Adaptive Low Power Listening for Wireless Sensor Networks

Abstract

Most sensor networks require application-specific network-wide performance guarantees, suggesting the need for global and flexible network optimization. The dynamic and nonuniform local states of individual nodes in sensor networks complicate global optimization. Here, we present a cross-layer framework for optimizing global power consumption and balancing the load in sensor networks through greedy local decisions. Our framework enables each node to use its local and neighborhood state information to adapt its routing and MAC layer behavior. The framework employs a flexible cost function at the routing layer and adaptive duty cycles at the MAC layer in order to adapt a node's behavior to its local state. We identify three state aspects that impact energy consumption: 1) number of descendants in the routing tree, 2) radio duty cycle, and 3) role. We conduct experiments on a test-bed of 14 mica2 sensor nodes to compare the state representations and to evaluate the framework's energy benefits. The experiments show that the degree of load balancing increases for expanded state representations. The experiments also reveal that all state representations in our framework reduce global power consumption in the range of one-third for a time-driven monitoring network and in the range of one-fifth for an event-driven target tracking network.

Full text

Adaptive Low Power Listening
for Wireless Sensor Networks
Raja Jurdak, Member, IEEE , Pierre Baldi, Senior Member, IEEE, and
Cristina Videira Lopes, Member, IEEE
Abstract—Most sensor networks require application-specific network-wide performance guarantees, suggesting the need for global
and flexible network optimization. The dynamic and nonuniform local states of individual nodes in sensor networks complicate global
optimization. Here, we present a cross-layer framework for optimizing global power consumption and balancing the load in sensor
networks through greedy local decisions. Our framework enables each node to use its local and neighborhood state information to
adapt its routing and MAC layer behavior. The framework employs a flexible cost function at the routing layer and adaptive duty cycles
at the MAC layer in order to adapt a node’s behavior to its local state. We identify three state aspects that impact energy consumption:
1) number of descendants in the routing tree, 2) radio duty cycle, and 3) role. We conduct experiments on a test-bed of 14 mica2
sensor nodes to compare the state representations and to evaluate the framework’s energy benefits. The experiments show that the
degree of load balancing increases for expanded state representations. The experiments also reveal that all state representations in
our framework reduce global power consumption in the range of one-third for a time-driven monitoring network and in the range of one-
fifth for an event-driven target tracking network.
Index Terms—Cross-layer, framework, cost function, routing, medium access control, state information, test-bed, energy, power.
Ç
1INTRODUCTION
W
IRELESS sensor networks consist of autonomous sensor
nodes that monitor physical indicators in their
environment and communicate with each other to report
the data, interfacing the physical world with the digital
domain. Sensor networks present both new application
opportunities and new design challenges. Sensor network
applications include tracking and intrusion detection for
military purposes, pollutant and habitat monitoring for
environmental purposes, traffic and location monitoring for
civilian use, and warehousing and toxic leak monitoring for
industrial automation applications.
Each application typically requires custom performance
guarantees, so the wide application space is a design
challenge in and of itself. Other challenges for sensor
network design include the limited resources at each node
and the lack of infrastructure of these networks.
1.1 Design Challenges
Strictly layered communication models, such as the OSI
reference model, are designed for conventional wired and
wireless networks, where individual nodes have relatively
large communication bandwidth, processing power, mem-
ory capacity, and energy resources. Unlike most conven-
tional networks, the limited resources at individual nodes in
sensor networks require more versatile cross-layer models
that enable fine-grained optimization of resource usage.
Several cross-layer models have been proposed to address
specific optimization variables in wireless sensor networks,
such as link scheduling and routing flow [2], [3], [4], [5].
Here, we adopt the cross-layer optimization framework for
sensor networks which we had proposed in [6], and we
tailor this framework for a data gathering sensor network.
Before presenting the framework, the following discussion
explores the challenges that motiv ate the framework’s
design choices.
