2010 IEEE 6th International Conference on Wireless and Mobile Computing.

Neworking and Communications

A New Algorithm for Backbone Formation in Ad
Hoc Wireless Networks of Nodes with Different
Transmission Ranges
Hossein Kassaei
Dept. of Computer Science, Concordia University
Montreal, Quebec, Canada H3G IM8
Email: h_kassae@cs.concordia.ca
Abstact-We consider the problem of backbone formation
in ad hoc wireless networks composed of heterogeneous nodes.
A virtual backbone in an ad hoc wireless network provides a
hierarchical infrastructure that can be used to address important
challenges such as efcient routing, muiticasting/broadcasting,
activity-scheduling, and energy efciency. We model a wireless
network in which nodes have diferent transmission ranges by a
dik graph. A virtual backbone in such a network can be modeled
by a Stongly Connected Domiating and Absorbent Set (SCDAS)
in the associated disk graph. For practical reasons, it is desirable
to minimize the size of this backbone. In this paper, we propose an
efcient distributed algorithm for the construction of an SCDAS
in ad hoc networks modeled by disk graphs. Extensive simulation
results show that the SCDAS constructed by our algorithm is
signifcantly smaller than those generated by the algorithms prior
to our work.
Index Ters-Wireless ad hoc networks, dominating set, ab
sorbent set, virtual backbone, directed graph.
I. INTRODUCTION
A wireless ad hoc network is an infastuctureless, peer
to-peer network of wireless nodes communicating with each
other in a multihop fashion. Such a network comes together
as the need arises and achieves a goal without relying on
any established infrastructure. The lifespan of an ad hoc
network depends on the application and may vary fom a few
hours to a possibly long period of time. Applications of such
networks include conferencing, home networking, personal
area networks, emergency services, and sensor networks. The
unique characteristics of ad hoc networking along with its
stringent requirements bring about many challenges that need
to be addressed before it can emerge as a fll-fedged technol
ogy. One such group of crucial problems are the intertwined
issues of efcient routing, broadcasting, scalability and energy
conservation. The use of a hierarchical infastructure in which
a small subset of nodes forms a virtual backbone in an
otherwise inherently fat network has been suggested in the
literature to address these problems. It is desirable to minimize
the size of this backbone because the smaller the size of the
backbone, the fewer the number of the nodes that need to keep
their transceivers on at all times and the smaller the size of
the routing tables that these nodes need to maintain.
978-1-4244-7742-5/10/$26.00 2010 Crwn 83
Lata Narayanan
Dept. of Computer Science, Concordia University
Monteal, Quebec, Canada H3G IM8
Email: lata@cs.concordia.ca
Unit Disk Graphs (UDGs) are typically used to model
wireless ad hoc networks with symmetric links. In such a
model, the virtual backbone described earlier corresponds to a
connected dominating set. In a UDG G " (V, E), a Connected
Dominating Set (CDS) is a subset V' V such that for every
node u, U is either in V' or has a neighbor in V' and the
subgraph induced by V'is connected. Minimizing the size
of the virtual backbone corresponds to fnding the smallest
possible CDS in the underlying graph; however fnding a
Minimum CDS (MCDS) in UDGs is known to be NP-hard
[4].
In reality, nodes in a network may not necessarily have the
same transmission range. This might simply occur when the
network consists of various kinds of wireless devices with
diferent powers and diferent functionalities. Even when the
network consists of similar nodes, these nodes may need
to adjust their transmission ranges for many reasons. For
example, in many power control schemes, nodes adjust their
transmission power to save energy, reduce collisions and so
on. All of these scenarios result in introducing asymmetric
links in the network. In the presence of unidirectional links
in the network, UDGs can no longer be used to model the
network. Instead, Disk Grphs (DGs) ae used to model ad
hoc networks of nodes with diferent transmission ranges. In a
DG G " (V, E), fnding a backbone translates to constructing
a Strngly Connected Dominating and Absorbent Set (SCDAS)
in G. An SCDAS is a subset of nodes V' V such that every
node u is either in V' or has an outgoing and an incoming
neighbor in V', and the subgraph induced by V'is strongly
connected. A directed graph is strongly connected if for any
two vertices U and v, there exists a directed path between u
and v. For the same practical reasons, we seek to minimize the
size of the SCDAS in DGs, however, that remains an NP-hard
problem.
CDS construction in ad hoc networks with symmetric links
has been extensively studied over the past decade. However,
not many algorithms have been proposed for networks with
asymmetric links and those few tat have been proposed
construct relatively large SCDASs or have high complexities
that make them impractical in the context of ad hoc networks.
In this paper, we propose a distributed algorithm for the
construction of SCDAS in ad hoc networks with asymmetric
