An Overview Of Channel Assignment Methods For Multi-Radio Multi-Channel Wireless Mesh Networks

ARTICLE IN PRESS
J. Parallel Distrib. Comput. ( ) –
Contents lists available at ScienceDirect
J. Parallel Distrib. Comput.
journal homepage: www.elsevier.com/locate/jpdc
An overview of Channel Assignment methods for multi-radio multi-channel
wireless mesh networks
Weisheng Si, Selvadurai Selvakennedy, Albert Y. Zomaya∗
Centre for Distributed and High Performance Computing, School of Information Technologies, University of Sydney, Australia
a r t i c l e i n f o
Article history:
Received 14 May 2009
Received in revised form
16 September 2009
Accepted 28 September 2009
Available online xxxx
Keywords:
Wireless mesh networks
Multiple radios
Multiple channels
Channel assignment
Survey
a b s t r a c t
Channel Assignment (CA) is an active research area due to the proliferating deployments of multi-
radio multi-channel wireless mesh networks. This paper presents an in-depth survey of some of the
CA approaches in the literature. First, the key design issues for these approaches are identified, laying
down the basis for discussion. Second, a classification that captures their essentials is proposed. Third,
the different CA approaches are examined individually, with their advantages and limitations highlighted;
furthermore, categorical and overall comparisons for them are given in detail, clarifying their sameness
and differences. Finally, the future research directions for CA are discussed at length.
© 2009 Published by Elsevier Inc.
1. Introduction
Wireless Mesh Networks (WMNs) have been adopted in many
municipal/enterprise area network deployments according to the
recent industry white papers [5,32,41]. They are used in a variety
of application scenarios such as the last-mile broadband Internet
access, campus networks, mobile telephony backhaul networks,
and public safety networks. With multiple hops and a mesh
topology, the WMN architecture generally consists of three levels
as described below.
• The top level comprises one or several gateways, which connect
to both the WMN and the wired Internet and forward traffic
between these two types of networks. This level can be absent
if no Internet access is required and the WMN is only used for
local communication.
• The intermediate level comprises numerous mesh routers,
which are the vertices in the mesh topology and relay the traffic
within the WMN. To enable wide and low-cost deployment, the
link and physical layer protocols adopted by mesh routers are
international standards such as the IEEE 802.11 protocol [16]
or the IEEE 802.16 protocol [17] according to the current
practice [5,32,41].
• The bottom level comprises many WLANs or mobile phone
cells depending on the usage of WMNs. Either a WLAN or a
mobile phone cell generally consists of an Access Point (AP)
∗ Corresponding author.
E-mail address: a.zomaya@usyd.edu.au (A.Y. Zomaya).
and a certain number of wireless clients (e.g., laptops or mobile
phones) that are the real producers and consumers of the
wireless traffic.
Note that, the gateways and the APs also have the full
functionality of the mesh routers besides interfacing with the
Internet and the wireless clients respectively, so they both play
double roles. And in the study of WMNs, researchers mainly focus
on the issues of mesh routers, leaving the Internet or the WLAN/cell
issues to other areas of research. In this paper, we also refer to the
mesh routers as ‘‘nodes’’ hereafter.
In the initial design of WMNs, the traditional wireless network
paradigm is followed, where only one radio (i.e., wireless interface
card1
) is equipped at each node and all nodes share a single
channel [2,7]. However, several research findings revealed that
the capacity per node in such solutions drops significantly with
the increase of the network size. For example, researchers in [15]
demonstrated that when n identical randomly placed nodes with
bandwidth W form a single-channel network, the throughput
obtained per node is Θ (W/
√
n log n) asymptotically; Ref. [45]
demonstrated that in a multi-hop network with all links running
the same IEEE 802.11 protocol, the end-to-end performance
suffers low throughput and unfairness problems. An intuitive
interpretation of the above phenomenon is as follows: in a
single-channel multi-hop network, interference can occur not
only between nearby flows (known as inter-flow interference)
1 In this paper, we use the terms ‘‘radio’’ and ‘‘interface’’ interchangeably, both of
which refer to the ‘‘wireless interface card’’.
0743-7315/$ – see front matter © 2009 Published by Elsevier Inc.
doi:10.1016/j.jpdc.2009.09.011
Please cite this article in press as: W. Si, et al., An overview of Channel Assignment methods for multi-radio multi-channel wireless mesh networks, J. Parallel Distrib.
Comput. (2009), doi:10.1016/j.jpdc.2009.09.011

ARTICLE IN PRESS
2 W. Si et al. / J. Parallel Distrib. Comput. ( ) –
Fig. 1. Inter-flow and intra-flow interference.
but also between the nearby hops in a single flow (known as
intra-flow interference), thus significantly degrading the network
performance. To illustrate these two kinds of interference, Fig. 1
shows two concurrent flows in a multi-hop network, where not
only flow 1 and flow 2 interfere with each other but also hop A → B
and hop B → C in flow 1 interfere with each other.
Due to the above observation, the single-radio single-channel
architecture is not appealing. How about the single-radio multi-
channel architecture then? It is also considered undesirable
because of the following two reasons. First, with only a single radio,
a node has to change its channel frequently with the dynamic
network traffic so as to fully exploit the multi-channel advantage.
Unfortunately, channel switching involves non-negligible delays.
For instance, in the IEEE 802.11 hardware, this delay is usually in
the order of several milliseconds [36], which is quite significant
compared with the transmission delays. Second, with the single
radio assigned to different channels to reduce interference,
the nodes suffer low connectivity and even disconnectedness.
Consequently, it is difficult to provide fault tolerance support or to
synchronize neighboring nodes to the same channel if they want
to communicate.
Therefore, almost all the current WMN deployments and pro-
posals adopt the multi-radio multi-channel (MR–MC) architecture,
where each node is equipped with multiple radios and can use
multiple non-overlapping channels. Fig. 2 illustrates the MR–MC
architecture with an example mesh network, where each node has
two radios and can use four non-overlapping channels that are
reused spatially.
As apparent from the current commercial deployments [5,32,
41], this MR–MC architecture is practical due to the following two
facts:
(1) The cost of wireless interface cards has dropped rapidly with
the proliferation of wireless networks, so the cost of multiple
radios is no longer a prohibitive factor.
(2) The current IEEE 802.11 and 802.16 standards both support
multiple non-overlapping channels. For instance, there are
twelve non-overlapping channels with 20 MHz center fre-
quency spacing in IEEE 802.11a [18] and three non-overlapping
channels with 25 MHz center frequency spacing in IEEE
802.11b [19]. Fig. 3 shows the three non-overlapping channels
regulated in US; in Europe, the regulation is only slightly differ-
ent [19]. For IEEE 802.16, it utilizes radio frequencies of both
licensed and unlicensed bands from 2 GHz to 66 GHz2
with
a flexible channel bandwidth, so it can support significantly
more non-overlapping channels than IEEE 802.11 [14]. There-
fore, the availability of multiple non-overlapping channels is
not an issue as far as the standardization is concerned.
As an aside, [13] recently demonstrated through extensive
real-world experiments that there actually exists significant
interference between these standard non-overlapping channels in
2 Not all are available due to the different regulations in different countries.
the current commodity IEEE 802.11 hardware. Nevertheless, Paul
et al. also pointed out that this problem can be resolved by using
better frequency filters in the hardware for multi-channel use.
The advantages of MR–MC WMNs are obvious. With each node
equipped with multiple radios, multiple transmissions/receptions
can happen concurrently, which multiplies the throughput. With
multiple channels, neighboring links assigned to different channels
can carry traffic free of interference, such that the link-layer delay
can be dramatically reduced. As a result, when compared with
the single-radio and single-channel solution, one of the proposed
solutions [36] shows that its multiple-radio and multiple-channel
solution improves the network throughput up to a factor of seven,
and an industry report [29] claims that its multiple-radio and
multiple-channel solution improves the network throughput by a
factor of five.
However, the above advantages cannot be fully realized unless
a number of issues related to MR–MC WMNs are handled
properly. In general, these issues include node deployment,
channel assignment, link scheduling, and routing. Among them,
the issue of channel assignment (CA), which aims to optimize the
MR–MC WMN performance by seeking a proper mapping between
the available channels and the radios at every node, has received
extensive attention. The CA issue is especially important due to the
following reasons:
• It constitutes a new research area for WMNs.
• It is a must-solve problem for the operation of MR–MC WMNs.
• It is challenging because many formulations of the CA problem
turn out to be NP-hard.
Therefore, we focus on the CA issue in this paper and present
an in-depth survey of the typical approaches addressing this
issue. The rest of this paper is organized as follows. The basis for
examining these CA approaches, Section 2 specifies the MR–MC
WMN model assumed by them and identifies the key design issues
that confront them. Section 3 proposes a classification of the CA
approaches. Section 4 describes the main idea and basic steps
of every surveyed CA approaches individually, points out their
advantages and limitations, and makes both categorical and overall
comparisons on them. Section 5 presents the future research
directions. Finally, Section 6 concludes this paper.
2. Network model and design issues
This section clarifies the network model considered by the
surveyed CA approaches and identifies the key design issues for
them to address. Besides, this section also gives the necessary
background used later in this paper.
2.1. Network model
All the approaches surveyed in this paper assume the following
characteristics on the WMNs. Though each approach may bear its
own distinct characteristics, the following ones are common.
(1) The nodes in the WMNs are not mobile.
(2) At least some of the nodes in the WMNs are equipped with
multiple radios.
(3) Multiple non-overlapping channels (free of inter-channel
interference) are available to the network.
2.2. Key design issues
The characteristics in the above model distinguish MR–MC
WMNs from other types of wireless networks. Given these
characteristics, we identify the following key design issues for the
CA approaches to address.
Comput. (2009), doi:10.1016/j.jpdc.2009.09.011

ARTICLE IN PRESS
W. Si et al. / J. Parallel Distrib. Comput. ( ) – 3
Fig. 2. The MR–MC architecture of WMNs.
Fig. 3. North America non-overlapping channel selection according to IEEE 802.11b.
2.2.1. Interference
Interference is the foremost factor that degrades the wireless
network performance, so the primary goal of CA is to minimize
interference within the MR–MC WMNs by utilizing multiple radios
and multiple channels. To address the interference issue, a model
describing the interference effect needs to be assumed. Currently,
there are two widely-adopted models: Protocol Model and Physical
Model, both of which are initially proposed in [15].
A very simple one, the Protocol Model can be described
as follows: (1) each radio has a transmission range and an
interference range, with the former less than the latter; and (2) a
transmission from radio X to radio Y is successful if Y is in the
transmission range of X and not in the interference range of radios
other than X that are currently transmitting. Compared with the
Protocol Model, the Physical Model is close to reality but quite
complex. It can be described as follows: (1) a transmission is
successful if the Signal to Interference and Noise Ratio (SINR) of the
transmitter’s signal at the receiver is larger than a threshold value;
and (2) the interference and noise power at the receiver consists
of the noises generated by other ongoing transmissions and the
ambient noise in the network. Due to the complexity of the Physical
Model, most of the CA approaches surveyed in this paper adopt the
Protocol Model (only except [30], which uses the Physical Model).
Besides the interference model, the following constraints
related to interference also need to be considered by the CA
approaches.
• The number of available channels: due to technical facts or
government regulations, the MR–MC WMNs actually cannot
use as many non-overlapping channels as needed. Thus,
interference cannot be completely eliminated from the MR–MC
WMNs.
• The number of radios at each node: it limits the number of non-
overlapping channels that can be assigned to each node.
• The node deployment: it determines the geometric distances
among the nodes, and hence influences the interference
relationship significantly. If the Protocol Model is assumed,
the area within which the nodes interfere with each other is
actually very large. This is because, to imitate the reality, the
interference range is generally considered to be 2–3 times the
transmission range. For instance, in both NS2 and QualNet (two
popular simulators in use today), the default interference range
is approximately twice the default transmission range [44].
So in addressing the interference issue, CA approaches need
to consider at least the aforementioned factors: the interference
model, the number of available channels, the number of radios at
each node, and the node deployment. As to be described later, the
joint consideration of these factors mostly results in NP-hard CA
problems.
2.2.2. Connectivity
Connectivity is a concept defined on graphs, which are
extensively used to model computer networks. To make our later
discussion of connectivity issue unambiguous, here we describe
and distinguish the following two graph concepts exploited by
most of the existing CA approaches: unit disk graph and network
topology. Note that, due to the terminological inconsistency of the
literature, these approaches may not use the same terms when
they refer to these two graphs.
The unit disk graph [9] is widely used to represent wireless
networks under the Protocol Model. Specifically, a unit disk graph
G (V, E) is defined as an undirected graph where (1) V is the set
of nodes in the network and (2) ∀v1, v2 ∈ V, (v1, v2) ∈ E if v1
is in the transmission range of v2 (also implying v2 is in the
transmission range of v1). Note that, as suggested by the name
of this concept, an identical transmission range for all nodes is
assumed here. This assumption is reasonable in wireless networks
because bidirectional communication is necessary in an unreliable
wireless link: the sender generally requires an acknowledgement
from the receiver after transmitting the data.
While the unit disk graph is only concerned with the geometric
distances among nodes, the network topology [28,34] further
models whether any two nodes actually share a common channel
as their communication link. Specifically, a network topology T
(V, E) is defined as an undirected graph where:
• V represents the set of nodes and E represents the set of actual
communication links in the network.
• ∀v1, v2 ∈ V, (v1, v2) ∈ E if one of the v1’s radios and one of
the v2’s radios share a common channel and they are within
the transmission range of each other. Moreover, if the multiple
radios at v1 and v2 share n common channels, there are n links
(v1, v2) ∈ E, where n is a positive integer.
According to the above two definitions, the unit disk graph is
independent of the channel assignment, but the network topology,
on the other hand, is known only when the channel assignment
Comput. (2009), doi:10.1016/j.jpdc.2009.09.011

