This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication.

IEEE TRANSACTIONS ON COMPUTERS

A Geometric Deployment and Routing Scheme for

Directional Wireless Mesh Networks
Weisheng Si, Student Member, IEEE, Albert Y. Zomaya, Fellow, IEEE, Selvadurai Selvakennedy

Abstract This paper first envisions the advent of the wireless
mesh networks with multiple radios and directional antennas
in future. Then, based on the observation that simplicity
induces efficiency and scalability, the paper proposes a joint
geometric deployment and routing strategy for such mesh
networks, and also gives a concrete approach under this
strategy. The main idea of this strategy is to deploy mesh
networks in certain kind of geometric graph, and then design
a geometric routing protocol by exploiting the routing
properties of this graph. The proposed concrete approach
comprises two parts: (1) a topology generation algorithm
based on Delaunay triangulations and (2) a geometric routing
protocol based on the greedy forwarding algorithm. Both
parts are characterized by simplicity and appealing properties,
with formal proofs provided when possible. The simulation
results validate our proposed approach.
Index Terms directional antennas, Delaunay triangulations,
geometric routing, network deployment, wireless mesh networks
I. INTRODUCTION
Wireless mesh networks (WMNs) typically consist of a
set of mesh routers that communicate with each other via
wireless links and form a mesh topology [1, 2, 3, 4, 5]. The
basic functionalities of these mesh routers are (1) providing
the backhaul connection for Wireless LANs (WLANs) and
(2) routing the traffic in the backhaul. These mesh routers
may also have additional functionalities, based on which
they can be classified into the following three categories:
• gateways: also interface with the Internet.
• APs: also serve as Access Points (APs) for WLANs.
• pure mesh routers: only have the aforementioned basic
functionalities.
In this paper, we solely focus on the basic functionalities of
these mesh routers, aiming to improve their performance in
the backhaul. For brevity, we also refer to these mesh
routers as “nodes” hereafter.
According to [1, 2, 3, 4, 5], the recent years saw the
following two new technologies for the WMNs:
• Multiple radios and multiple channels: each node is
equipped with multiple radios, each of which uses a
————————————————
• The authors are with the Centre for Distributed & High Performance
Computing, School of Information Technologies, University of Sydney,
NSW 2006, Australia.
• Email: weisheng@it.usyd.edu.au, albert.zomaya@sydney.edu.au,
and drskennedy@acm.org
Manuscript received (insert date of submission if desired). Please note that all
acknowledgments should be placed at the end of the paper, before the bibliography.
distinct non-overlapping channel. By enabling multiple
channels to carry network traffic simultaneously, this
trend essentially multiplies the available bandwidth for
network nodes.
• Directional antennas: these antennas are used in the
backhaul connections, enabling the nodes to
communicate in a point-to-point fashion. Thus, the
interference among links sharing identical channels is
basically eliminated. Note that, the directional antennas
are still not perfect in eliminating interference, and the
effects of interference between directional antennas are
discussed in [6, 7].
Until now, not all WMN deployments have employed the
above two technologies, and only some of them did (e.g., [1,
3]). However, with the decreasing hardware cost, we
envision that the WMNs in future will employ both.
Accordingly, in this paper, we focus on such WMNs with
multiple radios/channels and directional antennas, and refer
to them as Directional WMNs (DWMNs). For a DWMN, if
we depict its nodes on a (x, y) coordinates system and draw
a straight line for each point-to-point communication link
among the nodes, we will get a geometric graph (see Figure
1).
Figure 1. Modeling a DWMN with a geometric graph
Based on the above model, we propose a geometric
scheme for the DWMNs, which advocates that the position
information (i.e., the (x, y) coordinates) of the nodes will
play a major role in the deployment and routing of the
DWMNs. The motivation for proposing this scheme is to
perform what can be planned in advance to simplify the
operation of WMNs, thus achieving efficiency and
scalability. In retrospect, an important lesson from Internet
deployment is the Simplicity Principle [8], which states that
‘complexity is the primary aspect which impedes efficient