links. We assume there is no topology change induced by
node mobility in the network and that in the face of mobility,
the recalculation of a new SCDAS is initiated once the
network stabilizes. Extensive simulation results show that our
algorithm outperforms the existing algorithms in the literature
and constructs SCDASs that are signifcantly smaller in size.
The remainder of this paper is organized as follows. In
section 2, we give an overview of the relevant work done
in this area. We then present our algorithm in Section 3.
The simulations conducted to evaluate the performance of
our algorithm are detailed in Section 4. Finally, Section 5
concludes the paper and points out possible directions for
fture work.
II. RELATED WORK
CDS formation in homogeneous wireless ad hoc networks
modeled by UDGs is a well-studied problem [ 1, 2, 3, 6, 7, 10].
However, the counterpat of this problem in networks of
heterogeneous nodes is relatively new. Wu [9] extended the
concept of a connected dominating set (CDS) in undirected
graphs to a connected dominating and absorbent set (SCDAS)
in directed graphs and gave a simple local algorithm for its
construction. Initially all the nodes are unmaked. A node u
is marked if there exists a node v in its dominating set and
a node w in its absorbent set, but v does not dominate w. In
other words, a node is marked only if it lies on the shortest
path fom one neighbor to another. Just as in the marking
algorithm of [ 10] , this making rule generates a large SCDAS
with a lot of redundant nodes. Therefore, Wu proposed two
rules to reduce the size of the constructed SCDAS. These
two rules are indeed the extended forms of the rules proposed
in [ 10] , which are applied to the nodes in the SCDAS. The
frst rule allows a node u to be removed fom the SCDAS
if its dominating (absorbent) neighbor set is covered by the
dominating (absorbent) neighbor set of node v with a higher
ID. The second rule allows a node u to be removed fom the
SCDAS if its dominating (absorbent) neighbor set is covered
by the union of the dominating (absorbent) sets of the two
connected nodes v and w provided that the ID of node u
is the smallest among the three nodes. This algorithm does
not provide a constant approximation ratio and its time and
message complexities are 9(m) and O(3) respectively.
In [8] , Park et al. proposed a centralized constant ap
proximation algorithm for the construction of a Minimum
SCDAS (MSCDAS) in wireless ad hoc networks with diferent
transmission ranges. In their work, it is assumed that the
ratio of the maximum to the minimum transmission range is
bounded. They also present two heuristics and evaluate their
performance through simulations. The heuristics are indeed
the counterpat of Guha and Khuller's algorithms [5] for
undirected graphs.
In the constant approximation algorithm, an outgoing span
ning tree and an incoming spanning tree rooted at an ar
bitrary node u are constructed. The non-leaf nodes of the
84
two trees form an SCDAS. Park et al. also presented two
centralized heuristics for the construction of the SCDAS in
directed graphs. Both of these heuristics rely on a Dominating
and Absorbent Set (DAS) as the input. The algorithm that
constructs the DAS consists of two stages: construction of a
Dominating Set (DS) and an Absorbent Set (AS). The union
of the two sets then forms the fnal DAS. In fnding the DAS,
initially all nodes are marked as uncolored. Then, at each
iteration, a node u with the highest degree is colored black
and all its uncolored neighbors with an incoming edge fom u
(its dominatees) become gray. This process terminates when no
uncolored node is left. At the end of this stage, the set of black
nodes forms a DS. Before proceeding to the construction of
AS, a preprocessing step is performed whose goal is to reduce
the number of nodes selected as AS by checking if any gray
node (a dominated node) is also absorbed by a black node. If
there exists such node(s), they are colored white.
Once the preprocessing phase is fnished, nodes in the graph
are either black, gray or white. Since white nodes are already
absorbed, the construction of AS is equivalent to fnding an
absorbent node for every gray node. In doing so, a greedy
approach is adopted: at each iteration, a gray node that absorbs
the highest number of gray nodes is marked black and its
absorbed gray nodes ae marked white. This stage terminates
when no gray node is lef, at which time the set of black nodes
forms a DAS.
Then two heuristics are proposed to make the above DAS
connected: (i) greedy spider contraction algorithm (G-SCA)
and (ii) greedy strongly connected component merging algo
rithm (G-CMA). The frst one uses a greedy approach to fnd
an approximation for the directed Steiner tree with minimum
Steiner nodes problem to minimize the number of white nodes
required to connect the black nodes. The second heuristic
iteratively fnds two strongly connected components which
can be merged at minimum cost among all the pairs of such
components, and merges them by coloring the white nodes
on the two directed paths between the two components black.
They conduct simulations that show G-CMA consistently out
