You are on page 1of 14

Multipath Dissemination in

Regular Mesh Topologies


Kranthi K. Mamidisetty, Minlan Duan, Shivakumar Sastry, Member, IEEE Computer Society, and
P.S. Sastry, Senior Member, IEEE
AbstractMesh topologies are important for large-scale peer-to-peer systems that use low-power transceivers. The Quality of
Service (QoS) in such systems is known to decrease as the scale increases. We present a scalable approach for dissemination that
exploits all the shortest paths between a pair of nodes and improves the QoS. Despite the presence of multiple shortest paths in a
system, we show that these paths cannot be exploited by spreading the messages over the paths in a simple round-robin manner;
nodes along one of these paths will always handle more messages than the nodes along the other paths. We characterize the set of
shortest paths between a pair of nodes in regular mesh topologies and derive rules, using this characterization, to effectively spread
the messages over all the available paths. These rules ensure that all the nodes that are at the same distance from the source handle
roughly the same number of messages. By modeling the multihop propagation in the mesh topology as a multistage queuing network,
we present simulation results from a variety of scenarios that include link failures and propagation irregularities to reflect real-world
characteristics. Our method achieves improved QoS in all these scenarios.
Index TermsWireless communication, network communications, packet-switching networks, routing protocols, mesh topology.

1 INTRODUCTION
T
HE use of novel devices for computing and communica-
tion in highly engineered networked embedded systems
suchas streetlight management [1], reconfigurable conveyors
[2], and critical infrastructures [3], presents new challenges
and opportunities for monitoring and diagnostics. These
systems contain a large number of nodes, with each node
incorporating a tiny microcontroller, sensors, actuators, and
anintegratedlow-power transceiver [4]. The nodes interact in
a peer-to-peer manner over low-bandwidth wireless links to
achieve the application objectives. To increase the number of
simultaneous interactions between the nodes in the system,
the transmission range of each node is limited so that it
communicates directly only with its set of immediate
neighbors; such an arrangement of nodes is referred to as a
mesh topology [5].
Multihop communications are necessary in such systems
to send messages from any source to any destination. For
example, intermediate nodes must forward messages to a
monitoring station from nodes that cannot communicate
directly with the monitoring station. Routing protocols are
used extensively in wired and wireless networks to support
multihop communication [6]. Such protocols construct and
maintain routing tables at each node by relying on system-
wide unique node identifiers. When the number of nodes is
very large, such as in sensor networks, it is not feasible to use
such identifiers. Several techniques, called dissemination
methods, were developed at the network layer to regulate
the flow of messages between nonadjacent nodes without
relying on unique node identifiers or constructing routing
tables using these identifiers [7], [8], [9].
In this paper, we consider highly engineered systems
comprising of nodes arranged in a regular mesh topology.
We focus on methods for effectively utilizing all the shortest
paths available between a pair of nodes and present results
to show that effective utilization of all the available paths
significantly improves the Quality of Service (QoS).
In many highly engineered systems, one can assume that
the nodes have fixedrelative locations. Often, the systems are
designed to overlay on an underlying grid. For example, in
automation systems, the regions demarcated by such a grid
are calledzones [10]; a zone is a commonlyusedabstractionto
support the design, operation, and maintenance activities.
Motivated by applications in such domains, we consider
regular meshtopologies that arise by embeddingthe nodes in
a 2D Basegrid. We show in Section 2 that several mesh
topologies arise when the location in the 2DBasegrid and the
transmission range of the nodes change. Because the grid
coordinates can specify the nodes, the shortest paths between
any pair of nodes can be locally computed. Most routing
protocols select only one of the shortest paths even when
multiple such paths exist. This results in reduced system-
level QoS [11]. We address the issue of how to effectively
utilize all the shortest paths available. Since the resulting
methods amount to a node making local decisions on howto
distribute messages among its immediate neighbors, without
having to dynamically construct any routing tables, we refer
to this method of forwarding messages as dissemination in
spite of the fact that nodes are identified by their global
coordinates in the underlying 2D Basegrid.
Because each node communicates directly with its im-
mediate set of neighbors, there are multiple shortest paths
betweenmanypairs of nodes inameshtopology. The number
of such paths is limited by the relative locations of the nodes.
For example, the number of shortest paths between certain
1188 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 20, NO. 8, AUGUST 2009
. K.K. Mamidisetty, M. Duan, and S. Sastry are with the Department
of Electrical and Computer Engineering, University of Akron, Akron,
OH 44325-3904.
E-mail: {km50, ssastry}@uakron.edu, MINLAN_DUAN@denso-diam.com.
. P.S. Sastry is with the Department of Electrical Engineering, Indian
Institute of Science, Bangalore 560 012, India.
E-mail: sastry@ee.iisc.ernet.in.
Manuscript received 11 Sept. 2007; revised 10 Apr. 2008; accepted 11 July
2008; published online 27 Aug. 2008.
Recommended for acceptance by T. Abdelzaher.
For information on obtaining reprints of this article, please send e-mail to:
tpds@computer.org, and reference IEEECS Log Number TPDS-2007-09-0312.
Digital Object Identifier no. 10.1109/TPDS.2008.164.
1045-9219/09/$25.00 2009 IEEE Published by the IEEE Computer Society
pairs of nodes is one, despite the mesh topology. We define a
Contour as the unionof all the shortest paths betweena pair of
nodes and present some results to precisely characterize the
structure of contours. Using this structure, we show that
when messages are spread in a round-robin manner, nodes
along one path in the contour will always handle more
messages than the nodes along other paths in the contour.
Consequently, the benefits of the multiple paths cannot be
fully realized. We then present a strategy for spreading
messages to neighboring nodes that effectively exploits the
available shortest paths andshowthat our rules for spreading
the messages result in a balanced loading of all the available
shortest paths. We refer to this approach as Contour Guided
Dissemination (CGD).
CGD improves the QoS by disseminating the messages
over all the available shortest paths. Typical application
scenarios in which CGD will be useful are: 1) a node
responding to a diagnostic query from an operator at a
monitoring station, and2) a node sendinga recordedincident
to a monitoring station in a surveillance application. CGD
equitably disseminates the messages over all the available
shortest paths in a manner that maximizes utilization of all
available shortest paths. We present simulation results to
demonstrate all these aspects.
Routing protocols used in traditional wired and wireless
networks are based on shortest path algorithms such as the
Bellman-Ford algorithm [12] and Dijkstras algorithm [6].
Similar protocols have been reported for ad hoc, wireless,
and mobile networks [13], [14], [15], [16]. The QoS achieved
in these systems has also been studied [17], [18], [19]. The
dissemination method we describe in this paper is some-
what similar to a gradient dissemination scheme [13] with
the cost metric being the deviation from evenness of load
distribution on all available shortest paths.
Recent efforts have focused on exploiting multipaths to
improve the QoS in systems [7], [20], [21] using constrained
node distribution models. For example, in [21], the nodes are
distributed randomly in a unit disk. In [7], the nodes are
distributed in a narrow strip no wider than 0.86 times the
transmission range of a node. The multipath, multihop,
approachwe present inthis paper also assumes a constrained
distributionof the nodes, namely, nodes are ona 2DBasegrid.
While there is no stochasticity in the node distribution, we
precisely characterize the geometry of the shortest paths and
show how to exploit it to achieve better QoS.
In the current methods, the motivating factors for
considering multipath routing include fault tolerance, high-
er aggregate bandwidth, and load balancing [22], [20], [23],
[24], [25]. The QoS aspects of multipath routing have been
addressed in [26], [27], [28], and [29]. The split multipath
routing protocol maintains maximally edge-disjoint paths
[30]. Braided multipaths discussed in [31] are useful when
the routing is coupled with diagnostic or prognostic
methods to select alternative paths. All such methods
essentially focus on the discovery and maintenance of
multiple paths that are useful under various constraints on
the node distribution. In contrast, we assume regular mesh
topologies and then precisely characterize the set of all
shortest paths between any pair of nodes. We then use this
geometric structure to propose rules for dissemination that
result in all available shortest paths being utilized effec-
tively so that QoS improves.
The rest of the paper is organized as follows: Section 2
presents embedding functions that characterize a class of
mesh topologies that we consider for this work. Section 3
discusses contours, their structure, and the significance of
the structure on dissemination methods. Section 4 presents
results to show that when the messages are spread in a
round-robin manner, nodes along one path in the contour
will always handle more messages than nodes along other
paths. Section 5 presents optimal spreading rules that
ensure that all the nodes, which are at a same distance from
the source in a contour, handle roughly the same number of
messages. After presenting simulation results to illustrate
the QoS of CGD in Section 6, we conclude in Section 7.
1
2 EMBEDDING FUNCTIONS
We consider mesh topologies that are obtained by embed-
ding nodes on a 2D Basegrid and adjusting the transmission
range of each node. Each location on a 2D Basegrid is
identified by a unique ordered pair i. ,. The distance
between two consecutive locations on the grid, i.e., between
i. , and i. , 1 or between i. , and i 1. ,, is /.
Let N denote a set of nodes. An embedding function
assigns a location on the 2D Basegrid and a transmission
range to each node, i
i
2 N:
: i
i
2 N ! ININIR. 1
The location to which a node is assigned is assumed to be
fixed for the application. Each node knows its own location
and the number of neighbors it has. Further, we assume that
all the neighbors of a node reliably receive messages sent by
the node; however, in our simulations, we empirically
explore robustness to temporary link failures. Each message
contains the locations of the source and destination nodes as
specified by the embedding function.
Different mesh topologies are obtained by specifying
different embedding functions. Suppose that the nodes in N
must be embedded along 1 rows and C columns, starting at
location A. Y of the 2D Basegrid.
2
Let