Traditionally, communication networks have not been
designed from scratch in a principled way to optimize
global network properties. As an emerging class of net-
works, sensor networks can greatly benefit from a prin-
cipled design approach that addresses highly dynamic node
states and that aims to optimize complex trade-offs between
use of resources, quality-of-service (QoS), and other
relevant parameters. We have identified three main design
goals for sensor network optimization mechanisms:
1. Flexible: The wide and diverse range of sensor
network applications highlights the need for a
flexible means to customize network behavior in
order to meet each applications’ performance goals.
2. Global: The nodes in sensor networks collaborate to
achieve a common network-wide goal. As such,
sensor network optimization mechanisms should
provide global network performance guarantees.
3. Adaptive: The local node states in sensor networks
are highly dynamic. An optimization strategy for
sensor networks should adapt to changes in local
node states.
Sensor network optimization methods should balance
the two conflicting design challenges of providing global
performance guarantees and of ensuring adaptability to
IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 6, NO. 8, AUGUST 2007 1
. R. Jurdak is with the PRISM Laboratory, School of Computer Science and
Informatics, Belfield, Dublin 4, Ireland. E-mail: Raja.Jurdak@ucd.ie.
. P. Baldi and C.V. Lopes are with the School of Information and Computer
Sciences, University of California, Irvine, Irvine, CA 92697-3435.
E-mail: {pfbaldi, lopes}@ics.uci.edu.
Manuscript received 13 Oct. 2005; revised 1 June 2006; accepted 28 Nov.
2006; published online 7 Feb. 2007.
For information on obtaining reprints of this article, please send e-mail to:
tmc@computer.org, and reference IEEECS Log Number TMC-0301-1005.
Digital Object Identifier no. 10.1109/TMC.2007.1037.
1536-1233/07/$25.00 ß 2007 IEEE Published by the IEEE CS, CASS, ComSoc, IES, & SPS

local node states. In our optimization framework, we
distinguish between three types of state information. Node
state is the vector of parameters that represents the cross-
layer state of a single sensor node. For a particular node N,
the neighborhood state is the vector of node states of the
direct neighbors of N. The global network state is the vector
of all node states in the network.
Optimizing network behavior is trivial if each node has
the global network state. However, in multihop sensor
networks, having a global network state at each node is not
a scalable solution, since it involves excessive communica-
tion overhead as the network size grows. To address these
challenges, several previous works have proposed greedy
locally optimal decisions [1], [2] based on 1-hop neighbor
states to make on-demand routing decisions. In our frame-
work, we also enable nodes to greedily optimize their
behavior based on the local state and neighborhood state,
but we adopt a proactive collaboration strategy. Our choice
for a proactive routing strategy stems from the time-driven
nat ure of our target ap plication, so nodes can easily
maintain their neighbors’ states by piggybacking state
information in periodic routing messages, minimizing
overhead communication. This collaboration strategy
strikes a balance between the two seemingly contradicting
goals of providing network-wide performance guarantees
and ensuring adaptability to local sensor node states.
1.2 Optimization Framework
Within the rationale of global optimization through greedy
local decisions, we consider our general sensor network
optimization framework from [6], which adopts a flexible
node state representation consolidating state variables from
several layers of the communication stack. The work in [6]
has applied the framework to three diverse ad hoc and
sensor network application scenarios with different com-
munication technologies. In this paper, we tailor the
framework for a monitoring sensor network application,
as shown in Fig. 1. At the routing layer, the framework
enables nodes to set their neighbors’ routing cost according
to the neighborhood state. The flexible and cross-layer state
representation ensures that routing behavior in the network
adapts to all relevant state parameters.
At the MAC layer, nodes can also adapt their behavior
according to the cross-layer state representation. Radio
power consumption is one of the main causes of power
consumption at sensor nodes. In monitoring networks, idle
listening is the main contributor to radio power consump-
tion. In order to minimize idle listening, each node can
adapt its radio duty cycle according to its current local and
neighborhood states.