Digital Authorized licensed

Object Indentifier use limited to: Dayananda Sagar College of Engineering.
10.1109/TC.2010.169 Downloaded
0018-9340/10/$26.00 on August
© 2010 IEEE 17,2010 at 06:26:18 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
IEEE TRANSACTIONS ON COMPUTERS
scaling, and is the primary driver of increases in both capital
expenditure and operational expenditure for the network
carriers’.
For this geometric scheme, we propose a joint geometric
deployment and routing strategy, and also give a concrete
approach to validate this strategy. The idea of this strategy is
to deploy the DWMNs with certain kind of geometric graph
as the network topology, and then design a geometric
routing protocol by exploiting this graph’s routing
properties. The justifications for this strategy are as follows:
• Feasibility: the (x, y) coordinates of mesh routers can be
obtained by the GPS devices, and are accurate enough
for use in a city area. Recall that the Internet uses the
prefix of IP addresses to perform routing; however, since
the Internet turns out to be deployed haphazardly, the
prefix of IP addresses cannot reflect the actual locations
of subnetworks or computers. However, the (x, y)
coordinates of mesh routers can achieve this purpose.
• Locality: geometric routing protocols are localized [9]
protocols in that the routing decision is solely based on a
constant amount of information stored in the packets and
the positions of the current forwarding node, its
neighbors and the destination. Thus, geometric routing
protocols are inherently efficient and scalable.
• Ease of maintenance: the localized property also
significantly eases the maintenance of WMNs, since the
reconfiguration to one part of the network does not need
to be notified to the other parts.
• Fixedness: with the nodes in WMNs not mobile, the
periodical exchange of position information among them
is no longer needed. In contrast, such exchanging
overhead is significant for the mobile wireless networks
[10, 11].
• Directionality: with the interference among backhaul
links eliminated by the directional antennas, the metrics
related to positions become practical for making routing
decisions. Otherwise, the interference-aware routing
metrics (e.g., WCETT in [12] and MGF in [13]) are
more reasonable.
• Dynamics: though the nodes in WMNs are not mobile,
there exist significant network dynamics such as link
failure and congestion, making the static routing
protocols unsuitable. However, since the geometric
routing protocols basically rely on the one-hop neighbor
information that can be updated quickly, they are well
suited for dynamic networks.
To validate this strategy, we present a concrete approach
that consists of the following two parts dealing with
deployment and routing respectively: the Pruned Delaunay
Triangulation (PDT) generation algorithm and the Limited
Backward Greedy Forwarding (LBGF) protocol.
The PDT generation algorithm uses Delaunay
triangulations (DT, see Section 3 for its definition and
routing properties) as the basis to generate the network
topologies. Basically, it produces first refined and then
pruned DTs. The PDTs have the advantages of (1) retaining
Authorized licensed use limited to: Dayananda Sagar College of Engineering. Downloaded on August 17,2010 at 06:26:18 UTC from IEEE Xplore. Restrictions apply.
the desirable routing properties of the complete DTs and (2)
being feasible for the WMN deployment and (3) having a
low computation complexity. Note that though the PDT
generation algorithm is the same as that presented in our
previous work [14], in this paper we prove a theorem stating
its low computation complexity and conduct more
evaluations on its performance.
The LBGF protocol extends the well-known greedy
forwarding algorithm [10] by allowing the delivery of a
packet farther from the destination in limited number of
times. Moreover, it proactively detects two kinds of loops to
reduce the packet drop ratio. Compared with other
geometric routing protocols, LBGF achieves more
simplicity by discarding the mechanism of face routing [15],
which are exploited by most other geometric protocols to
recover from local failure when the greedy forwarding does
not work. In addition, LBGF provably causes no loops even
under the network dynamics and non-planar network
topologies, while the geometric routing protocols with the
face routing mechanism cannot achieve this. We note here
that LBGF is a significant modification to the Backward-
Enabled Greedy Forwarding protocol presented in our
previous work [14], which incurs a larger packet drop ratio.
To the best of our knowledge, we are the first to
advocate such a combined geometric deployment and
routing strategy for WMNs. We show by extensive
simulations that:
• The PDT Generation algorithm can generate topologies
that not only enable LBGF to find paths very close to the
actual shortest paths, but also reduce the installation cost
for deployers.
• The LBGF protocol can scale to a large number of
nodes. Moreover, it performs well in terms of packet
drop ratio and path qualities.
The rest of this paper is organized as follows. Section 2
reviews related work and distinguishes our work from
others. Section 3 introduces the theoretical background for
our concrete approach. Section 4 presents the PDT
generation algorithm and its evaluation. Section 5 presents
the LBGF protocol and its evaluation. Finally, section 6
concludes this paper and discusses the future work.
II. RELATED WORK
As our work covers two areas, node deployment and
geometric routing, the related work in these two areas is
discussed respectively in this section. Generally speaking,
our work has its uniqueness in both areas.
A. Node deployment
The node deployment issue exists in the context of
WMNs, wireless sensor networks (WSNs), and cellular
mobile networks. Generally, the node deployment solutions
aim at optimizing different criteria for these three types of
networks.
Since WMNs just emerged in the recent years and the
deployment issue has not received enough attention from the
academic community, we only discovered one study [16]
that investigates the impact of several factors such as the
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
IEEE TRANSACTIONS ON COMPUTERS
number of radios and backhaul connectivity on the
deployment of WMNs. As a simulated study on network
performance, this work is based on extensive Monte Carlo
simulations, which reveal several practical guidelines on the
WMN deployment.
For WSNs, the study on deployment mainly considers
the criteria of coverage area, connectivity, and installation
cost [17, 18, 19]. It is worth noting that in WSNs, coverage
area and connectivity are generally optimized instead of
routing, since the WSNs are more concerned with gathering
data or detecting events. A similar work to our PDT
generation algorithm is the Adaptive Triangular Deployment
algorithm (ATRI) [19], which also exploits the Delaunay
triangulations to conduct deployment. The difference is that
the ATRI algorithm is designed to maximize the coverage
area and minimize the coverage gaps, while our PDT
generation algorithm is designed to preserve the routing
property of the complete DTs and save the installation cost.
For cellular mobile networks, the study on the
deployment of base stations (BS) is called cellular BS
planning, which is a major research topic in this area. For
cellular BS planning, the main criteria to consider include
coverage area, traffic distribution, signal quality, and
installation cost [20, 21]. Note that, routing is not considered
here, since routing is basically done at the mobile switching
centers instead of BSs in cellular networks. The current BS
planning approaches generally use the hexagonal
deployment strategy.
In brief, the uniqueness of our approach for WMN
deployment lies in the following:
• The main factors considered are routing and installation
cost, since we believe that routing is the foremost
functionality provided by the WMNs, while installation
cost is a primary concern of deployers.
• A novel kind of geometric graph, PDT, is proposed as
the network topology.
B. Geometric routing
According to the two surveys [15, 22], there are plenty
of approaches that apply geometric routing (also called
geographic routing and position-based routing) on wireless
networks. The geometric graphs adopted by those
approaches mainly include the relative neighborhood graph
(RNG), Gabriel graph (GG), Yao graph (YG), and
Delaunay triangulation. Despite using different kinds of
graphs, most of these approaches have the following
common characteristics.
First, the wireless network environment is modeled by
the unit disk graph (UDG), in which all network nodes use
omni-directional antennas with an identical transmission
range and two nodes have a link between them if their
distance is no more than the transmission range. Thus, there
exists significant interference among links that are near each
other, which makes the routing protocols solely based on
position information not appealing.
Second, the geometric graphs used as the network
topology needs to be established and maintained by
Authorized licensed use limited to: Dayananda Sagar College of Engineering. Downloaded on August 17,2010 at 06:26:18 UTC from IEEE Xplore. Restrictions apply.
exchanging control messages among the nodes, which
incurs significant overhead.
Finally, in routing a packet to the destination, if the
processing node does not have a neighbor closer to the
destination than itself, face routing is used to overcome this
communication void phenomenon and to guarantee the
packet delivery [15]. Though face routing can not generate
loops in static networks, it can do so in dynamic networks.
For example, face routing will fall in a loop in the
dynamic network scenario depicted in Figure 2, where the
source s has a packet P destined for t, and the link ac is
temporarily broken. We assume that a combined greedy-
face routing algorithm is used here. In the beginning, s will
send P to u according to the greedy forwarding algorithm.
At u, since no neighbor is closer to t than u, u will start the
face routing process. Without loss of generality, here we
assume that u uses the left-hand rule to forward P. Thus, P
will traverse nodes a, b, c sequentially. When P arrives at c,
suppose the link ac becomes available. In this new topology,
P will be trapped in the loop bcab, since in a face routing
algorithm, loops are detected by remembering the first edge
of the current face being traversed [9]. In this example, the
face routing algorithm will remember the edge ua, so it will
not be able to discover the loop bcab.
a
c u
s t
b
d e
Figure 2. A dynamic network scenario where face routing
generates loops
Compared with these previous approaches, the
uniqueness of our routing approach lies in the following:
• The UDG model is replaced with the point-to-point link
model due to the use of directional antennas, thus the
interference from nearby links is avoided.
• With the fixedness of mesh routers and the planning in
advance, the overhead of maintaining network topology
is obviated.
• Since no fixed faces exist in dynamic networks and no
mechanism can guarantee packet delivery in dynamic
networks (which can be disconnected anyway), we do
not see the need to exploit faces to assist routing. Instead,
we use a mechanism of imposing a limit to the number
of routing backward to overcome the communication
void phenomenon. This mechanism provably causes no
loops even when the network topology is dynamic or
non-planar. Moreover, we also employ two loop
detection techniques to proactively break the loops. As
to be described later, our approach is simpler than the
face routing approach in two ways: (1) less source codes
are used in implementation and (2) less number of fields
is added in the packet header.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
IEEE TRANSACTIONS ON COMPUTERS
III. THEORETICAL BACKGROUND
This section describes the concepts and properties of the
greedy forwarding (GF) algorithm and the Delaunay
triangulation (DT) graph, and also introduces some other
definitions used in this paper.
GF is a geometric routing algorithm that finds a path
from a source node s to a destination node t in the following
method [10]: at each node (say u) along the path to t, u
chooses the neighbor v that has the smallest d(v, t) as the
next hop, where d(x, y) denotes the Euclidean distance
between node x and y; and ties are broken arbitrarily. GF is
very simple, since it is not only localized, but also makes the
routing decision by just one search of the neighbor list.
• on the same node need to use distinct channels. Besides,
•
• the angle between any two links incident on a node
• should be no less than a threshold value min, so as to
• •
• lmax. Hereafter, a link longer than lmax is called a bad link.
Figure 3. An example of DT
DT is widely used as network topologies by geometric
routing protocols. It is defined as a triangulation graph that
no fourth node lies inside the circumcircle of any its triangle
[23]. Figure 3 shows an example of the DT graph. A DT has
the following appealing properties for geometric routing:
• For any triangulation graph, let n denote the number of
nodes, e the number of edges, k the number of convex
hull edges, we have e = 3n – 3 – k [23]. This implies that
in a DT, the total number of links in a DT is less than 3n
and the average node degree is less than 6, thus
simplifying the operation of routing.
• In a DT, for any node u to any destination t, there always
exists a neighbor v of u satisfying that d(v, t) < d(u, t),
thus GF can always find a path between any two nodes
[24].
• In a DT, the Euclidean length of the shortest path
between any two nodes u and v is guaranteed to be less
than c·d(u, v), where c is proved to be between 1.5846
and 2.42 [25, 26]. As an aside, determining c exactly is
one of the most challenging problems in the area of
computational geometry.
In this paper, we introduce the following definitions. A
node u is said to forward a packet if u routes this packet to a
neighbor v with d(v, t) < d(u, t); otherwise, u is said to
backward this packet. Correspondingly, we say that a packet
is always transmitted in two modes: the forward mode or the
backward mode. And we say that a graph is backward-free,
if in this graph from any node u to any node t, u always has
a neighbor v such that d(v, t) < d(u, t). Obviously, a DT
graph is backward-free according its properties described
above.
Authorized licensed use limited to: Dayananda Sagar College of Engineering. Downloaded on August 17,2010 at 06:26:18 UTC from IEEE Xplore. Restrictions apply.
IV. THE PDT GENERATION ALGORITHM
A. Problem formulation
In formulating our topology generation problem, we
make the following assumptions on the DWMNs.
• The positions of the AP nodes are decided by the sites of
hotspots (e.g., restaurants and hotels), and are taken as
given inputs.
• The positions of the other two types of nodes (gateways
and pure mesh routers) and the existence of links
between any pair of nodes can be decided by the
deployers.
• Due to the imperfection of directional antennas, radios
fully avoid the inter-channel interference. Hereafter, an
angle less than min is called a bad angle in this paper.
• Due to the regulation on transmitting power, the length
of a point-to-point link cannot exceed a threshold value
Since a DT has the routing properties mentioned in
Section 3, it is straight-forward to use the DT obtained by
simply triangulating the given AP nodes as the network
topology (denoted by DTAP hereafter). However, there
usually exist bad angles and links in DTAP, making it
impractical for deployment. Figure 4 illustrates a DTAP of
39 given AP nodes, where ∠ 1,∠ 2,∠ 3 are examples of bad
angles, and l1, l2 are examples of bad links. To remove these
bad angles and links, and also to keep the backward-free
property, a possible way for deployers is to add pure mesh
routers into the topology to change the triangulation. Thus,
we formulate the following topology generation problem:
Given a set of AP nodes and the thresholds min and lmax,
find a graph T as the network topology by adding pure mesh
routers into the DTAP and recalculating the triangulation,
such that:
1. T has no bad angles or bad links.
2. T is backward-free.
3. The number of pure mesh routers added is as small as
possible.
Note that in this formulation, objective 2 is to guarantee the
support to GF algorithm, and objective 3 is to reduce the
installation cost for deployers.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
IEEE TRANSACTIONS ON COMPUTERS