ARTICLE IN PRESS
(a) Unit disk graph. (b) Network topology with
discrepancy (1).
(c) Network topology with
discrepancy (2).
Fig. 4. Two kinds of discrepancies between unit disk graph and network topology.
is done. This difference arises from the communication constraint
that, in a multi-channel environment, two neighbor radios must
share a common channel to communicate. Due to this constraint,
altogether two kinds of discrepancies can occur between the unit
disk graph and the network topology: (1) a link between two nodes
in the unit disk graph is absent in the network topology if the
radios on these two nodes are not assigned a common channel;
(2) multiple links exist between two nodes in the network topology
if multiple common channels are assigned to the radios on these
two nodes. These two kinds of discrepancies are illustrated in Fig. 4,
where (1) there are three non-overlapping channels available
(numbered 1, 2, and 3 respectively), (2) nodes A, C each have one
radio and nodes B, D each have two radios, and (3) the channels
assigned to these radios are shown inside the square brackets
nearby the nodes.
Though the unit disk graph and network topology differ in their
structures, they are both important for the CA approaches. The
former is usually used as the basis to perform channel assignment,
since it is known initially and gives the distance relationship
between the network nodes. And the latter is usually used to
specify the connectivity requirement for channel assignment, since
it embodies the actual communication links.
Therefore, in addressing the connectivity issue in MR–MC
WMNs, the CA approaches need to be aware that the CA decisions
can actually change the network topology, which is a key difference
between the single-channel and multi-channel networks. If the CA
approaches ignore this difference, the network topology may be
in danger of being disconnected, even though the interference is
reduced by distributing channels to different radios. Thus, there
is essentially a trade-off between interference and connectivity:
the more radios that are assigned to the same channel, the more
connectivity is achieved, but the more interference is induced.
As to be detailed later, all the approaches surveyed in this paper
at least require that the network topology is connected after
the channel assignment, while some approaches have stronger
requirements. Consequently, how to minimize interference while
satisfying the connectivity requirement is an important issue for
the CA approaches to address.
2.2.3. Stability
The CA operation can cause two phenomena that undermine the
network stability: ripple effect and channel oscillation. The ripple
effect, first described in [36], also arises from the communication
constraint described previously. The following example is given
to exemplify the ripple effect. Assume interface I2 originally at
channel x wants to communicate with interface I1 at channel y,
thus I2 switches to channel y. At same time, assume interface I3
is currently communicating with I2 using x, so I3 has to switch to
y to maintain the communication. And such channel change will
continue to propagate, if another I4 is currently communicating
with I3 using channel x. Another problem associated with the
ripple effect is that, say in the above example, when I2 switches
to channel y, some packets may be lost in the communication
between I3 and I2 on channel x, before I3 switches to channel y.
A similar phenomenon to oscillation in routing, channel oscil-
lation means that the channel assignment does not converge and
changes back and forth among several choices. This phenomenon
usually happens when the CA is based on a dynamic metric. For ex-
ample, when two nodes discover that a channel is under-utilized
according to such a dynamic metric, they may simultaneously
switch to this channel and both begin transmission on it, and then
switch back because this channel is now overloaded, as indicated
by this dynamic metric. Since channel switching involves signif-
icant overhead such as switching delay and traffic interruption,
frequent channel switching resulted from oscillation will severely
impair the network performance.
2.2.4. Throughput/Latency
Throughput and latency are the two most important measures
for network performance. Having a deterministic relationship,
they are generally addressed together. To obtain optimal through-
put/latency in MR–MC WMNs, the CA approaches need to consider
the following two basic strategies. First, reducing interference is
almost the most effective method in achieving the optimality, and
it is better to make this method adaptive to the dynamic network
traffic. Second, links should be treated differently when assigning
channels, since different links in MR–MC WMNs impact through-
put/latency to different extents. For example, the backbone links
carry much more network traffic than the stub links, so they should
be given more bandwidth either by assigning more number of con-
current channels or by assigning less interfered channels. In this
sense, an MR–MC WMN is analogous to a wired enterprize/campus
network consisting of a hierarchy of Ethernet switches, where the
ports on an upper-level switch usually have much more bandwidth
than the ports on a lower-level switch.
2.2.5. Routing
As shown in Section 2.2.2, the network topology of an MR–MC
WMN – a basic factor for making the routing decisions – can be
changed by the CA decisions. Thus, routing is dependent on CA. On
the other hand, routing can change the traffic load distribution in
the network, which is a primary factor considered by CA to reduce
the interference dynamically. So in this case, CA is also dependent
on routing.
To handle such a relationship between CA and routing, two
ways are proposed in the surveyed CA approaches. In one way, CA
is viewed as a lower-layer mechanism and does not consider the
traffic load, while routing is viewed as an upper-layer mechanism
and is fully responsible for distributing the traffic load. Thus, being
independent of CA, any routing protocol should be supported by
the CA mechanism. In the other way, CA and routing are viewed to
be mutually dependent, so they are combined in order to obtain
optimal network performance. In the later survey of individual
approaches, the routing issue will be discussed further.
2.2.6. Fault tolerance
Though the nodes in MR–MC WMNs are stationary, they can fail
because of software or hardware problems. Moreover, the wireless
links can also fail due to unexpected scenarios such as external in-
terference or temporary obstacles. So it is necessary for a CA ap-
proach to support fault tolerance such that the network can op-
erate in a self-healing fashion. Furthermore, though the multiple
Comput. (2009), doi:10.1016/j.jpdc.2009.09.011

ARTICLE IN PRESS
Fig. 5. Classification of CA approaches for MR–MC WMNs.
radios and multiple channels in MR–MC WMNs provide abundant
choices for recovering from faults, a selection among these choices
has to be made to obtain the optimal results. Therefore, the fault
tolerance issue is not an easy task for the CA approaches to address.
2.2.7. Fairness
To obtain the optimal overall network performance, fairness
among the nodes is sometimes sacrificed in designing a CA
approach. But in many scenarios such as the mobile phone
backhaul networks, fairness is a necessary property of the network
services. A basic fairness criterion for CA is the capability to
avoid that the traffic of some nodes only has access to crowded
channels shared by many links, while the traffic of other nodes
has access to only partly-occupied channels shared by a few links.
Though guaranteeing fairness to certain level introduces additional
difficulty to the CA design, it is worth achieving in practice.
2.3. Summary to the key design issues
All the CA approaches discussed in this paper aim to solve
some of the above issues. Though these issues are listed separately,
they are inevitably related to each other. Some of the issues
have mutually-beneficial relationships such as ‘less interference
results in more throughput’ and ‘richer connectivity offers better
fault tolerance’. However, there also exist trade-off relationships
among these issues. For example, reducing interference gives
rise to less connectivity, and supporting fairness sacrifices overall
throughput. Generally, if both issues with a trade-off relationship
are to be addressed, the balance needs to be made according to the
requirements of the concrete CA approaches.
We note that channel assignment (or allocation) is also an active
research area in cellular networks, where the basic components
are the hexagonal cells covering the whole network area. Simply
put, the channel allocation problem in cellular networks is how to
assign channels to the cells such that adjacent cells use different
channels to avoid the interference. Since the cells are connected
to the mobile switching center and then to the public switch tele-
phone network via wired links, there is no connectivity issue for
the CA in cellular networks. Due to the same reason, the wireless
part in cellular networks is not multi-hop, so the CA in cellular net-
works does not concern the issue of routing. We believe these are
the two key differences between the CA in cellular networks and
the CA in mesh networks. Limited by the scope of this paper, please
refer to [6] for the design issues of CA in cellular networks.
3. Classification of CA approaches
Despite being a nascent research area starting around 2004, CA
for MR–MC WMNs has seen many approaches up to now. While
possessing considerable similarity, these approaches also exhibit
significant diversity. In this section, we present a classification of
the typical approaches surveyed in this paper (see Fig. 5). Note
that there can be different classifications according to the different
criteria used, and our taxonomy is what we believe the best-suited
for capturing the essentials of the CA approaches.
In general, all the CA approaches are classified into two cate-
gories: centralized or distributed. For the centralized approaches,
a central control is assumed and it has the complete knowledge
about the mesh networks. Thus, the formulated CA problems can
be solved at a single place. After the result of CA is calculated, it is
distributed to the nodes to accomplish. Altogether, we have seen
three types of such problems being formulated: the graph-based
problem, the network flows problem, and the network partition-
ing problem. Accordingly, the centralized approaches are classified
into three categories using the above problem formulations respec-
tively. The details of the problem formulations and the algorithms
proposed to solve them will be detailed in Section 4.
For the distributed approaches, no central control is assumed
and each node runs its own copy of the algorithm to assign the
channels. They are further classified into two categories according
to the traffic pattern being considered: gateway-oriented CA
approaches and peer-oriented CA approaches. The former assumes
that the main network traffic is to or from the gateways, so the
CA can exploit the heuristic that the near-gateway links should
be given a relatively high bandwidth. The latter assumes that the
network traffic can occur between any pair of nodes with no fixed
pattern, so the CA approaches have to be as general as possible to
accommodate various kinds of network traffic. The details of these
approaches will also be discussed in Section 4.
4. Description and comparison of CA approaches
In this section, we provide a description and comparison of
CA approaches based on our classification presented in Section 3.
The centralized approaches are examined first, followed by the
distributed approaches. For these approaches, we extract their
basic ideas, identify their advantages and limitations, and make
comparisons in each category. Also in the end, a comprehensive
table that compares all the CA approaches based on their basic
properties is given.
4.1. Centralized CA approaches
As mentioned in Section 3, the centralized approaches are clas-
sified into three categories according to their problem formula-
tions: the graph-based approaches, the network flows approaches,
and the network partitioning approaches. To solve these problems,
only a centralized algorithm and no communication protocol is
needed. Thus, our classification on the centralized approaches is
actually a classification on the centralized algorithms. We observe
Comput. (2009), doi:10.1016/j.jpdc.2009.09.011

ARTICLE IN PRESS
that the inputs, the objectives of outputs, and the heuristic or ap-
proximation methods are the three characterizing aspects within
each category of algorithms. For the inputs, there are many pa-
rameters from the networks to consider. Specifically, all the ap-
proaches discussed in this section consider some, if not all, of the
following parameters as the inputs.
• Node deployment: the geometric position of each node in the
network.
• Numbers of radios and channels: the number of radios at each
node and the number of non-overlapping channels available at
each radio.
• Interference model: all the centralized approaches surveyed
in this paper adopt the Protocol Model, with some of them
claiming that their schemes can also be generalized to support
the Physical Model.
• Traffic profile: the bandwidth of each link and the end-to-end
traffic rate of each flow.
• Connectivity constraint: the level of network connectivity
to be achieved. All the approaches at least require the
network topology to be connected after the channel assignment,
while some approaches have stronger requirements such as
k-connectedness in [40].
• Fairness constraint: specifying a certain fairness requirement to
be satisfied. For example, [3] requires that λ times the traffic
rate of each source node can be routed over the network after
the channel assignment, where λ is a constant for all the nodes.
For the objectives of outputs, some approaches aim to maximize
the overall throughput, while others aim to minimize the overall
interference, etc. Despite different problem formulations, different
sets of inputs, and different objectives of outputs, most centralized
approaches (only except [8]) surveyed in this paper state or
formally prove [28,35,37] that their formulated problems are
NP-hard. Consequently, they resort to heuristic or approximation
algorithms to get suboptimal results such that the computation
complexity is polynomial. Thus, we consider the heuristic or
approximation methods as a third characterizing aspect within
each category of approaches.
Therefore, when these three categories of approaches are
examined in the following subsections, special attention is
paid to the three aspects (inputs, objectives of outputs, and
heuristic or approximation methods) in contrasting them within
each category. The choice of these aspects is actually natural
since a well-accepted definition of ‘algorithm’ is ‘a sequence of
computational steps to transform inputs into outputs’ [10], which
exactly includes these three aspects.
4.1.1. Graph-based approaches
In graph-based approaches, an MR–MC WMN is modeled by
a graph with a vertex set and an edge set, and the CA problem
is formulated into the problem of assigning channels to vertices
or edges in this graph. Three types of graph theoretic concept are
generally exploited by this category of approaches: unit disk graph,
network topology, and conflict graph. With the first two concepts
already described in Section 2, we next describe the concept of a
conflict graph GC (VC , EC ), which was first introduced to the MR–MC
WMN context in [20].
The basic ideas of the conflict graph concept are as follows:
(1) it is derived from the network topology T(V, E) and models
whether two links in E interfere with each other; (2) any link in E is
represented by a vertex in VC ; and (3) if two links interfere, an edge
connecting the two corresponding vertices in VC is included in EC
to show this conflict. With the above said, the definition of conflict
graph GC (VC , EC ) as an undirected graph under the Protocol Model
is given as follows:
(a) Network topology. (b) Conflict graph.
Fig. 6. An example of network topology and its derived conflict graph.
• ∀ (vp, vq) ∈ E of T, there is a vertex vp,q ∈ VC .
• ∀ (vp, vq), (vs, vt ) ∈ E of T (i.e., vp,q, vs,t ∈ VC ), (vp,q, vs,t ) ∈ EC
if (1) vp or vq is in the interference range of vs or vt , or vice versa
and (2) (vp, vq), (vs, vt ) use the same channel.
According to the second condition of this definition, bidirec-
tional communication is assumed here: if either direction of the
communication is interfered with, the two links are considered to
conflict with each other. To simplify our description, when this sec-
ond condition is referred to hereafter, we simply say ‘two links in-
terfere with each other’. An example of network topology and its
derived conflict graph is illustrated in Fig. 6, where (1) two chan-
nels (numbered 1 and 2 respectively) are available, (2) node A has
two radios and nodes B, C, D have one radio, and (3) the chan-
nels assigned to these radios are shown inside the square brack-
ets nearby the nodes. As shown in Fig. 6(a), there are three links
in the network topology, so three corresponding vertices exist in
the conflict graph depicted in Fig. 6(b). Since only links AB and AC
(both using channel 2) interfere with each other, there is one edge
connecting the two corresponding vertices in the conflict graph.
As an aside, in most of the literature, the elements in V
and VC are called nodes and vertices respectively, and the
elements in E and EC are called links and edges respectively. For
clarity, this naming convention is followed throughout this paper.
Next, four approaches belonging to this category are discussed:
(1) Connected Low Interference Channel Assignment (CLICA) [28],
(2) minimum INterference Survivable Topology Control (INSTC)
[40], Centralized Tabu-based Algorithm (CTA) [39], and (4) Breadth
First Search Channel Assignment (BFS-CA) [34], all of which base
their algorithms on the above graph concepts.
(1) Connected Low Interference Channel Assignment (CLICA)
To motivate their problem formulation, Marina et al. propose
a new concept called base channel assignment, which has the
following three characteristics: (1) resulting in a network topology
that any link in the unit disk graph is preserved and (2) distributing
the non-overlapping channels among different links such that
the interference is reduced and (3) independent of the traffic
pattern and intended for use during the network deployment or
a maintenance stage [28].
With the base channel assignment problem as their focus,
Marina et al. formulate it into the problem of assigning channels
to the links in the unit disk graph (called the connectivity graph
in [28]). Specifically, the problem formulation in [28] has the
following two output objectives: (1) the two adjacent radios of any
link in the unit disk graph are assigned a common channel, and this
objective is referred to as the ‘connectivity preservation’ hereafter;
(2) the interference over the whole network is minimized in the
sense of minimizing the maximum link conflict weight among all
links in the resultant network topology. Note that the link conflict
weight for a link is defined as the sum of the number of edges
incident to the vertex representing this link in the conflict graph.
After proving that the above formulated problem is NP-hard,
Marina et al. propose a polynomial time heuristic algorithm named
Connected Low Interference Channel Assignment (CLICA) to solve
it. The basic ideas of CLICA are summarized as follows:
Comput. (2009), doi:10.1016/j.jpdc.2009.09.011