denote the
embeddinginwhicheachnode i 2 N communicates directly
with neighbors. For 4, the embedding function

4
: i
i
2 N ! A i 1 mod C. Y
i 1
C

. /

2
yields a mesh topology. Since the transmission range for
each node is /, it can only communicate with its four
immediate neighbors. Similarly, for 8, the embedding
function

8
: i
i
2 N ! A i 1 mod C. Y
i 1
C

. /

2
p

. 3
where the transmission range is /

2
p
, yields the overlapping
mesh topology shown in Fig. 1a. Similarly, other embed-
dings such as
6
shown in Fig. 1b arise by changing the
location and transmission radius of the nodes [33].
In the remainder of this paper, nodes that are assigned to
location i. , on the 2D Basegrid are referred to as i
i.,
. In
the next section, we define contours and their structure
precisely. The results for contours in the
8
embedding are
presented in detail. Contours in other embeddings are
briefly discussed at the end of the section.
MAMIDISETTY ET AL.: MULTIPATH DISSEMINATION IN REGULAR MESH TOPOLOGIES 1189
1. A preliminary version of this work was reported in [32].
2. Our method depends only on the relative locations of the source and
destination nodes, and hence, the specific A. Y is not important.
3 CONTOURS AND THEIR STRUCTURE
Definition 1. A contour is the union of all the shortest paths
between a pair of nodes.
Fig. 2 shows two example contours between a pair of
nodes i
i.,
and i
.i
.
3
The shape of a contour depends on the
relative locations of the source and the destination. For
example, Fig. 2a shows a contour in
8
that is a rectangle
enclosed by the four corner nodes: i
.i
, c1, i
i.,
, and c2. Fig. 2b
shows another contour that is a hexagon enclosed by six
corner nodes: i
.i
, c1, c2, i
i.,
, c4, and c3.
Since the contour depends on the relative position of
the source and destination nodes, its structure can be
characterized by using only the differences in the r and
y locations of the pair of nodes. To this end, we define
the following quantities:

r
i
i.,
. i
.i
j ij. 4

y
i
i.,
. i
.i
ji ,j. 5

ry
i
i.,
. i
.i

r
i
i.,
. i
.i

y
i
i.,
. i
.i

. 6
ui
i.,
. i
.i


ry
2

. 7
For example, in Fig. 2a,
r
7,
y
3, and
ry
4, and
in Fig. 2b,
r
8,
y
3, and
ry
5. When the context is
clear, we use
r
,
y
,
ry
, and u without specifying the
source and the destination. The number of shortest paths in
a contour increases as
ry
increases.
In the
8
embedding, the length of the shortest path
between two nodes i
i.,
and i
.i
is
di
i.,
. i
.i
maxf
r
.
y
g. 8
When viewed as the union of shortest paths, a contour is
not symmetric withrespect tothe source andthe sinkbecause
the paths have direction. In addition, a contour can also be
viewedas aset of nodes. Anode i
:.t
is ina contour if andonlyif
it isonashortest pathfromthesourcetothedestination. Inthis
view, acontour issymmetricwithrespect tothesourceandthe
destination. Fromthewell-knownprincipleof optimality[12],
a node, i
:.t
, is in the contour of i
i.,
and i
.i
if and only if
di
i.,
. i
:.t
di
:.t
. i
.i
di
i.,
. i
.i
. 9
3.1 Structure of Contours in
8
We now present results to characterize the contour between
two nodes i
.i
and i
i.,
.
To simplify the notation, we only consider the case when i,
i ,, and, whenever
ry
0, i i ,. All other cases
can be handled in a similar manner.
1190 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 20, NO. 8, AUGUST 2009
3. We consistently use i
.i
to represent the source and i
i.,
to represent
the destination in a contour.
Fig. 1. Other embedding examples. (a)
8
. (b)
6
.
Fig. 2. Different contours in the
8
embedding. Depending on the relative
positions of the source and the destination, the contour can be a
rectangle as in (a) or a hexagon as in (b). (a) Contour in
8
enclosed by
a Rectangle. (b) Contour in
8
enclosed by a Hexagon.
Lemma 1. In
8
, the contour of i
i.,
and i
.i
is a single shortest
path if and only if
ry
0.
Proof. When
ry
0, we know that i i ,. The path
< i
i.,
. i
i1.,1
. . . . . i
1.i1
. i
.i
is a shortest path in the
embedding
8
between i
i.,
and i
.i
. All other nodes i
:.t
violate the criterion in (9) and are hence are not in the
contour.
Next, we need to show that
ry
0 when the
contour is a single path. Suppose
ry
6 0, and without
loss of generality, let
r

y
. By our notation, this
means that the length of the shortest path is i. Let
1 0 be such that i i,1. Using (8), by direct
calculation, we have di
i.,
. i
i1.,
di
i1.,
. i
.i

di
i.,
. i
1.i
di
1.i
. i
.i
di
i.,
. i
.i
. This implies
that nodes i
1.i
and i
i1.,
have to be on optimal
paths. However, using the distance criterion (9), it is
easily seen that a path from i
i.,
to i
.i
through both
these nodes is not a shortest path. This implies we
have to have at least two shortest paths. But since we
are given the fact that the contour is a single path, we
must have 1 0, which is same as
ry
0. tu
Lemma 2. In
8
, the contour of i
i.,
and i
.i
comprises two
shortest paths that are enclosed by the nodes i
i.,
, i
i1.,
, i
1.i
,
and i
.i
if and only if
ry
1.
Proof. Given the nodes as stated, we can use the distance
criterion in (9) to verify that node i
i1.,
and node i
1.i
are on different shortest paths from i
i.,
to i
.i
. It now
suffices to note that nodes i
i2.,
and i
2.i
do not satisfy
the distance criterion in (9) and, hence, these are the only
two shortest paths possible.
Given two shortest paths between i
i.,
and i
.i
, we now
show that the paths are enclosed by the four nodes as
stated. We know that a path from i
.i
to i
i.,
is a shortest
path only if maxf
r
.
y
g is reduced along each step of the
path. Since by our notation, i i ,,
r
must reduce
along each step of the path. In the
8
embedding, there are
only three neighbors of i
.i
with a lower
r
, namely,
i
1.i1
, i
1.i
, and i
1.i1
. i
1.i1
does not satisfy the
distance criterion in (9). Using similar arguments in the
neighborhood of i
i.,
, we conclude that only nodes i
i1.,1
andi
i1.,
can be on any shortest path between i
i.,
andi
.i
.
Finally, we note that only the nodes on the two shortest
paths satisfy the distance criterion in (9), and hence, the
four enclosing nodes are as stated. tu
Lemma 3. In
8
, given i
i.,
and i
.i
with
y
0, the contour is a
rectangle if and only if
r
is even. The four nodes at the
corners of this rectangle are i
i.,
, i
iu.,u
, i
u.,u
, and i
.,
.
(The proof is available in [32].)
Theorem 1. In
8
, the contour of i
i.,
and i
.i
is a rectangle
bounded by the nodes i
i.,
, i
iu.,u
, i
u.iu
, and i
.i
if and
only if
ry
is even and
ry
0.
Proof. Given that
ry
is even, we first show that the corner
nodes of the contour are as stated. Since
ry
is even,
either
r
and
y
are both even or they are both odd. Let