When a bad triangle is near the boundary of a DT, its

circumcenter or off-center can possibly lie outside the
boundary of this DT. In this case, the software Triangle will
not add a new node at the circumcenter or off-center.
Instead, several new nodes will be added on the boundary of
this DT to remove this bad triangle. This phenomenon is
illustrated in Figure 5, where the DTAP in Figure 4 is the
input, and bold points depict those new nodes added on the
boundary. It can be seen that the number of such new nodes
is large and these new nodes are not in the critical positions
for relaying network traffic.
Based on this observation, we use the options provided in
Triangle to prohibit it from adding new nodes on the
boundary when dealing with the bad triangles near the
boundary. After this stage 2 of refinement, for those bad
triangles left, we propose a DT pruning algorithm as stage 3
to remove them by eliminating some relevant edges. Thus,
Figure 4. A DTAP of given 39 AP nodes considerable cost for installing new nodes is saved for
B. Algorithm description deployers. In addition, we want this pruning algorithm to
keep the backward-free property of DTs. To illustrate this
To solve the problem formulated above, we propose the
pruning idea, Figure 6 (using the DTAP in Figure 4 as the
PDT generation algorithm, which has the following three
input) shows the actual DTR obtained by stage 2 with some
stages:
bad triangles left near the boundary (e.g., ΔA - ΔE ), and
1) DT construction: construct the DTAP of the given AP Figure 7 shows the PDT obtained by our pruning algorithm.
nodes. For both figures, the bold points depict the pure mesh
routers added within the boundary of the DTAP during the
2) DT refinement: add pure mesh routers to remove bad refinement.
angles and links in the DTAP, producing a refined DT
(denoted DTR hereafter) that has bad angles and links
only near the boundary.
3) DT pruning: remove the remaining bad angles and links
near the boundary in DTR by eliminating some relevant
edges, while guaranteeing that the resulting graph
(called PDT) is still backward-free.
Note that, both DTAP and DTR are DTs, but the PDTs are the
subgraphs of DTs. Since there is extensive literature in
computational geometry on DT construction and refinement
by adding new nodes, stages 1 and 2 are accomplished using
the existing state-of-the-art algorithms, which are
implemented in the open-source software Triangle [27].
Concretely, for stage 1, we use the well-known divide-
and-conquer algorithm [23] to construct the DTAP. For stage
2, the refinement algorithm implemented in Triangle is a
hybrid of three algorithms [28, 29, 30] ever proposed in the
Figure 5. The DTR with new nodes added on the boundary
literature, and it can refine an input DT into another DT that
satisfies various specified criteria (e.g., having no bad angles
or bad links). Software Triangle calls a triangle that does not
satisfy the specified criteria a bad triangle (e.g., a triangle
with a bad angle or a bad link), and its basic refinement
operation is to add a new node at either the circumcenter
[28] or the off-center [30] of each bad triangle to split it,
until there are no bad triangles.
In using Triangle to do the refinement, we set min = 30°
and lmax = 1000m respectively. The 30° is used because
Triangle is proved to terminate when min ≤ 30° [28, 30],
otherwise it may infinitely split the triangles. Actually, it can
usually terminate at a larger min in practice [30]. For lmax,
Triangle can terminate at any value.