ARTICLE IN PRESS
• Randomly assign a node v the highest priority, then assign other
nodes priorities decreasing in the order obtained by depth-first
searching the unit disk graph starting from the node v.
• While traversing the nodes in the decreasing order of their
priorities obtained above, assign channels to the incident links
of these nodes. The operation of assigning a channel to a link
includes assigning this channel to both a radio at this node and
a radio at the neighbor node. Then, the priorities of unvisited
nodes are adjusted according to their degree of flexibility, which
is the number of channels that a node can choose from without
breaking the connectivity preservation. Essentially, the nodes
with a lower degree of flexibility will have their priorities
increased so that they are visited earlier in the later steps.
• When picking a channel in the above step, a node v1 picks
a channel for its incident link (v1, v2) in a greedy manner: a
locally optimal choice is made by selecting the channel that
minimizes the maximum link conflict weight among all links
that can interfere with link (v1, v2). After a channel is assigned
to a link, the conflict graph is updated to reflect the new link
conflict weights.
The proposed CLICA algorithm mainly has two advantages:
(1) the paper proves that it achieves the connectivity preservation
objective specified in the problem formulation; (2) it is a simple
algorithm with a locally greedy heuristic, such that it can be
implemented efficiently. The CLICA algorithm mainly has two
limitations: (1) by using the unit disk graph to assign channels,
the number of radios at each node is difficult to model, so an
additional step in the end of the algorithm is needed to deal with
these unassigned radios; (2) fairness may be sacrificed because it
picks a channel for a link in a locally optimal fashion.
(2) Minimum INterference Survivable Topology Control (INSTC)
INSTC [40], proposed by Tang et al., is similar to CLICA in that
it also formulates the CA problem into the problem of assigning
channels to links in the unit disk graph. To introduce a metric
for minimizing the interference, Tang et al. define a concept
called Link Co-channel Interference (LCI) for a link in the network
topology, which means the number of links that interfere with
this link in the network topology. Though this definition does not
leverage the concept of conflict graph, it is equivalent to the concept
of link conflict weight defined in CLICA. It is worth noting that
this equivalence exemplifies that two parallel works sometimes
coincidentally define an identical concept using different ways.
With the LCI metric, Tang et al. formulate the CA problem into
the ‘minimum INterference Survivable Topology Control (INSTC)’
problem, which is described as follows. Given the unit disk graph
G, the number of radios Q at each node, the total number of
channels C, and an integer k, seek a channel assignment such
that the resultant network topology T satisfies the following two
objectives: (1) T is k-connected so that the network is survivable
to node/link failures; (2) the interference over the whole network
is minimized in the sense that the maximum LPI for all links in T is
minimized.
It can be seen that the second objective regarding interference
in this formulation is the same as the second objective in CLICA’s
problem formulation, since the LPI concept is equivalent to the
link conflict weight concept. The differences between these two
problem formulations are that (1) INSTC requires all nodes to have
an identical number of radios Q , while CLICA allows a variable
number of radios at each node and (2) INSTC imposes a stronger
connectivity requirement by k-connectedness than CLICA, which
only requires that any link in the unit disk graph is preserved.
Due to the NP-hardness of the INSTC problem, Tang et al.
propose a heuristic polynomial time algorithm (named the INSTC
algorithm) to solve it. The basic idea of the INSTC algorithm is
to assign the channels by traversing the links in a k-connected
subgraph of the given unit disk graph in a predetermined order.
To determine this order, Tang et al. introduce another concept
called Link Potential Interference (LPI) for a link in the unit disk
graph. Simply put, the LPI for a link is defined to be the number
of links that interfere with this link in the unit disk graph by
only considering the interference range. Then, the links in the unit
disk graph is sorted in a non-increasing order of their LPIs, thus
obtaining the predetermined order used by the link traversal. As
a note, LPI is used instead of LCI to determine the traversal order
because the network topology is not known initially and hence the
LCI is not available at the beginning of the algorithm.
The behaviors of the INSTC and CLICA algorithms are similar
in that they both pick a channel for a link in a locally optimal
manner, and are different in that INSTC uses a predetermined
order to traverse the links, while CLICA dynamically adjusts its
traversal order according to the degree of flexibility obtained from
the constantly updated conflict graph.
The main advantage of the INSTC algorithm is that it is proved
to be a polynomial time algorithm that achieves k-connectedness,
thus the algorithm is efficient and the obtained network topology
is robust. Due to its close similarity with CLICA as revealed above,
its limitations are the same as those of CLICA.
(3) Centralized Tabu-based Algorithm (CTA)
While CLICA and INSTC formulate the CA problem into the
problem of assigning channels to links in the unit disk graph,
CTA [39], proposed by Subramanian et al., formulates the CA
problem into the problem of assigning channels to the vertices of
the conflict graph. In the basic version of CTA, its inputs are the
number of radios at each node, the number of channels, and the
conflict graph; and its objective of output is to minimize the total
network interference, which is defined as the total number of edges
connecting the vertices in the conflict graph that are assigned
the same channels. To consider the multiple radios at each node,
the so-called interface constraint is enforced in CTA: the number
of channels assigned to each node cannot exceed the number of
interfaces it has. In the generalization of CTA, the traffic load is also
supported as input.
To solve the above CA problem, CTA translates it into the Max
K-cut problem, which is generally defined as follows in graph
theory: given a graph G, partition the vertices of G into K partitions
such that the number of edges whose endpoints belong to different
partitions is maximized. Thus, if the vertices assigned a certain
channel in the conflict graph is viewed to lie in one partition, the
total network interference is actually the number of edges whose
endpoints belong to the same partitions. In addition, CTA enforces
the interface constraint to the Max K-cut problem.
Stating that the Max K-cut problem with the added interface
constraint is NP-hard, CTA resorts to the Tabu search, one of the
typical heuristic techniques in genetic computing, to solve it in
polynomial time. The main advantage of CTA is that it obtains very
close results to the lower bound of the total network interference as
shown by the authors, and its main limitation is that by assigning
channels to the vertices of the conflict graph, in which only one link
can be modeled between each node pair, CTA fails to consider the
multiple-link case as discussed in Section 2.2.2.
(4) Breadth First Search Channel Assignment (BFS-CA)
As described above, all the previous three algorithms have
difficulty in modeling the number of radios at each node by using
the unit disk graph or the conflict graph. BFS-CA [34], proposed
by Ramachan et al., overcomes this difficulty by introducing the
concept of Multi-radio Conflict Graph (MCG), which explicitly
includes the number of radios at each node by extending the
traditional conflict graph concept. The MCG differs from the
traditional concept in two ways. First, it considers links between
radios as vertices instead of considering links between nodes as
vertices. Second, it models the interference between links based on
Comput. (2009), doi:10.1016/j.jpdc.2009.09.011

ARTICLE IN PRESS
(a) Unit disk graph. (b) Traditional conflict
graph.
(c) Multi radio conflict
graph.
Fig. 7. A comparison of the traditional conflict graph and the MCG.
the unit disk graph instead of the network topology. As BFS-CA works
by assigning channels to the vertices of MCG, the MCG needs to be
available at the beginning of the algorithm. The definition of MCG
(VM , EM ) under the Protocol Model is given below.
• For each pair of radios that is adjacent to a link in the unit disk
graph, there is a vertex representing this link in VM .
• If two links interfere with each other, there is an edge in EM
connecting the two vertices that represent these two links to
indicate this conflict.
To understand this concept, Fig. 7 gives a comparison of the
traditional conflict graph and the MCG. Since the conflict graph
is actually based on the network topology, we assume here that
radios A1, C1, B1 use an identical channel and radio C2 uses a
different channel in (a). Thus, only links AC and BC interfere with
each other in (b). The MCG for (a) independent of the channel
assignment is given in (c), where each pair of radios within
transmission range constitutes a vertex, thus there are altogether
four vertices representing links and four edges representing
conflicts in this MCG. Because radio A1 can only be assigned one
channel, links A1C1 and A1C2 cannot exist simultaneously, so they
do not conflict with each other, and the same applies to links B1C1
and B1C2.
The problem formulation in [34] makes the following assump-
tions on the MR–MC WMN environment: (1) there is a gateway
in the MR–MC WMN and the major network traffic is to or from
the Internet via this gateway; (2) there is a Channel Assignment
Server (CAS) running on the gateway, and its task is to collect in-
formation from all nodes in the network, calculate the CA results,
and distribute the results to all nodes; (3) each node at least has a
default radio operating on a common default channel, which can
carry both control and data packets; (4) there exist co-located ex-
ternal wireless networks that interfere with this MR–MC WMN,
since the MR–MC WMN today mostly use the unlicensed RF bands.
With these assumptions, Ramachan et al. formulate the CA prob-
lem into the problem of assigning channels to the vertices in MCG
such that the following two objectives are satisfied: (1) minimizing
the interference among the mesh routers and (2) minimizing the
interference between the MR–MC WMN and the co-located exter-
nal wireless networks.
The advantages of the BFS-CA algorithm mainly include the
following: (1) a novel concept of Multi-radio Conflict Graph is
proposed, such that it is straightforward to consider the number
of radios at each node in designing a CA algorithm; (2) it is
claimed in [34] that BFS-CA is the first algorithm considering the
external interference, which is a practical problem in the current
MR–MC WMN deployments; (3) the practicality of this algorithm
is demonstrated by a prototype implementation in a multi-radio
IEEE 802.11b testbed. On the other hand, the BFS-CA algorithm
has the following two limitations: (1) as seen from step 4 of
the BFS-CA algorithm, its heuristic to reduce both internal and
external interference by combining the channel ranking and the
MCG constraint is intuitive, providing no known bound for the
worst-case performance; (2) it is only suitable for the MR–MC
WMNs where a gateway acts as the central point of the network
traffic.
(5) Summary to the graph-based approaches
As seen above, the four graph-based algorithms all use a certain
kind of graph to model MR–MC WMNs and assign channels with
a heuristic algorithm. We compare them based on the inputs,
objectives of outputs, and heuristic methods in Table 1.
Besides the above comparison table, we also point out the
advantages and limitations of the graph-based approaches as a
category below.
• Advantage: graph models are intuitive and convenient for
designing the CA algorithms.
• Limitation: by solely modeling the network by vertices and
edges, it will be difficult to consider traffic load, an important
factor for a CA algorithm to improve the network performance.
4.1.2. Network flows approaches
In network flows approaches, an MR–MC WMN is modeled by
a flow network, thus overcoming the aforementioned limitation
of the graph-based approaches. Network flows is a discipline
originated from the 19th century that has applications in
many fields. Here we only describe two basic definitions in
this discipline: flow network and flow, which are essential for
understanding later discussions. For the details of network flows
theory, two classical books [1,10] are recommended.
Specifically, a flow network is a directed graph G (V, E) in which
each link (u, v) ∈ E has a capacity c (u, v) ≥ 0. If (u, v) 6∈ E,
then c (u, v) = 0. There are two special subsets of V: source
set and destination set, denoted by Vs and Vd respectively. A flow
on the flow network G (V, E) is defined as a real-valued function
f : V × V → R that satisfies the following properties:
1. capacity constraint: for all (u, v) ∈ E, 0 ≤ f (u, v) ≤ c (u, v).
This implies that the traffic rate at a link should not exceed the
capacity of this link.
2. skew symmetry: if (u, v) ∈ E and (v, u) 6∈ E, define f (v, u) =
−f (u, v); if (u, v) 6∈ E and (v, u) 6∈ E, define f (u, v) = 0.
3. flow conservation constraint: for all u ∈ V − Vs ∪ Vd,
P
v∈V
f (u, v) = 0. This implies that, for a node u that is neither in
Vs and nor in Vd, the aggregate traffic rate entering u should be
equal to the aggregate traffic rate leaving u.
To formulate the CA problem into a network flow problem, all
the approaches in this category assume the traffic rates (either
end-to-end or at each source) are known. In this subsection,
we examine three such approaches: (1) Load Aware Channel
Assignment (LA–CA) [38], (2) Balanced Static Channel Assignment
& Packing Dynamic Channel Assignment (BSCA&PDCA) [22], and
(3) Joint Routing and Channel Assignment and Link Scheduling
(RCL) [3].
To ease the examination of these three approaches individually,
we first provide some of their similarities and differences. Briefly,
they have the following two common characteristics.
• Their formulated problems are NP-hard, so all of them propose
heuristic or constant-factor approximation methods to solve
their problems.
Comput. (2009), doi:10.1016/j.jpdc.2009.09.011