performs G-SCA in terms of the size of the resulting set.
III. OUR ALGORITHM
In this section, we give an algorithm to construct an SCDAS
in wireless ad hoc networks with asymmetric links. Our
algorithm uses a pruning-based approach proposed in [7] ,
adapted to asymmetric networks modeled by directed graphs.
A. Defnitions and preliminaries
In our algorithm, we use Nd(U) to denote the dominating
neighbor set of node u, i.e. Nd(U) " {vl(v, u) E E}. A node
v E Nd(U) is also referred to as an incoming or ingress
neighbor of node U in the literature. Likewise, Na(u) is used
to denote the absorbent neighbor set of node u; i.e Na(u) "
{vl(u, v) E E}. A node v E Na(u) is also referred to as an
outgoing or egress neighbor of u. Figure 1 illustrates the
dominating and absorbent neighbor sets of a node. Note that
these two sets may overlap. In other words, a neighbor of node
u can be both a dominator and an absorbent of node u. We
defne the degree of a node u to be INd(U) 1 + INa(u) l.
(/
v
.
.
........
dominating neighbor set
Ndu)
...
..... .
..
)'<\
, .
, .
: 1
. '
. '
, ,
.... >1: ...
u
w
Absorbent neighbor set
Na(u)
Fig. 1. Dominating and absorbent neighbor sets of node u
Every node u has a rank (o(u) , id(u)) which is an ordered
pair of its efective degree and id, where the efective degree of
node u is the number of u's neighbors in SCDAS, i.e o(u) "
I{vlv E Na(u) 1 v E SCDAS}I + I{vlv E Nd(U) 1 v E
SCDAS}I. Since the membership of nodes in the SCDAS
changes during the algorithm, so does the efective degree
of a node. Assigning a unique id to every node provides a
mechanism to break ties.
The rank of a node in our algorithm can be considered a
general weight that is defned based on the goal function. Since
we intend to minimize the size of the constucted SCDAS,
we use a node's effective degree in the defnition of its rank.
Alternatively, if we seek to prolong the node's lifetime, we
can defne its weight as the node's remaining energy. In the
following section, we present the centralized description of
our algorithm for ease of exposition, and then we will discuss
the distributed implementation.
B. Centralized Description
We adopt a pruning-based approach in this algorithm.
Initially, every node is in the SCDAS and has pending status.
At each step, we select the node u with the smallest rank
fom the set P of the nodes with pending status. We remove
node u fom P and apply two local tests to see if it can
be removed from the fnal SCDAS. The two tests are called
(i) the domination and absorbency test and (ii) the strng
connectivit test. If u passes both tests, we change its status to
out; otherwise, its status becomes in and we will keep it in the
fnal SCDAS. Node u passes the domination and absorbency
test (DAT) if:
(a) all the nodes that are dominated by u have at least one
other dominator.
(b) all the nodes that are absorbed by u have at least one other
absorbent.
Node u passes the strong connectvity test if the subgraph in
duced by its dominators and absorbents is strongly connected.
The formal description of this extended algorithm is given in
Algorithm 1.
It is easy to see that the removal of a node u that passes both
tests fom the fnal SCDAS neither leaves any of its neighbors
without a dominator/absorbent, nor disconnects the SCDAS,
and that both tests are needed to guarantee this. Therefore,
the set of the nodes that form the fnal solution is indeed an
SCDAS of the original graph. Also, since at each round one
85
Algorithm 1 Centralized Strongly Connected Dominating and
Absorbent Set (SCDAS) Algorithm
SCDAS fV
PfV
while P = 0 do
u fargmin{(o(v), id(v) ) lv E P}
pfp - {u}
DAT ftrue
I For ever node v dominated by u, check i there is some
lother node that dominates v
for all v E Na(u) do
if ((Nd(V) n SCDAS) - {u} " 0) then
DAT ffalse
end if
end for
IFor ever node v absorbed by u, check i there is some
lother node that absorbs v
for all v E Nd(U) do
if ((Na(v) n SCDAS) - {u} " 0) then
DAT ffalse
end if
end for
if DAT then
G fGraph[((Na(v) U Nd(V)) n SCDAS) - {u}]
if G is strongly connected then
SCDAS fSCDAS - {u}
for all v E (Na(u) n SCDAS) do
o(v) fo(v) - 1
end for
for all v E (Nd(U) n SCDAS) do
o(v) fo(v) - 1
end for
end if
end if
end while
Retur SCDAS
node is removed fom the set P, the algorithm terminates in
n steps.
C Distributed Implementation
In order to run the above algorithm in a distributed fashion,
every node needs to perform a preliminary setup which
involves obtaining some information from its neighbors and
setting some variables that will be used later during the
execution of the algorithm.
Initially, every node u sets its status to pending and ex
changes its rank with all its neighbors (incoming and outgoing)
and stores the set of its neighbors in Nu,SCDAS, a vaiable
holding the set of neighbors in the SCDAS. Also, it maintains
the list of its lower rank neighbors in Lower _Ranku. Note
that in this section, whenever we talk about neighbors in
general, we mean both incoming and outgoing neighbors.
When a node fnishes running the algorithm, it sends a
Finished_M sg( Status) to all its neighbors with the Status
Algorithm 2 Distributed Connected Dominating and Ab
sorbent Set Algorithm, executed by node u
when Lower _Ranku " 0
