TDMA Scheduling With Optimized Energy Efficiency and Minimum Delay in Clustered W

TDMA Scheduling with Optimized Energy
Efficiency and Minimum Delay in Clustered

Wireless Sensor Networks
Liqi Shi, Student Member, IEEE, and Abraham O. Fapojuwo, Senior Member, IEEE
AbstractIn this paper, we propose a solution to the scheduling problem in clustered wireless sensor networks (WSNs). The
objective is to provide network-wide optimized time division multiple access (TDMA) schedules that can achieve high power efficiency,
zero conflict, and reduced end-to-end delay. To achieve this objective, we first build a nonlinear cross-layer optimization model
involving the network, medium access control (MAC), and physical layers, which aims at reducing the overall energy consumption. We
solve this problem by transforming the model into two simpler subproblems. Based on the network-wide flow distribution calculated
from the optimization model and transmission power on every link, we then propose an algorithm for deriving the TDMA schedules,
utilizing the slot reuse concept to achieve minimum TDMA frame length. Numerical results reveal that our proposed solution reduces
the energy consumption and delay significantly, while simultaneously satisfying a specified reliability objective.
Index TermsWireless sensor networks, cross-layer optimization, power efficiency, TDMA scheduling, slot reuse.
1 INTRODUCTION
S
CHEDULING of medium access plays an important role in
the performance of wireless sensor networks (WSNs). In
the literature, time division multiple access (TDMA) and
carrier sensing multiple access (CSMA) are two major
medium access approaches in WSNs. This work will only
focus on the TDMA approach because the scenario
specification of our research is a static network in which
TDMA is said to be more effective than CSMA, especially
under medium to high traffic load [1].
In this paper, we aim at deriving TDMA schedules with
optimized power consumption and minimum latency in
clustered WSNs. Energy efficiency is a major concern in
WSNs, since the batteries are often impossible to be
replaced or recharged in many cases [2]. In TDMA, a node
can be active only if it is scheduled to send or receive data,
which give it advantages in power efficiency. In the
literature, the medium access control (MAC) protocol of
many WSN and ad hoc network proposals is TDMA based,
e.g., [3], [4], [5]. Recently, cross-layer design is further
adopted to achieve least energy consumption in TDMA-
based WSNs, as summarized in [6].
In addition to energy efficiency, quality of service
(QoS) metrics such as end-to-end delay needs to be taken
into account in some applications or under certain
scenarios, for instance, delivering real-time data in radio
harsh environments.
To achieve the papers objective, we propose a two-step
approach to derive TDMA schedules supporting both high
energy efficiency and minimum delay in WSNs. In the first
step, we formulate the problem via cross-layer optimiza-
tion, aiming at deriving the most energy-efficient flows on
every link. Based on the calculated per-link flows, in the
second step, we propose an algorithm to obtain a TDMA
schedule with the least frame length. From the analysis
presented in Section 5, the least frame length guarantees
minimum delay for the derived TDMA schedules.
We consider the application of the proposed two-step
approach in clustered WSNs. It is widely known that
clustering technique can provide scalability for large-size
WSNs, since most of the operations can be accomplished at
cluster heads (CHs) and gateways (if present), whose
number is much less than the number of sensing nodes.
Instead of only addressing intracluster slot assignment, as is
mostly done in the literature (e.g., BMA [7]), we focus on the
more challenging intercluster slot assignment.
The main contributions of this paper are twofold. First,
we build a cross-layer nonlinear optimization model to
achieve energy efficiency with specified link reliability and
bandwidth constraints. Instead of solving this nonlinear
optimization problem directly by heuristic algorithms, we
transform the problem into two simpler subproblems at less
complexity, which facilitates the application of our ap-
proach in large-size WSNs. Second, we propose a schedul-
ing algorithm for slot assignment in clustered WSNs. This
scheduling algorithm incorporates the slot reuse concept
(from cellular networks) in calculating the schedules based
on the optimal flows derived from the proposed optimiza-
tion model. We show that the slot reuse concept signifi-
cantly reduces the end-to-end latency without a penalty in
the energy efficiency. The above two contributions distin-
guish our work from previous related works in the
literature, for example, [8], [9].
The paper is organized as follows: Section 2 provides an
overview of the related work in the literature. In Section 3,
IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 9, NO. 7, JULY 2010 927
. The authors are with the Department of Electrical and Computer
Engineering, The University of Calgary, 2500 University Drive N.W.,
Calgary, AB T2N 1N4, Canada. E-mail: {lishi, fapojuwo}@ucalgary.ca.
Manuscript received 29 Aug. 2008; revised 4 June 2009; accepted 12 Oct.
2009; published online 9 Mar. 2010.
For information on obtaining reprints of this article, please send e-mail to:
tmc@computer.org, and reference IEEECS Log Number TMC-2008-08-0348.
Digital Object Identifier no. 10.1109/TMC.2010.42.
1536-1233/10/$26.00 2010 IEEE Published by the IEEE CS, CASS, ComSoc, IES, & SPS
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 26,2010 at 10:10:33 UTC from IEEE Xplore. Restrictions apply.
we describe the system model. Section 4 presents the
proposed cross-layer nonlinear optimization model to
achieve the energy efficiency. This is followed by the
description of the proposed scheduling algorithm, which
forms the topic of Section 5. Numerical results are presented
and discussed in Section 6. We conclude the paper in
Section 7.
2 RELATED WORK
Earlier works on TDMA scheduling for WSNs mainly focus
on obtaining the shortest schedules [10], [11] or distributed
implementation [3], [4], [5], [10]. In [10], nonconflicted
schedules are obtained first, and a following algorithm then
goes through all nodes and slots, in turn, to produce
maximal broadcasting sets, thus reducing the schedule
length. The authors of [11] further take into account the
number of packets being sent at every node, and provide
the shortest schedules by eliminating the nodes without
packets to send at each loop of the proposed algorithm.
Both of the algorithms in [10] and [11] need global topology
information, which may not be scalable for very large-size
networks. To overcome the difficulty of obtaining global
topology information in very large-size networks, many
distributed slot assignment schemes have been proposed,
such as DRAND [3], PACT [4], TRAMA [5], and the depth
first search (DFS) scheme used in [10]. These approaches
obtain local topology and interference information at each
node, and compose schedules by exchanging messages
between local nodes within a certain range (i.e., the
interference range). Compared to the approaches demand-
ing global topology information, distributed scheduling is
more flexible, but at a cost of increased schedule length.
The focus of the works in [3], [4], [5], [10], [11] is at the
MAC layer. Working on a single layer only may lead to
inefficiency in utilizing the network resources. Recently,
cross-layer design approach is combined with TDMA
scheduling in order to obtain prolonged network lifetime.
In [12], joint routing, link scheduling, and power control are
considered to support high data rate for broadband wireless
multihop networks. A framework for cross-layer design
toward energy-efficient communication is presented in [13].
The authors address joint link scheduling and power
control with the objective of energy efficiency subject to
QoS guarantees in terms of bandwidth and bit error rate
(BER). A heuristic solution is proposed.
Interference-free TDMA schedules are calculated in [8]
for a small-scale network by joint optimization of the
physical, MAC, and network layers. The authors use convex
optimization to solve the cross-layer-based network lifetime
optimization problem, employing the interior point method
[14]. A single frame without slot reuse for the whole
network guarantees no interference. However, this also
leads to significant end-to-end delay, which makes this
approach unsuitable for large-size WSNs. In [9], the authors
consider both joint layer optimization and slot reuse to
derive energy-efficient schedules. A convex cross-layer
optimization model is proposed and solved iteratively to
maximize the network lifetime. The link schedules evolve at
each iteration until a specific energy consumption goal is
reached or no more optimal solution can be found.
For consistency with previous works, such as [8], [9], [12],
[13], the energy efficiency formulation presented in this
paper is also based on cross-layer design, and includes the
QoS constraints of packet loss rate (PLR) [15], [16]. Our work
also calculates interference-free TDMA schedules by incor-
porating the slot reuse concept, similar to what was done in
[9]. However, compared to these previous works, our work
exhibits three differences: First, the proposed framework
(i.e., cross-layer-based energy optimization model and
scheduling algorithm) achieves energy efficiency for a
specified PLR objective while minimizing the delay. For
example, Cui et al. [8], [9] have focused on lifetime
maximization, Kwon et al. [15] considered only the end-to-
end PLR, and Yuan et al. [16] included both PLR and delay
objectives, treated in the context of a specific virtual multiple
input multiple output cooperative communication. Second,
we incorporate the slot reuse concept for the purpose of
reducing the delay of the optimal flows obtained from the
cross-layer optimization model. Our intent of incorporating
the slot reuse concept is, therefore, quite different fromits use
in [9] whose goal is that of maximizing the networks
lifetime. Third, the proposed scheduling heuristic is flexible,
applicable to both clustered and nonclustered WSN archi-
tectures, unlike the approaches in previous works that are
specific to a given network architecture.
3 SYSTEM MODEL
In this section, we present the wireless sensor network
architecture and models of the physical, MAC, and network
layers. These models serve as input to the cross-layer
optimization model.
3.1 Network Architecture
The nodes in the network are divided into multiple
clusters, each comprising a CH and cluster members that
communicate via one hop to the CH. For example, the
clustering scheme proposed in [17] can be used to form
clusters, as development of clustering algorithms is outside
the scope of this paper. However, the solution presented in
this paper applies to any other clustering schemes. Using
the clustering scheme presented in [17], gateways are
selected to connect neighboring CHs. A gateway is a
cluster member belonging to the cluster represented by one
of the CHs it connects. As seen from Fig. 1a, note that we
do not constrain the transmission range of CHs and
gateways within one hop though. Only the intracluster
communication (i.e., communication from cluster members
to the CH) is restricted to one hop. To simplify the
presentation of the network model (Fig. 1a), the sensor
nodes inside each cluster are not shown explicitly. Instead,
we represent the total traffic generated by all the active
members of a cluster and sent to the CH by a virtual
link, as illustrated in Fig. 1b.
We assume that the operation of the network is divided
into rounds, each composed of an initialization phase
followed by a data relay phase, as shown in Fig. 2. A new
round is initiated when a WSN is deployed or the available
power of a CH or gateway reaches a specified threshold, or
after a fixed time interval. A contention-based protocol,
such as CSMA, is used in the initialization phase of each
928 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 9, NO. 7, JULY 2010
round to form clusters, and allows the only sink node in the
network to obtain the topology of the backbone network
composed of CHs and gateways. The sink then calculates
the optimized schedule (i.e., number of slots required on
each link and the ID number of the slot in a TDMA frame to
be assigned to each required slot) and informs all nodes
(e.g., broadcasts the schedule directly to all nodes), after
which the network enters the data relay phase. The duration
of the initialization phase, t1, is far less than that of the data
relay phase, t2, so that the energy consumption during the
initialization phase can be neglected. This operation round
repeats until a certain percentage of nodes run out of energy
or their energy is below a certain threshold. In this paper,
we only focus on the data relay phase and investigate
TDMA scheduling during this phase, under optimized
power consumption and minimum latency.
A graph G \ . 1 is used to denote the backbone
network (composed of CHs and gateways, including the
sink), where \ is the set of nodes and 1 the set of links.
Henceforth, in this paper, the term ictnoi/ refers to the
backbone network with node set of \ and link set of 1. The
total number of nodes in the network is ` j\ j, in which the
sink is taken as node #1, without loss of generality. If a
maximum transmission power is specified, we use G
\ . 1
0
insteadof G \ . 1, where 1
0
, a subset of 1, is the set
of links that can be formed according to the specified
maximum transmission power 1
tr.ior
. A link i. , is valid
only if i. , 2 1
0
. The links are unidirectional so that link i. ,
and link ,. i are two different links. We assume that the
sensor nodes in the network are stationary, for simplicity.
3.2 Physical Layer
To some extent, the physical layer model determines the
accuracy of the cross-layer optimization model. A realistic
physical layer model considers the propagation model,
modulation, encoding, demodulation, and decoding tech-
niques, thus linking major parameters such as power, signal
to noise ratio (SNR), and bit error rate together. Examples of
such a physical layer model can be found in [18].
In this paper, we use the following path loss model [19]:
1|d 1|d
0
10 log 10d,d
0
. 1
where d is the distance between the transmitter and receiver,
d
0
is the reference distance, and is the path loss exponent.
The parameter can be obtained empirically from propaga-
tion measurements. By setting the distance between nodes
assigned the same slot large enough such that the
interference is far less than thermal noise, the interference
between these nodes is considered to be negligible. The bit
error rate 1
c
can then be expressed as a function of SNR,
given the modulation-demodulation (MODEM) scheme. In
this paper, we suppose that Mica2 motes [20] are used. For
the noncoherent FSK MODEM scheme used in Mica2 motes,
the bit error rate is given by [21]:
1
c