ARTICLE IN PRESS
Table 1
Comparison of the four graph-based algorithms.
Inputs Objectives of outputs Heuristic methods
CLICA • The unit disk graph Channel assignment such that any link in the
unit disk graph is preserved and the maximum
link conflict weight among all links is
minimized
• Assign channels to the links in
unit disk graph
• Number of radios at each node and total
number of channels
• DFS traversal, considering the
degree of flexibility to adjust the
order
• Interference information described by the
conflict graph
• Pick channels greedily according
to link conflict weight
INSTC • The unit disk graph Channel assignment such that k-connectedness
is achieved and the maximum LCI among all
links is minimized
• Assign channels to the links in
unit disk graph
• Fixed number of radios Q at a node and total
number of channels C
• A k-connected subgraph traversal,
using a predetermined order based
on LPI
• Interference information reflected by LCI and
LPI
to LPI
CTA • Number of radios at each node and total
number of channels
Channel assignment such that the total network
interference is minimized and the interface
constraint is considered
• Assign channels to vertices in the
conflict graph
• The conflict graph • Translated to the Max K-cut
problem with the interface
constraint. And a Tabu search is
used to solve it.
BFS-CA • The unit disk graph Channel assignment such that both internal
interference among mesh routers and external
interference between mesh and co-located
wireless networks are reduced
• Assign channels to vertices in
MCG
• Location of the gateway • BFS traversal, using the distance
from the gateway as the guideline
number of channels
to the MCG and the external
interference level
• Internal interference reflected by the MCG
• External interference measured by mesh
routers
• All of them combine their CA algorithms with routing
algorithms. As explained in Section 2, channel assignment and
routing are mutually dependent. Specifically for the network
flows approaches, the routing algorithms decide the traffic
rate at each link, which affects the CA decisions. On the other
hand, the CA decisions determine the capacity of each link
in the network, which affects the routing decisions. Thus, by
combining these two algorithms, the network flow approaches
aim to improve the network performance.
Though these three approaches all combine the CA and routing
algorithms, they differ in that the first approach uses a loop
consisting of the CA algorithm followed by the routing algorithm to
make these two algorithms gradually converge to the final results,
while the last two approaches jointly address the CA and routing
problems by solving a Linear Program (LP).
(1) Load Aware Channel Assignment (LA–CA)
Pioneering the research on centralized CA approaches, Raniwala
et al. propose two algorithms for solving the channel assignment
problem: Neighbor Partitioning Scheme (NPS) and Load-Aware
Channel Assignment (LA–CA) [37]. The NPS is a very simple
algorithm that partitions the neighbors at each node into n groups,
where n is the number of radios at each node and can vary at
each node. Its goal is that a node communicates with each group
of its neighbors using only one radio and one channel, and the
way of assigning channels reduces the interference in the network.
On the other hand, the LA–CA is a complex algorithm that uses
the network flows formulation. We first briefly describe the NPS
algorithm because, as an early work, it gives the intuition for
solving the CA problem, and then discuss the LA–CA algorithm as
our focus.
In NPS, the CA problem is viewed as two subproblems:
neighbor-to-interface binding and interface-to-channel binding.
Neighbor-to-interface binding determines through which inter-
face a node communicates with each of its neighbors, and
interface-to-channel binding determines which channel is as-
signed to each of the interfaces. To complete these two bindings
at each node, NPS randomly starts with one node, say v1, and
partitions v1’s neighbors into n groups, where each group has an
approximately equal number of neighbors. Then, each group is as-
signed to one of v1’s radios and each radio is assigned a currently
least-used channel in terms of the number of radios using this
channel within the interference range. In turn, each of v1’s neigh-
bors performs the same procedure, while maintaining the follow-
ing constraint: the radio used to communicate with v1 must be
assigned the same channel as the one used by v1 to communicate
with this neighbor. This process is iterated until all nodes have par-
titioned their neighbors. Fig. 8 gives an example channel assign-
ment of a grid mesh network using this algorithm. In this figure,
each node has two radios and four channels are used, numbered 1,
2, 3, and 4 respectively.
Next, the LA–CA algorithm is discussed in-depth, with its
problem formulation summarized as follows:
• The unit disk graph of the network is known.
• In the unit disk graph, the capacity of each link l, which can
use multiple channels concurrently, is determined by (1) the
number of non-overlapping channels assigned to l and (2) l’s
bandwidth share at each of these non-overlapping channels,
where l’s bandwidth share at a channel equals to the bandwidth
of this channel divided by the number of links sharing this
channel within l’s interference range.
Comput. (2009), doi:10.1016/j.jpdc.2009.09.011

ARTICLE IN PRESS
Fig. 8. An example channel assignment of a grid mesh network by neighbor
partitioning.
• The end-to-end traffic rates (i.e., traffic loads) for a set of node
pairs are known a priori.
• The output objective for CA is to ensure that each link’s capacity
is no less than the aggregate traffic load on it from every node
pair.
After formally proving the above formulated CA problem is
NP-hard, the LA–CA algorithm uses a heuristic approach to solve
it. As mentioned previously, channel assignment and routing
are mutually dependent. To break this mutual dependency,
LA–CA starts with an estimation of the initial load on each link
based on the a priori end-to-end traffic rates regardless of the
channel assignment, and then iterates the channel assignment step
followed by the routing step until the output objective stated above
is achieved (see Fig. 9). The basic idea of CA in LA–CA is summarized
as follows: the links are visited in the decreasing order of link
criticality, which is measured by the amount of traffic load on a
link. When a link l is visited, it is greedily assigned a channel that
has the least degree of interference, which is measured by the sum
of the traffic loads on the links that share the same channel within
l’s interference range.
The main advantages of the LA–CA algorithm are that (1) it uses
a simple greedy heuristic such that it is efficient to implement and
(2) its practicality is demonstrated in an IEEE 802.11b testbed. Its
main limitation is that, as acknowledged in [37], it may fail to find a
feasible CA result due to the adoption of a simple greedy heuristic.
(2) Common backgrounds for the next two approaches
Since the next two approaches in this category both resort to an
LP to solve the formulated network flows problems, their common
backgrounds are presented together here.
First of all, both of them incorporate the various CA inputs
and the objective of CA output into an LP. The CA inputs mainly
include two parts: (1) the inputs related to the definition of a flow
such as the end-to-end traffic rates and the capacity constraint and
(2) the inputs arising from the MR–MC WMN environment such
as the node deployment, the number of radios at each node, the
number of channels available to each radio, and the interference
model. And the objective of CA output, such as maximizing the
overall network throughput or minimizing the overall network
interference, is expressed by the objective function of the LP.
Moreover, both of them propose algorithms for jointly solving
the routing, channel assignment, and link scheduling problems in
MR–MC WMNs based on their formulated LPs. Two clarifications
on this statement are given here. First, ‘jointly’ means that the three
problems of routing, CA, and link scheduling are solved within
a single formulation, not that they are solved simultaneously.
Actually it can be seen from the later description that, for both
of them, the routing problem is solved first to produce the
routes for the input flow, and then the channel assignment and
link scheduling problems are solved based on these routes. For
the channel assignment and link scheduling problems, however,
Fig. 9. Flowchart of the combined LA–CA and routing algorithms.
they may be solved either simultaneously or sequentially as to
be described further. Second, in their LP contexts, these three
problems have the following specific meanings:
• Routing: finding routes for the given flow in the given flow
network G (V, E) such that the constraints included in the LP
are satisfied.
• Channel Assignment: assigning channels to each link in E
such that the obtained routes are fulfilled under the capacity
constraint.
• Link Scheduling: finding the schedule of transmissions for links
assigned to the same channel within the interference range,
such that no interference occurs and the traffic rates of all routes
can be accommodated. Note that, to make link scheduling
possible, both approaches assume that all nodes in the network
operate in a time-synchronized manner.
(3) Balanced Static Channel Assignment (BSCA) & Packing Dynamic
Channel Assignment (PDCA)
Kodialam and Nandagopal [22] formulate the joint routing, CA,
and scheduling problem in MR–MC WMNs into a classic network
flows problem — the multi-commodity flow problem [1]. In their
formulation, a set of source nodes transmits to a set of destination
nodes and the traffic between each source/destination pair is
regarded as a commodity, with its end-to-end rate specified by an
element in a rate vector. The LP framework proposed in [22] for
solving this problem is flexible. It can not only specify a variety
of network inputs such as channels with different bandwidth, but
also specify many linear objective functions such as determining
the feasibility of a rate vector, maximizing the overall network
throughput, or imposing certain fairness constraint.
Under this LP framework, the basic steps of the approach
proposed in [22] to jointly obtain the routing, channel assignment,
and link scheduling results are described as follows.
1. Determine the feasibility of the given rate vector and solve the
routing problem.
2. Based on the routes obtained, both static and dynamic
algorithms are proposed to solve the channel assignment and
link scheduling problems. Here ‘static’ means the channel
assignment is only decided once and remains unchanged in
satisfying the rate vector, while ‘dynamic’ means the channel
assignment can be changed at the beginning of certain time
slot to respond to the traffic changes. Because calculating the
optimal CA results is NP-hard, both algorithms use a greedy
heuristic.
Comput. (2009), doi:10.1016/j.jpdc.2009.09.011

ARTICLE IN PRESS
3. The static algorithm is named Balanced Static Channel Assign-
ment (BSCA), which introduces a concept called ‘constraint set’
and tries to minimize the maximum traffic rates on a constraint
set when assigning a channel to a link. After the channel assign-
ment, different time slots are assigned to the links that use the
same channel within the interference range to avoid interfer-
ence.
4. The dynamic algorithm is named Packing Dynamic Channel
Assignment (PDCA), which defines a time period consisting of
n time slots (n >= 1) and performs the channel assignment
at the beginning of each period. PDCA uses a greedy packing-
based heuristic, in which the links are assigned channels in the
descending order of their traffic loads, with each assignment
picking the channel that can currently provide the highest
bandwidth share. Unlike BSCA, where channel assignment and
link scheduling problems are solved sequentially, PDCA solves
these two problems simultaneously.
The main advantage of the BSCA&PDCA algorithms is that the
adopted LP framework is flexible so that it can incorporate a variety
of inputs and objective output functions. And their main limitation
is that both of them use a greedy heuristic, such that there is no
performance bound in the worst case.
(4) Joint Routing, Channel Assignment, and Link Scheduling (RCL)
Overcoming the above limitation of the heuristic BSCA&PDCA
algorithms, Alicherry et al. propose a constant-factor approxima-
tion algorithm named ‘‘RCL’’ to jointly solve the routing, channel
assignment, and scheduling problems [3]. However, the LP frame-
work of RCL cannot incorporate as flexible inputs and objectives
of outputs as that in BSCA&PDCA. Specifically, it considers the fol-
lowing inputs: (1) each node has its own number of radios and
there are altogether K non-overlapping channels; (2) all the net-
work traffic is to or from the Internet, which is represented by a
single node t ∈ V in its flow network G (V, E); (3) each mesh router
u ∈ V has an individual traffic rate to t, represented by l(u). And
its objective of output is to maximize an identical coefficient λ for
every node u such that the multiplied traffic rate λ · l(u) can be
routed through the flow network. Note that λ is used to impose
a fairness constraint, since the main concern of RCL is to achieve
fairness among all the nodes rather than to maximize the overall
network throughput.
While the background for RCL already described together with
BSCA&PDCA, we next summarize the basic steps of RCL as below.
1. Solve LP: To overcome the NP-hardness, this step solves a LP
relaxation of the formulated problem, obtaining the routes in
the form of a flow f : V × V → R.
2. Channel Assignment: This step is based on the observation that
the channel assignment is easy to perform if the number of
available channels is equal to the minimum number of radios
of every node (represented by Imin), since in this case each
node can then be assigned all the channels. Accordingly, this
step exploits a flow network transformation technique: it first
transforms a flow network G into G′
where each node has
only Imin radios by creating multiple copies of each node; then,
it assigns the channels on G′
using only Imin channels out of
the total K channels available; next, it revises the channel
assignment by using the remaining K − Imin channel; finally,
it maps the channel assignment back from G′
to G.
3. Post Processing: Based on CA from the last step, this step
readjusts the flow on the flow network, making efforts to
minimize the maximum interference over all channels.
4. Flow Scaling: This step readjusts the flow and CA to make the
interference-free link scheduling possible.
5. Interference-Free Link Scheduling: This step schedules the links
sharing the same channel within the interference range to use
different time slots, while still capable of carrying the flow.
The main advantage of RCL is that it is the only constant-
factor approximation algorithm proposed among the centralized
approaches that strive to solve an NP-hard CA problem, while
all other approaches adopt a heuristic algorithm providing no
performance bound. And its main limitation is that the adopted LP
framework is inflexible, which can only consider the simple inputs
where the traffic rate from each node to the Internet is specified,
and a single output objective of maximizing λ such that λl(u) for
each node u has a feasible route.
(5) Summary to the network flows approaches
First, we compare the three algorithms surveyed above based
on their inputs, objectives of outputs, and heuristic/approximation
methods in Table 2.
Moreover, the advantages and limitations of the network flows
approaches as a category are summarized as follows.
• Advantages: the traffic load information, which is neglected by
graph-based approaches, is inherently included in the network
flows formulation.
• Limitations: all the three surveyed approaches assume constant
traffic rates, which is not realistic in most practical networks
where the traffic pattern is usually bursty and characterized by
random on/off sources.
4.1.3. Network partitioning approaches
In network partitioning approaches, the CA problem is viewed
as the problem of partitioning the radios and links in an MR–MC
WMN into disjoint subnetworks such that each subnetwork
uses a single channel and different channels are assigned to
these subnetworks to reduce interference. Of all the surveyed
centralized approaches, two approaches fall in this category:
Matroid Cardinality Intersection Channel Assignment (MCI-CA) [8]
and Maxflow-based Centralized Channel Assignment (MCCA) [4].
(1) Matroid Cardinality Intersection Channel Assignment (MCI-CA)
Brzezinski et al. propose MCI-CA in [8], which makes the
following assumptions on the MR–MC WMNs:
• The unit disk graph GN (VN , EN ) of the network is given, called
network graph in [8].
• Each node v ∈ VN has a certain number of radios, and there are
k channels available to the network.
• Interference constraint is given by the conflict graph GI (VI , EI ),
also called interference graph in [8].
• For each link (i, j) ∈ EN , packets arrive in a stochastic pro-
cess with an average rate λij, which is specified in an arrival
rate vector Λ = (λij, (i, j) ∈ EN ). Note that different from
previously-surveyed approaches, MCI-CA uses stochastic pro-
cesses instead of fixed traffic rates to describe the traffic load
information.
As to the network partitioning output, MCI-CA requires the
network topology of each obtained subnetwork to satisfy the
following objective: the interference within this subnetwork due
to the single channel can be avoided by a stable distributed link-
scheduling algorithm. It is known that not all network topologies
can have a stable scheduling algorithm. To find out what kinds of
topologies can have such an algorithm, MCI-CA resorts to a recent
work [12], which gives the sufficient conditions for the ‘maximal
weight scheduling algorithm’ to be a stable algorithm and refers to
these sufficient conditions as the ‘Local Pooling (LoP)’. Note that
the maximal weight scheduling algorithm can be implemented
as a distributed link-scheduling algorithm. To use the maximal
weight scheduling algorithm as the stable scheduling algorithm for
a single-channel subnetwork, MCI-CA proves that a forest topology
satisfies the LoP. Thus, MCI-CA formulates its network partitioning
problem (i.e., the CA problem) as follows: given the above network
model, assign channels to each (i, j) ∈ EN (i.e., partition the MR–MC
Comput. (2009), doi:10.1016/j.jpdc.2009.09.011