Authorized licensed use limited to: Dayananda Sagar College of Engineering. Downloaded on August 17,2010 at 06:26:18 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
IEEE TRANSACTIONS ON COMPUTERS
Figure 6. The DTR without new nodes added on the boundary
Figure 7. The PDT obtained by our pruning algorithm
According to [28, 30], a bad triangle left in stage 2
satisfies the following two properties: (1) it is a right or
obtuse triangle with its circumcenter or off-center outside
the boundary of the DTR and (2) it either has a vertex on the
boundary of DTR (e.g., Δ A in Figure 6), or is recursively
adjacent to another bad triangle that has a vertex on the
boundary of DTR (e.g., Δ B in Figure 6). Based on these two
properties, the basic idea of our pruning algorithm is as
follows: keep traversing around the boundary of DTR to
remove those bad triangles that have a vertex on the
boundary, until no such triangles are found; to remove a bad
triangle, delete the edge that opposites the right or obtuse
angle of this bad triangle. The details of our pruning
algorithm are given in Figure 8, and comments are inserted
for explanation.
Authorized licensed use limited to: Dayananda Sagar College of Engineering. Downloaded on August 17,2010 at 06:26:18 UTC from IEEE Xplore. Restrictions apply.
1. // Initialization
graph T = DTR;
v = an arbitrary node on the boundary of T;
2. // Traversing each node on the boundary of T clockwise
do {
// examining triangles at node v
for (each triangle with v as a vertex) {
// to keep the triangulation structure during the
// examination, only mark for deletion here
if (it is a bad triangle)
mark the edge opposite to the right or
obtuse angle (also the longest edge in it)
for deletion;
}
T = T – those marked edges;
// advance v in clockwise order
v = next node on the boundary of T ;
}
until (no bad triangles are found on a boundary node)
3. Output T as the final PDT;
Figure 8. The DT pruning algorithm
To justify the backward-free property of PDTs and the
simplicity of the PDT generation algorithm, we prove the
following two theorems.
Theorem 1: The PDTs obtained by our proposed topology
generation algorithm are backward-free.
Proof: In stage 1 and stage 2 of our algorithm, the graphs
obtained are complete DTs, so they are backward-free. Thus,
we only need to prove that our pruning algorithm preserves
the backward-free property.
Since the basic operation of this algorithm is to delete the
edge that opposites the right or obtuse angle of a bad
triangle, the proof can be accomplished by proving that if
the graph before deleting such an edge is backward-free, it
remains backward-free after this deletion. Consider Figure 9,
where Δ xyz depicts a bad triangle left in a DTR, and ∠ xyz
is its right or obtuse angle, and o is the circumcenter of
Δ xyz , and lxy, lxz, and lyz are the perpendicular bisectors of
edges xy, xz, and yz respectively, and t is an arbitrary
destination node. According to [30, 31], the circumcenter o
of Δ xyz lies outside the boundary of DTR, so an assisting
line lb can be drawn by extending one of the boundary edges
of this DTR such that all nodes in this DTR is in one side of lb
and o is in the other side.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
IEEE TRANSACTIONS ON COMPUTERS
Figure 9. The proof of backward-free property
Since the deletion of edge xz only affects the
neighborhood of x and z, to prove the graph remains
backward-free only requires to prove that for both x and z
(without loss of generality, we prove it for x here), given any
destination t, there exists a neighbor v satisfying d(v, t) < d(x,
t). As shown in Figure 9, lb, lxy, and lyz divide the half plane
where t resides into three areas: I, II, and III. Suppose t is in
I, which completely lies in the left side of the bisector lxz, so
we have d(x, t) < d(z, t). Moreover, z is the only neighbor
that x loses when xz is deleted, so x must have another
neighbor v, such that d(v, t) < d(x, t). Suppose t is in area II
or III, which completely lies above the bisector lxy, so we
have d(y, t) < d(x, t), thus y is the nearer neighbor of x to t.
And similar proof can be done for node z.
Theorem 2: The PDT generation algorithm has a time
complexity of O(nÂlog(n)) + O(N), where n is the number of
given AP nodes and N the total number of nodes at the end
of the algorithm.
Proof: The time complexity of the PDT generation
algorithm is the sum of the time complexities of its three
stages. For stage 1, according to [23], the construction of DT
costs O(nÂlog(n)), where n is the number of the given AP
nodes. For stage 2, according to [30, 31], the refinement
algorithm costs O(N), where N is the total number of nodes
after the refinement. For stage 3, according to our pruning
algorithm described in Figure 8, it at most processes all the
triangles in the DTR, so it costs O(NT), where NT is the
number of triangles in the DTR. Since the number of
triangles has a linear relationship with the number of nodes
in a triangulation [23], the stage 3 also costs O(N).
Combining the costs of the three stages, the total time
complexity of the PDT generation algorithm equals
O(nÂlog(n)) + O(N) + O(N) = O(nÂlog(n)) + O(N).


As to the relationship between n and N, it is empirically
shown that the software Triangle adds very limited number
of new nodes during the refinement [30], so N will be not
significantly larger than n. Theoretically, it is proved that the
minimum distance between any two nodes in DTR is no less
than the parameter dmin, which denotes the minimum
distance between any two nodes in the original input DT
[30, 31]. Thus, in DTR, each node at least occupies a circular
Authorized licensed use limited to: Dayananda Sagar College of Engineering. Downloaded on August 17,2010 at 06:26:18 UTC from IEEE Xplore. Restrictions apply.
area of π d min , so N is bounded by the total area of DTR
2
divided by πd 2
min
. In light of this, the parameter dmin is
critical in determining the upper bound of N. In our
evaluations below, we use dmin =50m (we believe this
minimum distance can be realized by the deployers), and
our experiments show that our algorithm at most adds n new
nodes to the network topology. Therefore, the PDT
generation algorithm actually has a time complexity of
O(nÂlog(n)) in our settings.
C. Algorithm evaluation
This subsection first describes the experiment setup, and
then gives experimental results on the following four
metrics: the ratio of added pure mesh routers, the ratio of
pruned edges, the link deviation ratio and Euclidean
deviation ratio of the PDTs. The first two metrics reflect the
installation cost for using the PDTs as the network
topologies, and the last two metrics reflect the path quality
of PDTs.
1) Experiment setup
The given AP nodes are assumed to be randomly
distributed in a square area with a constant density, and this
density is set to 10 nodes/km2. For instance, if the number of
nodes is 20, the side length of the square area is
approximately 1414m. Since a DT exists for any set of
points on a plane [23], our PDT generation algorithm works
for any kind of distribution of AP nodes, and we only use
the random distribution here to conduct experiments. Our
PDT generation algorithm is implemented by modifying the
aforementioned software Triangle. With this
implementation, experiments are conducted on network
sizes of 50, 100, 200, 400, 600, 800, and 1000 given AP
nodes respectively. For each network size, 200 random
topologies are generated and the average result of them is
obtained.
2) The ratio of added pure mesh routers
The ratio of added pure mesh routers for a topology
equals the number of added pure mesh routers divided by
the total number of nodes after the topology generation.
Figure 10 plots the average ratio of 200 topologies for each
experimented network size. From this figure, we see that (1)
this ratio increases with the growth of the network size, and
the increasing slope levels out gradually; and (2) this ratio is
less than 0.5 for all plotted network sizes. Thus, we can infer
that the number of added pure mesh routers is below half the
number of all mesh routers, and tends to increase linearly in
relation to the network size.
0.48
Ratio of pure mesh routers
0.46
0.44
0.42
0.4
0.38
0.36
0 100 200 300 400 500 600 700 800 900 1000
Number of given AP nodes
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
IEEE TRANSACTIONS ON COMPUTERS

Figure 10. The ratio of added pure mesh routers GF tend to deviate more from the shortest paths when the
network size grows; (2) PDT improves the link deviation
ratio significantly over DTAP; and (3) for PDTs, the link
3) The ratio of pruned edges deviation ratio remains below 1.1 for all experimented
The ratio of pruned edges for a topology equals the network sizes. Thus, it can be seen that the PDTs provide
number of pruned edges divided by the total number of very high path quality in terms of number of links.
edges after the topology generation. Figure 11 plots the
average ratio of pruned edges of 200 topologies for each 1.3
experimented network size. From this figure, we see that (1) DT_AP PDT

Avg link deviation ratios

1.25
this ratio decreases with the growth of the network size,
since the pruned edges are only located near the boundary of 1.2

PDTs; and (2) this ratio is quite small (below 0.1) for all 1.15
plotted network sizes. 1.1

1.05
0.1
0.09 1
Ratio of pruned edges

0.08 0 100 200 300 400 500 600 700 800 900 1000
0.07
0.06 Number of give n AP nodes
0.05
0.04 Figure 12. The link deviation ratios for DTAP and PDT
0.03
0.02
0.01
5) The Euclidean deviation ratio of PDTs
0 For the definition of Euclidean deviation ratio, please
0 100 200 300 400 500 600 700 800 900 1000 see the previous subsection. Figure 13 plots the average
Number of given AP nodes
Euclidean deviation ratio of 200 topologies for each
experimented network size. Also for the comparison
Figure 11. The ratio of pruned edges purpose, the deviation ratios for both DTAP and PDT are
According to the formula ‘e = 3n – 3 – k’ mentioned in plotted. From this figure, we see that (1) the Euclidean
Section 3, the total number of edges in a PDT is less than 3 deviation ratio is lower than its link counterpart, because the
times the total number of nodes. Since Figure 11 shows that structure of DTs is more helpful for finding short paths in
the edge pruning does not reduce the number of edges the Euclidean metric than the link metric; (2) for the
significantly, the installation cost of network links remains a Euclidean deviation ratio, PDT also improves significantly
linear relationship with the network nodes. Though the over DTAP; and (3) for PDTs, this deviation ratio remains
pruning of the edges does not help significantly in saving the below 1.05 for all plotted network sizes. Thus, it can be seen
installation cost, it does remove the bad angles and links, that the PDTs also provide very high path quality in terms of
thus making the PDTs feasible for deployment. Euclidean distance.