1
2
c
o`1
2
1
`
1
. 2
where 1 is the data rate in bits per second (bps) and 1
`
is
the noise bandwidth. The SNR at a receiver which is located
at distance d from the transmitter can be expressed in an
average manner, given by
o`1d
d1
1
tr d1i
1|d
d1
1
i d1i
. 3
in which the parameters on the right-hand side are,
respectively, the transmission power, average path loss at
distance d, and the noise floor, all expressed in decibel units.
Suppose the length of a packet is 1 bytes, the encoding
rate is ,, then the packet loss rate j on each link, defined as
the probability that at least one transmitted bit in a packet is
corrupted, is given by
j 1 1 1
c
8,1
. 4
After linking all these parameters together, we still need
to build an energy consumption model. The major energy
consumption is spent on CPU, radio, and sensors, etc. It is
hard to estimate the CPU cycles in analysis, but in most
cases, it can be neglected [22]. The energy consumed by
sensing task is determined by the environments being
monitored and the sensing scheme (e.g., only some
designated sensors are active to achieve a certain coverage
reliability objective). Given the environments being mon-
itored and the predetermined monitoring scheme, over a
given period, we can assume that the sensing energy
consumption is fixed or can be estimated, and this energy
consumption can then be deducted from the total available
energy in the given period. Thus, we can now focus only
on the radio power, which can be adjusted to attain less
energy consumption. The average transmission energy
consumption by node r, in transmitting a data bit to
SHI AND FAPOJUWO: TDMA SCHEDULING WITH OPTIMIZED ENERGY EFFICIENCY AND MINIMUM DELAY IN CLUSTERED WIRELESS... 929
Fig. 2. Operation rounds.
Fig. 1. Network model.
node y with transmit power 1
tr.ry
1
cii.tr
1
tr
d
ry
, can
be expressed as
11
tr.ry

1
1
1
cii.tr
1
tr
d
ry
. 5
where 1
cii.tr
is the power consumed by the transmitter
circuits (excluding the power amplifier circuit), which is a
constant. The term 1
tr
d
ry
is the power consumed by the
power amplifier circuit at node r separated from node y by
distance d
ry
, and is calculated using (3) (expressed in linear
units) in concert with (1) and (2). Clearly, 1
tr.ry
is variable
depending on the transmission distance d
ry
and the packet
loss rate objective. For a given receiver circuit, the per bit
average receiving energy consumption at receive power 1
ir
is fixed, calculated by 11
ir
1
ir
,1, where 1
ir
is set to
1
cii.ir
, the power consumed by the receiver circuits, which
is also a constant.
3.3 MAC Layer
As stated in Section 1, the TDMA scheme is the MAC layer
protocol employed during the data relay phase. A TDMA
frame consists of a number of slots, each with fixedlength t.
We assume that NACKs and retransmissions are used at the
MAC layer. However, to simplify the analysis and the
scheduling algorithm, we neglect the cost of NACKs by
assuming that a NACK is only generated once in each fixed
feedback period to inform the source nodes of packets not
correctly received in the last period. By properly setting the
feedback period, energy cost of NACKs is negligible
compared to that of data packets. Given the packet loss
rate j, the per-hop average number of total transmissions for
a packet to be successfully received is 1,1 j. The
retransmissions denote the extra cost caused by bit errors.
On the other hand, increasing transmission power (which
means more energy consumptionper bit) can reduce bit error
rate, as seen from (2). This implies that an optimal transmis-
sion power exists for a link to achieve power efficiency.
3.4 Network Layer
Multihop, multipath routing is performed at the network
layer. Every CH or gateway has knowledge of its next hops
and the proportion of data traffic to every next hop. The
sink provides this information to every CH or gateway after
deriving the optimal schedules at the end of the initializa-
tion phase of each operation round.
4 CROSS-LAYER OPTIMIZATION MODEL
Based on the information provided in Section 3, a cross-
layer optimization model is formulated as follows:
iiiiii:c
_
`
i2
`
,1
11
tr.i,
1
1 j
i,
)
i,