ARTICLE IN PRESS
Table 2
Comparison of the three network flows approaches.
Inputs Objectives of outputs Heuristic/Approximation
LA–CA • Node set and link set Channel assignment such that each link’s
capacity is no less than its traffic load
• The links are visited in the decreasing order of their
amount of traffic loads
• Number of radios at each node and
total number of channels
• End-to-end traffic rates of a set of
node pairs
• Greedily assign a channel that has the least degree of
interference
• Protocol Model as interference
model
BSCA & PDCA • Node set and link set Flexible CA objectives: many linear objective
functions related to traffic load and link
capacity can be achieved
• BSCA greedily picks a channel that minimizes the
maximum traffic rates on a constraint set
• PDCA greedily picks a channel that can currently provide
the highest bandwidth share
• A vector of end-to-end traffic rates
model
RCL • Node set and link set, with a
special node t representing the
Internet
Channel assignment that maximizes λ such
that λl(u) for each node u can be routed
• A constant-factor approximation method
• Exploit a flow network transformation technique
• Source traffic rates from each node
to the Internet
model
WMN network into subnetworks), such that each subnetwork is
assigned a single channel and has a forest topology.
MCI-CA solves this problem by reducing it to the matroid
intersection problem. A concept in combinatorics, a matroid is
defined as a pair (E, I), where E is a finite set and I is a set of
subsets of E, with each element in I called an independent set of
E that satisfies the following properties:
• The empty set is independent, i.e., at least one subset of E is
independent.
• Every subset of an independent set is independent.
• If A, B ∈ I and A has more number of elements than B, then
there exists an element e ∈ A and e 6∈ B such that {e} ∪ B is an
independent set.
And the matroid intersection problem is stated as follows: given
two matroids M1 = (E, I1) and M2 = (E, I2), find the largest
independent set A ∈ I1 ∩ I2 of E. For detailed information on
matroids, the text [33] is recommended.
For the matroid intersection problems, they can be solved
optimally by a known polynomial time algorithm called ‘Matroid
Cardinality Intersection (MCI)’ [26]. However, the forest topologies
obtained by MCI vary considerably in size, thus degrading
the overall network throughput. To balance the sizes of the
forests, MCI-CA proposes the following three algorithms to
improve the topologies obtained by MCI and still retain the
forest topology: greedy reallocation algorithm, maximum degree
reallocation algorithm, and average degree reallocation algorithm,
all of which are polynomial time algorithms. Due to the above
solution methodology, we refer to the algorithms proposed
in [8] collectively as the Matroid Cardinality Intersection Channel
Assignment (MCI-CA) for the convenience of reference.
We point out the advantages and limitations of MCI-CA as
follows.
• Advantages: (1) it formulates the CA problem into the matroid
intersection problem, which is a polynomial-time problem,
not an NP-hard problem as in most other approaches, so no
heuristic/approximation method is needed; (2) each forest
obtained by partitioning can use a distributed link-scheduling
algorithm to achieve the stability region.
• Limitations: (1) in case the number of channels is not large
enough to assign each partition a unique channel, the stability
region in some partitions can only be partially achieved; (2) the
partitioning of network topology into forests may not be
optimal in terms of overall network throughput; (3) it needs
to assume that the network operates in a time-synchronized
manner such that the links in a subnetwork can be scheduled
with the maximum weight scheduling algorithm.
(2) Maxflow-based Centralized Channel Assignment (MCCA)
Avallone and Akyildiz consider the MR–MC WMNs where the
gateways are used to relay all network traffic to the Internet [4].
They model such a WMN by an undirected graph GI (V, EI ),
in which V consists of two types of nodes: mesh routers (no
connection to the Internet) and gateways, and EI consists of the
links connecting any two nodes that are within the transmission
range of each other. The objectives of their channel assignment
approach are three-fold: (1) maximizing the throughput flowing
from the mesh routers to the gateways and (2) ensuring the
network connectivity by assigning a common channel on both the
end nodes of every link and (3) the number of channels assigned to
each node does not exceed the number of radios at that node.
The MCCA algorithm is proposed to fulfill these three objectives.
For realizing the first one, MCCA first determines the criticality
of each link, and then gives the links with higher criticalities
more protection against interference when assigning channels. To
decide the link criticality, MCCA constructs a single commodity flow
problem on the initial graph GI . The purpose of solving this problem
is to compute (1) the maximum throughput flowing from the mesh
routers to the gateways and (2) the amount of flow that each
link must carry to achieve this maximum throughput. After the
computation, the amount of flow carried by a link is used as the
metric for the criticality of this link.
For realizing the last two objectives, MCCA splits the channel
assignment into two stages: link-group binding and group-channel
assignment. In the first stage, all links in the network are partitioned
into different groups based on their criticalities. While doing this
partition, it is ensured that the number of groups containing the
links at a node is not larger than the number of radios at this
node. In the second stage, the links of each group are assigned a
Comput. (2009), doi:10.1016/j.jpdc.2009.09.011

ARTICLE IN PRESS
Table 3
Comparison of the two network partitioning approaches.
Inputs Objective of outputs Exploited algorithms
MCI-CA • The network graph (i.e., the unit disk graph) Partition the WMN into subnetworks, such that each
subnetwork is assigned a single channel, and links in
each subnetwork can be scheduled by a stable
algorithm causing no interference
The Algorithm for Matroid
Intersection Problem
number of channels
• The packet arrival rate vector
• The conflict graph
MCCA • The unit disk graph, with the set of gateways (1) Maximizing the throughput flowing from the
mesh routers to the gateways
The Algorithm for Single
Commodity Flow Problem
number of channels
(2) Ensuring the network connectivity
• Protocol Model as interference model (3) The number of channels assigned to each node
does not exceed the number of radios at that node
single channel. The basic idea of this stage is to sort the groups in
the descending order of the maximum criticality of all links in a
group, and then assign the channels to the groups in such an order.
Thus, the groups with high criticality links are considered first and
are most likely assigned channels with less interference. Note that
(1) since every link is considered in the first stage and every group
is assigned a channel in the second stage, the second objective is
automatically fulfilled; and (2) since the number of groups at a
node is not larger than its number of radios and each group is only
assigned a single channel, the third objective is also satisfied.
The main advantage of MCCA is that it naturally realizes its
second and third objectives for CA by splitting the CA task into two
stages. Its main limitation is that it may not achieve the maximum
throughput claimed in its first objective for CA, since it assigns
channels simply based on the criticalities of the links.
(3) Summary to the network partitioning approaches
Due to the simple methodology of the network partitioning
approaches, they do not generate NP-hard problems, so no
heuristic/approximation method is needed. However, they can still
be different in the type of known polynomial-time algorithm being
exploited. Accordingly, we compare the two algorithms surveyed
above based on their inputs, outputs, and exploited algorithms in
Table 3.
The advantages and limitations of the network partitioning
approaches as a category are summarized as follows.
• Advantage: the network partitioning methodology is straight-
forward and simple, resulting in polynomial-time problems in-
stead of NP-hard problems.
• Limitation: the approaches in this category are not flexible,
since all links in a partition are fixed to a common channel;
consequently, they cannot optimally achieve certain objectives
such as maximizing the throughput or minimizing the interfer-
ence in the network.
4.2. Distributed CA approaches
Since the distributed CA approaches involve communication
and coordination among multiple parties, they are more challeng-
ing to design than their centralized counterparts. In all the eight
distributed CA approaches surveyed in this paper, each node mea-
sures local channel information and exchanges it with other nodes
to calculate the channel assignment. However, these approaches
differ in their choices of the local channel information. For instance,
these choices can be the number of links sharing a common chan-
nel within the interference range, the traffic load on a channel, the
signal to interference and noise ratio on a channel, or a combi-
nation of them. For convenience, we refer to these choices as CA
metrics hereafter. Analogous to the routing metrics that serve as
the basis for routing, the CA metrics also play an equivalent role
in making the CA decisions. Therefore, when each CA approach is
examined below, special attention is paid to its CA metrics.
According to Section 3, the distributed CA approaches are
classified into two categories: gateway-oriented approaches and
peer-oriented approaches, which are in turn discussed below.
4.2.1. Gateway-oriented approaches
Gateway-oriented approaches assume a certain number of
gateways as the central sinks of the major network traffic. In this
subsection, three such approaches are examined: Hyacinth [36],
DMesh [11], and CoMTaC [31].
(1) Hyacinth
Pioneering the research in distributed CA approaches, Rani-
wala and Chiueh propose an MR–MC WMN architecture called ‘Hy-
acinth’ [36], where the existence of a gateway is assumed and most
network traffic is to or from the Internet via this gateway. In Hy-
acinth, both distributed CA and routing protocols are proposed, and
the CA protocol is based on the topology established by the routing
protocol, which finds paths for all network nodes to the gateway by
establishing a tree topology rooted at the gateway.
The distributed CA protocol originates from the centralized
NPS algorithm (see Section 4.1.2) proposed by the same authors,
and also views the CA problem as two subproblems: neighbor-to-
interface binding and interface-to-channel binding. In neighbor-
to-interface binding, each node’s network interface cards (NICs)
are divided into two classes: (1) UP-NICs that are used to connect
to the parent and (2) DOWN-NICs that are used to connect to the
children. With this division, the binding of neighbors and interfaces
is determined. In interface-to-channel binding, each node only
needs to assign channels to its DOWN-NICs, and mandates each
of its UP-NICs to use the same channel as the one used by its
parent’s corresponding DOWN-NIC (see Fig. 10). In order to select
a channel for a DOWN-NIC, each node measures the traffic loads
on all its channels periodically and exchanges this information
via the CHNL-USAGE packets with other nodes in its interference
range. Then, each node can calculate the metric for making the
CA decisions — the total load of a channel, which is a weighted
sum of two measures: (1) the total number of links using this
channel within the interference range and (2) the aggregate traffic
load on this channel from all links using this channel within
the interference range. After calculating the total loads of all its
channels, a node orders these channels according to their total
loads, thus obtaining the least-used channel.
Besides determining the least-used channel, every node is also
given a priority to differentiate their privileges to use channels.
Specifically, the priority of a node is defined as its hop distance
to the gateway, with a smaller number of hops indicating a
higher priority. Combining the total load metric and the priority
mechanism, the CA operation in Hyacinth works as follows: in
Comput. (2009), doi:10.1016/j.jpdc.2009.09.011

ARTICLE IN PRESS
Fig. 10. An example of CA in Hyacinth.
assigning a channel to a DOWN-NIC, a node always chooses the
least-used channel that is not currently used by a higher priority
node within the interference range. The objectives of this CA
operation are twofold: (1) reducing the local interference at a
node and (2) offering the near-gateway links higher priorities for
choosing channels to obtain higher bandwidth.
As a note, the rationale for the two-step bindings in Hyacinth
is to prevent the ripple effect. By mandating that each node is
only responsible for the channel assignment at the DOWN-NICs,
the propagation of consecutive channel switching is blocked. We
view this neat solution of a ripple effect as a major advantage of
Hyacinth. And the major limitation of Hyacinth is that when a new
node joins the network, a long period of a channel scanning process
is needed to discover the DOWN-NIC channels of its potential
parents. This process has a length from 5 to 10 s as mentioned in
Hyacinth, during which a new node broadcasts HELLO packets at
all channels to perform the discovery.
(2) DMesh
While Hyacinth pioneers the gateway-oriented CA approaches,
DMesh [11], proposed by Das et al., constitutes a further
advance within this category by introducing directional antennas.
Specifically, DMesh is an MR–MC WMN architecture where each
node has two types of radios: (1) one radio with an omni-
directional antenna to transmit control packets as well as data
packets in case of fault recovery and (2) multiple additional radios
with directional antennas to transmit data packets. The goal of
DMesh is to exploit both spatial separation by directional antennas
and frequency separation by non-overlapping channels to improve
the network capacity.
To be cost-effective, DMesh does not adopt the directional an-
tennas with perfect spatial separation due to their high price. In-
stead, the directional antennas in DMesh have an interference
range in the shape of a cone, and a 45◦
degree of beamwidth is
assumed in the simulations. Thus, a CA protocol is still needed
to reduce the interference within the interference cones. In
general, the CA in DMesh requires every node to maintain a
‘channel map’ composed of the following three state vectors:
(1) Channel Vector: records whether it uses each channel, (2) Rate
Vector: records its traffic rates at each channel, and (3) Destination
Vector: records the directional information of the neighbor nodes
that it communicates with at each channel. And every node trans-
mits this channel map to its interfering neighbors restricted by di-
rectionality, thus exchanging both the traffic and the directional
information.
Similar to Hyacinth, the CA in DMesh also stipulates that each
node is only responsible for assigning channels to its children. In
performing such an assignment, four directional CA schemes are
proposed. We give their common basic idea as follows. When a
node X attempts to assign a channel to one of its children Y, it
first looks for a free channel that is not used by any node whose
interference cone contains X or Y. If such a channel exists, this
channel is assigned to Y. Otherwise, X looks for a channel that is the
least-loaded in terms of the aggregate traffic rate on this channel
(or simply, aggregate number of traffic flows) and assigns it to Y.
The advantages and limitations of DMesh are pointed out
below.
• Advantages: (1) both spatial separation and frequency separa-
tion are exploited to reduce interference, such that a significant
throughput gain, contrasted with an omni-directional solution
(termed ‘OMesh’ in DMesh), is achieved as shown in their simu-
lations; (2) the practicality of DMesh is validated with a 16-node
testbed that uses commercial directional antennas.
• Limitations: as mentioned in DMesh, the directions of the
antennas are manually positioned during network deployment
and remain unchanged during network operation. Thus, a CA
protocol that considers steerable directional antennas is more
appealing.
(3) Cluster-based Multipath Topology control and Channel assignment
(CoMTaC)
As described in Section 2.2.2, the change of channel assignment
can change the network topology, thus the network may become
disconnected. To solve this issue, Naveed et al. propose an all-in-
one scheme called the CoMTaC [31] to address the topology control
and channel assignment in a two-phase fashion.
The first phase of CoMTaC focuses on the topology control and
runs during the network startup. It consists of two steps. In the first
step, the network nodes are grouped into clusters based on their
hop distances to the set of gateway nodes in the network. Within
each cluster, a default channel is used by all constituent nodes on
one of their interfaces called the default interface to provide the
intra-cluster connectivity. And a node bordering several clusters
has its second interface tuned to the default channel of a second
cluster, thus providing the inter-cluster connectivity. In the second
step, certain links other than those established in the first step are
also established on the non-default interfaces of the nodes, such
that the network connectivity is enhanced.
The second phase of CoMTaC focuses on the channel assignment
and runs periodically during the network operation. Basically,
this phase assigns channels to the links in the network topology
established in the first phase, aiming to minimize the interference.
It has two parts: (1) assigning channels to the default interfaces
and (2) assigning channels to the non-default interfaces. For both
parts, the channel assignment is preceded by a process of local
interference estimation, which associates a cost with each channel.
The cost is used as the metric to minimize the interference locally.
In assigning channels to the default interfaces, only the inter-
ference from external sources is considered. For this external in-
terference estimation, each node passively monitors the following
two parameters of a channel in a periodical manner: channel uti-
lization and channel quality. Basically, the former is measured by
capturing the traffic load on all available channels, and the latter
is measured by accessing the bit error rate and the received signal
strength from the lower-layers. Then, each node sends these two
parameters to the cluster head, which calculates the cost for each
channel by combining these two parameters. Finally, the cluster
head selects the channel with the least cost as the default channel,
and informs it to the other nodes in this cluster.
In assigning channels to the non-default interfaces, the
interference of each channel is estimated by the average link-layer
queue length of the interfaces on this channel. This method is
motivated by the following two facts: (1) a large link-layer queue
length of an interface is indicative of the high interference on
the channel being used and (2) the operation of accessing the
packet queue length from the link-layer is simple and efficient.
To use this estimation method, each node periodically retrieves
Comput. (2009), doi:10.1016/j.jpdc.2009.09.011