4) The link deviation ratio of PDT 1.1 DT_AP PDT

Avg Euclidean deviation

For a given routing algorithm, the link deviation ratio of 1.08

a source/destination pair (s, t) in a graph is defined as the
number of links (or hops) in the path found by the given 1.06
ratios

algorithm versus the minimum number of links from s to t. 1.04

Further, the link deviation ratio of a graph is defined as the
average link deviation ratio for all (s, t) pairs in the graph. 1.02
Similarly, with a given routing algorithm, the Euclidean 1
deviation ratio for a (s, t) pair or for a graph can be defined 0 100 200 300 400 500 600 700 800 900 1000
in terms of Euclidean distance instead of the number of Number of given AP nodes
links.
The implication of the deviation ratio is that the closer the Figure 13. The Euclidean deviation ratios for DTAP and PDT
ratio is to 1, the less deviation from the actual shortest paths
a given routing algorithm can achieve on a given network
topology. As to be described later, since our LBGF protocol V. THE LBGF PROTOCOL
reduces to the GF protocol if no network dynamics exist on A. Protocol overview
PDTs, we can use GF here to evaluate the path quality
provided by PDTs. Hereafter, we will not explicitly mention Given PDTs as network topologies, LBGF is designed
GF when discussing the deviation ratios. with the following main ideas. First, when no network
dynamics exist, it reduces to the GF algorithm and fully
Figure 12 plots the average link deviation ratio of 200 exploits the backward-free property of PDTs. Second, it
topologies for each experimented network size. To see the extends GF by considering two kinds of network dynamics:
effect of our refinement and pruning algorithms, both the link failure and link congestion. The former can be notified
deviation ratios for DTAP and PDT are plotted. From this by the link-layer functionalities, and the latter can be
figure, we see that (1) the link deviation ratios for both DTAP obtained by examining the occupancy of link packet buffers.
and PDT increase slowly, reflecting that the paths found by If the packet buffer for a link is full, the LBGF simply does

Authorized licensed use limited to: Dayananda Sagar College of Engineering. Downloaded on August 17,2010 at 06:26:18 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
IEEE TRANSACTIONS ON COMPUTERS
not send packets to this link. Thus, both kinds of network
dynamics can be monitored locally by nodes, without the
overhead of exchanging control packets in the network-
layer. Finally, with the presence of network dynamics, the
underlying network topology may not be backward-free. To
address this issue, LBGF allows backward a packet to
reduce the packet drop ratio. For loop prevention, LBGF
imposes a limit to the number of times that a packet can be
backwarded (denoted by BW_LIMIT hereafter). Later, we
will prove that this technique guarantees no routing loops.
Besides guaranteeing no loops by imposing the
BW_LIMIT, LBGF also detects the following two kinds of
loops to break them proactively.
• two-hop loops: the loops between two neighboring nodes.
• backward-then-forward loops: the loops that consist of a
series of consecutive transmissions in backward mode
and then a series of consecutive transmissions in forward
modes. Figure 14 gives two examples of such backward-
then-forward loops with node t as the destination.
b c
c a t d b a
d e
(1) Loop abcda (2) Loop abcdeba
Figure 14. Two examples of the backward-then-forward loops
We believe these two kinds of loops are the most
common loops that occur in our network settings. By
breaking these two kinds of loops, the packet drop ratio shall
be significantly reduced, which is verified in our simulations.
B. Protocol description
Required by LBGF, the following three fields are added
to a packet’s network-layer header:
• Dst Coords: the (x, y) coordinates of the destination
node.
• Backward Number (BN): the number of times that this
packet has been backwarded. Moreover, we use the
most significant bit (msb) of BN to indicate the
transmission mode of a packet, with ‘0’ indicating the
forward mode and ‘1’ the backward mode. When
creating a packet, the msb of BN is initialized to ‘0’.
• Backward Head (BH): for each series of consecutive
backward transmissions, this field records the first node
that starts this backward series. For comparison purpose,
when creating a packet, the BH field is initialized to a
non-existing value.
Note that, LBGF uses less number of fields in the packet
header than the combined greedy-face routing mechanism,
which has several variants. All these variants typically need
the following five fields in their packet header: the
destination coordinates, the forwarding mode, the node
where the greedy forwarding fails, the first edge of the
current face being traversed, and the nearest node or point
on an edge ever reached toward the destination [9, 10].
Authorized licensed use limited to: Dayananda Sagar College of Engineering. Downloaded on August 17,2010 at 06:26:18 UTC from IEEE Xplore. Restrictions apply.
Exploiting these three fields in the packet header as well
as the neighbor lists at nodes, the LBGF protocol is detailed
in Figure 15, where the processing node is denoted u, the
packet being processed P, P’s destination node t. For
explanation, comments are also inserted in Figure 15.
1. // search the neighbor list
u looks for a neighbor v with the shortest d(v, t) that
satisfies:
(1) link uv is up;
(2) packet buffer of link uv is not full;
// the two-hop loop detection:
(3) if (u is not a dangling node)
v is not the neighbor that just sent P to u ;
// the backward-then-forward loop detection:
(4) v cannot be the node recorded in BH;
2. // all neighbors are not available
if (v is not found) {
u drops P;
return;
}
3. // when v is closer to t than u
if (d(v, t) < d(u, t)) {
u puts 0 to the msb of BN;
u forwards P to v;
t return;
}
4. // check whether the BW_LIMIT is exceeded
if ( BN in P < BW_LIMIT ) {
// the previous transmission mode is forward?
if ( the msb of BN == 0 ) {
u puts its node ID into BH;
u puts 1 to the msb of BN;
}
u increases BN by 1;
u forwards P to v;
return;
}
5. // BN in P >= BW_LIMIT
u drops P;
Figure 15. The LBGF protocol
As seen from Figure 15, to break the two-hop loops, a
node u is prohibited from returning P to its predecessor (i.e.,
u’s neighbor that sent P to u in the last step), unless u is a
dangling node. Here a dangling node is defined as a node
that only has one neighbor, excluding those neighbors
connected by broken or congested links. Note that, the
reason for permitting a dangling node to return P to its
predecessor is to enable P to come out of the dangling node
situation. And to break the backward-then-forward loops, a
node u is prohibited from sending P to the node recorded in
BH. Since a BH node will definitely appear in a backward-
then-forward loop, avoiding sending P to the BH node again
breaks this kind of loops.
To exemplify the LBGF protocol, we next give two
network scenarios involving two-hop loop and backward-
then-forward loop respectively. For the two-hop loop, in
Figure 16, the source s has a packet P to transmit to the
destination t, and node a constitutes a dangling node. In this
scenario, P will be sent to b and then a. Since a is a dangling
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
IEEE TRANSACTIONS ON COMPUTERS
node, it is allowed to return P to b. Not a dangling node, b is
prohibited to send P to a, thus has to send P to s. Similarly, s
will send P to c. Finally, P will follow the path cdet to reach
t, thus overcoming the two-hop loop and the dangling node
situation.
s b a loop-free property of LBGF:
c t
d e
Figure 16. An example scenario for the two-hop loop with a
dangling node
For the backward-then-forward loop scenario depicted by
Figure 17, the source s also has a packet P to transmit to the
destination t. Initially, s will send P to node a, set the
transmission mode to backward, and record itself in BH of P.
Since a is prohibited from sending P back to s, a will send P
to b. Then, P will follow the path bcda to reach a again, with
the transmission mode now changed to forward. Since s is
recorded in BH of P, a is not allowed to send P to s. Thus, a
will send P to b, set the transmission mode to backward, and
record itself in BH of P. Afterwards, the packet will follow
the path bcd to reach d. At d, since a is recorded in BH of P,
d cannot send P to a. Instead, d will send P to e, thus
breaking the two backward-then-forward loops sabcdas and
abcda.
b in its header for LBGF. Since both types of space costs are
c a s t
f bounded by 12 due to min = 30°.
d e
Figure 17. An example scenario for the backward-then-
forward loop
We next prove the following two theorems regarding the
loop-free property and the complexity of LBGF
respectively. These two proofs also give some insights to
this protocol.
Theorem 3: For any network topology, LBGF is loop-free
in that it either delivers a packet P to the destination t or
drops P.
Proof: This proof is done in two cases. First, suppose P is
never backwarded. Then in each transmission, P gets closer
to the destination, thus each node visited is different from
all the previously-visited ones. Since there are finite
number of nodes in the network, P will definitely reach the
destination in finite transmission steps or get dropped due
to no available neighbors.
Second, suppose P is ever backwarded. Because the
number of backward transmissions is bounded, and also
because the number of consecutive forward transmissions
between any two backward transmissions cannot be infinite
as revealed in the previous paragraph, after finite
transmission steps, P will be backwarded for the last time.
10
Authorized licensed use limited to: Dayananda Sagar College of Engineering. Downloaded on August 17,2010 at 06:26:18 UTC from IEEE Xplore. Restrictions apply.
After this final backward transmission, all transmissions of
P will be forward transmission, thus P will reach the
destination or get dropped in light of the same argument in
the previous paragraph.