`
i2
`
,2
11
ir
1
1 j
i,
)
i,
`
i2
`
i
11
ir

`
i2
`
o`.i
/1
1
:ci:./
11
tr./i
`
i
,`
o`.i
_
.
i. , 2 1
0
. i 6 ,
6a
s.t. )
i,
! 0. i 2 2. `. , 2 1. `. i 6 ,. 6b
)
i,
0 i)) i. , 62 1
0
oi i ,. 6c
`
,1.,6i
)
i,
)
,i
`
i
,C
i
. i 2 2. `. 6d
`
i2
`
,1
1
1 j
i,
)
i,
1. i 6 ,. 6e
`
,1.,6i
_
11
tr.i,
)
i,
1
1 j
i,
11
ir
)
,i
1
1 j
,i
_
`
i
11
ir
c
i
i 2 2. `.
6f
where 11
tr.i,
is the per bit average transmission energy
consumed at node i when transmitting to node , using
transmission power 1
tr.i,
, j
i,
is the packet loss rate on
link i. ,, )
i,
is the data flow rate (in bits per second) from
node i to , without accounting for the retransmissions, `
i
is
the new generated traffic (in bits per second), C
i
is the traffic
aggregation factor at node i if node i is a source CH, 1
:ci:./
is
the sensing energy expended by each sensor node / in a
cluster to capture andprocess the data, `
o`.i
is the number of
active sensor nodes in cluster i, and c
i
is the allowed
maximum consumed energy per unit time at node i (e.g.,
c
i
c
i.o.oi|
,T, where c
i.o.oi|
is the initial available energy in
millijoule (mJ) at node i and T is a fixed lifetime in seconds).
The objective function in (6a) gives the total consumed
energy of the whole network per unit time (i.e.,
iJ,:cc i\), which is composed of transmission and
receiving mode energy consumption at all backbone nodes
excluding the sink (the energy consumed at the sink is
neglected because its power source is replaceable), and the
energy consumed by each active sensor node (cluster
member) in sensing and transmitting the data to its CH.
Note that retransmissions are ignored for the intracluster
transmissions because of the close proximity of the sensor
nodes in a cluster to their cluster head, this is not the case for
intercluster communicationwhere the intercluster distance is
typically much larger than the intracluster distances.
Now, as noted in Section 3.2, for a given WSN
application (i.e., given monitoring environment), 1
:ci:./
is
constant. Also, if we assume, for simplicity, that the sensor
nodes in a cluster are at approximately the same average
distance from their CH and each generating the same
amount of traffic, then 11
tr./i
`
i
,`
o`.i
is also a constant.
From these preceding statements, we can consider the last
double summation term in (6a) as a constant. Based on the
foregoing, along with the fact that the energy optimization
is being done with respect to the backbone network nodes,
(6a) then reduces to
_
`
i2
`
,1
11
tr.i,
1
1 j
i,
)
i,

`
i2
`
,2
11
ir
1
1 j
i,
)
i,
`
i2
`
i
11
ir
_
. i. , 2 1
0
. i 6 ,.
6a
0
The first two constraints (6b) and (6c) state that the flow
rate should be nonnegative and flows can only exist on
valid links. The third constraint (6d) indicates that the
output data rate at each node should be the sum of input
data rate and data generation rate at that node (the data are
generated by cluster member nodes, aggregated at the
corresponding CH, and then, relayed to the sink). The
fourth constraint (6e) is the data rate or bandwidth limit of
the network, showing that the summed data rate with
retransmissions at all links should not be greater than the
maximal allowed data rate. This is a conservative con-
straint, since slots can be reused in our model. The fifth
constraint (6f) is the per unit time energy consumption
constraint at every node.
The optimization model depicted by (6) is cross-layer
based because it jointly combines network layer traffic loads,
MAC layer retransmission scheme, and physical layer
modulation scheme and bit error rate together, in order to
derive appropriate transmission power 1
tr.i,
andflowrate )
i,
at every link i. ,, i. , 2 1
0
. i 6 ,, to achieve network-wide
optimal power efficiency. Applying (2)-(5), this model is
obviously nonlinear due to the terms 11
tr.i,
1
1j
i,
and
11
ir
1
1ji,
. i. , 2 1
0
. i 6 ,, which are, respectively, the
transmit and receive energy consumed in relaying a bit from
node i to ,. Solving this nonlinear optimization problem
directly is a nontrivial task, so we transformthe probleminto
two simpler subproblems. We, therefore, propose to solve
the following subproblem #1 first.
Sub_Problem #1:
min. 11
tr.i,
11
ir
1
1 j
i,
i. , 2 1
0
s.t. |/ < 1
tr.i,
< n/ i 2 2. `. , 2 1. `. i 6 ,.
7
where lb and ub are the lower and upper bounds of the
transmission power, respectively. The purpose of subpro-
blem #1 is to find the optimal transmission power on
link i. ,, 1
ojt.tr.i,
, in order to consume the least energy
1
ojt.i,.trir
for transmitting a bit from node i to ,. Subpro-
blem #1 is actually a simple problem of deriving the
minimum value of a function. Numerical methods such as
golden section search [23] can be adopted to solve this
problem. The solution to (7) then serves as input to (6).
However, before being able to use the results of (7) as
input to (6), we need to prove that the optimal transmission
power on each link obtained from solving (6) is equivalent
to that obtained from solving (7).
Lemma 1. For the optimal energy consumption model proposed
in (6) on any link i. ,, the model must use the optimal
transmission power of this link 1
ojt.tr.i,
to achieve network-
wide optimal energy consumption.
Proof. Assume that the network-wide per bit optimal relay
energy consumption is 1
ojt.toto|
A 1
i,.trir
, where
1
i,.trir
is the per bit relay energy consumption on
link i. , and A is the per bit relay energy consumed by
the other links in the network. Suppose the transmis-
sion power at link i. ,, 1
tr.i,
, is not equal to the
optimal transmission power 1
ojt.tr.i,
of this link, we then
have 1
i,.trir
1
ojt.i,.trir
; thus, 1
ojt.toto|
A 1
i,.trir

1
0
ojt.toto|
A 1
ojt.i,.trir
, which contradicts with the
statement that 1
ojt.toto|
is the network-wide per bit
optimal relay energy consumption. tu
Now, we can solve (7) and input the results to (6). This
transforms the nonlinear optimization problempresented by
(6) into a simpler linear one, which we call subproblem #2.
Sub_Problem #2:
Define the parameters c1
i,
1,1 j
i,
and c2
i,

11
tr.i,
, i. , 2 1
0
. i 6 ,, both are constants that can be
calculated from solving (7). Subproblem #2 is then for-
mulated as
min.
_
`
i2
`
,1
c1
i,
c2
i,
)
i,