DAT f true
Send Dominator _Query to all nodes in Na(u)
Send AbsorbenCQuery to all nodes in Nd(U)
for all (v E (Na(u) U Nd(U))) do
Wait for Dominator _Reply( v, Dv) and
AbsorbenCReply( v, Av)
if ((Dv - {u} " 0)or(Av - {u} " 0)) then
DAT f false
end if
end for
if DAT then
if G[Nu,SCDAS - {u}] is strongly connected then
Statusu f out
Send Finished_Msg(out) to (Na(u) U Nd(U))
else
Statusu f in
Send Finished_Msg(in) to (Na(u) U Nd(U))
end if
else
Statusu f in
Send Finished_Msg(in) to (Na(u) U Nd(U) )
end if
Upon reiving (DominatorIAbsorbentLQueT_Msg from
v:
if !(Replyjn_Transit) then
Replyjn_Transit f true
Send (Dominator/Absorbent)_Reply(u, Du/Au) to v
else Enqueue (Dominator / Absorbent )_Query( v) in
DQQ
end if
Upon reiving Finihed_Msg(sttus) from v:
if status " out then
if ((v E Na(u) ) &(v E Nd(U)) ) ten
8(u) f 8(u) - 2
else
8(u) f 8(u) 1
end if
Nu,SCDAS " Nu,SCDAS - {v}
end if
if (Rank(v) < Rank(u)) then
Lower _Ranku " Lower _Ranku - {v}
end if
if DQQ = 0 then
v f Dequeue DQQ
Send Dominator/AbsorbenCReply(u, Du/Au) to v
else Reply_In_Transit f false
end if
86
being in or out depending on the decision it has made afer
running the tests.
Afer the initial setup, a node u with the lowest rank among
its neighbors becomes the initiator and runs Algorithm 2. Since
the ranks assigned to nodes ae unique, there exists at least
one such initiator. In fact, the experimental results show that
in random scenarios, on the average, the algorithm exhibits a
desirable degree of parallelism.
Four messages are used in conjunction with the
domination and absorbency test. Dominator _Query and
Dominator _Reply messages are used to verify if the nodes
dominated by u have other dominators or not. Likewise,
AbsorbenCQuery and AbsorbenCReply messages are used
to verif if the nodes absorbed by u have other absorbents.
An outgoing neighboring node v includes the list of its
dominators, Dv, in Dominator _Reply (v, Dv) and an incom
ing neighbor v includes the list of its absorbents, Av, in
AbsorbenCReply( v, Av). In order to rule out the possibility
of the simultaneous dopout of two nodes that dominate/absorb
the same node v, node v also reserves two fags; Domina
tor_Reply jn_Transit and AbsorbenCReply jn_Transit fags
that are initially set to false. These fags ae set to true
whenever v has sent a reply to a Dominator _Reply or an
AbsorbenCReply message, but has not received the corre
sponding Finished_M sg( Status) message fom that node.
As long as the fags ae true, node v enqueues any frther
queries that it might receive fom other neighbors and replies
to them only when it becomes aware of the decision made by
the node u to which it has a reply in transit.
Once all the replies are received fom the neighbors, node
u proceeds to the strong connectivity test if all the neighbors
dominated by u have at least one other dominator and all
the nodes absorbed by u have at least one other absorbent.
Otherwise, node u changes its status to in and sends a
Finished_M sg( in) to all its neighbors.
The stong connectivity test at node u examines if the
subgraph induced by Nu,SCDAS is strongly connected. Note
that this test does not require any message passing and can
be done locally by u, using the local variables which it
maintains. If it also passes the second test, node u drops out
of the SCDAS and sends a Finished_M sg( out) to all its
neighbors. Otherwise, it stays in the fnal SCDAS and sends
a Finished_Msg(out) to its neighbors.
Upon receiving a Finished_M sg( Status) fom a neighbor
v, node u removes v fom the list of its lower rank neighbors
Lower _Ranku. Additionally, if the status of the node v is
out, it removes it fom the set of its neighbors in the SCDAS,
Nu,SCDAS. The formal description of this algorithm is given
in Algorithm 2.
Before proceeding to the experimental results, we explain
how a node's efective degree is computed and updated in the
above algorithm. Initially, every node u sets its efective degree
to 0 (8(u) " 0). Then it increments 8(u) for each incoming
or outgoing neighbor v. If node u has a bidirectional link to
node v; i.e. node v is both in Na(u) and Nd(u) , then node
v causes 8 ( u) to be incremented by 2. Correspondingly, we
decrease 8 ( u) by 2 during the execution of the algorithm if a
bidirectional neighbor v drops out of SCDAS in Algorithm 2.
D. k-Hop Extension
The stong connectivity test in Algorithm 2 can be extended
to a k-hop neighborhood. It is easy to see that a node
can make more informed decisions about the role it plays
in maintaining the strong connectivity of te SCDAS if it
considers a larger neighborhood. By doing so, it is able to
decide if it is redundant and if so, removes itself fom the
fnal SCDAS, thereby resulting in reducing the size of the
fnal solution. However, it should be noted tat this comes at
the expense of message overhead. In order to implement the
k-hop extension, every node should exchange its rank with
its k-hop neighbors and adjust its local variables accordingly.
Also, when a node fnishes running the tests, it should send
the Finished_Msg(Status) to its k-hop neighborhood.
E Perorance Analysis
The analysis of time and message complexities of Algorithm
2 for the computation of an SCDAS is complicated by the