1.3 Adaptive Mechanisms
Enabling individual nodes to adjust their routing and MAC
behavior according to their local state requires mechanisms
that are adaptive, flexible, and modular. Fig. 2 shows the
communication mechanisms that drive the optimizations in
this paper. At the network layer, our model proposes a
flexible cost function for routing optimizations, inspired by
the work of Baldi et al. [7]. In this paper, the cost function
metrics include power terms related to the duty cycle of
node radios and sensor activity, although additional QoS
metrics, such as delay or reliability, can be included too.
Minimization of the cost function determines the routing
strategies of the network.
At the MAC layer, to validate our framework, we choose
BMAC [8], a modular and flexible sensor network MAC
protocol which aims at reducing idle listening at sensor
nodes. BMAC provides interfaces that enable services and
applications to set low power listening modes and transmit
modes on a per-packet basis if needed. The creators of
BMAC also suggest that further power savings could be
achieved by using the interfaces to set listening and
2 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 6, NO. 8, AUGUST 2007
Fig. 1. Cross-layer optimization framework.
Fig. 2. Communication mechanisms.

transmit modes according to additional information on the
application and operation of a sensor network.
Building on BMAC, we had proposed in earlier work [9],
[10] a cross-layer mechanism called Adaptive Low Power
Listening (ALPL) to provide a seamless basis for locally
setting the listening mode while ensuring reliable data
delivery. ALPL also ensures consistency for joint optimiza-
tion at the MAC and routing layers. In the original BMAC
protocol, setting a network-wide listening mode disregards
the nonuniform and dynamic local states of individual
nodes. Enabling each node to set its own listening mode
according to its local state is more energy-efficient, but it is
difficult to predict the state of each node prior to deploy-
ment. To address these challenges, ALPL supports the
adaptation of listening modes in BMAC to local sensor node
states, and it enables a node to learn the listening mode of its
neighbors in order to ensure correct data delivery.
Our earlier versions of ALPL have represented node
state as a combination of two aspects: number of descen-
dants in the routing tree and radio duty cycle. A node’s
number of descendants is a network layer aspect that
indicates the node’s forwarding load. For instance, leaf
nodes have no forwarding load. The radio duty cycle
incorporates knowledge about the duty cycles of neighbors
into the decision of optimal listening mode, quantifying
how busy a node has been in the recent time window. The
benefits of the ALPL’s state representations were evaluated
through deployment experiments for a time-driven mon-
itoring sensor network test-bed.
In this paper, we expand the node state representation to
include the role of each node. Certain events or queries may
trigger nodes to assume different roles during deployment,
causing increased resource usage at these nodes. Including
role information into node state representation accounts for
heterogeneous roles and enables nodes to better adapt their
listening mode and routing costs to dynamic network
conditions. To assess the benefits of the expanded state
representation, this article presents four new sensor net-
work test-bed deployment experiments for an event-driven
network application. This article also highlights the cross-
layer design aspects of earlier ALPL optimizations, and it
presents our tailored c ross-layer framework for this
purpose. Finally, this article revisits earlier ALPL results
with a clearer representation and a more involved analysis
of stability issues.
1.4 Overview of the Paper
In sum, the novel contributions in this paper are threefold:
. formalization and tailoring of a flexible optimization
framework for sensor networks,
. expansion of previous ALPL state representations
within the framework and analysis of the resulting
energy benefits, and
. validation of the framework through experiments on
a test-bed of sensor nodes running time-driven and
event-driven applications and evaluation of the
framework’s energy benefits on a global network
basis and on a node-specific basis
The rest of the paper is organized as follows: Section 2
discusses related work, including BMAC. Section 3 presents
our cross-layer optimization framework and its building
blocks. Section 5 presents our experiments to validate the
approach on mica2 motes. Section 6 discusses the results
and concludes the paper.