`
i2
`
,2
c1
i,
11
ir
)
i,
`
i2
`
i
11
ir
_
i. , 2 1
0
. i 6 ,.
8a
s.t. )
i,
! 0. i 2 2. `. , 2 1. `. i 6 ,. 8b
)
i,
0 i)) i. , 62 1
0
oi i ,. 8c
`
,1.,6i
)
i,
)
,i
`
i
,C
i
. i 2 2. `. 8d
`
i2
`
,1
c1
i,
)
i,
1. i 6 ,. 8e
`
,1.,6i
c1
i,
c2
i,
)
i,
c1
,i
11
ir
)
,i
`
i
11
ir
c
i
i 2 2. `.
8f
This linear optimization problem only has )
i,
as the
variables. Let 1
0
j j be the number of links in the backbone
network constrained by the maximum transmit power, this
linear optimization problem has O 1
0
j j variables and
O` 1
0
j j constraints. The problem is similar to the
traditional minimum cost flow problem [24], [25] and can
be solved accordingly. Thus, we significantly reduce the
complexity of solving the original nonlinear optimization
problem based on Lemma 1.
Some applications may have a predetermined per-link
packet transmission success rate (PTSR) objective (where
PTSR is the complement of PLR). In this case, we can
calculate the bit error rate and transmission power on every
link according to (2)-(4). We then derive the constants c1
i,
and c2
i,
, and directly input the results to (8) to solve this
linear optimization problem.
5 MINIMUM DELAY SCHEDULING ALGORITHM
Solving the optimization model in (6) gives the flow
distribution and transmission power on every link, which
achieves the energy-efficient data relay. We can, of course,
calculate the TDMA slots assigned to every link accordingly
without slot reuse. However, this results in very long
TDMA frame length, and thus, unacceptable delay in large-
size WSNs. Our goal then is to derive a relationship
between the delay incurred by a data packet at each
backbone network node (along its path to the sink) and the
TDMA frame length M so that, by reducing the frame
length through slot reuse, the delay is minimized. For this
analysis, we determine the node delay from the instant the
first bit arrives at the node until the time the last bit exits the
node. With the TDMA MAC adopted, the nodal delay can
be approximated by the time interval between the input
time slot where the data are read and the output time slot
from which the data are read out. Suppose the data packet
is received at a node in slot i1, i1 2 1. `, and relayed out
in slot i2, i2 2 1. `, the delay experienced by the relayed
data packet at this node is
1i1. i2
i2 i1. i2 i1.
` i1 i2. i2 < i1.
_
i1. i2 2 1. `.
9
Assuming that i1 and i2 are independent of each other,
the average delay experienced by the data packet at this
node is then
`
i21.i26i1
1i1. i2
` 1

`
2
. i1 2 1. `. 10
from which we conclude that reducing the TDMA frame
length translates to a reduction in nodal delay. The end-to-
end delay is the sum of the nodal delays along a path to the
sink, plus the associated propagation delays which are
negligible due to short distances between the backbone
nodes. Hence, minimizing the TDMA frame length also
minimizes the end-to-enddelay. Next, weapplythe slot reuse
concept toachieve areductioninTDMAframe length. That is,
a slot already assigned to a link can be reused at another link
provided that the interference between the two links is below
a specified threshold. In the following, a scheduling scheme
that implements the slot reuse concept is presented.
The first step is to determine the number of slots
required by each link in the backbone network and the
virtual link in each cluster. For a backbone link i. , with
flow )
i,
, maximum data rate 1 (both in bits per second), and
per-hop packet loss rate j
i,
, the slots needed (each of
length t) per second are:
`
i,

)
i,
1t1 j
i,
_ _
. i 2 2. `. , 2 1. `. 11
where r d e is the smallest integer greater than or equal to r.
Note that the slot length t should be selected as small as
possible to reduce the error introduced by the ceiling
function in (11). Also, note that the flow distribution )
i,
derived from the optimization model in (6) need not be
uniform for all the links; hence, the `
i,
s can be different for
different links. Let `
..i
denote the number of slots required
by the virtual link directed to CH i, which is determined by
CH i according to the aggregate intracluster traffic, and
then, sent to the sink. The sink uses the virtual link slot
requirements, along with the corresponding requirements
for the backbone links to determine the slots needed per
second at each node i (CH or gateway):
`
i