Note that, his proof reveals the following facts about the
• It does not require the network topology to be backward-
free. Hence, the purpose for PDTs to be backward-free is
not to guarantee the loop-free property of LBGF, but to
reduce the length of the routing paths and the packet
drop ratio.
• It holds for arbitrary network topologies, not necessarily
planar ones as in the face routing.
• It holds under the conditions of link failures or
congestions.
Theorem 4: For any network topology with its maximum
node degree bounded by a constant, LBGF runs with O(1)
complexity in both time and space at a node.
Proof: For the time complexity, the only non-trivial
computation in LBGF is one search of the neighbor list. If
the maximum node degree is bounded by a constant, this
search can be completed in O(1) time.
For the space complexity, a node needs to store its
neighbor list in the memory, which costs O(1) space due to
the bounded maximum node degree. Moreover, a packet,
which also costs a node’s memory, only carries three fields
constant, LBGF consumes O(1) space at a node. ͚
Note that, this theorem certainly holds when the network
topologies are PDTs, in which the maximum node degree is
C. Protocol evaluation
We implement the LBGF protocol in ns-2 [32], into
which two parts of codes are added: a routing module that
realizes the steps in Figure 15 and a link object that
simulates the point-to-point wireless links. Besides, to
simulate these point-to-point links, the WirelessChannel
object in ns-2 is modified, so that it only delivers the
wireless signal to the two neighboring nodes in a point-to-
point link. To compare with other works, we implement
another well-known geometric routing protocol called
GOAFR [33] under ns-2. GOAFR claims itself to be worst-
case optimal and average-case efficient in terms of path
qualities. Also, a theoretical work [9] shows that GOAFR
guarantees delivery under arbitrary static network topologies,
while other well-known geometric routing protocols such as
GOAFR+ [34] and GPVFR [35] cannot.
Using the network topologies produced by our PDT
generation algorithm, we conduct experiments on LBGF and
GOAFR with network sizes of 100, 200, 400, 600, 800, and
1000 nodes. To simulate the dynamic networks, the
Exponential Model included in ns-2 is used to generate link
failures. In this model, the up time and down time for a link
are exponentially distributed, and their means can be set by
the command-line arguments up-interval and down-interval
respectively. Using this model, experiments are conducted
on the following three scenarios for each network size:
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
IEEE TRANSACTIONS ON COMPUTERS

• Static: no link failures

0.0045 N100 N400 N900
• Dyn1: for every link, up-interval = 10s, down-interval = 0.004
1s 0.0035

Packet Drop Ratio

0.003
• Dyn2: for every link, up-interval = 10s, down-interval =
0.0025
2s
0.002
For each scenario of a given network size, 200 experiments 0.0015
with different random topologies are conducted and the 0.001
average result is obtained. In each experiment, each node is 0.0005
found a peer randomly, and a CBR flow is generated from 0
each node to its peer. In the later evaluation figures, the data 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
lines depicting LBGF and GOAFR performances in these Backward Limit Coefficient
three scenarios are labeled LBGF0, GOAFR0, LBGF1,
GOAFR1, LBGF2, and GOAFR2. Figure 18. Impact of BW_LIMIT on the Packet Drop Ratio
In the following, we first present evaluation results solely Figure 18 shows the following impact of BW_LIMIT to
related to LBGF: (1) the impact of BW_LIMIT to the the packet drop ratio: (1) the packet drop ratio decreases
packet drop ratio and the link deviation ratio and (2) the with the increasing of BW_LIMIT and levels out gradually;
effectiveness of the two loop detection techniques. Then, we (2) around c = 1.5, the decreasing slope becomes small,
present the evaluation results on the following three metrics implying that a larger BW_LIMIT value does not help
for both LBGF and GOAFR: (1) packet drop ratio, (2) link significantly in reducing the packet drop ratio.
deviation ratio, and (3) Euclidean deviation ratio. Note that
the definitions of link/Euclidean deviation ratios here are 1.7
N100 N400 N900
slightly different from those when evaluating PDTs by being
Average link deviation ratio

defined on a packet instead of being defined on a (s, t) pair. 1.6

Specifically, we define the link deviation ratio of a packet as
the number of links traversed by this packet from a source s
to a destination t versus the minimum number of links from
s to t. Further, we define the link deviation ratio of a routing
protocol as the average link deviation ratio of all packets
routed by this protocol in an experiment. Similarly, the
Euclidean deviation ratio of a routing protocol is defined in
terms of Euclidean distance instead of the number of links.
1) Impact of BW_LIMIT
Our choice of BW_LIMIT is based on the concept of
Graph diameter, which is defined as the length (number of
links) of the longest shortest path between any two nodes of
a graph [36]. For random DTs, the mean of the graph
diameters is O( n ) , where n is the number of nodes in a
DT [37]. Since the number of backward transmissions is
directly related to the length of the path traversed by a
packet, we conjecture that BW_LIMIT should be set to
c ⋅ n , where c is a configurable coefficient. Below, we
present our experimental results on the packet drop ratio and
link deviation ratio of LBGF when the c value ranges from
0.5 to 5 with a step of 0.5. From these experimental results,
we will select an appropriate c value for our WMN
environment. All these experiments are conducted in the
Dyn2 scenario, since we believe that a sufficiently large c
value for Dyn2 will be also sufficient for Dyn1 and Static
scenarios. To make our experiments encompass different
network sizes, the networks of 100, 400, 900 nodes are
tested, with their data lines labeled N100, N400, and N900,
respectively.
1.5
1.4
1.3
1.2
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Backward Limit Coefficient
Figure 19. Impact of BW_LIMIT to the link deviation ratio
Figure 19 shows the following impact of BW_LIMIT to
the link deviation ratio of LBGF: the link deviation ratio
increases slowly with the growth of BW_LIMIT, reflecting
that when the BW_LIMIT value increases, packets can
travel longer paths, thus more network capacity will be
consumed.
Based on the above observations, we use c = 1.5 in
conducting all the subsequent experiments. Note that for the
network environment other than ours, a different c value
may be more appropriate.
2) Effectiveness of the two loop detection techniques
To evaluate the effectiveness of our two loop detection
techniques, we show the ratio of number of loops detected to
the number of packets dropped. Since when a loop is
detected, a packet drop is avoided, so this ratio reflects how
successful our loop detection techniques are in avoiding
loops. Specifically, we measure the ratio of the number of
two-hop loops detected versus the number of packet
dropped (denoted by R_TwoHop) and the ratio of the
number of backward-then-forward loops detected versus the
number of packet dropped (denoted by R_BthenF) in each
experiment. The experiments are conducted in the Dyn2
scenario with network sizes of 100, 200, 600, and 1000
nodes. These two ratios are summarized in Table I, which