existence of unidirectional links in the network. If node u has
a unidirectional link to node v, then v can directly receive
packets fom u and is therefore aware of the existence of
its incoming neighbor (dominator); however, node u cannot
directly hear fom node v and thus is not aware of its existence.
In other words, the main issue is that a node may not be able
to directly identif its outgoing (absorbent) neighbor(s). One
solution is to have each node in the network emit a beacon,
with its I D appended to it, at regular intervals. Any node
that receives a beacon appends its own I D and forwads it.
Since we assume the underlying graph is strongly connected,
every node will eventually hear fom its absorbent neighbor(s)
and can detect them using the chain of IDs appended by
forwarding nodes. Using this solution, the messages exchanges
when a node wants to identify the set of its absorbents, or
when it asks them about their other dominators during the
domination and absorbency test, is not necessaily restricted
to a fxed-size neighborhood. The reply messages fom an
absorbent node v to a dominator u may be forwarded along a
path containing O(n) nodes. With this complication in mind,
the analysis of Algorithm 2's time and message complexity is
as follows.
Assuming is the degree of the underlying graph that mod
els the network, the size of u's k-hop neighborhood is bounded
by k. During the initial setup phase, every node sends
a constant number of messages to its k-hop neighborhood
inquiring about the ranks of the nodes in this neighborhood.
If all the links were bidirectional, this would translate to
O(nk) messages, however, since a reply message might need
to be forwarded along a path of O(n) nodes, the message
complexity of this phase is 0 ( n 2 k). During the execution
of the algorithm, 0 ( n 2 k) messages ae required for the
domination and absorbency test and no message passing is
required for the sting connectivity test. Since every node sends
(nk) Finished_Msg(Status) messages once it completes
87
the execution of the tests, the message complexity of this last
round of the Algorithm is 0 ( n 2 k). Thus, the overall message
complexity is O(n2k).
The execution time of Algorithm 2 depends on the longest
chain of dependency in the underlying graph where each
node has to wait for nodes of smaller rank in the chain to
fnish running the algorithm before it can run the algorithm.
Since this chain could be of length O(n) in the worst case,
and each node may have to wait for O(n) time to receive
the information about its absorbents, the time complexity of
Algorithm 2 is O(n2).
However, for practical reasons, we are interested in net
works in which there is a bound on the maximum length of
the directed reverse path between any pair of nodes with a
directed edge. We call a directed graph a-reciprocal if for
every directed edge (u, v) E E, there exists a directed path
fom v to u of length at most a hops. Under this assumption,
the bound for the time and message complexities of Algorithm
2 would be O(an) and O(ank), respectively.
IV. EXPERIMENTAL RESULTS
We conducted extensive simulations to evaluate the perfor
mance of our algorithm in networks with asymmetric links
in our own simulator developed using Java SE 6. To study
the impact of the perentage of unidirectional links (PUL) as
well as diferent node densities on te size of the constructed
SCDAS, we experimented with varying number of nodes and
transmission ranges. We considered a geographic area of 200m
by 200m and varied the number of nodes fom 50 to 300 with
increments of 50. Also, each node was assigned a transmission
range randomly selected fom the range [r _min, r _maxl.
The only algorithms proposed in the literature to construct
an SCDAS in networks with diferent tansmission ranges
are the ones in [8] and [9] , which were discussed in detail
in the related work. Thus, we compared the performance of
our proposed algorithm with the localized making algorithm
in [9] , hereafer referred to as Wu after the name of the
author and the two centralized algorithms in [8] , namely
Dominating-Absorbent Spanning Tree (DAST) and Greedy
Strongly Connected Component Merging Algorithm (G_CMA).
In our performance compaison, we focused on te impact of
node density and the percentage of unidirectional links on the
size of the SCDAS constructed by the four algorithms.
Clearly, the [r _min, r _maxl range afects the percentage of
unidirectional links in the network. Therefore, we created fve
diferent scenaios in which we experimented with diferent
ranges to generate diferent PULs. The transmission ranges
in the frst four scenarios were selected from [10m,50ml,
[20m,50ml, [30m,50ml, and [40m,50ml respectively. In the
last set, all the nodes were assigned the fxed transmission
range of 50m to make it possible to also compare diferences
between UDGs and DGs as input graphs. For each value of
n in each scenario, we generated as many random graphs as
required until we had 1000 strongly connected graphs. The
graphs were stored and used across diferent simulations using
diferent algorithms.
80 -10-50
--20-50
'
70 Min-Max transmission range:
..
30-50
-
:
.40-50
Ii
60 -50
c
0
B 50

