research challenges in en-
of IEEE Infocom. 1258–1267. vironmental observation and forecasting sys-
RAPPAPORT, T. 2002. Wireless Communications: tems. In Proceedings of ACM Mobicom. 292–
Principles and Practice, 2nd Ed. Prentice Hall, 299.
Upper Saddle River, NJ. SZEWCZYK, R., MAINWAIRING, A., POLASTRE, J., AND
RODOPLU, V. AND MENG, T. 1999. Minimum energy CULLER, D. 2004. An analysis of a large scale
mobile wireless networks. IEEE J. Select. Areas habitat monitoring application. In Proceedings
Comm. 17, 8, 1333–1344. of ACM SenSys. 214–226.
ROYER, E., MELLIAR-SMITH, P., AND MOSER, L. 2001. WAN, P., CALINESCU, G., LI, X., AND FRIEDER, O.
An analysis of the optimum node density for 2002. Minimum energy broadcasting in static
ad hoc mobile networks. In Proceedings of the ad hoc wireless networks. ACM/Kluwer Wirel.
IEEE International Conference on Communica- Netw. 8, 6, 607–617.
tions. 857–861. WAN, P. AND YI, C. 2004. Asymptotical critical
SADLER, C., ZHANG, P., MARTONOSI, M., AND LYON, transmission radius and critical neighbor num-
S. 2004. Hardware design experiences in ze- ber for k-connectivity in wireless ad hoc net-
branet. In Proceedings of ACM SenSys. 227–238. works. In Proceedings of ACM MobiHoc. 1–8.
SANCHEZ, M., MANZONI, P., AND HAAS, Z. 1999. De- WANG, W., LI, X., MOAVENINEJAD, K., WANG, Y.,
termination of critical transmitting range in ad AND SONG, W. 2003. The spanning ratio of β-
hoc networks. In Proceedings of Multiaccess, Mo- skeletons.
bility and Teletraffic for Wireless Communica- WANG, Y., LI, X., AND FRIEDER, O. 2002. Distributed
tions Conference. spanners with bounded degree for wireless ad
SANTI, P. 2005. The critical transmitting range for hoc networks. Int. J. Found. Comput. Sci. To ap-
connectivity in mobile ad hoc networks. IEEE pear.
Trans. Mobile Comput. 4, 3, 310–317. WATTENHOFER, R., LI, L., BAHL, P., AND WANG, Y. 2001.
SANTI, P. AND BLOUGH, D. 2002. An evaluation of Distributed topology control for power efficient
connectivity in mobile wireless ad hoc networks. operation in multihop wireless ad hoc networks.
In Proceedings of IEEE Dependable Systems and In Proceedings of IEEE Infocom. 1388–1397.
Networks (DSN). 89–98. WATTENHOFER, R. AND ZOLLINGER, A. 2004. Xtc: A
SANTI, P. AND BLOUGH, D. 2003. The critical trans- practical topology control algorithm for ad hoc
mitting range for connectivity in sparse wire- networks. In the 4th International Workshop on
Algorithms for Wireless, Mobile, Ad Hoc and Sen- less ad hoc networks with bernoulli nodes.
sor Networks (WMAN). In Proceedings of IEEE Wireless Communica-
WIESELTHIER, J., NGUYEN, G., AND EPHREMIDES, A. tions and Networking Conference (WCNC) To
2000. On the construction of energy-efficient appear.
broadcast and multicast trees in wireless net- YI, C., WAN, P., LI, X., AND FRIEDER, O. 2003. Asymp-
works. In Proceedings of IEEE Infocom. 585– totic distribution of the number of isolated nodes
594. in wireless ad hoc networks with bernoulli
XUE, F. AND KUMAR, P. 2004. The number of neigh- nodes. In Proceedings of IEEE Wireless Commu-
bors needed for connectivity of wireless net- nications and Networking Conference (WCNC).
works. ACM/Kluwer Wirel. Netw. 10, 2, 169– 1585–1590.
181. YUKICH, J. 2000. Asymptotics for weighted mini-
YI, C. AND WAN, P. 2005. Asymptotic critical mal spanning trees on random points. Stochastic
transmission ranges for connectivity in wire- Proc. Appl. 85, 123–128.