r
and
y
be both odd without loss of generality.
Consider the node i
i,.,
. By the criterion of (9),
this node would be in the contour, and no node i
:.,
with : i , would be in the contour. The only
shortest path from node i
i,.,
to i
.i
is through the
nodes i
i,1.,1
. i
i,2.,2
. . . . . i
1.i1
(by Lemma 1).
Because
y
is odd, the length of the path from i
.,
to
i
i,.,
must also be odd. Hence, the length of the path
from i
i.,
to i
i,.,
is even because
r
is odd. It follows
from Lemma 3 that the contour of i
i.,
and i
i,.,
is a
rectangle with u b
i,i
2
c and is enclosed by the
corner nodes: i
i.,
, i
iu.,u
, i
iu.,u
, and i
i,.,
. Let us
refer to this contour as 1. Since the four corner nodes
of 1 satisfy the distance condition in (9), we can
conclude that all the nodes in contour 1 are also in the
contour of i
i.,
and i
.i
.
Now, consider the node i
ii,.i
. By the criterion of (9),
this node would be in the contour, and no node i
:.i
with
: < i i , would be in the contour. We can see that
i
i.,
, i
i1.,1
. . . . . i
ii,.i
is the unique shortest path
between i
i.,
and i
ii,.i
. Since
y
is odd, the length of
this shortest path must be odd. Consequently, the length
of the shortest path from i
ii,.i
to i
.i
via the nodes
i
ii,1.i
. i
ii,2.i
. . . . . i
1.i
. i
.i
must be even. From
Lemma 3, the contour of i
ii,.i
and i
.i
is a rectangle
with u b
i,i
2
c and is enclosed by the nodes i
.i
,
i
u.iu
, i
u.iu
, and i
.i
. Let us refer to this contour as
T. Since the four corner nodes of T also satisfy the
distance condition in (9), we conclude that all the nodes
in contour T are also in the contour of i
i.,
and i
.i
.
Consider the path <i
i.,
. i
i1.,1
. i
i2.,2
. . . . . i
ii,u.
i u . Let us refer to this path as line-L. From the
definition of u (7), i i , u u, and hence, line-L
starts at node i
i.,
and ends at i
u.iu
. Similarly, let us
refer to the path < i
.i
. i
1.i1
. i
2.i2
. . . . . i
iu.,u
as
line-R. Every node i
:.t
that is not in contour T or
contour 1 which is in the contour of i
i.,
and i
.i
must lie
between line-L and line-R. Any node not in the contours
T and 1, which is not between line-L and line-R, does not
satisfy the distance criterion in (9). We can therefore
conclude that the contour of i
i.,
and i
.i
is a rectangle
with line-L and line-R as two edges, and hence, the four
corner nodes are at either ends of these lines.
Next, given a contour of i
i.,
and i
.i
with four corner
nodes as stated in the theorem, we showthat
ry
i
i.,
. i
.i

is even.
ry
i
i.,
. i
.i
is even only when
r
i
i.,
. i
.i
and

y
i
i.,
. i
.i
are both even or are both odd. By definition,
ui
i.,
. i
.i

jijji,j
2
, and ui
i.,
. i
i,.,

ji,ijj,,j
2
.
Thus, ui
i.,
. i
.i
ui
i.,
. i
i,.,
. i
iu.,u
is a corner
node in the contour of i
i.,
and i
i,.,
.
r
i
i.,
. i
i,.i

2 ui
i.,
. i
i,.,
by the structure of the 2DBaseGrid and
hence is even by Lemma 3. Since
ry
i
i.,
. i
i,.i
0, the
distance di
i,.,
. i
.,
is always equal to di
.i
. i
.,
.
Therefore,
r
i
i.,
. i
.i
and
y
i
i.,
. i
.i
are either both
even or both odd. Hence,
ry
i
i.,
. i
.i
must be even. tu
Lemma 4. In
8
, given i
i.,
and i
.i
with
y
0, the contour is
enclosed by a hexagon if and only if
r
is odd and
r
1. The
six nodes at the corners of this hexagon are i
i.,
. i
iu. ,u
.
i
iu1. ,u
. i
u.,u
. i
u1.,u
and i
.,
. (The proof is avail-
able in [32].)
Theorem 2. In
8
, the contour of i
i.,
and i
.i
is a hexagon
bounded by the nodes i
i.,
, i
iu.,u
, i
iu1.,u
, i
u.iu
,
i
u1.iu
, and i
.i
if and only if
ry
is odd and
ry
1.
Proof. The proof is similar to that of Theorem 1 by using
Lemma 4 instead of Lemma 3. tu
MAMIDISETTY ET AL.: MULTIPATH DISSEMINATION IN REGULAR MESH TOPOLOGIES 1191
In summary, the results in this section completely
characterize the structure of the contour (or the set of shortest
paths) between any two nodes in the
8
embedding.
3.2 Contours in Other Embeddings
The structure of contours inthe embeddings
3
,
4
, and
6
are
reported in [33]. In the
4
embedding, the contour of i
i.,
and
i
.i
is a single path when either
r
or
y
0. Further, when

r
0and
y
0, thecontour is always arectanglebounded
bythenodesi
i.,
, i
i.i
, i
.i
, andi
.,
. Fig. 3bshowsafewexample
contours inthe
4
embedding. Thestructure of contours inthe