ARTICLE IN PRESS
Table 4
Comparison on the CA metrics of the gateway-oriented approaches.
Metric name Basic channel information included Objective
Hyacinth Total load (1) The number of links using this channel within this node’s
interference range
Choose the least-used channel that is not currently
used by a higher priority node within the
interference range
(2) The traffic load on this channel within this node’s
interference range
DMesh Channel map (1) channel vector Choose the channel that is the least-loaded in
terms of the aggregate traffic rate on this channel
(2) Rate vector
(3) Destination vector
CoMTaC Average link-layer queue
length
The average link-layer queue length of all interfaces operating on
this channel within a cluster
Assign the currently least-cost channel that is still
available to the interfaces according to the order of
their priorities
(a) Initially. (b) After switching.
Fig. 11. An example of protocol operation with fixed interfaces and switchable interfaces.
and transmits the link-layer queue lengths of all its interfaces to
the cluster head. Then, the cluster head calculates the average
queue length within the cluster for each available channel and
associates a cost to that channel based on the average value. With
these channel costs, the cluster head assigns the channels to the
non-default interfaces of all cluster nodes. The basic operation
of this assignment is to prioritize these non-default interfaces,
and then assign the currently available least-cost channel to these
interfaces according to the order of their priorities. When deciding
the priorities for these interfaces, the following three types of
interfaces are given higher priorities: (1) interfaces of the border
nodes, (2) interfaces with a greater number of neighbors, and
(3) interfaces with a smaller hop-distance to the gateway.
The main advantages of CoMTaC are that (1) the network
connectivity is ensured by dealing with topology control and
channel assignment separately in a two-phase fashion and (2) the
interference estimation for a channel by accessing the link-layer
queue length is highly efficient. And the main limitation of CoMTaC
is that the channel assignment is restricted to the links in the
network topology established during the network startup, so it fails
to consider all the possible links in the network.
(4) Summary to the gateway-oriented approaches
As seen in the above discussion of the three gateway-oriented
CA approaches, the CA metrics play an important role in assigning
channels, so we compare the CA metrics adopted by them in
Table 4.
Moreover, we highlight the advantages and limitations of the
gateway-oriented CA approaches as follows.
• Advantage: the opportunity to utilize the gateway nodes to
simplify the CA approach.
• Limitation: the incapability of accommodating other kinds of
traffic patterns.
4.2.2. Peer-oriented approaches
Unlike the gateway-oriented approaches, the peer-oriented ap-
proaches do not make any assumptions on the traffic pattern, so
they are general to accommodate various kinds of network traffic.
In this subsection, five such approaches are examined: (1) Proba-
bilistic Channel Usage based Channel Assignment (PCU-CA) [25],
(2) Joint Optimal Channel Assignment and Congestion Control
(JOCAC) [30], (3) Superimposed Code based Channel Assignment
(SC-CA) [43], (4) Distributed Greedy Algorithm (DGA) [39], and
(5) Self-Stabilizing Channel Assignment (SS-CA) [23].
(1) Probabilistic Channel Usage based Channel Assignment (PCU-CA)
To prevent the ripple effect, PCU-CA [25], proposed by Kyasanur
et al., classifies the interfaces at a node into two classes: fixed in-
terfaces and switchable interfaces. The channels used by the former
are changed periodically by the CA protocol but stay fixed during
each period, while the channels used by the latter can change at any
time according to the network traffic. Note that the channels used
by the fixed interfaces are simply called ‘fixed channels’ in PCU-CA.
The basic idea of this interface classification is that the fixed
interfaces are used to receive network traffic by staying at fixed
channels, while the switchable interfaces are used to transmit
network traffic by switching to the channels of the fixed interfaces.
Of course, if the fixed interfaces at two nodes happen to share the
same channel, they communicate with this fixed channel. Because
the channels of every node’s fixed interfaces stay unchanged in
each period, the ripple effect is blocked. An example of protocol
operation with the above idea is illustrated in Fig. 11, where each
of the three depicted nodes A, B, and C exactly has one fixed
interface (indicated by a gray shade) and one switchable interface.
Initially, the fixed interfaces of A, B, and C use channel 1, 2, and
3 respectively. At a certain time, assume A has traffic to send to
C via B. To accomplish this transmission, A’s switchable interface
(originally at channel x) switches to channel 2 so as to transmit to B,
and then B’s switchable interface (originally at channel y) switches
to channel 3 so as to transmit to C, resulting in Fig. 11(b).
Corresponding to the two classes of interfaces, the CA protocol
proposed in [25] has two parts: assigning channels to the fixed
interfaces and switching channels at the switchable interfaces. The
CA metric used in the first part is called the channel usage, which
is measured by the number of nodes using this channel within a
two-hop neighborhood (i.e., the assumed interference range). In
the protocol operation of the first part, each node maintains two
data structures: a NeighborTable that records the fixed channels
of its one-hop neighbors, and a ChannelUsageList that records the
channel usage of each channel. With these two data structures, the
basic ideas of CA for the fixed interfaces are summarized as follows.
1. Each node periodically broadcasts Hello packets on all channels
to its one-hop neighbors. A Hello packet mainly contains the
fixed channels and the current NeighborTable of the sender.
2. Based on the Hello packets received from its neighbors, each
node updates its NeighborTable and ChannelUsageList.
3. According to its ChannelUsageList, a node decides whether the
number of other nodes using one of its fixed channels is large (a
Comput. (2009), doi:10.1016/j.jpdc.2009.09.011

ARTICLE IN PRESS
configurable criterion). If so, it changes this channel to a less-
used channel with a probability 0.4 (also configurable), thus
preventing the channel oscillation, an undesirable phenomenon
described in Section 2.
The second part of the CA protocol requires each node to maintain
packet queues for each channel with packets to transmit. When
a node has packets to transmit to a neighbor, it looks up the
NeighborTable to obtain the fixed channel of this neighbor, and
then inserts the packets to the corresponding channel queue.
The fixed interface is responsible for transmitting packets in the
queue of the fixed channel, while the packets in other queues are
transmitted by the switchable interfaces. Thus, the channels used
in transmission actually depend on the channels assigned to the
fixed interfaces, so the first part of the CA protocol determines
the channel assignment. Since the first part CA protocol assigns
channels based on the channel usage metric in a probabilistic
manner, we refer to the entire CA protocol, which is not named
in [25], as Probabilistic Channel Usage based Channel Assignment
(PCU-CA) for the convenience of later reference.
The advantages and limitations of PCU-CA are given as follows.
• Advantages: (1) both ripple effect and channel oscillation are
tackled, thus stabilizing the network; (2) the ‘channel usage’,
the number of nodes sharing this channel within two hops, is
easy to obtain, thus simplifying the protocol.
• Limitations: (1) despite prevention of channel oscillation, the
probabilistic channel switching may not be optimal for the
network performance; (2) the traffic load information is not
considered, so the criticality of links is ignored in assigning
channels.
(2) Joint Optimal Channel Assignment and Congestion Control (JOCAC)
Different from all other CA approaches surveyed in this pa-
per, JOCAC [30], proposed by Mohsenian et al., views the CA prob-
lem from a novel perspective: optimizing the performance of TCP
congestion control. The rationale for this idea is as follows: if CA
does not perform well, interference will cause severe congestion in
wireless links, which degrades the TCP performance significantly
due to the AIMD algorithm adopted by TCP congestion control. Ac-
cordingly, JOCAC formulates the CA problem into the Joint Optimal
Channel Assignment and Congestion Control problem and names
its proposed algorithm to solve this problem the JOCAC algorithm.
Basically, the JOCAC problem is a distributed utility maximization
problem for congestion control with the interference modeled by
the Physical Model. Below, some background on the JOCAC prob-
lem is given first, and then the JOCAC algorithm is summarized.
The distributed utility maximization for congestion control is a
general analytic model proposed in [21], which views the network
as a set of L links, with each link having capacity cl (l ∈ 1, 2, . . . , L).
There are S sources in the network, and each source s has a
transmission rate rs (s ∈ 1, 2, . . . , S) and a utility function Us(rs).
To model the congestion control feedback, each link is associated
with a congestion price λl, and each source is assumed to have
access to the aggregate price qs(rs) of all the links in its route
to the destination. Each source s tries to maximize its profit:
Us(rs) − qs(rs), while the role of congestion price is to adjust
the behavior of each source such that the aggregate utility of all
sources (
P
s∈S Us(rs)) is maximized, subject to the link capacity
constraint that the aggregate traffic rate from all sources at each
link does not exceed the capacity of this link. A key restriction of
the above model is that each source or link only needs to measure
its local information, such that the model can be implemented
in a distributed way. Up to now, many analytical works on TCP
congestion control have been done using this model. Us(rs) and λl
for well-known TCP congestion control versions Reno and Vegas
are well defined, as summarized in the review paper [27].
Based on the above analytic model and existing definitions
of Us(rs) and λl for TCP Vegas, JOCAC extends these results by
incorporating into them the MR–MC WMN parameters such as the
number of available channels, number of radios at each node, and
SINR at each channel. Thus, the new mathematical expressions
of Us(rs) and λl in the MR–MC WMN environment are obtained,
and the JOCAC algorithm that maximizes
P
s∈S Us(rs) subject to
the link capacity constraint is derived within the new formulation.
The basic behavior of JOCAC’s distributed implementation is
summarized as follows: each node periodically measures and
exchanges with other nodes the local information that mainly
includes the SINR and λl, and then calculates the CA result using
the JOCAC algorithm based on the information measured locally
and received from other nodes.
As mentioned in [30], the distinguishing advantage of JOCAC
is that it can utilize not only the non-overlapping channels but
also the partially overlapping channels.3
This is due to its modeling
of interference using SINR, which means that if the SINR is larger
than the capture threshold, the packets can be received correctly.
That is, the partially overlapping channels can be used, as long as
the interference caused by the overlapping frequency bands is not
significantly strong. And its limitation is that it introduces a large
amount of control traffic in its distributed implementation: each
node periodically exchanges information such as SINR and λl with
all other nodes, since the calculation of CA at each node requires
this information along the route for each TCP flow. In this sense,
the distributed implementation of JOCAC is similar to the link state
routing protocols, where the link states of each node are flooded to
the entire network.
(3) Superimposed Code based Channel Assignment (SC-CA)
Xing et al. propose the SC-CA [43], which applies the
superimposed code theory to the channel assignment for mesh
networks. Established in 1964, the theory of superimposed code
has been widely applied in computer science since then. Basically,
a superimposed code is a special kind of binary matrix in which
the Boolean sum of its columns (called codewords) satisfies certain
properties, and an s-disjunct code is one category of superimposed
code that the Boolean sum of any s codewords in it does not cover
any codeword not in that set of s codewords.
The network model of SC-CA is as follows: (1) the N available
channels are classified into two categories: primary channels and
secondary channels; (2) a binary N-element vector cu (called a
channel codeword or simply a codeword) is associated with a node
u to label its channels, with value 1 indicating a primary channel
and value 0 indicating a secondary channel; (3) the MR–MC WMN
is represented by G(V, E, C), where V is the set of nodes, E is the
set of links, and C is a N ×|V| binary matrix (called a channel code or
simply a code) in which each column corresponds to the codeword
of a node in V. We next summarize the basic ideas of SC-CA as
below:
• Given a network G, an s-disjunct code C can be found for G.
Before starting, each node in V is allocated a unique codeword
from C.
• Each node broadcasts its codeword to its neighbors, and
calculates the multiple channels for itself based on the
codewords of its own and received from others.
• By exploiting the properties of s-disjunct code, a node always
tries to select a channel that minimizes the interference in its
neighborhood. The main heuristic for achieving this is to only
choose channels from its primary channels that are not being
used by its interferers.
3 For instance, in FCC regulation for IEEE 802.11b, there are 3 non-overlapping
channels out of the 11 partially overlapping channels.
Comput. (2009), doi:10.1016/j.jpdc.2009.09.011