2RELATED WORK
2.1 Cross-Layer Approaches
Several previous models [3], [4], [5] have adopted a cross-
layer design approach for sensor networks. Most of these
models propose cross-layer optimizations involving specific
mechanisms, such as optimal power control at the physical
layer, link scheduling at the MAC layer, routing flow at the
network layer, and QoS constraints at the transport layer.
Our model generalizes these approaches by adopting a
flexible cross-layer representation of node states which
enables any combination of cross-layer variables, including
application state variables, to affect network behavior.
Woo and Culler [11] propose another cross-layer scheme
that couples transmission control with media access. Their
approach enables nodes to adapt their data origination rate
according to their position within the routing tree in order
to promote fairness among downstream and upstream
flows. ALPL also considers a node’s location within the
routing tree, along with duty cycle and role information, in
order to choose the node’s listening mode that optimizes
local energy consumption.
Another cross-layer proposal by Culler et al. [12] outlines
the guidelines for establishing a sensor network architecture
that enables interoperability among different components.
Their recommendation is to have a Sensor Protocol (SP)
abstraction layer, similar to IP in the Internet, over which all
new sensor network protocols and services could reside.
They also propose cross-layer visibility and management of
several aspects, such as power and security. ALPL takes a
step in this direction by providing higher layer services
with a seamless platform for setting listening modes in
BMAC according to their requirements.
Within a proposed implementation of the SP abstraction
layer, Polastre et al. [13] present an adaptive mechanism
for adjusting the preamble length in BMAC. They propose
the infrequent transmission of packets with long preambles
to enable neighboring nodes to learn each other’s wake-up
schedules. Nodes can then send packets with short
preambles to communicate with their neighbors, thereby
reducing BMAC’s preamble overhead. Broadcast, discov-
ery, and retransmitted packets still use long preambles for
more aggressive transmissions. ALPL [10] independently
proposes setting preambles according to the intended
receiver’s current listening mode, which is determined
by the node’s cross-layer local and neighborhood state
information.
2.2 Local State
A central challenge in sensor networks is that no node has a
global view of the entire network. Romer and Mattern [14]
examine the effectiveness of event-notification mechanisms
in supporting the detection of real events in sensor
networks by communicating the local state changes in
individual nodes. Most event-driven mechanisms are
reactive in nature. Liu et al. [15] realize the need for
JURDAK ET AL.: ADAPTIVE LOW POWER LISTENING FOR WIRELESS SENSOR NETWORKS 3

collaboration and state-sharing among sensor nodes in
order to achieve application requirements. They propose
collaboration among groups of nodes with a flexible
grouping strategy depending on application needs. Here,
we consider a proactive collaboration among sensor nodes
in the same communication neighborhood in the form of
state information sharing within periodic routing messages.
2.3 Cost Optimization
Previous efforts [7], [16] have also realized the need for
optimizing global cost in wireless networks, which deter-
mines the routing strategies of the network. Baldi et al. [7]
develop a global cost function for wireless ultra wideband
radio networks, based on the cost of individual links. Their
cost function is additive and considers transmission power,
link setup cost, interference, delay, and reliability. Mhatre
et al. [16] optimize the hardware cost of heterogeneous
sensor networks with a lifetime constraint. Our framework
focuses on dynamic cost optimization, so it disregards the
hardware cost which is static once the network is deployed.
2.4 Energy-Efficiency
In sensor networks, energy efficiency and load balancing
are the primary metrics of interest. Many protocols have
been designed to provide energy-efficient behavior at both
the MAC layer [17], [19] and the routing layer [21].
2.4.1 Routing Protocols
Sensor network routing protocols are either proactive [22],
[23] or reactive [24], [25]. Proactive protocols maintain
network or neighborhood routing state tables at each node.
Reactive protocols compute optimal routes on-demand.
Proactive protocols are advantageous for networks that
require limited mobility, low latency, or high throughput.