6
embedding is similar to the structure of contours in
4
. In
the
3
embedding, contours have zero, one, or two pendant
(i.e., adjacent only with a single edge) nodes [33].
3.3 Impact of Contour Structure on Dissemination
In a multihop routing scheme, each node receives messages
from its upstream neighbor along some path and forwards
this message to its downstream neighbor. In this paper, our
interest is inmechanisms that exploit all the shortest paths. To
utilize multiple paths, eachnode must receive messages from
a set of upstreamneighbors and forward these messages to a
set of downstream neighbors.
To utilize the results in this paper inpractical applications,
it is assumed that each message packet contains the location
of both the source and the destination. As the messages
propagate through the network, every intermediate node
could easily compute (using, e.g., the distance criterion)
which of its immediate neighbors are on a shortest path to the
destination. (This is explained further in the next section). In
addition, the results of this section allow the intermediate
node to knowits relative position in the contour. In Section 5,
we show how this location information can be used to
effectively exploit all the available shortest paths ina contour.
4 EFFECTS OF UNIFORM SPREADING
The first strategy we consider for spreading messages over
all shortest paths will be called Uniform Spreading. This is a
straightforward strategy where the source, as well as each
intermediate node along every path in the contour, sends
successive messages in a round-robin fashion to all its
immediate neighbors in the contour. We present this
algorithm and show that the nodes along one of the paths
will always handle more messages than the nodes along
other paths whenever this strategy is used.
The following results characterize the nodes that are
inside a contour.
Observation 1. For any node i
:.t
that is in an
8
contour of
shortest paths from i
.i
to i
i.,
, its neighbors i
:.t1
, i
:.t1
, and
i
:1.t1
cannot be in the contour.
This follows from the distance criterion in (9). (Recall that
we are considering the case where i, i ,, and,
whenever
ry
0, i i ,).
Lemma 5. A node i
:1.t1
is in the
8
contour of shortest paths
from i
:.t
to i
i.,
only if
ry
i
:.t
. i
i.,
1.
Proof. For i
:1.t1
to be on a shortest path, we require
maxfj: 1 ij. jt 1 ,jg < maxfj: ij. jt ,jg. Clearly,
j: i 1j < j: ij and, hence, this is possible only
when jt 1 ,j j: 1 ij. Rearranging terms, we get
2 j: ij jt ,j, and hence, from the definition (7),
we conclude that
ry
i
:.t
. i
i.,
1. tu
When a message travels along a shortest path, the
distance to the destination must be reduced along each step
of the shortest path. In view of the above observation and
lemma, a node i
:.t
can send the messages it handles to at
most three of its neighbors, i.e., i
:1.t1
, i
:1.t
, and i
:1.t1
.
Thus, the structure of shortest paths in a contour and the
location of a node in the contour together constrain the
number of shortest paths from the node to the destination.
Specifically, a node i
o./
cannot appear on a shortest path
from i
:.t
to i
i.,
if
ry
i
o./
. i
i.,
!
ry
i
:.t
. i
i.,
.
Given below is the algorithm for uniform spreading. We
note that in the algorithm both i:qConit and iq/i: are
dependent on the source-destination pair. We need separate
counters for each source-destination pair. If the first packet
in a batch contains the total number of packets in that batch,
then we know how long to maintain the counters. The set of
neighbors, iq/i:, of a node in the contour can be computed
by the node as explained above.
Uniform Spreading Algorithm
In the rest of this section, we analyze the properties of
this algorithm.
4.1 Loading under Uniform Spreading in
8
Uniform spreading is achieved when all the nodes in the
contour execute the Uniform Spreading algorithm. To
1192 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 20, NO. 8, AUGUST 2009
Fig. 3. Examples of different contour structures for various values of
ry
.
(a)
8
. (b)
4
.
characterize the effects of this algorithm, we first define what
we call rows in a contour.
Definition 2. A row in a contour is a set of nodes that are at the
same distance (8) from the source.
If the length of the shortest path is |, then there will be
| 1 rows in a contour. Note that the source and destination
nodes aretrivial rowswithasinglenodeineach. Fig. 4shows a
contour witheight rows that has a shortest pathof 7 hops. The
numbers inthecircles showthenumber of packets handledby
different nodes when all nodes use the uniform spreading
algorithmandthesourcesends 99packets. Eventhoughevery
node spreads the messages it handles over all the available
neighbors, many nodes closer to the destination handle more
messages than other nodes. This is the loading phenomenon
that we will characterize precisely in this section.
In the embedding
8
, notice in Fig. 4 that under uniform
spreading, all the nodes i
:.t
in rows 1. 2. . . . . u (where u is as
defined by (7)) of the contour divide the messages they
handle into three parts and send each part along one of their
three neighbors that are closer to the destination. Thus, we
can conclude that the ratio of the messages handled by a
node in a row to the smallest number of messages handled
by another node in the same row follows the trinomial
coefficients.
4
The trinomial coefficient Tc. is the coeffi-
cient of r

in the expansion of 1 r r
2

c
. We use the
following properties of trinomial coefficients in this section:
1. Tc. c, the central trinomial coefficient, is the
largest coefficient in the expansion of 1 r r
2

c
and Tc. c 0.
2. Suppose c
P
c1
i0
Tc. i
P
2c
,c1
Tc. ,; then,
c
3
c
<
1
2
. This follows from the symmetry of the
trinomial coefficients and Tc. c 0.
Lemma 6. Givena
8
contour with
ry
i
.i
. i
i.,
1, if i
.i
sends
` messages to i
i.,
, nodes alongone pathhandle ` 1messages
after log
2
` steps along the shortest path from i
.i
to i
i.,
.
Proof. Since
ry
1, there are two paths in the contour from
Lemma 2. All the nodes along one path have
ry
1 with
respect toi
i.,
; the nodes alongthe other pathhave
ry
0.
Let us refer to these paths as Path-1 and Path-0,
respectively. Since all the nodes on Path-1 have a higher

ry
than the nodes on Path-0, no messages can go from a
node on Path-0 to a node on Path-1. However, every node
i
:.t
on Path-1 divides the messages it handles equally
between its neighbor i
:1.t
on Path-0 andneighbor i
:1.t1
on Path-1. Hence, if the source i
.i
generates ` messages,
the node onPath-1 that is at a distance log
2
` fromi
.i
will
handle at most one message. This means that the nodes
starting at node i
log
2
`1.ilog
2
`
on Path-0 handle at least
` 1 messages. tu
To show that uniform spreading results in loading one
path of the contour in general, let us consider subpaths in a
contour labeled as shown in Fig. 5. All the nodes on Path-/
have
ry
/ with respect to the destination, i
i.,
. Using
arguments similar to ones used in the proof of Lemma 6, we
note that a node in a lower numbered subpath cannot
forward messages to a node in a higher numbered path
because it will not be on a shortest path.
Lemma 7. Under uniform spreading in a contour in
8
, in every
row, nodes along the subpath with
ry
0 with respect to the
destination will always handle more messages than the nodes
along other paths.
Proof. The proof is similar to that of the previous lemma. tu
Lemma 8. Under uniform spreading in a contour in
8
, nodes
along Path-0 and Path-1 always handle more than
`
2
messages.
Proof. We show that the nodes i
2u1.i
and i
2u1.i1
handle more than
`
2
. Consider the nodes in a row at a
distance 2u 1 from the source in the contour. Notice
that the number of messages handled by node i
2u1.i
corresponds to the central trinomial coefficient
T2u 1. 2u 1. The sum of all the messages in
nodes handled by the nodes in row 2u 1 is
22u 1 T2u 1. 2u 1. The lemma follows
from the property that 2u 1 < 1,2. tu
Theorem 3. Under uniform spreading in a contour in
8
, when

r
6 0 and
y
6 0, nodes on the subpath with
ry
0, with
respect to the destination, will always handle more messages
than the nodes along other paths.
MAMIDISETTY ET AL.: MULTIPATH DISSEMINATION IN REGULAR MESH TOPOLOGIES 1193
Fig. 4. Uniform spreading in
8
.
Fig. 5. Labeling paths in a contour.
4. This is true to the extent that the number of messages can be equally
divided among the nodes in a row.
Proof. First, consider nodes along Path-2u. Since each node
sends the messages it handles to three downstreamnodes,
after i log
3
` steps on Path-2u, the node i
i.ii
will
handle at most one message. This node must be inrowiof
the contour. From the properties of trinomial coefficients,
the node i
i.ii1
, on Path-2u 1 path, will handle
imessages. After i
0
log
3
log
3
` more steps alongPath-
2u 1, node i
ii
0
.iii
0
1
will handle at most
i
0
messages. Extending this argument, we conclude that
in Path-/, 2. . . . . 2u, some node i
Olog
3
`. iOlog
3
`2u/
must handle at most one message. Since the messages
always move from nodes in higher numbered layers to
nodes in lower numbered layers, such messages must
eventually arrive at the nodes on the Layer-0 path. Hence,
we have the theorem. tu
4.2 Effects of Embedding on Loading
We observed similar loading effects under uniform spread-
ing in
4
and
6
. Recall that the structure of the contours in