`
..i

`
,1
`
i,
. i 2 CH:.
`
,1
`
i,
. i 2 Gotcnoy:.
_
_
i 2 2. `. i 6 ,. 12
Without slot reuse, the TDMA frame length is
`
i2
`
i
,
which can become too long (i.e., too long delay) for a large-
size backbone network. A TDMA frame with short length is,
therefore, desired. Previous work has shown that in TDMA-
based radio networks, the problem of finding the shortest
TDMA frame is NP-hard [10], regardless of traffic distribu-
tion across the links. In this paper, to circumvent the
solution complexity, we, therefore, adopt a simple incre-
mental heuristic to calculate the TDMA schedule that has
the shortest length. The heuristic can start from any of the
backbone nodes (CH or gateway), assigning slots first to the
virtual link (if the node is a CH), and then, to all the
outgoing links of the node. Before assigning slots to any
succeeding nodes, the algorithm checks all the links that are
previously assigned with slots to make sure no conflicts
exist. As illustrated in Fig. 3, three criteria that guarantee
the conflict free feature must be simultaneously met before
assigning a slot : to any link (outgoing links or the virtual
link) of node i, i 2 2. `.
Criterion 1. Node i is not the sender or receiver of a previously
scheduled link using slot :, and the link of node i to be
scheduled with slot : does not have the same receiver as that of
a previously scheduled link currently assigned slot :.
Criterion 2. Scheduling : as a slot used by one of node is links
causes negligible interference to the receiver of a previously
scheduled link also using slot :.
Criterion 3. The sender of a scheduled link using slot : causes
negligible interference to the current receiver if it is using
slot : on one of node is links.
In this paper, we assume that the interference at a node is
negligible only if the interference is less than 10 percent of
the noise floor. This explains why we use SNR instead of
signal to interference plus noise ratio (SINR) in the system
model presented in Section 3.
Our proposed algorithm can be applied to both uniform
and nonuniform traffic. In addition, the abstraction of
intracluster communication by virtual links makes it easy to
apply this algorithm to clustered WSNs, which is said to
Fig. 3. Criterion 1: (a) Node i is a sender of slot :; thus, slot : cannot be
assigned to any other links of node i. (b) Node i is a receiver of slot :; thus,
slot : cannot be assigned to any other links of node i. (c) Slot : is assigned
to link i. i; thus, cannot be assigned to link i. i. Criteria 2 and 3:
(d) Transmission from node i to i causes negligible interference to ,,
and transmission from node i to , causes negligible interference to i.
have advantage in large-scale deployment compared to
nonclustered WSNs [26].
A description of the proposed Algorithm 1 for conflict
free TDMA frame slot assignment is given as follows:
Algorithm 1. Incremental heuristic for conflict free TDMA
frame slot assignment (implemented at the sink)
Input: backbone network graph G \ . 1. `
i
. d
i,
& 1
ir.i,
Output: Conflict-free schedule with frame length `
1: for each node ii 6 :ii/ of \ do
2: for each required slot i in `
i
do
3: Try assigning slot : 1;
4: while any of the 3 interference criteria is not
satisfied do
5: Try assigning the next slot : : 1;
6: end while
7: Assign slot : to required slot i of node i;
8: end for
9: Store the Idof the last slot : assignedto node i.
i
:;
10: end for
11: Calculate the frame length ` maxf
i
g. 8i 2 f2. `g
After deriving `, which is the maximum slot ID being
assigned in Algorithm 1, one must verify that `t 1.
This means that the frame duration should be less than
1 second in accordance with the definition of `
i
. Otherwise,
smaller values of t are needed to regenerate ` until
`t 1. In this algorithm, and also in the optimization
model described in Section 4, we assume that the distance
between every two nodes is known, and the sink can obtain
the distance information during the deployment phase of a
WSN, or by GPS devices. Using the propagation model and
transmission power on each link, and with knowledge of
the slots already assigned to the links, the interference
mentioned in criteria 2 and 3 can be calculated. We claim
that a conflict-free schedule is produced when this algo-
rithm terminates, since all the three criteria must be
satisfied before assigning any slot.
Proof of the conflict-free schedule. Suppose the derived
schedule is not conflict-free. Then, there exists at least one
slot : 2 1. `, which is scheduled on links ZZ f|
1
.
|
2
. . . . . |
7
g, such that conflict exists between a link |
r
(1 r 7) and at least one of the remaining links |
y
(1 y 7. y 6 r) in ZZ. Without loss of generality, assume
that |
y
has already been assigned with slot : prior to |
r
.
When we try to assign slot : to |
r
, from lines 4 and 5 in
Algorithm 1, due to the conflict between |
r
and |
y
,
Algorithm 1 will assign slot :
0
(:
0
:) to |
r
. This contra-
dicts with our assumption that |
r
and |
y
are both in ZZ
using slot :. Thus, we conclude that the derived TDMA
schedule by Algorithm 1 must be conflict-free. tu
Note that Algorithm 1 can also be used with any values
of the link flow )
i,
s (aside from the optimal )
i,
s determined
by the optimization model). The complexity of Algorithm 1
is estimated as follows: In the worst case, the number of
iterations from the three loops in Algorithm 1 are ` 1,
max`
i
. i 2 2. `, and ` 2 max`
i
. respectively. In a
large-size WSN, max`
i
is usually far less than `. Thus,
the complexity of Algorithm 1 is approximated as O`
2
.
Finally, from knowledge of the Id of the slots assigned to
the virtual link in each cluster i, CH i then assigns the
virtual link slots to its cluster members on an as needed
basis. Existing intracluster slot assignment approaches (e.g.,
[7]) can be adopted for this purpose.
We now illustrate the slot assignment procedure. Con-
sider a linear network shown in Fig. 4a in which per-hop
distance andeach cluster radius are both set to 15 meters. The
number on each link in Fig. 4a is the number of slots required
per second by that link. As stated earlier, this is calculated
using (11). Before implementing Algorithm1, we assume, for
illustration purpose, the following input parameter values:
f`
2
. `
3
. `
4
. `
5
. `
6
. `
7
. `
8
. `
9
g f4. 1. 4. 3. 5. 2. 4. 2g,
f1
tr.23
. 1
tr.34
. . . . . 1
tr.78
. 1
tr.89
. 1
tr.91
. 1
tr.24
. 1
tr.61
. 1
tr.81
g
f5.34. 5.34. . . . . 5.34. 5.34. 5.34. 85.45. 1.367.2. 85.45g i\.
and d
i,
, which is calculated according to the per-hop
distance specification. `
i
and 1
tr.i,
are obtained from
solving the optimization model built from this network
scenario. We also assume the slot length t, which is used
in calculating `
i
, is set to 50 milliseconds. Suppose the
heuristic Algorithm 1 starts from node 2, progressing
incrementally to node 9:
1. Node 2 (i 2): `
2
4
m 1. s 1: Slot 1 is never used before; thus, all
criteria are satisfied. Schedule slot 1 as a slot of the
virtual link of node 2.
m 2. s 1: Criterion1 is not satisfiedsince node 2
is the receiver of the assigned slot 1.
m 2. s 2: Slot 2 is never used before. Schedule
slot 2 as a slot of the virtual link of node 2.
slot 3 as a slot of link (2, 3).
is the sender of the assigned slot 3.
Fig. 4. Slot assignment using Algorithm 1.
The final slots assigned to the virtual and
outgoing links of node 2 are shown in Fig. 4b.
2. Node 3 (i 3), `
3
1
m 1, s 1: The average interference caused by
1
tr.34
5.34 mW at node 2, the receiver of previously
assigned slot 1, is calculated to be 94.77 dBm
according to d
32
15 m and the path loss model in
(1). Suppose the noise floor is 105 dBm [18], this
interference is far greater than 10 percent of the noise
floor. Criterion 2 is not satisfied.
m 1, s 2: Same as m 1. s 1.
m 1, s 3: Criterion1 is not satisfiedsince node 3
m 1, s 4: Criterion 1 is not satisfied since
link3. 4 andlink2. 4 havethesamereceiver, node4.
m 1, s 5: Slot 5 is never used before. Schedule
3. Nodes 4 and 5 (i 4 and 5, respectively)
Perform similar operations as shown in steps 1
and 2, and the slots assigned are shown in Fig. 4b.
4. Node 6 (i 6), `
6
5
m 1, s 1: Consider the virtual link of node 6
first. Slot 1 is assigned to the virtual link of node 2. In
the worst case, the sender of slot 1, as a cluster member
of node 2, causes worst interference to node 6 if it is at
the location of node 3 (closest to node 6). According to
1
tr.32
5.34 mW (one hop transmit power) and d
36

45 m, the interference caused by 1
tr.32
at node 6 is
113.85 dBm, which is not less than 10 percent of the
noise floor. Criterion 3 is not satisfied.
m 1, s 2-12: Similar reuse criteria evaluation
is performed, but none of the slots (2-12) can be
reused.
m 1, s 13: Slot 13 is never used before.
Schedule slot 13 as a slot of the virtual link of node 6.
m 2, s 1: The interference caused by 1
tr.67