In contrast, reactive protocols are suitable for high mobility,
low throughput, high latency networks. In this paper, we
consider a typical stationary time-driven monitoring sensor
network in which the periodic nature of data transmission
fits well with periodic state exchanges in proactive routing
protocols. As a result, we consider a proactive and periodic
routing protocol in which nodes store the routing state of
their one-hop neighbors.
Although many cost metrics have been proposed for
routing in sensor networks, we focus the discussion here on
metrics for optimizing energy consumption. Previous work
has examined energy optimizing routing strategies, where
cost metrics include residual battery energy [26], [27]. The
residual battery capacity strategy is intuitively valid, but the
discharge of real batteries becomes nonlinear or unpredict-
able at a certain voltage level, which effectively cancels
battery considerations from routing decisions at some point
during the deployment [28]. In our work, we consider the
radio, sensor, and processor duty cycles in the recent time
window as the energy metric for routing. We believe that
this metric is more expressive of each sensor node’s energy
profile since it is independent of the battery technology and
it does not rely on unpredictable battery discharge models.
2.4.2 MAC Protocols
At the MAC layer, idle listening constitutes a large portion
of power consumption because data is sent infrequently.
This effect is even more pronounced in monitoring sensor
networks [29]. Thus , energy-efficien t MAC protocol
proposals have focused on minimizing idle listening at
sensor nodes.
IEEE 802.15.4 [18] is a standard with physical and
MAC layer specifications for low rate, low power, short
range networks, including sensor networks. IEEE 802.15.4
specifies a plethora of functionality choices, many of
which may never be used. As a result, several researchers
have proposed new MAC protocols on top of the 802.15.4
PHY layer.
SMAC [30] is a heavyweight MAC protocol for sensor
networks that relies on time synchronization and schedul-
ing among nodes to enforce periodic sleep and listen
schedules. SMAC reduces energy consumption and pro-
vides scalability at the cost of per-hop fairness, throughput,
and latency. A more recent version of SMAC [32] has
introduced adaptive duty cycles by enabling a node to
snoop on neighbors’ RTS and CTS messages in order to
schedule its own wake up time. Because nodes have to
maintain neighbors’ schedules, SMAC is not sufficiently
scalable for large-scale networks of resource-limited nodes.
T-MA C [31] also proposes adap tive duty cycle s to
address the nonuniform traffic patterns in sensor networks.
T-MAC is similar to SMAC in its essence, but it introduces
early sleeping to enable nodes that are scheduled to be
active to go into sleep mode if they are idle. T-MAC suffers
from similar complexity and scaling problems as S-MAC
because it trades off a short active time for reduced
adaptivity to changing network conditions.
El-Hoiydi et al. propose Wireless Sensor MAC (WiSe-
MAC) [20], an asynchronous MAC protocol that relies on
the preamble sampling technique. Sensor nodes periodi-
cally sample the channel for activity. Packets are sent with
preambles of equal duration to the channel sampling period
to ensure that the receiver will detect the packets. Although
WiSeMAC nodes wake up with the same period, the
individual node offsets are independent. To address the
inherent trade-off between listening power consumption
(for period channel sampling) and transmission power (due
to preamble size), WiSeMAC proposes a mechanism for a
node to learn the sampling schedule of its neighbors in
order to send packets with short preambles during the
intended receiver’s wake-up time.
The recent similar work by Polastre et al. independently
proposes BMAC [8], an asynchronous and lightweight
sensor network MAC protocol that aims at providing
versatile medium access while keeping the MAC function-
ality as simple as possible. Because it is an asynchronous
protocol, BMAC eliminates the communication and pro-
cessing overhead for schedul ing and synchronization,
which reduces energy consumption. BMAC enables each
node to wake up periodically to check for channel activity.
The wake-up period is referred to as the check interval.