6
closely follows that in
4
. In these embeddings, only the
nodes along paths on either the A or Y axis get loaded. If

r

y
, then the path along the A-axis in the contour will
be loaded. Proofs of loading, similar to the ones presented
for
8
in the preceding section, for contours in
4
and
6
follow from the properties of binomial coefficients.
5 OPTIMAL SPREADING
We now present an algorithm for spreading the messages so
that all the available paths are effectively utilized. Recall
that a row is a collection of nodes in the contour that are at
the same distance from the source. Let n be the number of
nodes in a row of a contour. We refer to n as the width of
the row. If the source sends ` messages and if every node
in every row handles
`
n
messages,
5
then we can say that the
spreading is the best in the sense that all available paths are
effectively used. This is the criterion of optimality that we
choose. We will show that the algorithm presented in this
section is optimal in this sense.
Fig. 6 shows examples of two contoursone in
8
and the
other in
4
. Notice that the width of a row, n, varies with the
row number in three ways in both the contours. n first
increases monotonically, then remains constant, and then
decreases monotonically. We refer to these three regions as
the Expansion region, Propagation region, and Contraction
region, respectively.
Observation 2. Inthe
8
embedding, the width nof any rowinthe
expansion region is odd, and the width of the next downstream
row is n 2. When
ry
0 and even, the width of the rows in
the propagation region is the same as the largest n in the
expansionregion; when
ry
0andodd, the widthof the rows in
the propagationregionis n 1. The widthof eachsuccessive row
in the contraction region reduces by two until the i
i.,
is reached.
Consider two adjacent rows i and i1 in the
expansion region with widths n and n 2, respectively.
Since there are more nodes in row i1, if all of them have
to handle the same number of messages, then at least some
of them have to receive messages from two nodes in row i.
The strategy we propose uses a special labeling of the nodes
in a row; every node can compute its own label using the
contour structure results discussed in Section 3.
5.1 Node Labeling
Inanyrow, nodes onbothsides of themiddle nodeare labeled
as1. 2. 3. . . . . bn,2c, asillustratedinFig. 7. Let i
i
j
represent the
node that is labeledj inthe ithrowof acontour. Note that i
i
j
is not the ID (or coordinates) of the node; in each row, two
nodeswithdifferent IDswouldhavethesamelabel. However,
for any node except the middle one in rowi, only one of the
two nodes with label j would be a neighbor. Thus, for all
nodesexcept themiddleoneinrowi, thenodelabel asper the
convention shown in Fig. 7 uniquely identifies a neighbor in
rowi1. Under this labeling in the
8
embedding, node i
i
j
can communicate with nodes i
i1
j
and i
i1
j1
. For the middle
node in row i, there would be two neighbors in row i1
with the same label.
1194 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 20, NO. 8, AUGUST 2009
Fig. 6. Expansion, propagation, and contraction regions.
(a) Contour in
8
. (b) Contour in
4
.
Fig. 7. Spreading in the expansion region.
5. This is only possible to the extent that ` is an integral multiple of n.
5.2 Optimal Spreading
Our algorithm for optimal spreading is given below. In
the algorithm, the array iq/i: keeps track of the relevant
neighbors for any node. As explained above, the middle
node in a row has two relevant neighbors in the next row
that have the same label.
Optimal Spreading Algorithm
Our algorithm for spreading messages is based on the
following strategy:
. In the expansion region, nodes i
i
j
send messages to
nodes i
i1
j
and nodes i
i1
j1
in different ratios.
. In the propagation region, node i
i
j
sends all the
messages it handles to node i
i1
j
.
. In the contraction region, nodes i
i
j
send messages
to node i
i1
j
and nodes i
i1
j1
in different ratios, as
shown in the algorithm.
Intuitively, the nodes in the expansion region spread
messages over the available paths, the nodes in the propaga-
tion region do not spread messages and propagate them
along a single path, and the nodes in the contraction region
coalesce messages from multiple paths. Our objective is to
ensure that all the nodes in every row of a contour handle
roughly the same number of messages. Theorem 4 below
shows that this objective is achieved when each node in the
expansionregionandthe contractionregionspreadmessages
to its neighbors in the ratios specified in the Optimal
Spreadingalgorithm. The empirical results inthe next section
showthat this notion of optimality significantly improves the
QoS.
Theorem 4. When the source in a contour sends ` messages and
there is no message loss, under the Optimal Spreadingalgorithm,
every node in each row of width n handles
`
n
messages.
6
Proof. We prove the result for the expansion region of the
contour using mathematical induction. As a basis, i
1
1
is in
a row with width n 1, and it handles
`
n
` messages.
This node sends messages to three nodes in the next row
in a specific ratio, as given in the algorithm. The middle
node in row 2 receives
`
n2

`
3
messages from i
1
1
. The
two nodes labeled i
2
1
(recall that the middle node in row
i has two neighbors with the same label in row i1)
receive
`
nn2

`
3
messages each. Thus, all the nodes in
row 2 that has width n 3 handle
`
3
messages.
Assume that the nodes in row i of the contour with
width n handle
`
n
messages. We show that all the nodes
in row i1 handle
`
n2
messages. As per our algorithm,
node i
i
j
sends messages to i
i1
j
and i
i1
j1
in the ratio
n 2 2j : 2j. Since this node has
`
n
messages, this ratio
implies that i
i
j
sends
`j
n2

`j1
n
messages to i
i1
j
and
`j
n

`j
n2
messages to i
i1
j1
. Following this rule, the
nodes i
i1
1
receive
1`
n2

11`
n

`
n2
messages from
the nodes i
i
1
.
The nodes i
i1
,
, 2 , b
n2
2
c 1, receive
.
,1`
n

,1`
n2
messages from nodes i
i
,1
, and
.
,`
n2

,1`
n
messages from nodes i
i
,
.
Hence, each node i
i1
,
handles
`
n2
messages.
Now, consider nodes i
i1
b
n2
2
c
. These nodes receive
1
2