5.34 mW at node 2 is 118.85 dBm, and the
interference caused by 1
tr.32
5.34 mW at node 7,
the receiver of link 6. 7, is also 118.85 dBm. All
criteria are satisfied; thus, slot 1 is reusedby link 6. 7.
m 3, s 2: Similarly, as m 2, s 1, slot 2 is
also reused by link 6. 7.
m 4, s 1-13: Consider link 6. 1 now. Similar
reuse criteria evaluation is performed, but none of
the slots (1-13) can be reused.
Schedule slot 14 as a slot of link 6. 1.
m 5, s 1-14: Similar reuse criteria evaluation
is performed, but none of the slots (1-14) can be
reused by link 6. 1.
Schedule slot 15 as a slot of link 6. 1.
5. Nodes 7, 8, and 9 (i 7, 8, and 9, respectively)
Perform similar reuse criteria evaluation as
shown in the above steps. The final schedules are
shown in Fig. 4b.
6. The frame length ` comes out to be 18. Verify that
`t 18 50 900 ms < 1 :. Otherwise, reduce
t, recalculate `
i
, and run Algorithm 1 again until
`t 1.
7. Each cluster head assigns virtual link slots to its
cluster members as needed.
6 NUMERICAL RESULTS
In this section, we present numerical results obtained from
the proposed optimization model and incremental heuristic
scheduling algorithm. We select the parameter values
assuming the Mica2 motes [20], and take 1 19.2 kbps,
1
`
30 kHz, , 2 (Manchester encoding), 1 50 bytes,
1
cii.tr
15 mW, and 1
ir
22.2 mW. We also set d
0
1 m,
1|d
0
55 dB, 4, and 1
i
105 dBm [18]. In order to
thoroughly study the transmission behavior under the
optimal scheme, we put no limit on the upper bound of
the transmission power, which means 1
0
1, implying that
all nodes in the network can reach each other, if necessary.
Without loss of generality, in all our numerical studies, we
assume that CH is traffic aggregation factor C
i
1. That is,
the traffic is not compressed at every CH. The numerical
studies are done using MATLAB [27] and a comprehensive
optimization tool LINGO [28].
6.1 Linear Backbone Network
We start first with a linear backbone network of size
7 nodes, as shown in Fig. 5, to observe the performance of
the optimization model and the scheduling algorithm
proposed in this paper. In the linear topology, the backbone
network nodes are arranged linearly and separated by the
same distance [9].
6.1.1 Scenario #1: Effect of Hop Distance
In the first scenario, we investigate the optimal transmission
schemes under various hop distance specifications. The
purpose is to discover how the values of transmission
power, transmitter circuit power consumption, and receiver
circuit power consumption affect the transmission behavior
in order to achieve power efficiency. In this scenario, we
assume that the application has a per-hop PTSR objective of
90 percent; thus, we only need to solve the linear optimiza-
tion problem presented by (8). As shown in Fig. 5, the traffic
generated at CHs 2, 4, and 6 are set to 0.5, 2, and 1 kbps,
respectively. The operation period is set to one day and the
residual energy at each node is 6,000 J. More information on
battery energy storage can be obtained from [29]. We vary
per-hop distance d
/oj
to 5, 10, and 15 meters. Accordingly,
Fig. 5. Optimal transmission schemes for: (a) d
/oj
5 m, (b) d
/oj
10 m,
and (c) d
/oj
15 m.
the transmission power needed in each case is shown in
Table 1. It is seen that at a fixed PTSR, the transmission
power increases significantly with the distance between the
transmitter and receiver, since the path loss increases
exponentially with the distance.
The flowdistributions derivedfrom(8) are showninFig. 5.
Verifying by (8e) and (8f), we realize that the above traffic
generation and per node residual energy settings can easily
satisfythe twoconstraints; thus, theydonot affect the optimal
results and we can focus on investigating the optimal
transmission behavior only. As seen in Figs. 5 and 6, using
the proposed cross-layer optimization model, the ratio of
transmission power to (transmitter + receiver) circuit power
consumption determines the transmission behavior. When
the total circuit power consumption is dominant (i.e., the
transmission range is not greater than 4 and 2 hops in Fig. 6
for per-hop distances 5 and 10 meters, respectively), the
nodes send data directly to nodes multihops away (by
increasing the transmission power which is smaller than the
total circuit power) withless number of intermediate nodes to
save energy consumed by transmitter and receiver circuits
(i.e., Figs. 5a and 5b). However, short-distance relay is
preferred if the cost of transmission is significantly higher
than the total energy cost of the transmitter and receiver
circuits. For example, in Fig. 5c, the power for transmission
beyond a distance of two hops away is significantly larger
than the total circuit power; thus, packet relay is done in a
hop-by-hop manner.
6.1.2 Scenario #2: Effect of the Residual Energy and the
Slot Assignment
In the second scenario, we investigate the effect of per node
residual energy constraint and look at the capability of
Algorithm 1 in reducing the TDMA frame length. We use
the same per-link PTSR objective (90 percent), topology, and
traffic source configurations of scenario #1, and set per-hop
distance to 15 meters. The residual energy at node 4 is
assumed to be 500 J, and set to 6,000 J at the other nodes for
one-day continuous operation. Due to the limited energy
constraint at node 4, we derive a flow distribution, as
shown in Fig. 7a, in which part of the traffic from node 3 has
to be transferred on a suboptimal path to node 5. Compared
to the flow distribution in Fig. 5c, this new flow distribution
consumes more power due to the residual energy constraint
at node 4. It can then be reasonably concluded that the
power cost for transmitting a packet to the sink in a WSN
increases as more and more nodes run out of energy, since
more and more suboptimal paths have to be adopted. On
the other hand, assume that some relay nodes are deployed
in a WSN such that these nodes consume negligible energy
when not in communication mode. We can say that the
optimization model achieves more power efficiency in such
a WSN (for example, if assuming that a backup CH near
node 4 exists in this scenario), since there are more choices
of selecting optimal paths toward the sink even if some
nodes run out of energy.
Next, we show how the slots are assigned and reused.
Let the slot size t be 10 milliseconds. Applying (11) to
every link including the virtual links in Fig. 7a, we have the
number of slots required by each link, as shown in Fig. 7b.
Now, we can assign slots to every link according to
Algorithm 1, and the slot assignment procedure is similar
as the example illustrated in Section 5. We finally have a
frame of 94 slots with duration of 940 milliseconds, as seen
in Fig. 7c. We can decrease the slot size t to mitigate the
error introduced by (11) thus reducing the frame duration.
For example, repeating the above procedure with t 5 ms
gives a frame with duration of 915 milliseconds. However,
this is at the cost of increased synchronization complexity
when t is set to a small value.
TABLE 1
Transmission Power (in Milliwatts)
at Various Distance (PTSR 0.9)
Fig. 6. Transmission power to (transmitter + receiver) circuit power.
Fig. 7. (a) The effect of per node residual energy constraints. (b) Number
of slots needed on each link. (c) Assigned slots and slot reuse.
In Fig. 7c, slots 1-3 are reused on link 6. 7 and slots 4-6
are reused on link 7. 1. Of the 94 slots per TDMA frame,
six are reused, resulting in a frame length reduction of
approximately 7 percent (6,94 6) of that without slot
reuse. We expect that the benefit of slot reuse will become
significant in large-size WSNs. This point is further
supported by the following scenario #3, where more nodes
are used for composing the linear topology.
6.1.3 Scenario #3: Comparison with Non-Cross-Layer
Approach
In the third experiment, we compare the following cross-
layer design (CLD) and non-cross-layer design (NCLD)
approaches:
1. CLD optimal: First, calculate per-link optimal trans-
mission power using (7), and then, input the results
to (6). Also, use Algorithm 1 to derive the TDMA
schedules.
2. CLDwithprefixedPTSRobjective: First, calculate per-
link transmission power using a prefixed PTSR
objective, andthen, input theresultsto(6). Algorithm1
is used for TDMA schedules.
3. NCLD, shortest path, without slot reuse: In this
NCLD approach, shortest path (in number of hops)
is used as a network layer routing scheme for a
traffic source (or CH) to send data hop by hop to the
sink. TDMA without slot reuse serves as the MAC
layer protocol. The distance of one hop for shortest
path routing is determined by the maximum
transmission power configuration.
We investigate the performance of the above three
approaches by first using a linear backbone network
topology consisting of 19 nodes (including the sink). The
distance between two neighboring nodes is 10 meters. As
shown in Fig. 8, CHs 2 and 4 are configured as data sources,
each with a data generation rate of 0.5 kbps. The residual
energy at each node is 6,000 J, which is enough for one days
operation under the given traffic configuration.