BMAC defines eight check intervals, and each check
interval corresponds to one of BMAC’s eight listening
modes. To ensure that all packets are heard by the nodes,
packets are sent with a preamble whose reception time is
longer than the check interval. BMAC therefore defines
eight different preamble lengths, referred to as transmit
4 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 6, NO. 8, AUGUST 2007

modes. Additionally, Polastre et al. analytically derive
optimal listening modes based on the number of neighbors
of a n ode. In their experiments, they determine the
maximum neighborhood size in the network, and they
set the optimal listening mode for that neighborhood size.
The experimental results yield significant energy savings
for BMAC over previous protocols, such as SMAC. BMAC
is also the standard MAC protocol in the communication
stack of TinyOS [35], the primary sensor network research
platform.
As in T-MAC and SMAC, our paper addresses the
nonuniform node states by adapting listening modes in
BMAC to enable adaptive duty cycles. We choose BMAC
because of its flexibility and scalability, which are two of our
main design goals. Instead of proposing a new MAC pro-
tocol, this paper’s focus is designing a mechanism (namely,
ALPL) that enables each node to adapt its duty cycle based
on flexible cross-layer node state representations.
3CROSS-LAYER OPTIMIZATION FRAMEWORK
The algorithm that characterizes our framework periodi-
cally runs the same three basic steps for any network
scenario: 1) gather neighborhood state information, 2) per-
form local calculations on gathered state, and 3) modify
local configuration accordingly.
The details of the algorithm and its implementation are
highly dependent on the network scenario. As such, this
section discusses the details of the algorithm tailored for a
monitoring sensor network application. We begin by
introducing the motivation for ALPL within the framework.
Next, we describe the node interaction that enables nodes to
set listening modes through ALPL and to exchange their
state information. We subsequently describe the three state
representations that are considered in this paper. Next, we
explain how each node uses its neighborhood state
information to calculate the routing cost of each neighbor
through the cost function. Finally, we present the routing
issues to ensure correct data delivery and route stability
with ALPL.
3.1 Adaptive Low Power Listening
Adaptive Low Power Listening [9], [10] is a cross-layer
mechanism that adapts the listening mode at each node
according to its local state, while ensuring correct data
delivery.
The motivation for ALPL is to address the unpredictable
and dynamic node states in sensor networks which are
affected by factors such as interference variations and
dynamic node membership. Network designers have to
make conservative assumptions in determining the network
configuration. In the case of energy optimization, conserva-
tive assumptions lead to setting a network-wide listening
mode in BMAC prior to network deployment. This causes
unnecessary idle listening to occur in less active portions of
the network. ALPL’s purpose is to reduce idle listening in
BMAC by allowing each node to set its own listening mode
depending on its local node and neighborhood states. The
rationale is that, in dynamic sensor networks, each node
always has the most up-to-date view of its own local state
[34]. Node states can be defined by the network designer or
operator depending on the applications’ goals and quality
of service requirements.
3.2 Node Collaboration
The collaboration strategy in our framework enables a node
to learn its neighborhood state information, to set its own
listening mode accordingly, and to adapt its transmit mode
to fit the listening mode of its routing parent. We note here
that ALPL requires no synchronization among node clocks
since it builds on BMAC listening modes. The only
requirement for correct communication is proper matching
of the sender’s preamble length and the receiver’s listening
mode. Our deployment experiments in Section 5 evaluate the
soundness of this approach through observation of long-
term data delivery.
Fig. 3 illustrates the node interaction to enable nodes to
set their listening mode adaptively. We assume a proactive
routing protocol in which nodes periodically send routing
update messages to declare their routing information to
their neighbors.
Initially, nodes are unaware of their neighborhood state,
so all nodes listen at an initial listening mode L
init
and use
the corresponding transmit mode T
init
, both of which are
known a priori to all nodes. Each node begins sending
periodic route update messages to declare its presence and
state. Once nodes learn of their neighbors’ presence, a
routing graph is formed and data flows toward the base
station (Fig. 3a). As a result, each node learns the state of its
direct neighbors. Before sending the next route update
message, a node A first sets the optimal listening mode L
A
for its local state. Then, node A sends a routing update
message that includes its new listening mode and state
information along with other routing information (Fig. 3b).