`
n

`
n2
messages from the middle node in row i and
bn,2c`
n

bn,2c`
n2
messages from nodes i
i
bn,2c
. Thus,
nodes i
i1
bn2,2c
also handle
`
n2
messages. Therefore,
since the middle node in row i1 receives
`
n2
messages from the middle node in row i, we conclude
that the spreading rules presented are optimal for the
expansion region.
It is easy to see that if the number of messages
handled by the nodes in a row at the beginning of the
propagation region are roughly the same, this number
will not change since the nodes do not spread messages.
The proof for the contraction region also follows from
similar induction, and the details are omitted. tu
In the above proof (and in specifying our algorithm), it is
assumed that the width of rows in the propagation region is
the same as the largest width in the expansion region.
When the contour is a hexagon (i.e., when
ry
0 and
odd), there are two special cases that must be considered
because if n is the maximum width of the expansion region,
the width of the propagation region is n 1, and the
MAMIDISETTY ET AL.: MULTIPATH DISSEMINATION IN REGULAR MESH TOPOLOGIES 1195
6. This is only possible to the extent that ` is an integral multiple of n.
maximum width of the contraction region is n. These rules
are also similar to the rules specified in our algorithm [33].
To implement optimal spreading in practice, in more
general situations, one can effect the spreading of messages
in the desired ratio by using a randomized strategy that
chooses the neighbors with the proportional probabilities.
Then, the above result of equitable distribution of load
holds in an expected sense.
6 RESULTS
To evaluate the QoS achieved by CGD, we designed and
implemented a discrete-event simulation using the OM-
Net++ framework [34]. The results in this section demon-
strate the performance of CGD. For all simulations, the nodes
were embedded on the 2D Basegrid using the
8
embedding
with
ry
4, one source, and one sink in the contour. In the
scenarios where the number of messages was not varied, the
source sent 5,000 messages, each of which is 36 bytes long. To
study the effects of an increased number of messages, the
number was changed between 5 and 50,000.
6.1 Simulation Approach
The multipath, multihop, dissemination of messages from
the source to the sink was modeled as a multistage queuing
network. The service time of each queue was exponentially
distributed with a mean time of 30 ms to represent 1) the
time required to propagate a message from the preceding
node, 2) the time required to receive a message in the node,
and 3) the time required to forward the message toward the
sink. Using the channel modeling capability in OMNet++,
the data rate (bandwidth) was set to 38.4 Kbps. Propagation
delay and bit error rates were assumed to be zero, and all
links between pairs of nodes were identical. The source
generated messages with an interarrival time that was
exponentially distributed with a mean time of 20 ms.
7
To ensure that the simulations reflected real-world
constraints on low-power wireless propagation and topol-
ogy, we weakened the regular topologies, which were used
in the analysis in the preceding sections, by modeling link
failures and wireless propagation irregularities. To model
link failures, the probability of success for each link was
varied between zero (failure) and one (success). Before
sending a message to a dowstream node, a random number
was drawn from a uniform distribution; the message was
sent only if this random number was larger than the link
failure threshold. To account for wireless propagation
irregularities [35], a probability of success was assigned to
each link in the neighborhood of a node that was drawn
from a Weibull distribution using the parameters reported
in [35]. We varied the percentage of such irregular nodes
from 0 (no nodes with irregular propagation) to 100 (all
nodes have irregular propagation). To investigate the
relationship between wireless propagation irregularity and
the structure of the contour, we also considered the case
where irregular nodes were localized to the expansion and
propagation regions of the contour.
6.2 QoS Metrics
The number of messages handledby each node, average end-
to-end latency, jitter, and message loss rate were the primary
QoS metrics. For each message, the time at which it was sent
fromthe source, t, andthe time at which this message arrived
at the sink, t
0
, were recorded. The end-to-end latency for this
message is t
0
t. The average end-to-end latency was the
average value of t
0
t over all the messages received at the
sink. Jitter was the standard deviation of the end-to-end
latencies. The ratio of the number of messages that were not
received at the sink to the number of messages that were sent
from the source was the message loss rate. To evaluate the
effects of congestion on the performance, we considered the
number of messages lost when 1) the service time in each
node increases and 2) the number of messages sent by the
source increases. In addition, throughputdefined as the
number of messages arriving at the sink per unit timewas
used as a metric to evaluate the scalability of CGD.
Since none of the multipath routing methods described in
Section 1 relies on effectively using all the available paths to
improve QoS, the metrics discussed above were used to
compare the performance of CGD with the performance of
1) flooding and 2) single-shortest-path routing (Unipath),
which is typically used in distance vector routing. Our focus
on regular topologies allowed us to hand-compute the
shortest paths. We used flooding with duplicate elimination
and only considered the communication costs incurred by
nodes in the contour. We expected the average latency and
jitter to be much better than flooding. Since CGD exploits
multiple paths, Unipathservedas a watermarkfor acceptable
performance. The performance of CGDwas investigatedboth
under uniformspreading (uniformCGD) and under optimal
spreading (optimal CGD).
6.3 Performance of CGD
Fig. 8 shows a histogram of the number of messages
handled in each node in uniform CGD and optimal CGD.
There were 56 nodes in the contour with
ry
4. As shown
1196 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 20, NO. 8, AUGUST 2009
7. We developed a Java program that generated the source files required
by the OMNet++ framework. The rules for dissemination, using uniform
and optimal spreading, were encoded in these files.
Fig. 8. In optimal CGD, all the nodes in a row of the contour handle
roughly the same number of messages. This balance is not maintained
in uniform CGD. (a) Uniform CGD. (b) Optimal CGD.
in Fig. 8b, all the nodes in a row handle roughly the same
number of messages in optimal CGD.
6.3.1 Effects of Link Failures
To weaken the regular topologies considered in the
analysis, we examined the effects of link failures. Fig. 9
shows the number of messages handled by the nodes in a
row when the probability of link failure is 0.2. Notice that
optimal CGD (Fig. 9b) is more balanced than uniform CGD
(Fig. 9a). A similar balance across the nodes was observed
when the link failure probability was varied over a range.
This balance improves the QoS achieved by CGD.
Fig. 10 compares the QoS achieved by optimal CGD and
uniform CGD with that achieved using Unipath and
flooding. The probability of link failure was varied from
zero to one, and each link in the multihop network was
assumed to fail independently with equal probability.
Notice in Figs. 10a and 10b that optimal CGDachieves the
lowest average latency and jitter. We believe that the almost
constant jitter achieved by optimal CGD is due to the better
load balancing across nodes. The performance of uniform
CGD follows the performance achieved using Unipath. As
expected, Fig. 10c shows a lowmessage loss rate for flooding.
6.3.2 Effects of Propagation Irregularity
Fig. 11 shows the QoS of CGD in the presence of nodes with
wireless propagation irregularity. As shown in Fig. 11a,
optimal CGD achieves the lowest average end-to-end
latency. Similarly, the Jitter (Fig. 11b) is also low. However,
as expected, the loss rate is lower for flooding.
By comparing Fig. 11 with Fig. 10, it can be noticed that
irregular wireless propagation effects are much less severe
than the effects of link failure. This is because unlike link
failure, when the wireless propagation is irregular, we can
still expect some communication to proceed.
We alsoexaminedhowthe locationof nodes withirregular
wireless propagation affects the QoS of CGD. Fig. 12 shows
the QoS achieved under optimal spreading when the
irregular nodes are localized either to the expansion region
or the propagation region of the contour. The average latency
(Fig. 12a) is worst when the nodes with irregular wireless
propagation occur in the expansion region of a contour. The
jitter followed a similar trend. Note that the message loss rate
(Fig. 12b) is low when the nodes with irregular wireless
propagation occur in the expansion region.
MAMIDISETTY ET AL.: MULTIPATH DISSEMINATION IN REGULAR MESH TOPOLOGIES 1197
Fig. 9. Optimal CGD maintains a balance among the nodes in a row of
a contour even in the presence of link failures. This result is obtained
when the probability of link failure was 0.2 for every link in the contour.
(a) Uniform CGD. (b) Optimal CGD.
Fig. 10. The average latency and jitter of optimal CGD is significantly
lower than that of uniform CGD, flooding, or Unipath. The nonzero
message loss rate for Unipath, Flooding, and Uniform CGD, when the
link failure probability is zero, is because of buffer overflow. (a) Average
latency. (b) Jitter. (c) Message loss rate.
6.3.3 Scalability
Weexaminedtheperformanceof optimal CGDwhenthescale
increases, i.e., when the number of nodes and the number of
messages increase. The results in the preceding section show
that uniform CGD performs better than Unipath and Flood-
ing, andoptimal CGDperforms better thanuniformCGD. For
this reason, the following results only compare the perfor-
mance of optimal CGDwith that of uniformCGD.
To investigate how QoS changes with scale, we increased
the number of nodes and the number of messages and
observed the average latency and throughput. Fig. 13 shows
the average latency and throughput achieved when the
number of nodes in the contour is increased. We maintained