Let the slot size t be 10 milliseconds. Table 2 gives the
backbone network power consumption and TDMA frame
length for the approaches of CLD optimal, CLD with
prefixed PTSR of 90, 80, and 70 percent, and NCLD with
maximum per-hop distance of 10, 20, and 40 meters and
unlimited (unlimited per-hop distance implies that each CH
sends data via one hop, directly to the sink). It is observed
that for the NCLD scheme, the TDMA frame length
decreases as the maximum per-hop distance increases,
since fewer hops are used for relaying the data. The shortest
frame length is achieved when there is direct communica-
tion (i.e., unlimited per-hop distance) from the data sources
to the sink. Even though the NCLD (unlimited) approach
achieves the smallest frame length, its high power cost is
significantly unacceptable. On the other hand, the NCLD
approach with specified per-hop distance reduces power
consumption, but at the expense of increased frame length.
Clearly, the CLD approach strikes a better balance in the
trade-off between power consumption and frame length,
compared to the NCLD approach. It is also seen from
Table 2 that the CLD optimal approach exhibits a better
performance in terms of energy consumption than the CLD
with prefixed PTSR approach because the former adopts
optimal energy consumption on a per-link basis. Fig. 9
depicts the per bit relay energy consumption (i.e., (7)),
plotted against the hop distance. Compared to the CLD
with prefixed PTSR approach, the CLD optimal approach
consumes the least per bit relay energy. However, it is
interesting to find that similar energy is consumed by the
approaches: CLD optimal and CLD (PTSR 90 percent)
(i.e., CLD with PTSR prefixed at 90 percent) at hop distance
of 40 and 50 meters. The explanation is obtained from
Table 3, which shows the values of PTSR obtained for the
CLD optimal approach, as the hop distance is varied.
Unlike CLD (PTSR 90 percent) where the PTSR is fixed at
90 percent for all the values of hop distance considered, in
the case of CLD optimal, the PTSR is adaptive to hop
distance. At a hop distance of 40 and 50 m, the PTSR for the
CLD optimal approach is approximately equal to 90 percent,
Fig. 8. Linear topology used in scenario #3.
Fig. 9. Per bit relay energy consumptions versus hop distance.
TABLE 2
Comparison of Different Relay Schemes (Linear Topology)
making the per bit relay energy similar to that of CLD
(PTSR 90 percent).
6.2 Grid Backbone Network
Now, consider a grid backbone network topology with
49 backbone nodes, as shown in Fig. 10a. As before, we only
show the CHs and gateways in this figure, and the virtual
links and cluster members are omitted for ease of
presentation. As a result of applying the clustering scheme
proposed in [17], there are no isolated nodes in the
network. The distance between two successive CHs is
either 20 meters or 28.3 meters, and the diameter of each
cluster is 30 meters. The sink is located at the top left corner,
but can be easily extended to the scenario in which the sink
is placed at the center.
We first set the traffic rate at each CH (or the
corresponding virtual link) to 0.2 kbps. The initial residual
energy at each CH or gateway is 6,000 J and the continuous
operation duration is one day. In this experiment, we
compare the performance of the NCLD and CLD ap-
proaches. The maximum one hop distance for the NCLD
approach is configured to be either of two values: 10 meters
and unlimited. For the CLD with prefixed PTSR approach,
the PTSR objective is set to 90 percent.
Fig. 10b gives the flow distribution which is identical
(due to the sparseness of the nodes) for the CLD
transmission schemes with or without a prefixed PTSR.
Even with the same flow distribution, the CLD optimal
scheme (i.e., without a prefixed PTSR), as seen from Table 4,
slightly outperforms the one with prefixed PTSR in terms of
energy efficiency, same as observed in Section 6.1. From
Table 4 and Fig. 10b, we conclude that the NCLD approach,
based on the shortest path routing, which is widely used in
the literature, is less energy efficient than the CLD scheme
with or without a prefixed PTSR.
We set the slot size t to 10 milliseconds. In Table 4, the
slots needed for each scheme are presented. Compared to
the NCLD scheme, the advantage of the CLD scheme is its
provision of the most power-efficient relay scheme (or flow
distribution), based on which a TDMA schedule with
minimum frame length (or delay) is derived.
In Fig. 11, we give an example schedule for the CLD
schemes (with or without a prefixed PTSR). Since Algo-
rithm 1 does not specify the order of the nodes and links
during scheduling, users can specify the order according to
the application needs. For example, starting slot assignment
at nodes farther away from the sink, or assigning successive
slots to a specific path. In Fig. 11, the scheduling order
is random. After applying Algorithm 1, the frame length is
98 slots with duration of 980 milliseconds. Even for this
moderate-size network (with 49 backbone nodes), 14 slots
TABLE 3
Per-Link PTSR for the CLD Optimal Approach
Fig. 10. (a) Grid backbone network topology. (b) Optimal flow distribution
with or without predetermined PTSR.
TABLE 4
Comparison of Different Relay Schemes (Grid Topology)
are reused, which means that the average delay is reduced
12.5 percent ( 14,98 14)) compared to using a TDMA
frame without slot reuse, given that the same optimal relay
scheme is adopted. Undoubtedly, more and more slots can
be reused as the network size increases. Thus, compared to
other solutions integrating network-wide traffic into one
single large frame [8], Algorithm 1 effectively reduces the
end-to-end latency.
6.3 Random Topology Backbone Network
We further consider a network with random topology. In
Fig. 12, 169 backbone nodes are uniformly distributed in a
120 m 120 m. Four nodes are randomly picked up as data
sources (which means that traffic is generated in the
corresponding clusters), each with data generation rate of
0.2 kbps.
The optimal relay scheme presented in Fig. 12 is
determined by the ratio of transmission power to (transmitter
+ receiver) circuit power, as concluded in the first scenario
using the linear topology.
Let the slot size t be 10 milliseconds. Running
Algorithm 1 gives the TDMA schedule with frame length
of 32 slots, as shown in Fig. 12. Note that without slot reuse
as done in [8], the TDMA frame would be 60 slots using the
same flows in Fig. 12. From (10), the delay is proportional to
the frame length. Thus, we conclude that compared to the
scheme in [8] integrating network-wide traffic into one
single TDMA frame, the average delay is 46.7 percent less in
this scenario using our approach. The percent delay
reduction is higher than that obtained for the grid topology,
because a larger size network is used. With nodes much
further apart, it is natural to expect that more slots can be
reused, translating to a higher reduction in the frame length
for large-size networks.
For this random topology, we also compare the
performance of the CLD optimal approach to that for
NCLD approach as defined in Section 6.1. Table 5 gives the
network-wide (backbone) power consumption and TDMA
frame length (t 10 milliseconds) under the CLD optimal
approach and the NCLD approach (with different max-
imum hop distance configurations). The NCLD approach is
not available (N/A) when the maximum hop distance is set
to 10 meters in this random topology, since some nodes
might not be able to find its next hop within this
transmission range specification. This does not occur in
the grid topology case where the distance between two
neighboring nodes in the same row (or column) is strictly
10 meters. When the maximum one hop distance increases
from 20 meters to unlimited for the NCLD approach, less
relay links are involved in forwarding the data to the sink
and the TDMA frame length is reduced. However, due to
the large per-hop transmission distance, the increase of
power cost is more significant and dominant compared to
the small gain in frame length reduction.
In Table 5, it is observed that the power consumption of
the NCLD approach closest to that of the optimal scheme
takes place when the maximum hop distance is 20 meters,
but its frame length is more than double of that obtained
using the proposed CLD optimal scheme, similar behavior
is observed earlier for the linear backbone topology.
Compared to the NCLD approach, the optimal scheme
gives acceptable performance in terms of both energy
efficiency and delay.
Fig. 11. TDMA schedule for a grid backbone network topology.
Fig. 12. TDMA schedule for a random backbone network.
TABLE 5
Comparison of Different Relay Schemes (Random Topology)
7 CONCLUSIONS
Realizing the relationship between the transmission power
and retransmissions on a link determining the optimal
transmission power, we build a cross-layer design-based
nonlinear optimization model which aims at minimizing
the network-wide energy consumption. We solve this
problem by transforming it into two subproblems with less
complexity. Based on the results obtained from the model,