ARTICLE IN PRESS
Table 5
Comparison on the CA metrics of the peer-oriented approaches.
Metric name Basic channel information included Objective
PCU-CA Channel usage The number of links using this channel within the two-hop
neighborhood of a node
According to the ChannelUsageList, a
node periodically switches to a
less-used channel with probability 0.4
(configurable).
JOCAC SINR and congestion price SINR and congestion price measured locally and received
from other nodes
Maximize the utility function for the
congestion control
SC-CA Codeword An N-element binary vector that labels a node’s primary or
secondary channels, where N is the total number of
available channels.
Minimize the network interference in
a node u’s neighborhood mainly by
choosing u’s primary channels not
being used by u’s interferers.
DGA local network interference The number of edges whose endpoints are assigned one
channel in the conflict graph within the m-hop
neighborhood of a node. (m is configurable)
Minimize the local network
interference without violating the
interface constraint
SS-CA Interference cost The interference level between two possibly overlapping
channels measured by their spectral distance
Minimize the sum of the interference
costs within a node’s interference
range
With the ideas above, the following two algorithms are
proposed in SC-CA.
• Algorithm 1: output the multiple channels assigned to a node
u. This algorithm is mainly designed for the broadcast from u,
but the channels obtained can also be used in the unicast from
u. SC-CA proves that if for the given s-disjunct code C, ∀u ∈ V,
s is not less than the number of interferers of u, this algorithm
guarantees interference-free communications.
• Algorithm 2: output the channel assigned to a directed link
(u → v). This algorithm is only designed for the unicast on a
link (u → v). SC-CA proves that if for the given s-disjunct code
C, ∀u ∈ V, s is not less than the number of one-hop neighbors of
u, this algorithm guarantees interference-free communications.
The main advantages of SC-CA are that (1) both Algorithm 1
and 2 only require codewords information from at most two-hop
neighbors, so they are localized and involve low communication
overhead, and (2) in certain conditions described above, both
algorithms can achieve interference-free communications, and
(3) the channel assignment for broadcast is addressed in algorithm
1, which is overlooked by most of other CA approaches. The main
limitations of SC-CA are that (1) though it is theoretically sound,
it models the interference basically by the Protocol Model, which
is not very realistic and thus limits its use in practice, and (2) the
performance of SC-CA is not compared with other CA approaches.
(4) Distributed Greedy Algorithm (DGA)
The problem formulation of DGA is the same as CTA, which
is surveyed in Section 4.1.1. Both algorithms are proposed by
Subramanian et al. in [39]. Recalling from Section 4.1.1, the CA
problem is translated to the Max-K-cut problem with the added
interface constraint in [39]. To solve it in a distributed manner, in
which the global information is not available at an individual node,
DGA resorts to a greedy heuristic that only uses local information.
The basic idea of this heuristic is as follows. Let u denote the vertex
in the conflict graph corresponding to the link (i, j), k denote a
channel in the set of all channels, and m denote the number of hops,
a configurable parameter. u is said to be owned by node i or node j,
whichever has the higher node ID. At an arbitrary node i, it selects
a (u, k) pair based on the channels used in its m-hop neighborhood,
such that (1) the pair (u, k) results in the largest decrease in the local
network interference of its m-hop neighborhood and (2) assigning
k to u does not violate the interface constraint. Also recalling from
Section 4.1.1, the local network interference is defined as the
number of edges whose endpoints are assigned one channel in the
conflict graph within the m-hop neighborhood of a node.
The main advantages of DGA are that (1) the distributed exe-
cution of DGA is proved to guarantee stabilization and (2) despite
the simple greedy heuristic with only local information, DGA still
obtains very close results to the lower bound of the total network
interference as shown by the authors. The main limitation of DGA
is that it is not very scalable, since the total number of iterations
for it to stabilize is in the order of O(VC · K), where VC is the total
number of vertices in the conflict graph and K is the total number
of available channels. Representing the total number of links in the
network, VC will become a very large number when the total num-
ber of nodes in the network increases.
(5) Self-Stabilizing Channel Assignment (SS-CA)
The design of SS-CA [23] is inspired by its authors’ previous
theoretical work on a distributed self-stabilizing protocol for
replica placement [24], of which the main idea is to place the
replicas ‘‘far’’ from each other. In adapting this replica placement
approach to the channel assignment problem, SS-CA views the
channels as the replicas, and more attention is given to the practical
issues that are ignored in [24].
Instead of assuming orthogonal channels, SS-CA supports
overlapping channels in the mesh network, and measures the
interference between any two channels by the interference cost
function f (a, b), or simply interference cost. Basically, f (a, b)is only
concerned with the spectral distance between a and b, and is not
affected by the spatial distance between the two nodes using a and
b, as long as they are within the interference range of each other.
In the testbed implementation of SS-CA, f (a, b) is defined to have
a linear relationship with the spectral distance of a and b, taking
value 0 when a and b does not overlap, and reaching its maximum
when a and b are identical.
Using the interference cost as the channel metric, SS-CA designs
a channel selection algorithm executed at a single node. In this
algorithm, a node greedily selects a channel that minimizes the
sum of the interference costs within this node’s interference range.
For the stabilization of this algorithm, when a node changes its
channels, this node should have the correct information about
other nodes’ current channels. However, this condition may not be
true in the real networking environment where there exist delays
in exchanging messages. Based on the aforementioned theoretical
work [24], SS-CA solves this coordination problem by designing a
stabilization-guarantee protocol that (1) runs the channel selection
algorithm at a node and (2) exchanges channel information among
the nodes within the running node’s interference range and
(3) exploits distributed mutual-exclusion operations to coordinate
the channel changes. For the details of this provably self-stabilizing
protocol, SS-CA refers the readers to [24].
The main advantages of SS-CA are that (1) it involves a low-
overhead since each node in it greedily selects channels using a
Comput. (2009), doi:10.1016/j.jpdc.2009.09.011

ARTICLE IN PRESS
simple local objective function — the interference cost function and
(2) the self-stabilization of the channel assignment is guaranteed
by the proposed algorithm and protocol and (3) a real 14-node
testbed implementation is conducted to validate it. The main
limitation of SS-CA is that in the interference cost function, only
the spectral distance and no spatial distance is considered in
calculating the interference among the two possibly overlapping
channels, which is not accurate in reality.
(6) Summary to the peer-oriented approaches
We compare the CA metrics adopted by the five peer-oriented
approaches in Table 5.
Moreover, we highlight the advantages and limitations of the
peer-oriented CA approaches as follows.
• Advantage: the capability of adapting to various kinds of traffic
patterns, making it applicable for most of the current WMNs,
where the peer-to-peer traffic is also a major part of the total
network traffic.
• Limitation: the difficulty of dealing with CA issues such as fault
tolerance, ripple effect, and channel oscillation without any
assumption on the traffic pattern.
4.3. Summary to all CA approaches
In the previous two subsections, we have surveyed state-of-the-
art centralized and distributed approaches, with their advantages
and limitations identified each. As an overall summary to these
approaches, we next present a comprehensive table that compares
and contrasts them based on ten basic properties (see Table 4).
Reflecting which approach is suitable or unsuitable in which
condition, these properties are listed as follows, with their labels
indicated within parentheses:
1. A default common channel is used to transmit control
messages. (DeCh)
2. Traffic load information is considered. (TrLd)
3. The existence of gateway nodes is required to facilitate CA.
(GW)
4. The Physical Model is adopted as the interference model.
(PhMo)
5. Ripple effect is addressed. (RpEf)
6. Channel oscillation is addressed. (ChOs)
7. Routing scheme is proposed in combination with CA. (Rtng)
8. Fault tolerance is supported. (FaTo)
9. Fairness is supported. (Fair)
10. Real testbed implementation is conducted. (Tbed)
The selection of these properties is justified as follows. For property
1, a default common channel provides a convenience to address
the connectivity, stability, and fault tolerance issues, but it incurs
hardware overhead and consumes a frequency band. For property
2, considering traffic load information enables the CA results to
adapt to the current network situation, but makes the CA operation
complex. For property 3, exploiting gateway nodes eases the CA
design at the price of sacrificing generality. Properties 4–9 are
selected from the key design issues identified in Section 2, with
their meanings described and their importance justified there.
Property 10 is concerned with real testbed implementation, which
shows the practicality of the corresponding CA approach. Note that
two key design issues identified in Section 2 are not included in the
above properties (i.e., connectivity and throughput/latency), since
they are imperative for any CA approach to address, thus omitted.
In Table 6, the names of the distributed approaches are shown
in italic, so as to be distinguished from the centralized ones. If an
approach satisfies/dissatisfies a property, the corresponding table
entry is marked by Yes/No respectively. If a property cannot be
applied on an approach, the corresponding table entry is marked
by N/A (Not Applicable). For example, due to the static nature of
the centralized approaches, there are no ripple effect and channel
oscillation issues in them, so all the corresponding entries are
marked by N/A.
This table provides the following additional understandings to
the current CA approaches besides the summarizations given in the
previous subsections.
• Properties 4 and 6 are only addressed by a few approaches,
which shows that the two corresponding issues – Physical
Model and channel oscillation – receive very limited attention
from the current CA research. Accordingly, these two issues
are further discussed in the next section about future research
directions.
• Property 8 is supported by all of the distributed approaches
and none of the centralized approaches, which shows that the
distributed approaches can adapt to the network dynamics
while the centralized ones do not have such capability.
• According to the criterion on fairness for CA described in
Section 2, if an approach explicitly shows the support of fairness
or considers the traffic load while performing CA, we regard it
as supporting fairness, otherwise not. As a result, more than
half the approaches provide such fairness support, showing
that fairness has received adequate attention from the current
approaches.
• Real testbed implementation is conducted by five approaches,
showing that the current research efforts are increasingly
attaching importance to the practicality of their proposals.
This trend is especially beneficial for the protocols in wireless
environment, where the modeling of physical channels is still
immature and only real implementations allow the protocols
to be tested upon various channel uncertainties.
For the two broad categories of CA approaches (centralized and
distributed), we also summarize their advantages, limitations, and
application scenarios in Table 7.
5. Future research directions
In this section, we first foresee the general research directions
for CA, and then pinpoint some specific research directions.
For the general research directions, we first argue that more
research effort should be put into the distributed CA approaches,
because (1) the distributed approaches are needed by the practical
MR–MC WMN networks, which exhibit considerable network
dynamics such as nodes/links failure, traffic load changes, and
external interference and (2) all of the specific future research
issues to be discussed below require further attention from the
distributed approaches. A second general research direction is to
consider a more realistic physical layer in the CA design. As shown
in Table 6, most current CA approaches only consider the Protocol
Model as the interference model, which is not very realistic. We
believe that using the signal strength instead of the predetermined
interference range to decide the channel status is a more desirable
method in future. Two signal strength metrics obtainable from
the physical layer are the RSSI (Receive Signal Strength Indicator)
suggested by the measurement study [13] and the SINR used in the
JOCAC [30].
Besides the above general directions, we also highlight the
following specific issues, which are not fully addressed or remain
unaddressed by the current CA approaches.
• External Interference: since the communication in MR–MC
WMNs uses the IEEE 802.11 and 802.16 standards that operate
in the unlicensed RF bands, there is no guarantee that other
external wireless sources do not use the same bands. In some
of the current MR–MC WMN deployments [5,32], the status of
all employed channels is constantly monitored by the radios at
Comput. (2009), doi:10.1016/j.jpdc.2009.09.011

ARTICLE IN PRESS
Table 6
Comparison of all CA approaches.
1. DeCh 2. TrLd 3. GW 4. PhMo 5. RpEf 6. ChOs 7. Rtng 8. FaTo 9. Fair 10. Tbed
CLICA No No No No N/A N/A No No No No
INSTC No No No No N/A N/A Yes No No No
CTA No Yes No No N/A N/A No No No No
BFS-CA Yes No Yes No N/A N/A No No No Yes
LA–CA No No Yes No N/A N/A Yes No No Yes
BSCA & PDCA No Yes No No N/A N/A Yes No Yes No
RCL No Yes Yes No N/A N/A Yes No Yes No
MCI-CA No Yes No No N/A N/A No No Yes No
MCCA No Yes Yes No N/A N/A No No No No
Hyacinth No Yes Yes No Yes No Yes Yes Yes Yes
DMesh Yes Yes Yes No Yes No Yes Yes Yes Yes
CoMTaC Yes Yes Yes No Yes No No Yes No No
PCU-CA No No No No Yes Yes Yes Yes Yes No
JOCAC No Yes No Yes No No No Yes Yes No
SC-CA No No No No Yes Yes No Yes Yes No
DGA No Yes No No No No No Yes No No
SS-CA Yes No No No No No No Yes No Yes
Table 7
Advantages and limitations of the centralized and distributed approaches.
Advantages Limitations Application scenario
Centralized Capable of getting the optimal or near-optimal CA
results, due to the availability of entire network
information.
Assume stable nodes/links and static traffic pattern,
due to the difficulty of gathering and distributing
information globally in an MR–MC WMN.
During network deployment
or maintenance stages
Distributed Capable of quickly adapting to network changes and
failures, because of only relying on the local
information.
CA results may be far from global optimality, due to
the use of only partial information.
During network operation
stage
each node, and if external interference is detected by a radio, the
channel assignment mechanism at the node can dynamically
switch this radio to a new channel. In the surveyed approaches,
only [34] considers the external interference. Unfortunately,
it can only detect the interference from external wireless
networks using the IEEE 802.11 standard, since its external
interference monitoring relies on the identification of external
MAC addresses. Therefore, more research effort should be
devoted to avoiding external interference in designing the
CA approaches in future. We believe that the basic means to
address this issue is to exploit the rich functionalities provided
by the physical layer to retrieve the channel status.
• Directional Antennas: in the surveyed approaches, only
DMesh [11] considers directional antennas, but imperfect direc-
tionality is assumed, so it still addresses the interference within
the interference cones. With the hardware price continuously
decreasing, directional antennas with strict directionality will
be prevalent in future. Therefore, designing CA with the inter-
ference issue simplified by the directional antennas should be
considered by future research to improve the network perfor-
mance.
• Channel Oscillation: to address this phenomenon, PCU-CA [25]
uses a probabilistic channel switching mechanism, which
prevents channel oscillation but may not be optimal for
the network performance. In light of this, we suggest that
the techniques used to solve the routing oscillation can
be borrowed with certain adaptation to tackle the channel
oscillation. To name some such techniques, a hysteresis factor
is used to avoid an overlay route flap in [46], and limiting route
advertisements by fixed timers is used to damp the BGP route
flap in [42].
• Quality of Service (QoS): QoS is needed by some major WMN
applications such as VoIP and video surveillance [5,32,41].
Unfortunately, no CA approaches surveyed in this paper include
QoS constraints in their problem formulations, though some
of them address QoS in their combined routing algorithms.
Similarly, the current MR–MC WMN deployments [5,32,41]
also only provide QoS support in layer-2 frame forwarding
or in layer-3 packet routing, where the frames/packets can
be classified and thereby treated differently. This leads to the
natural question whether CA is the proper place to implement
QoS functionality. To this question, our answer is that CA
should provide QoS support since it decides the bandwidths
of network links, which are the fundamental resources for
providing QoS. Though CA does not deal with the packet
forwarding directly, it should be designed to have QoS in
mind, thus supporting the bandwidth allocation requirement
needed by the packet forwarding components in the layer-2/3
protocols. For example, one kind of QoS support CA can offer is
to guarantee that the bandwidth of each link is no less than a
threshold value.
6. Conclusion
With the fast advance of wireless technologies nowadays,
the WMNs have evolved from a single-radio single-channel
architecture to an MR–MC architecture. Adopted in many WMN
deployments, this MR–MC architecture launches a new research
area — channel assignment, which seeks to find the optimal
mapping between the channels and radios at each node to improve
the network performance.
In this paper, considerable effort is made to provide insight into
the state-of-the-art approaches proposed in the CA research area.
Our contributions are fourfold. First, the key design issues for CA
approaches are identified, with the their importance elucidated.
Second, a classification of the CA approaches is proposed, where
the CA approaches are classified into two categories in general:
centralized and distributed. The centralized approaches are fur-
ther classified according to their problem formulations, and the
distributed approaches are further classified according to their as-
sumed traffic patterns. Third, each CA approach is treated with a
summary of its underlying idea, followed by the remarks on its pros
and cons. Also, both categorical comparisons and overall compar-
isons on these approaches are made, manifesting their sameness
and differences. Finally, both general and specific directions for fu-
ture CA research are pointed out, with solution hints given.
Comput. (2009), doi:10.1016/j.jpdc.2009.09.011