11

Authorized licensed use limited to: Dayananda Sagar College of Engineering. Downloaded on August 17,2010 at 06:26:18 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
IEEE TRANSACTIONS ON COMPUTERS

shows that both ratios have very large values, so our loop the network size is large; and (2) LBGF achieves low link
detection techniques are very effective. deviation ratios, especially in the Static scenario, which
indicates that LBGF performs well in finding paths with
Table I. The two loop detection ratios with respect to the
short lengths; and (2) LBGF performs slightly better than
number of packets dropped
GOAFR.
R_TwoHop R_BthenF
LBGF0 GOAFR0
N100 56.7 26.0 LBGF1 GOAFR1
1.6

Average link deviation ratios

N200 70.2 35.7 LBGF2 GOAFR2
1.5
N600 98.6 47.6
1.4
N1000 120.4 57.8 1.3

1.2

1.1
3) Packet drop ratio
1
Packet drop ratio is defined as the total number of packets
0 100 200 300 400 500 600 700 800 900 1000
dropped by all nodes versus the total number of packets
Number of mesh routers
generated by all nodes during an experiment. Figure 20 plots
our experimental results on it, showing that (1) both LBGF
and GOAFR exhibit very low packet drop ratios (less than Figure 21. Average link deviation ratios
0.017 even under Dyn2) and (2) in the Static scenario, where
the only network dynamics is network congestion, both 5) Average Euclidean deviation ratios
protocols achieve a drop ratio of nearly zero, and (3) in the Figure 22 plots the average Euclidean deviation ratios of
scenarios Dyn1 and Dyn2, LBGF outperforms GOAFR 200 experiments for both protocols in all the three scenarios.
considerably. In addition to what are already reflected in Figure 21, this
figure mainly shows that for both protocols in all the three
LBGF0 GOAFR0 scenarios, the Euclidean deviation ratio has a smaller range
0.018
LBGF1 GOAFR1 than its hop counterpart, which is because the structure of
0.016
LBGF2 GOAFR2 DTs is more helpful for finding short paths in the Euclidean
metric than the link metric.
Packet drop ratio

0.014
0.012
0.01
0.008 LBGF0 GOAFR0
0.006 LBGF1 GOAFR1
0.004 LBGF2 GOAFR2
Avg Euclidean deviation ratios

0.002 1.2
0 1.18
0 100 200 300 400 500 600 700 800 900 1000
1.16
Number of mesh routers
1.14
1.12
Figure 20. Packet drop ratios 1.1

Note that, the drop ratio of LBGF decreases with the 1.08
growth of the network size. This reflects that a larger 1.06

BW_LIMIT (set to 1.5 ⋅ n ) tends to offer a packet more 0 100 200 300 400 500 600 700 800 900 1000
Number of mesh routers
chance to reach the destination, despite that the path
traversed by a packet becomes longer. On the contrary, the
drop ratio of GOAFR increases with the growth of the Figure 22. Average Euclidean deviation ratios
network size. This is because GOAFR drops a packet when
(1) the packet encounters again the first edge that it traverses VI. CONCLUSION
on the current face such that a loop is discovered or (2) the In this paper, we first foresee the advent of the DWMNs
packet falls into a loop that cannot be discovered by the face with multiple radios and directional antennas in future. Then,
routing mechanism but the TTL (set to 128) is exceeded. based on the observation that simplicity brings the efficiency
And when the path traversed by a packet is longer, the and scalability, we propose a joint geometric deployment
packet has a larger chance to fall into a loop. and routing strategy, and then give a concrete approach
4) Average link deviation ratios under this strategy for the DWMNs. The proposed concrete
Figure 21 plots the average link deviation ratios of 200 approach has the following two parts: the PDT generation
experiments for both protocols in all the three scenarios. It algorithm and the LBGF routing protocol. We prove
shows that: (1) for both protocols in all the three scenarios, theoretically or show empirically the following merits of
the link deviation ratios increase with the growth of network these two parts.
size, reflecting that the paths found by the geometric routing • The PDT generation algorithm (1) has a low time
protocols are more likely to deviate from the optimal when complexity of O(nÂlog(n)), where n is the number of AP