we further propose an algorithm that provides a conflict-
free schedule utilizing the slot reuse concept. The schedul-
ing algorithm can be applied to WSNs with nonuniformly
distributed traffic and different link transmission power. In
addition, we introduce the virtual link concept, which treats
the whole intracluster communications of a cluster as a
virtual link, and makes our approach feasible for both
clustered and nonclustered WSNs. This feature significantly
increases the applicability of our approach.
Numerical studies are done for the linear, grid, and
random backbone network topologies. Results reveal the
advantage of the cross-layer optimization model in energy
conservation, and the effectiveness of the scheduling
algorithm in reducing the TDMA frame length, and thus,
the end-to-end latency.
Finally, it should be noted that in this paper, a slot is
reused only if the interference introduced is negligible.
Consider slot reuse with non-negligible interference, i.e.,
two links using the same slot and interfering with each
other. In this case, the trade-off between energy consump-
tion and frame length (or delay) needs to be investigated.
Also, the current work does not impose a limit on the frame
length; hence, there are no assignment failures during the
execution of the proposed scheduling algorithm. In practice,
the frame length is fixed so that, given a desired frame
length, a design objective is to derive the least consumed
energy in a WSN while also minimizing the slot assignment
failures. Future work is needed to extend the proposed
framework to the scenarios of slot reuse with non-negligible
interference and fixed frame length.
ACKNOWLEDGMENTS
This work was supported in part by a grant from the
Natural Sciences and Engineering Research Council
(NSERC) Canada. The authors also acknowledge the
valuable comments and suggestions from the anonymous
reviewers which have enhanced the quality of this paper.
REFERENCES
[1] I. Rhee, A. Warrier, M. Aia, J. Min, and M.L. Sichitiu, Z-MAC: A
Hybrid MAC for Wireless Sensor Networks, IEEE/ACM Trans.
Networking, vol. 16, no. 3, pp. 511-524, June 2008.
[2] I.F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci, A
Survey on Sensor Networks, IEEE Comm. Magazine, vol. 40, no. 8,
pp. 104-112, Aug. 2002.
[3] I. Rhee, A. Warrier, J. Min, and L. Xu, DRAND: Distributed
Randomized TDMA Scheduling for Wireless Ad-Hoc Networks,
Proc. Seventh ACM Intl Symp. Mobile Ad Hoc Networking and
Computing, pp. 190-201, May 2006.
[4] G. Pei and C. Chien, Low Power TDMA in Large Wireless Sensor
Networks, Proc. IEEE Military Comm. Conf. (MILCOM 01), Comm.
for Network-Centric Operations: Creating the Information Force, vol. 1,
pp. 347-351, Oct. 2001.
[5] V. Rajendran, K. Obraczka, and J.J. Garcia-Luna-Aceves, Energy-
Efficient, Collision-Free Medium Access Control for Wireless
Sensor Networks, Proc. First ACM Conf. Embedded Networked
Sensor Systems (Sensys 03), pp. 181-192, Nov. 2003.
[6] T. Melodia, M.C. Vuran, and D. Pompili, The State of the Art in
Cross-Layer Design for Wireless Sensor Networks, Lecture Notes
in Computer Science, Springer, 2006.
[7] J. Li and G.Y. Lazarou, A Bit-Map-Assisted Energy-Efficient
MAC Scheme for Wireless Sensor Networks, Proc. Third Intl
Symp. Information Processing in Sensor Networks, pp. 55-60, Apr.
2004.
[8] S. Cui, R. Madan, A.J. Goldsmith, and S. Lall, Cross-Layer Energy
and Delay Optimization in Small-Scale Sensor Networks, IEEE
Trans. Wireless Comm., vol. 6, no. 10, pp. 3688-3699, Oct. 2007.
[9] R. Madan, S. Cui, S. Lall, and A.J. Goldsmith, Cross-Layer Design
for Lifetime Maximization in Interference-Limited Sensor Net-
works, IEEE Trans. Wireless Comm., vol. 5, no. 11, pp. 3142-3152,
Nov. 2006.
[10] R. Ramaswami and K.K. Parhi, Distributed Scheduling of
Broadcasts in a Radio Network, Proc. IEEE INFOCOM, pp. 497-
504, Apr. 1989.
[11] S.C. Ergen and P. Varaiya, TDMA Scheduling Algorithms for
Sensor Networks, technical report, UC Berkeley, July 2005.
[12] R.L. Cruz and A.V. Santhanam, Optimal Routing, Link Schedul-
ing and Power Control in Multi-Hop Wireless Networks, Proc.
IEEE INFOCOM, vol. 1, pp. 702-711, Mar. 2003.
[13] U.C. Kozat, I. Koutsopoulos, and L. Tassiulas, A Framework for
Crosslayer Design of Energy-Efficient Communication with QoS
Provisioning in Multi-Hop Wireless Networks, Proc. IEEE
INFOCOM, vol. 2, pp. 1446-1456, Mar. 2004.
[14] S. Boyd and L. Vandenberge, Convex Optimization. Cambridge
Univ. Press, 2003.
[15] H. Kwon, T.H. Kim, S. Choi, and B.G. Lee, A Cross-Layer
Strategy for Energy-Efficient Reliable Delivery in Wireless Sensor
Networks, IEEE Trans. Wireless Comm., vol. 5, no. 12, pp. 3689-
3699, Dec. 2006.
[16] Y. Yuan, Z. He, and M. Chen, Virtual MIMO-Based Cross-Layer
Design for Wireless Sensor Networks, IEEE Trans. Vehicular
Technology, vol. 55, no. 3, pp. 856-864, May 2006.
[17] M. Gerla, T.J. Kwon, and G. Pei, On Demand Routing in Large
Ad Hoc Wireless Networks with Passive Clustering, Proc. IEEE
Wireless Comm. and Networking Conf. (WCNC 00), pp. 100-105,
Sept. 2000.
[18] M. Zuniga and B. Krishnamachari, Analyzing the Transitional
Region in Low Power Wireless Links, Proc. IEEE Ann. Conf.
Sensor and Ad Hoc Comm. and Networks (SECON 04), pp. 517-526,
2004.
[19] T.S. Rappaport, Wireless Communications: Principles and Practice,
second ed. Prentice Hall, 2001.
[20] Mica2 Wireless Measurement System, http://www.xbow.com/
Products/Product_pdf_files/Wireless_pdf/MICA2_Datasheet.
pdf, 2010.
[21] J.G. Proakis, Digital Communications, fourth ed. Prentice-Hall,
2001.
[22] V. Shnayder, M. Hempstead, B.-r. Chen, and M. Welsh, Power-
TOSSIM: Efficient Power Simulation for TinyOS Applications,
http://www.eecs.harvard.edu/~shnayder/ptossim, 2010.
[23] G.E. Forsythe, M.A. Malcolm, and C.B. Moler, Computer Methods
for Mathematical Computations. Prentice-Hall, 1976.
[24] R.K. Ahuja, T.L. Magnanti, and J.B. Orlin, Network Flows: Theory,
Algorithms, and Applications. Prentice-Hall, 1993.
[25] M.S. Bazaraa, J.J. Jarvis, and H.D. Sherali, Linear Programming and
Network Flows, third ed. Wiley, 2004.
[26] O. Younis, M. Krunz, and S. Ramasubramanian, Node Clustering
in Wireless Sensor Networks: Recent Developments and Deploy-
ment Challenges, IEEE Network, vol. 20, no. 3, pp. 20-25, May/
June 2006.
[27] MathWorksMATLAB and Simulink for Technical Computing,
http://www.mathworks.com, 2010.
[28] LINDO Systems, http://www.lindo.com, 2010.
[29] Battery Energy Storage in Various Battery Sizes, http://www.
allaboutbatteries.com/Energy-tables.html, 2010.
Liqi Shi received the BEng degree in commu-
nication engineering from the National University
of Defense Technology, China, in 1996, and the
MSc degree in electrical and computer engineer-
ing in 2007 from the University of Calgary,
Canada, where he is currently working toward
the PhD degree in electrical and computer
engineering. His previous work experience was
mainly focused on protocol development and
implementation in embedded systems. Cur-
rently, his research interests lie in the area of protocol design with
QoS support in mobile ad hoc networks and wireless sensor networks.
He is a student member of the IEEE.
Abraham O. Fapojuwo received the BEng
degree (with first class honors) from the Uni-
versity of Nigeria, Nsukka, Nigeria, in 1980, and
the MSc and PhD degrees in electrical engineer-
ing from the University of Calgary, Alberta,
Canada, in 1986 and 1989, respectively. From
1990 to 1992, he was a research engineer with
NovAtel Communications, Ltd., Calgary, where
he performed numerous exploratory studies on
the architectural definition and performance
modeling of digital cellular systems and personal communications
systems. From 1992 to 2001, he was with Nortel Networks, where he
conducted, led, and directed system-level performance modeling and
analysis of wireless communication networks and systems. In January
2002, he joined the Department of Electrical and Computer Engineering,
University of Calgary, where he is currently a full professor. He is also an
adjunct scientist with TRLabs, Calgary. His current research interests
include protocol design and analysis for future-generation wireless
communication networks and systems, and best practices in software
reliability and requirements engineering. He is a registered professional
engineer in the Province of Alberta. He is a senior member of the IEEE.
> For more information on this or any other computing topic,
please visit our Digital Library at www.computer.org/publications/dlib.

TDMA Scheduling With Optimized Energy Efficiency and Minimum Delay in Clustered W

Enviado por

Dados do documento

Descrição original:

Direitos autorais

Formatos disponíveis

Compartilhar este documento

Compartilhar ou incorporar documento

Opções de compartilhamento

Você considera este documento útil?

Este conteúdo é inapropriado?

Direitos autorais:

Formatos disponíveis

TDMA Scheduling With Optimized Energy Efficiency and Minimum Delay in Clustered W

Enviado por

Direitos autorais:

Formatos disponíveis

TDMA Scheduling with Optimized Energy

Efficiency and Minimum Delay in Clustered

Você também pode gostar