:
40
c
:
.

^
=~^
-
M"^

.
'
30
Q
t

*
c
20
Q
u
(
"
10
..................................................................................................
0
50 100 150 200 250 300
Number of nodes
Fig. 2. Relationship between maximum-to-minimum transmission range ratio
and percentage of unidirectional links
We present the results in the four following sections. In
Section IV-A, we look at the relationship between the ratio
of the maximum to minimum transmission range on the per
centage of unidirectional links in the input graphs. In Section
IV-B, we investigate the impact of the degree of locality with
which nodes run the connectivity test in Algorithm 2 on the
size of the constucted SCDASs. Then, we will compare the
performance of our algorithm with that of its competitors in
Sections IV-C and IV-D under varying node densities and
percentages of unidirectional links.
A. Impact of Trnsmission Range on The Percentage of Uni
directional Links
As illustrated in Figure 2, it can be generally seen that the
percentage of unidirectional links in the network is a function
of the ratio of maximum to minimum tansmission range and is
almost independent of the node density in the network. For ex
ample, in our simulations, when the transmission range varies
between 40m and 50m; i.e. _ " 1.25, the percentage of
unidirectional links varies between 11.2% to 12% for different
values of n. When the
_
: ratio is increased to 1.67, 2.5
and 5, the percentage of unidirectional links rises to 25%, 40%
and 52%, respectively. The only cases in which graphs exhibit
slightly unpredictable behavior is when the ratio of maximum
to minimum transmission range is high ( : " 5) and the
network is very sparse (n " 50,100). Finally, as expected,
when all the nodes have the identical transmission range of
50m, there are no unidirectional links in the network.
B. Impact of Localit on The Size of The SCDAS Constructed
by Our Algorithm
We also investigated the efect of the degree of locality
k with which the connectivity test is run on the size of the
generated SCDAS. Our goal was to experimentally determine
the best tradeof between the degree of locality in this test and
the number of nodes that can be pruned. One sample of our
results is shown in Figure 3. According to the results obtained
88
70
60
'

0
so u
'
.S

"
40
L
L
'
"
30

C
"
I
20 "
L
10
-..Num_of_nodes: 50
--Num_oCnodes: 100
..Num_of_nodes: 150
"Num_oCnodes: 200

.~^~~~
^^W~~=:

"P
.

^^^*************
:
~
. ^~q&q@,,,,,,, @,,,,,, .g...^..~ @ ~~~~~~.
, .
........
.................................... ! ............. ! ............. .............. ! .............. ! ............... ! ............... !
10
Locality of the connectivity test (k)
Fig. 3. Impact of the locality of connectivity test on the size of the SCDAS
when tansmission ranges vary between 20 m and 50 m
fom the experiments, on the average, k " 4 strikes a good
compromise; while in sparser networks the size of the set can
still be reduced by increasing k to 5 or 6, in denser instances
the gain is typically negligible for k's beyond 2 or 3. This
can be seen in Figure 3 where the curves gradually fatten out
beyond k " 5 in relatively sparse networks (n " 50,100) and
beyond k " 3 in denser networks. Therefore, in comparing
our algorithm with its competitors, we selected the I-hop and
4-hop implementations of our distributed algorithm; namely
PInOuCUnidirectional Distributed (k " 1) (PInouCUDl) and
PnOuCUnidirectional Distributed (k " 4) (PInOuCUD4).
C Impact of Node Densit
In this section we will investigate the impact of node density
on the performance of the four selected SCDAS construction
algorithms. Figures 4 through 6 show three diferent node
densities, in ascending order, in the presence of different PULs
in the network. Note that the graphs for n " 150, n " 200, and
n " 250 are not shown because they exhibit the same trends.
As can be seen in Figure 4, Wu performs better than DAST
in spase networks, especially when PULis higher. However,
as PU L drops below 12% and the underlying graph tends to
a UDG, DAST catches up and outperforms Wu. PInOucUDI
and PInOucUD4 consistently outperform the other three al
gorithms while G_CMA is closer to PInOucUDl, standing
in the middle. As the number of nodes increases to 100
and 150 (moderate densities), DAST consistently outperforms
Wu, and the gap between DAST and Wu as one group and
G_CMA, PInOucUDI and PInOucUD4 as the other group
widens, especially in the presence of PULs of 12% and higher.
Another noticeable tend is that DAST almost maintains the
same distance from the algorithms in the second group up to
PLUs of around 12%, but then considerably narrows down the
gap as PUL tends to zero.
90
80
V