12

Authorized licensed use limited to: Dayananda Sagar College of Engineering. Downloaded on August 17,2010 at 06:26:18 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
IEEE TRANSACTIONS ON COMPUTERS
nodes in the network, and (2) guarantees that the
generated PDTs are backward-free, and (3) incurs low
installation cost for the deployers.
• The LBGF protocol (1) has a constant complexity in
both time and space at nodes, (2) guarantees that the
route of a packet is loop-free, (3) incurs a low packet
drop ratio, and (4) finds high quality paths with both low
link deviation ratio and low Euclidean deviation ratios.
Being the first to propose the geometric deployment and
routing strategy for WMNs, we believe many problems
under this strategy are worth exploring, especially the
follows.
• The backward-free property is newly introduced for the
geometric graphs in this paper. Since this property
enables the geometric routing protocols to produce high
quality paths, finding the sufficient and necessary
conditions for a geometric graph to be backward-free
constitutes an interesting open problem.
• Our proposed PDT is only an example of the backward-
free graphs, and other kinds of such graphs are worth
exploring to suit different requirements of network
deployments.
• Since the face routing mechanism does not function
properly in dynamic networks as manifested in Section II,
designing other kinds of geometric routing protocols
capable of tackling network dynamics becomes a
challenging problem. While our LBGF protocol attempts
to solve this problem by imposing a backward limit and
detecting loops proactively, other different mechanisms
can be further explored.
Finally, we note that no current WMN deployments use
such a geometric deployment and routing strategy yet. This
is because the implementation of this strategy requires the
modification to the TCP/IP protocol stack at the mesh
routers (i.e., enabling these mesh routers to insert, exploit,
and then remove the position information from the headers
of network packets). Though this modification is not
difficult, no WMN deployers will be tempted to do this if
the benefit of this geometric strategy is not apparent. Thus,
we believe that when the scale of the WMNs becomes large,
the scalability and efficiency advantages of this geometric
strategy will become evident, and this geometric strategy
will be adopted and prevail.
REFERENCE
[1] BelAirNetworks, "BelAir300 Converged Multi-service
Wireless Node," http://www.belairnetworks.com, Data
Sheet,2007.
[2] MeshDynamics, "Technology - Performance Analysis,"
http://www.meshdynamics.com, Online,2006.
[3] Nortel, "Wireless Mesh Network Basics,"
http://www.nortel.com, Standard Document Release 3.0,2007.
[4] StrixSystems, "Solving the wireless mesh multi-hop
dilemma," http://www.strixsystems.com, White Paper,2006.
[5] TroposNetworks, "Tropos Networks MetroMesh
Architecture," http://www.tropos.com, White Paper,2005.
[6] V. Angelakis, N. Kossifidis, S. Papadakis, V. Siris, and A.
Traganitis, "The effect of using directional antennas on
adjacent channel interference in 802.11a: Modeling and
13
Authorized licensed use limited to: Dayananda Sagar College of Engineering. Downloaded on August 17,2010 at 06:26:18 UTC from IEEE Xplore. Restrictions apply.
experience with an outdoors testbed," in International
Symposium on Modeling and Optimization in Mobile, Ad
Hoc, and Wireless Networks and Workshops (WiOPT) 2008,
pp. 24-29.
[7] T. Ireland, A. Nyzio, M. Zink, and J. Kurose, "The Impact of
Directional Antenna Orientation, Spacing, and Channel
Separation on Long-distance Multi-hop 802.11g Networks: A
Measurement Study," in International Symposium on
Modeling and Optimization in Mobile, Ad Hoc and Wireless
Networks and Workshops (WiOpt), 2007, pp. 1-6.
[8] R. Bush and D. Meyer. (2002, Some Internet Architectural
Guidelines and Philosophy. RFC 3439, Informational.
[9] H. Frey and I. Stojmenovic, "On delivery guarantees of face
and combined greedy-face routing in ad hoc and sensor
networks," in ACM MobiCom, 2006, pp. 390-401
[10] B. Karp and H. T. Kung, "GPSR: Greedy Perimeter Stateless
Routing for Wireless Networks " in ACM MobiCom, 2000,
pp. 243-254.
[11] J. Gao, L. J. Guibas, J. Hershberger, L. Zhang, and A. Zhu,
"Geometric spanners for routing in mobile networks," IEEE
Journal on Selected Areas in Communications, vol. 23, pp.
174-185, 2005.
[12] R. Draves, J. Padhye, and B. Zill, "Routing in multi-radio,
multi-hop wireless mesh networks," in ACM Mobicom, 2004,
pp. 114 - 128.
[13] P. Kyasanur and N. H. Vaidya, "Routing and Link-layer
Protocols for Multi-Channel Multi-Interface Ad Hoc Wireless
Networks," ACM SIGMOBILE Mobile Computing and
Communications Review, vol. 10, 2006.
[14] W. Si and S. Selvakennedy, "A Position-Based Deployment
and Routing Approach for Directional Wireless Mesh
Networks," in The International Conference on Computer
Communications and Networks (ICCCN) 2008, pp. 1-8.
[15] D. Chen and P. K. Varshney, "A Survey Of Void Handling
Techniques For Geographic Routing In Wireless Networks,"
IEEE Communications Surveys & Tutorials, vol. 9, pp. 50-67,
2007.
[16] J. Robinson and E. W. Knightly, "A Performance Study of
Deployment Factors in Wireless Mesh Networks," in IEEE
Infocom, 2007, pp. 2054 - 2062.
[17] J. Pan, Y. T. Hou, L. Cai, Y. Shi, and S. Shen, "Topology
control for wireless sensor networks," in ACM MobiCom, San
Diego, CA, USA, 2003, pp. 286-299.
[18] C.-F. Huang and Y.-C. Tseng, "The Coverage Problem in a
Wireless Sensor Network," Mobile Networks and
Applications, vol. 10, pp. 519-528, 2005.
[19] M. Ma and Y. Yang, "Adaptive Triangular Deployment
Algorithm for Unattended Mobile Sensor Networks," IEEE
Transactions on Computers vol. 56, pp. 946-847, 2007.
[20] E. Amaldi, A. Capone, and F. Malucelli, "Planning UMTS
Base Station Location: Optimization Models with Power
Control and Algorithms," IEEE Transactions on Wireless
Communications, vol. 2, pp. 939-952, 2003.
[21] R. Mathar and T. Niessen, "Optimum positioning of base
stations for cellular radio networks," Wireless Networks, vol.
6, pp. 421-428, 2000.
[22] M. Mauve, J. Widmer, and H. Hartenstein, "A survey on
position-based routing in mobile ad hoc networks," IEEE
Network, vol. 15, pp. 30-39, 2001.
[23] M. d. Berg, M. v. Kreveld, M. Overmars, and O.
Schwarzkopf, Computational geometry: algorithms and
applications, 2nd ed. New York: Springer, 2000.
[24] P. Bose and P. Morin, "Online Routing in Triangulations," in
International Symposium on Algorithms and Computation,
1999, pp. 113-122.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
IEEE TRANSACTIONS ON COMPUTERS

[25] J. M. Keil and C. A. Gutwin, "Classes of graphs which Council Member of the Institute for Computer Sciences, Social–
approximate the complete euclidean graph," Discrete and Informatics, and Telecommunications Engineering (in Brussels). He
Computational Geometry, vol. 7, pp. 13-28, 1992. received the 1997 Edgeworth David Medal from the Royal Society
of New South Wales for outstanding contributions to Australian
[26] P. Bose, L. Devroye, M. Loffler, J. Snoeyink, and V. Verma, Science. Professor Zomaya is also the recipient of the Meritorious
"The spanning ratio of the Delaunay triangulation is greater Service Award (in 2000) and the Golden Core Recognition (in
than Pi/2," in Canadian Conference on Computational 2006), both from the IEEE Computer Society. He is a Chartered
Geometry, 2009, pp. 1-3. Engineer (CEng), a Fellow of the American Association for the
[27] J. R. Shewchuk. Triangle version 1.6, Advancement of Science, the IEEE, the Institution of Electrical
http://www.cs.cmu.edu/~quake/triangle.html. Available: Engineers (U.K.), and a Distinguished Engineer of the ACM. His
research interests are in the areas of high performance computing,
http://www.cs.cmu.edu/~quake/triangle.html, 2005.
parallel algorithms, mobile computing, and bioinformatics.
[28] P. Chew, "Guaranteed Quality Mesh Generation for Curved
Surfaces," in The Annual Symposium on Computational Selvadurai Selvakennedy obtained his PhD in computer
Geometry, 1993, pp. 274-280. networking in 1999 from the University of Putra, Malaysia. He
[29] J. Ruppert, "A Delaunay Refinement Algorithm for Quality 2- currently works as a consulting engineer with a professional
Dimensional Mesh Generation," Journal of Algorithms, vol. services company, and is also associated with the University of
18, pp. 548-585, 1995. Sydney, Australia. His research interests lie in developing protocols
[30] A. Ungor, "Off-centers: A new type of Steiner points for and algorithms for radio resource management, MAC and network
protocols, middleware and topology control issues in wireless mesh
computing size-optimal guaranteed-quality Delaunay
networks. He has served on the technical program committees of
triangulations," in Proceedings of LATIN, 2004, pp. 152-161. many international conferences. Dr. Selvakennedy is a professional
[31] L. P. Chew, "Guaranteed-Quality Triangular Meshes," member of ACM.
Department of Computer Science, Cornell University,
Technical Report TR-89-983,1989.
[32] NS-2. Network Simulator 2, http://www.isi.edu/nsnam/ns.
2008.
[33] F. Kuhn, R. Wattenhofer, and A. Zollinger, "Worst Case
Optimal and Average Case Efficient Geometric AdHoc
Routing," in ACM MobiHoc, 2003, pp. 267 - 278.
[34] F. Kuhn, R. Wattenhofer, Y. Zhang, and A. Zollinger,
"Geometric ad-hoc routing: of theory and practice," in ACM
PODC, 2003, pp. 63--72.
[35] B. Leong, S. Mitra, and B. Liskov, "Path Vector Face
Routing: Geographic Routing with Local Face Information,"
in ICNP, 2005, pp. 147--158.
[36] F. Harary, Graph Theory: MA: Addison-Wesley, 1994.
[37] P. Bose and L. Devroye, "On the stabbing number of a
random Delaunay triangulation," Computational Geometry:
Theory and Applications, vol. 36, pp. 89-105, 2007.

Weisheng Si received his bachelor and master degree in computer

science from Peking 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 PhD
student at the Centre for Distributed and High Performance
Computing, School of Information Technologies, University of
Sydney. His research interests include wireless mesh networks,
geometric routing algorithms, and graph theory.

Albert Y. Zomaya is currently Chair Professor of High Performance

Computing and Networking and an Australian Research Council
Professorial Fellow 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 19 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

14

Authorized licensed use limited to: Dayananda Sagar College of Engineering. Downloaded on August 17,2010 at 06:26:18 UTC from IEEE Xplore. Restrictions apply.