ry
4 as the number of nodes increased from around
50 to 500. While the average latency increased linearly,
because the length of the shortest path increased, the
throughput also decreases. However, the throughput of
optimal CGD is better than that of uniform CGD.
Figs. 14 and 15 show a comparison of the average delay
and throughput achieved when the queues in the intermedi-
ate nodes build up. In Fig. 14, the service time is increased
from a mean value of about 3 ms to about 80 ms. The rate at
which messages were generated at the source was not
changed. The average latency and throughput of optimal
CGD is better than that of uniform CGD. In Fig. 15, the
number of messages sent by the source increasedfrom5 to50,
from 50 to 500, from 500 to 5,000, and from 5,000 to 50,000.
Optimal CGD performs much better than uniform CGD.
6.4 Summary and Discussion
The preceding results demonstrate that despite the avail-
ability of multiple shortest paths between a pair of nodes,
these paths cannot be exploited using a straightforward
strategy such as uniform spreading. While the results in
Figs. 10 and 11 showthat uniformspreading performs better
than methods based on Unipath, optimal CGD effectively
exploits all the available shortest paths and improves QoS.
1198 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 20, NO. 8, AUGUST 2009
Fig. 11. The average latency and jitter achieved using optimal CGD is
significantly lower even when a large number of nodes in the contour
have irregular propagation. (a) Average latency. (b) Jitter.
Fig. 12. When the nodes with irregular wireless propagation can be
localized to the expansion region of a contour, the message loss rate
improves. (a) Average latency. (b) Message loss rate.
Fig. 13. The average latency and throughput achieved using optimal
CGD are significantly better than those of uniform CGD. (a) Average
latency. (b) Throughput.
Fig. 8 shows that the nodes in the same row of a contour
handle roughly the same number of messages when using
optimal CGD. Optimal CGD maintains this balance across
nodes even in the presence of link failures (Fig. 9). The
performance of CGD remains robust in the presence of
nodes with irregular wireless propagation. Interestingly, as
shown in Fig. 12, when the nodes with irregular wireless
propagation are localized to the expansion region of the
contours, the effect of irregularity on latency and through-
put is less pronounced. This implies that it is important for
designers and system installers to focus on minimizing
nodes with irregular wireless propagation in certain parts of
the system. The performance improvement achieved by
optimal CGD is robust with respect to the increase in
service time, the number of messages sent from the source,
and the number of nodes in the contour, as shown by the
results in the preceding section.
The practical use of CGD will be impacted by physical-
layer effects and application dynamics. In the future, we
expect to investigate the performance of CGD in application
scenarios, through simulations and experiments, that
account for Media Access Control (MAC) and interference
across multiple paths.
7 CONCLUSIONS
Many future engineered systems that are based on peer-to-
peer-connected mesh topologies are likely to have multiple
paths between a pair of nodes. We defined a contour as the
union of all shortest paths between a pair of nodes. Using a
regular topology, we proved that when the messages are
spread uniformly over the paths in a contour, nodes along
one path handle more messages than other messages. We
presented an optimal strategy for spreading messages in
such systems, and our results demonstrate the effectiveness
of the spreading strategy.
Although the results are based on regular topologies,
they represent upper bounds on what can be achieved in
general topologies. These results reveal that to achieve
optimal dissemination, some nodes must disseminate the
messages over the available paths, and other nodes use only
one of the available paths. Identifying these sets of nodes in
general topologies is an interesting problem. In the future,
the optimal dissemination techniques can be enhanced to
improve QoS, mitigate interference, reduce hotspot effects,
and design next-generation monitoring and surveillance
systems based on wireless mesh topologies.
APPENDIX
Proof (Lemma 3). Since
y
0, we have , i. We first
show that when
r
is even, the contour is enclosed by
the four corner nodes. Because
y
0, <i
i.,
. i
i1.,
. . . . .
i
1.,
. i
.,i
, shown in Fig. 16, is a shortest path. Since

r
is even, there are an odd number of nodes in this
path. Let us call this path a Layer-0 path. (This
nomenclature will become clear shortly).
Any subpath < i
:.,
. i
:1.,
. i
:2.,
in the Layer-0
pathcanbe replacedbya path< i
:.,
. i
:1.,1
. i
:2.,

o r < i
:.,
. i
:1.,1
. i
:2.,1
. . . . . i
:/. ,1
. i
:/1. ,
,
where i 1 / 1 without increasing the length of
the shortest path. We call such a path as a Layer-1 path.
Shortest paths that go up to , 2 are called Layer-2 paths,
and so on, until a Layer-u path. Similarly, we can define
Layer-1, Layer -2, . . . , Layer-u paths. Fig. 16 shows such
shortest paths.
MAMIDISETTY ET AL.: MULTIPATH DISSEMINATION IN REGULAR MESH TOPOLOGIES 1199
Fig. 14. When the service times increase, the queues in intermediate
nodes buildup. Notice that when uniform CGD is used, the effects on the
average latency and throughput are more pronounced because of
loading. (a) Average latency. (b) Throughput.
Fig. 15. As the number of messages sent from a source increase,
optimal CGD delivers higher throughput and lower latency than uniform
CGD. (a) Average latency. (b) Throughput.
Fig. 16. Layered paths when
r
is even.
The node i
i.,1
cannot be on any shortest path because
any path between i
i.,
and i
.,
that passes through i
i.,1
will have length !
r
1. Similarly, node i
.,1
cannot
be on any shortest path between i
i.,
and i
.,
.
In each layer /, 1 / u, the number of nodes that
can be on any shortest path between i
i.,
and i
.,
reduces by 2 /. For example, the Layer-1 path will
have
r
1 2 nodes, the Layer-2 path will have