C
70
u
V
: 60
:
-
50
g
' 40
Q
30
c

20
:;
L
10
Fig. 4.
80
V
70

C
u
60
V
-:

50
Q
-
0
c
'
40
Q

30
c
Q
20

Q
0.
10
Fig. 5.
70
V

60
C
u
V
50
.=

Q
-
0
40
c
'
30
Q

c
20
Q

Q
0.
10
Fig. 6.
-DAST
^

-G_CMA
.Wu
M.""..""..""..""."""."'".'"M'

-PlnOut_UDl
.
"'
.
"
""'":
"
:
":
:

: "

... . .
,.
.
:
.

""..."

Min-Max transmission range

Impact of the node density - number of nodes 50

"

"""..
p
^
$_

"
"
<
:':
'

Min-Max transmission range
.... wu
_PlnOut_UDl
" PlnOut_UD4

Impact of the node density - number of nodes 00

_

"
.
".""
_
"""".
""""..
""
_

"

"""

..wu
_PlnOut_UDl
" PlnOut_UD4
^
' ^'

B
Min-Max transmission range
C
U
Impact of the node density - number of nodes 300
As the number of nodes increases to 200, 250, and 300
(extremely dense networks), PInOucUDl and PInOucUD4
take the lead more conspicuously and PInOucUDl increases
its distance from G_CMA. On average, PInOucUDl and
PInOucUD4 construct SCDASs which are 24% and 35%
smaller than those constructed by G_CMA respectively. The
89
efciency of our schemes becomes even more noticeable when
we take into account the very high complexity of G_CMA
and the fact that it is a centralized algorithm. Finally, the last
consistent trend that can be seen from the fgures is that as the
node density increases, the difference between PInOut_UDl
and PInOuc UD4 decreases and in very dense networks, they
perform almost equally well, especially for PULs of 40% and
lower.
D. Impact of Unidirectional Links
In order to better analyze the impact of PUL on the perfor
mance of algorithms which we simulated in our experiments,
we give a diferent presentation of our results in this section.
Figures 7 and 8 show two diferent [Tmin, Tmaxl ranges and
the size of the SCDAS constructed by each algorithm under
varying number of nodes. As we discussed earlier, each
[Tmin, Tmaxl range corresponds to an almost fxed PUL. More
specifcally, if we ignore the slightly diferent trends in very
sparse networks in the presence of high PULs and round up the
average PULs for different [Tmin, Tmaxl transmission ranges,
the transmission ranges [ 10,50] and [30,50] correspond to
PULs of 52% and 25%, respectively. As depicted in Figure
9, we also considered the scenario in which PUL is zero to
make it possible to more accurately predict the trends as PUL
drops below 12% and tends to zero.
As illustrated in Figure 7, although Wu initially outperforms
DAST when n " 50, it does not improve much as the node
density grows in the presence of high PULs. In other words,
when the ratio of maximum to minimum transmission range
is very high in the network, Wu cannot take advantage of
the increase in node density whereas the other algorithms all
beneft from increased node density and reduce the relative
size of the SCDAS which they construct. The reason for Wu's
inability to use increased density in its favor is that its pruning
rules (Rules 1 & 2) are not efcient when PUL is high.
By comparing Figures 7 and 8 with Figure 9, it can be
seen that the existence of unidirectional links in the network
adversely affects the performance of DAST As can be seen
in Figure 9, when there are no unidirectional links, DAST's
performance is much closer to that of G_CMA compared
to Figures 7 and 8; it constructs SCDASs which are 128%
larger than those generated by G_CMA on average. However,
The performance of DAST degrades as the percentage of
unidirectional links increases. The size of the SCDAS built
by DAST is 153%, 175%, 189%, and 191% lager than that
built by G_CMA when the minimum transmission range is 40,
30, 20, and 10 respectively. This shows that DAST is more
sensitive to the presence of unidirectional links compaed to
G_CMA, PInOucUD1, and PInOucUD4.
The last interesting observation is about the relationship be
tween the locality of the connectivity test (k) in our algorithm
and the PUL. As shown in the fgures, as the PUL decreases,
so does the improvement in the SCDAS size as a result of
increasing k. The reason is that detecting strong connectivity
is more diffcult in smaller neighborhoods (e.g. k " 1,2) when
a large percentage of links are unidirectional. In other words,
the lower the PUL, the smaller the neighborhood required to
detect strong connectivity in the graph.
90
80
'P' wu
V
<
70
Q
U
V
.
60
^
^ _
-
_
- -- .-
@
7`} '-
P "
^
^
.''-
''*-
'.'"" '`."-
' '."'
".
-PlnOut_UDl
"MPlnOut_UD4
*

" 50
0
c
M
40 0

W
! 30
L

Q
20
L
10
50 100 150 200 250 300
Number of nodes
Fig. 7. Percentage of unidirectional links 52% [rmin, rmax] [10,50]
80
70
%
.wu
V
<
.....