ARTICLE IN PRESS
References
[1] R.K. Ahuja, T.L. Magnanti, J.B. Orlin, Network Flows: Theory, Algorithms, and
Applications, Prentice Hall, Englewood Cliffs, NJ, 1993.
[2] I.F. Akyildiz, X. Wang, W. Wang, Wireless mesh networks: A survey, Computer
Networks 47 (2005) 445–487.
[3] M. Alicherry, R. Bhatia, L.E. Li, Joint channel assignment and routing for
throughput optimization in multi-radio wireless mesh networks, in: ACM
MobiCom, Cologne, Germany 2005, pp. 58–72.
[4] S. Avallone, I.F. Akyildiz, A channel assignment algorithm for multi-radio
wireless mesh networks, in: ICCCN, 2007, pp. 1034–1039.
[5] BelAirNetworks, Building inclusive municipal wireless mesh networks, http://
www.belairnetworks.com, White Paper 2007.
[6] A. Boukerche, S. Hong, T. Jacob, A distributed algorithm for dynamic channel
allocation, Mobile Networks and Applications 7 (2002) 115–126.
[7] R. Bruno, M. Conti, E. Gregori, Mesh networks: Commodity multihop ad hoc
networks, IEEE Communications Magazine 43 (2005) 123–131.
[8] A. Brzezinski, G. Zussman, E. Modiano, Enabling distributed throughput
maximization in wireless mesh networks: A partitioning approach, in: ACM
MobiCom, 2006, pp. 26–37.
[9] B. Clark, C. Colbourn, D. Johnson, Unit disk graphs, Discrete Mathematics 86
(1990) 165–177.
[10] T.H. Cormen, Introduction to algorithms, 2nd ed., MIT Press, Cambridge,
Boston, MA, 2001.
[11] S.M. Das, H. Pucha, D. Koutsonikolas, Y.C. Hu, D. Peroulis, DMesh: Incorporating
practical directional antennas in multichannel wireless mesh networks, IEEE
Journal on Selected Areas in Communications 24 (2006) 2028–2039.
[12] A. Dimakis, J. Walrand, Sufficient conditions for stability of longest queue first
scheduling: Second order properties using fluid limits, Advances of Applied
Probabilities 2 (2006) 505–521.
[13] P. Fuxjager, D. Valerio, F. Ricciato, The myth of non-overlapping channels:
Interference measurements in IEEE 802.11, in: Fourth Annual Conference on
Wireless on Demand Network Systems and Services, WONS ’07 2007, pp. 1–8.
[14] A. Ghosh, D.R. Wolter, J.G. Andrews, R. Chen, Broadband wireless access
with WiMax/802.16: Current performance benchmarks and future potential,
Communications Magazine, IEEE 43 (2005) 129–136.
[15] P. Gupta, P.R. Kumar, The capacity of wireless networks, IEEE Transactions on
Information Theory 46 (2000) 388–404.
[16] IEEE, IEEE standard for Local and metropolitan area networks — Part 11:
Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY)
Specifications, in: ANSI/IEEE Std 802.11, 1999 Edition (R2003), pp. 1–513,
2003.
[17] IEEE, IEEE standard for local and metropolitan area networks — Part 16: Air
Interface for Fixed Broadband Wireless Access Systems, in: IEEE Std 802.16-
2004 (Revision of IEEE Std 802.16-2001), pp. 1–857, 2004.
[18] IEEE-STD, Supplement to IEEE standard — Part 11: Wireless LAN Medium
Access Control (MAC) and Physical Layer (PHY) specifications: High-speed
physical layer in the 5 GHz band, IEEE Std 802.11a-1999, 1999.
[19] IEEE-STD, Supplement to IEEE standard — Part 11: Wireless LAN Medium
Access Control (MAC) and Physical Layer (PHY) Specifications: Higher-speed
physical layer extension in the 2.4 GHz band, in: IEEE Std 802.11b-1999,
pp. 1–90, 2000.
[20] K. Jain, J. Padhye, V.N. Padmanabhan, L. Qiu, Impact of interference on multi-
hop wireless network performance, in: ACM MOBICOM, San Diego, CA, USA,
2003, pp. 66–80.
[21] F.P. Kelly, Charging and rate control for elastic traffic, European Transactions
on Telecommunication 8 (1997) 33–37.
[22] M. Kodialam, T. Nandagopal, Characterizing the capacity region in multi-radio
multi-channel wireless mesh networks, in: ACM MobiCom, 2005, pp. 73–87.
[23] B.-J. Ko, V. Misra, J. Padhye, D. Rubenstein, Distributed channel assignment in
multi-radio 802.11 Mesh Networks, in: WCNC, 2007, pp. 3981–3986.
[24] B.-J. Ko, D. Rubenstein, Distributed self-stabilizing placement of replicated
resources in emerging networks, IEEE/ACM Transactions on Networking 13
(2005) 476–487.
[25] P. Kyasanur, N.H. Vaidya, Routing and link-layer protocols for multi-channel
multi-interface ad hoc wireless networks, ACM SIGMOBILE Mobile Computing
and Communications Review 10 (2006) 31–43.
[26] E.L. Lawler, Combinatorial Optimization: Networks and Matroids, Holt,
Rinehart and Winston, New York, 1976.
[27] S.H. Low, F. Paganini, J.C. Doyle, Internet congestion control, control systems
magazine, IEEE 22 (2002) 28–43.
[28] M.K. Marina, S.R. Das, A topology control approach for utilizing multiple
channels in multi-radio wireless mesh networks, in: The 2nd International
Conference on Broadband Networks, 2005, pp. 381–390.
[29] MeshDynamics, Technology — Performance analysis, http://www.
meshdynamics.com, Online 2006.
[30] A.H. Mohsenian Rad, V.W.S. Wong, Joint optimal channel assignment and
congestion control for multi-channel wireless mesh networks, in: IEEE ICC’06,
2006, pp. 1984–1989.
[31] A. Naveed, S.S. Kanhere, S. Jha, Topology control and channel assignment
in multi-radio multi-channel wireless mesh networks, in: IEEE MASS, 2007,
pp. 1–9.
[32] Nortel, Nortel municipal wireless solutions, http://www.nortel.com, Solution
Brief 2006.
[33] J.G. Oxley, Matroid Theory, Oxford University Press, Oxford, New York, 1992.
[34] K. Ramachandran, E. Belding, K. Almeroth, M. Buddhikot, Interference
aware channel assignment in multi-radio wireless mesh networks, in: IEEE
INFOCOM, 2006, pp. 1–12.
[35] K. Ramachandran, E. Belding, K. Almeroth, M. Buddhikot, Interference
aware channel assignment in multi-radio wireless mesh networks, in: IEEE
INFOCOM, 2006.
[36] A. Raniwala, T.-C. Chiueh, Architecture and algorithms for an IEEE 802.11-
based multi-channel wireless mesh network, in: IEEE INFOCOM, 2005,
pp. 2223–2234.
[37] A. Raniwala, K. Gopalan, T.-C. Chiueh, Centralized channel assignment
and routing algorithms for multi-channel wireless mesh networks, ACM
SIGMOBILE Mobile Computing and Communications Review 8 (2004) 50–65.
[38] A. Raniwala, K. Gopalan, T.-C. Chiueh, Centralized channel assignment
and routing algorithms for multi-channel wireless mesh networks, ACM
SIGMOBILE Mobile Computing and Communications Review 8 (2004) 50–65.
[39] A. Subramanian, H. Gupta, S. Das, Minimum interference channel assignment
in multi-radio wireless mesh networks, in: IEEE SECON, 2007, pp. 481–490.
[40] J. Tang, G. Xue, W. Zhang, Interference-aware topology control and QoS routing
in multi-channel wireless mesh networks, in: ACM MobiHoc, 2005, pp. 68–77.
[41] TroposNetworks, Tropos networks MetroMesh architecture, http://www.
tropos.com, White Paper 2005.
[42] C. Villamizar, R. Chandra, R. Govindan, BGP route flap damping, in: RFC 2439,
Standards Track, 1998.
[43] K. Xing, X. Cheng, L. Ma, Q. Liang, Superimposed code based channel
assignment in multi-radio multi-channel wireless mesh networks, in: ACM
MobiCom, 2007, pp. 15–26.
[44] K. Xu, M. Gerla, S. Bae, Effectiveness of RTS/CTS handshake in IEEE 802.11 based
ad hoc networks, Ad Hoc Networks 1 (2003) 107–123.
[45] S. Xu, T. Saadawi, Does the IEEE 802.11 MAC protocol work well in multihop
wireless ad hoc networks?, IEEE Communications Magazine 39 (2001)
130–137.
[46] B.Y. Zhao, H. Ling, J. Stribling, A.D. Joseph, J.D. Kubiatowicz, Exploiting routing
redundancy via structured peer-to-peer overlays, in: Proceedings of 11th IEEE
International Conference on Network Protocols, ICNP’03, 2003, pp. 246–257.
Weisheng Si received his bachelor and master degree
in Computer Science from Beijing University, China and
University of Virginia, USA respectively. He then worked
as a lecturer in School of Software, Dalian University of
Technology, China, where he won the University Teaching
Award for two times. He is currently a Ph.D. student
at the School of Information Technologies, University of
Sydney, Australia. His research interests include wireless
mesh networks, geographic routing algorithms, and graph
theory.
S. Selvakennedy obtained his Ph.D. in Computer Network-
ing in 1999 from the University of Putra, Malaysia. He
currently works as a consulting engineer with a profes-
sional services company, and is also associated with the
University of Sydney, Australia. His research interests lie
in developing protocols and algorithms for radio resource
management, MAC and network protocols, middleware
and topology control issues in wireless mesh networks. He
has served on the technical program committees of many
international conferences. Dr. Selvakennedy is a profes-
sional member of ACM.
Albert Y. Zomaya is currently the Chair Professor of High
Performance Computing and Networking in the School
of Information Technologies, The University of Sydney.
He is also the Director for the newly established Sydney
University Centre for Distributed and High Performance
Computing. Prior to joining Sydney University he was a
Full Professor in the Electrical and Electronic Engineering
Department at the University of Western Australia, where
he also led the Parallel Computing Research Laboratory
during the period 1990–2002. He is the author/co-author
of seven books, more than 350 publications in technical
journals and conferences, and the editor of eight books and eight conference
volumes. He is currently an associate editor for 16 journals, the Founding Editor
of the Wiley Book Series on Parallel and Distributed Computing and a Founding Co-
Editor of the Wiley Book Series on Bioinformatics. Professor Zomaya was the Chair
the IEEE Technical Committee on Parallel Processing (1999–2003) and currently
serves on its executive committee. He also serves on the Advisory Board of the
IEEE Technical Committee on Scalable Computing and IEEE Systems, Man, and
Cybernetics Society Technical Committee on Self-Organization and Cybernetics
for Informatics and is a Scientific Council Member of the Institute for Computer
Sciences, Social–Informatics, and Telecommunications Engineering (in Brussels). He
received the 1997 Edgeworth David Medal from the Royal Society of New South
Wales for outstanding contributions to Australian Science. Professor Zomaya is
also the recipient of the Meritorious Service Award (in 2000) and the Golden Core
Recognition (in 2006), both from the IEEE Computer Society. He is a Chartered
Engineer (CEng), a Fellow of the American Association for the Advancement of
Science, the IEEE, the Institution of Electrical Engineers (UK), and a Distinguished
Engineer of the ACM. His research interests are in the areas of high performance
computing, parallel algorithms, mobile computing, and bioinformatics.
Comput. (2009), doi:10.1016/j.jpdc.2009.09.011

An Overview Of Channel Assignment Methods For Multi-Radio Multi-Channel Wireless Mesh Networks

More Related Content

Similar to An Overview Of Channel Assignment Methods For Multi-Radio Multi-Channel Wireless Mesh Networks (20)

More from Darian Pruitt (20)

Recently uploaded (20)

An Overview Of Channel Assignment Methods For Multi-Radio Multi-Channel Wireless Mesh Networks