r
1 4 nodes, and so on. Hence, the Layer-u path
must have a single node, i.e., i
iu.,u
since the Layer-0
path has an odd number of nodes because
r
is even.
Similar arguments apply to the Layer -1, Layer -2, . . . ,
Layer-u paths, and the Layer-u path must have a single
node, i.e., i
u.,u
. Hence, the four corners of the
rectangle are as required.
Next, given a contour with four corner nodes i
i.,
,
i
iu.,u
, i
u.,u
, and i
.,
, we show that
r
is even.
Consider the path from i
i.,
to i
iu.,u
. The length of this
path is the same as the length of the path from i
i.,
to
i
iu.,
because of the structure of the
8
embedding.
r
is
2 |ciqt/i
i.,
. i
iu.,
and is hence even. tu
Proof (Lemma 4). The arguments are similar to the ones
used in the proof of Lemma 3. The only difference is that
the Layer-u path will have two nodes instead of one
because there are an even number of nodes in the Layer-0
path, and the number of nodes on each successive layer
reduces by two. tu
ACKNOWLEDGMENTS
The authors thank the anonymous reviewers whose com-
ments on an earlier draft of this paper have helped to
improve the presentation. This material is based upon work
supported by the US National Science Foundation (NSF)
under Grant 0720736. Any opinions, findings, and conclu-
sions or recommendations expressed in this material are
those of the authors and do not necessarily reflect the views
of the NSF.
REFERENCES
[1] R.N. Murthy, A. Gosain, M. Tierney, A. Brody, A. Fahad, J. Bers,
and M. Welsh, CitySense: A Vision for an Urbanscale Wireless
Networking Testbed, Proc. ACM Sixth Workshop Hot Topics in
Networks (HotNets 07), Nov. 2007.
[2] N. Hayslip, S. Sastry, and J. Gerhardt, Networked Embedded
Automation, Assembly Automation, vol. 26, no. 3, pp. 235-241, 2006.
[3] J.Z. Hecker, PORT SECURITYNation Faces Formidable Chal-
lenges in Making New Initiatives Successful, technical report, US
General Accounting Office, 2002.
[4] J. Hill, M. Horton, R. Kling, and L. Krishnamurthy, The Platforms
Enabling Wireless Sensor Networks, Comm. ACM, vol. 47, no. 6,
pp. 41-46, June 2004.
[5] I.F. Akyldiz, X. Wang, and W. Wang, Wireless Mesh Networks:
A Survey, Computer Networks, vol. 47, pp. 445-487, 2005.
[6] G. Coulouris, J. Dollimore, and T. Kindberg, Distributed Systems
Concepts and Design, fourth ed. Addison-Wesley, 2005.
[7] J. Gao and L. Zhang, Load-Balanced Short-Path Routing in
Wireless Networks, IEEE Trans. Parallel and Distributed Systems,
vol. 17, no. 4, pp. 377-388, Apr. 2006.
[8] C. Intanagonwiwat, R. Govindan, D. Estrin, J. Heidemann, and
F. Silva, Directed Diffusion for Wireless Sensor Networking,
IEEE/ACM Trans. Networking, vol. 11, no. 1, pp. 2-16, 2003.
[9] J. Kulik, W. Heinzelman, and H. Balakrishnan, Negotiation-
Based Protocols for Disseminating Information in Wireless Sensor
Networks, Wireless Networks, vol. 8, pp. 169-185, 2002.
[10] J. Agre, L. Clare, and S. Sastry, A Taxonomy for Distributed
Real-Time Control Systems, Advances in Computers, vol. 49,
pp. 303-352, 1999.
[11] K. Russell, Maximizing Performance for Wireless Mesh Net-
works, EE Times, Feb. 2007.
[12] R.E. Bellman, Dynamic Programming. Princeton Univ. Press, 1957.
[13] K. Akkaya and M. Younis, A Survey on Routing Protocols for
Wireless Sensor Networks, Ad Hoc Networks, vol. 3, pp. 325-349,
2005.
[14] D.B. Johnson and D.A. Maltz, Dynamic Source Routing in Ad Hoc
Wireless Networks, Mobile Computing, Imielinski and Korth, eds.,
vol. 353, pp. 153-181, Kluwer Academic Publishers, 1996.
[15] S. Murthy and J.J. Garcia-Luna-Aceves, An Efficient Routing
Protocol for Wireless Networks, Mobile Networks and Applications,
vol. 1, no. 2, pp. 183-197, 1996.
[16] C.E. Perkins and P. Bhagwat, Highly Dynamic Destination-
Sequenced Distance-Vector Routing (DSDV) for Mobile Compu-
ters, Proc. ACM SIGCOMM 94, pp. 234-244, 1994.
[17] D. Chen and P.K. Varshney, QoS Support in Wireless Sensor
Networks: A Survey, Proc. Intl Conf. Wireless Networks
(IWCN 04), June 2004.
[18] T. He, J.A. Stankovic, C. Lu, and T. Abdelzaher, SPEED: A
Stateless Protocol for Real-Time Communication in Sensor Net-
works, Proc. 23rd Intl Conf. Distributed Computing Systems
(ICDCS), 2003.
[19] C. Lu, B.M. Blum, T.F. Abdelzaher, J.A. Stankovic, and T. He,
RAP: A Real-Time Communication Architecture for Large-Scale
Wireless Sensor Networks, Proc. Eighth IEEE Real-Time and
Embedded Technology and Applications Symp. (RTAS), 2002.
[20] S. Muller, R.P. Tsang, and D. Ghosal, Multipath Routing in
Mobile Ad Hoc Networks: Issues and Challenges, Lecture Notes in
Computer Science, Springer, 2004.
[21] E. Hyytia and J. Virtamo, On Traffic Load Distribution and Load
Balancing in Dense Wireless Multihop Networks, EURASIP
J. Wireless Comm. and Networking, vol. 2007, pp. 1-15, 2007.
[22] R. Draves, J. Padhye, and B. Zill, Routing in Multi-Radio,
Multi-Hop Wireless Mesh Networks, Proc. ACM MobiCom,
Sept./Oct. 2004.
[23] P.P. Pham and S. Perreau, Performance Analysis of Reactive
Shortest Path and Multi-Path Routing Mechanism with Load
Balance, Proc. IEEE INFOCOM, 2003.
[24] A. Valera, W. Seah, and S.V. Rao, Cooperative Packet Caching
and Shortest Multipath Routing in Mobile Ad Hoc Networks,
Proc. IEEE INFOCOM, 2003.
[25] L. Wang, Y. Shu, M. Dong, L. Zhang, and W.W.O. Yang,
Adaptive Multipath Source Routing in Ad Hoc Networks, Proc.
IEEE Intl Conf. Comm. (ICC 01), vol. 3, 2001.
[26] Y.S. Chen, C.Y. Tseng, J.P. Shue, and P.H. Kuo, An On-Demand,
Link-State, Multi-Path QoS Routing in a Wireless Mobile Ad-Hoc
Network, Proc. European Wireless Conf. (EW), 2002.
[27] A. Fallahi, E. Hossain, and A.S. Alfa, QoS and Energy Trade
Off in Distributed Energy-Limited Mesh/Relay Networks: A
Queuing Analysis, IEEE Trans. Parallel and Distributed Systems,
vol. 17, no. 6, pp. 576-592, June 2006.
[28] R. Leung, J. Liu, E. Poon, A. Chan, and B. Li, MP-DSR: A QoS-
Aware Multi-Path Dynamic Source Routing Protocol for Wireless
Ad-Hoc Networks, Proc. 26th Ann. IEEE Conf. Local Computer
Networks (LCN 01), pp. 132-141, 2001.
[29] W.H. Liao, C.Y. Tseng, S.L. Wang, and J.P. Sheu, A Multi-Path
QoS Routing Protocol in a Wireless Mobile Ad-Hoc Network,
Proc. First Intl Conf. Networking (ICN), 2001.
[30] S.J. Lee and M. Gerla, Split Multipath Routing with Maximally
Disjoint Paths in Ad Hoc Networks, Proc. IEEE Intl Conf. Comm.
(ICC 01), vol. 10, 2001.
[31] D. Ganesan, R. Govindan, S. Shenker, and D. Estrin, Highly-
Resilient, Energy-Efficient Multipath Routing in Wireless
Sensor Networks, Mobile Computing and Comm. Rev., vol. 1,
no. 2, 2002.
[32] I.-H. Chu, M. Duan, and S. Sastry, Contour Guided Dissemina-
tion for Networked Embedded Systems, Proc. Intl Symp.
Innovative Real-Time Applications of Distributed Sensor Network
(IRADSN 06), Oct. 2006.
[33] K.K. Mamidisetty, Generalizing Contour Guided Dissemina-
tion in Mesh Topologies, MS thesis, Univ. of Akron, May
2008.
[34] A. Vargas, Omnet++Discrete Event Simulation System, Proc.
European Simulation Multiconf., pp. 319-324, June 2001.
1200 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 20, NO. 8, AUGUST 2009
[35] G. Zhao, T. He, S. Krishnamurthy, and J.A. Stankovic, Impact
of Radio Irregularity on Wireless Sensor Networks, Proc. Second
Intl Conf. Mobile Systems, Applications, and Services (MobiSys 04),
pp. 125-138, 2004.
Kranthi K. Mamidisetty received the bachelors
degree in electronics and communication en-
gineering from Jawaharlal Nehru Technological
University and the masters degree in electrical
engineering from the University of Akron, Ohio,
in 2008. He is currently pursuing his PhD degree
at the Department of Electrical and Computer
Engineering, University of Akron.
Minlan Duan received the bachelors degree in
electrical engineering from the Beijing Institute of
Technology, China, in 2005 and the MSdegree in
electrical engineering from the University of
Akron, Akron, Ohio, in 2007. She is currently with
Denso International America, Inc.
Shivakumar Sastry received the bachelors
degree in electronics engineering from Banga-
lore University in 1984, the masters degree in
electrical engineering from Indian Institute of
Science, the masters degree in computer
science from the University of Central Florida,
and the PhDdegree fromCase Western Reserve
University. Prior to his transition into academics,
he held industry positions with GE Consulting
Services and Rockwell Automation. Since 2002,
he has been a member of the faculty in the Department of Electrical and
Computer Engineering, University of Akron, Akron, Ohio. He is the
recipient of several awards and holds two US patents. He is a member of
the IEEE Computer Society.
P.S. Sastry received the BSc(Hons) degree in
physics from the Indian Institute of Technology
(IIT), Kharagpur, in 1978 and the BE degree in
electrical communications engineering and the
PhD degree from the Department of Electrical
Engineering, Indian Institute of Science (IISc),
Bangalore, in 1981 and 1985, respectively. Since
1986, he has been a faculty member in the
Department of Electrical Engineering, IISc, where
he is currently a professor. He has held visiting
positions at the University of Massachusetts, Amherst, at the University of
Michigan, Ann Arbor, and at General Motors Research Laboratories,
Warren. He is an associate editor of IEEETransactions on Systems, Man,
and Cybernetics and IEEE Transactions on Automation Science and
Engineering. He is a fellowof the Indian National Academy of Engineering
and a senior member of the IEEE. He is a recipient of the Sir C.V. Raman
YoungScientist Award fromthe Government of Karnataka, the Dr. Vikram
Sarabhai Research Award from PRL, and the Most Valued Colleague
Award from General Motors.
> For more information on this or any other computing topic,
please visit our Digital Library at www.computer.org/publications/dlib.
MAMIDISETTY ET AL.: MULTIPATH DISSEMINATION IN REGULAR MESH TOPOLOGIES 1201