'.< ... Q
60
u
V

-.

-PlnOut_UDl
"^ PlnOut_UD4
.! W
%

-
:
50
&
%
%
%
" & %
%
0
^
c 40
%
M

30
e
c

20

L
10

50 100 150 200 250 300

Number of nodes
Fig. 8. Percentage of unidirectional links 25% [rmin, rmax] [30,50]
50
45
V
40 <
Q
U
V 35
.

30
"
0
c 25
'
20
W
!
c 15

10
L

''' ''".'"
'''
&

'
A

'
l, ',
....
&
A
50 100
'
.---*.." `' '
M
'.
"Wu
_PlnOut_UDl
"^PlnOut_UD4

'---
-:-.---:--.:
150 200 250 300
Number of nodes
Fig. 9. Percentage of unidirectional links 0% [rmin, rmax] [50-50]
Our simulations show that when PUL is around 52%
(transmission range " [ 10,50]), there is an average reduction
of 21 % in the size of the SCDAS as k is increased from 1 to
4. However, this gain is reduced to 17%, 13%, and 9% when
90
PUL is 40%, 25% and 12% respectively. As seen in Figure
9, when there ae no unidirectional links in the network, and
the network is extremely dense, this gain is only around 3%.
In summary, using a larger neighborhood in the connectivity
test is helpful when PUL is relatively high. For lower PULs,
using a larger neighborhood helps only when the network is
not very dense .
V. CONCLUSIONS
In this paper, we studied the problem of virtual back
bone formation in wireless ad hoc networks with asymmetric
links by constructing a Strongly Connected Dominating and
Absorbent Set (SCDAS) of the underlying directed graph.
We proposed an efcient distibuted pruning-based algorithm
for the construction of the SCDAS. Extensive simulations
were conducted to study the impact of various node and link
densities on the performance of the proposed algorithm. The
results show that our algorithm consistently outperforms the
other algorithms in the literature in terms of the size of the
constructed sets. Our algorithm also provides the fexibility to
adjust the tradeof between the size of the generated set and
the cost of its construction; i.e. time complexity and message
overhead.
REFERENCES
[1] K. Alzoubi, P.-J. Wan, and O. Frieder. New distrbuted algo
rithm for connected dominating set in wireless ad hoc networks.
In Proceedings of the 35th Hawaii Interational Conerence on
System Sciences, pages 3849-3855, Jan. 2002.
[2] S. Basagni, M. Mastrogiovanni, and C. Petrioli. A performance
comparison of protocols for clustering and backbone formation
in large scale ad hoc network. In Proceedings of The 1st
IEEE Interational Conference on Mobile Ad Hoc and Sensor
Systems, MASS 200, pages 70-79, October 25-27 2004.
[3] S. Butenko, X. Cheng, C. A. Oliviera, and P. Pardalos. A new
heuristic for the minimum connected dominating set problem
on ad hoc wireless networks. In Recent Developments in
Coopertive control and optimization, pages 61-73, 2004.
[4] B. N. Clark, C. 1. Colboum, and D. S. Johnson. Unit disk
graphs. Discrete Math., 86(1-3):165-177, 1990.
[5] S. Guha and S. Khuller. Approximation algorithms for con
nected dominating sets. Algorithmica, 20(4):374-387, 1998.
[6] B. Han. Zone-based virtual backbone formation in wireless ad
hoc networks. Ad Hoc Netw., 7(1):183-200, 2009.
[7] H. Kassaei, M. Mehrandish, L. Narayanan, and 1. Opatmy.
Efcient algorithms for connected dominating sets in ad hoc
networks. In Proceedings of the IEEE Wireless Communications
and Networking Conference (WCNC), 2010.
[8] M. A. Park, J. Willson, C. Wang, M. Thai, W Wu, and
A. Farago. A dominating and absorbent set in a wireless ad
hoc network with different transmission ranges. In Proceedings
of the 8th ACM international symposium on Mobile ad hoc
networking and computing (MobiHoc '07), pages 22-31, 2007.
[9] 1. Wu. Extended dominating-set-based routing in ad hoc wire
less networks with unidirectional links. IEEE Trns. Parllel
Distrib. Syst., 13(9):866-881,2002.
[10] J. Wu and H. Li. On calculating connected dominating set for
efcient routing in ad hoc wireless networks. In DIAL '99:
Proceedings of the 3rd interational workshop on Discrete algo
rithms and methods for mobile computing and communications,
pages 7-14, 1999.