All of A’s neighbors hear the routing update message,
and they learn A’s current listening mode L
A
and A’s state
information. Each neighbor of A records A’s listening
mode and state information in its local neighbor table.
JURDAK ET AL.: ADAPTIVE LOW POWER LISTENING FOR WIRELESS SENSOR NETWORKS 5
Fig. 3. (a) Nodes form their routing tree. (b) Each node periodically
announces its current state, including its listening mode and battery
state, enabling neighbors to use the appropriate transmit mode. (c) A
high duty cycle at A causes neighbors to increase A’s routing cost and
choose a new parent B. Node A listens less often as it has fewer
packets to forward.

Consequently, each node in the network always has up-to-
date information on the state of its neighbors. Whenever a
node D chooses A as a routing parent, it simply checks its
neighbor table for A’s listening mode L
A
. D then sends its
data packets using the transmit mode T
A
that matches L
A
.
Similarly, nodes that receive a routing update message
from their current parent indicating that the parent has a
new listening mode adapt their transmit mode accordingly.
For a particular node in ALPL, the only concern for data
delivery is whether the routing parent is reachable. If so,
then the node can transmit with the proper preamble
length. Otherwise, the node detects the unreachability of the
routing parent through missed route update packets and
attempts to find another suitable routing parent. If the node
hears a routing update packet from the original routing
parent at some later poi nt, the node ca n rerun the
algorithms for choosing the routing parent and the
appropriate listening mode.
3.3 State Representations
Our state representations consist of three aspects: 1) number
of descendants, 2) duty cycle, and 3) role. In this section, we
describe each of these three aspects in detail.
3.3.1 Number of Descendants
Our first representation of state considers a node’s number
of descendants in the routing tree, which has significant
impact on a node’s energy profile in data gathering and
monitoring applications. In monitoring applications, data
flow is typically toward a single data sink. The basic
requirement for correct data delivery is for each node to
listen often enough to hear all the packets that it must
forward toward the data sink. The number of packets that a
node forwards depends on the number of its descendants in
the routing tree.
In order to determine its optimal listening mode, each
node N learns how many descendants it has in the routing
tree by counting the number of packets  that it forwards
during a route update interval. The number  indicates how
busy N was during the last interval. When it is time to send
the next routing update message, N first sets its listening
mode to the optimal listening mode L
N
for a traffic load of
 packets. Then, N sends a routing update message that
contains L
N
along with other routing information.
Including the number of descendants into ALPL’s state
representation optimizes the overall power consumption in
the network. Each node optimizes its local power con-
sumption through the selection of an optimal listening for
its current forwarding load. The distributed optimization of
local node power consumption leads to an optimization of
network power consumption. The number of descendants
only exploits the load imbalance in the network to
opportunistically optimize power consumption at lightly
loaded nodes. The next two sections present additional
node state aspects that promote load balancing.
3.3.2 Duty Cycle
Our second state representation combines the number of
descendants and the node’s duty cycle. A high radio duty
cycle indicates that the node’s radio has been highly active
up till the present point in time and vice versa. Given the
inherently nonuniform energy consumption in sensor net-
works, enabling nodes to adapt their behavior according to
their radio duty cycle can help balance the power
consumption in the network.
Sharing duty cycle information among nodes involves
minimal overhead communication. Each node can piggy-
back its duty c ycle value within its routing update
messages in order to declare its power state to neighbors.
As a result, nodes learn the radio duty cycles of all their
one-hop neighbors and store this information in their local
neighbor table.
Including the radio duty cycle into ALPL’s state
representation enables local load balancing of network
traffic. Each node can consider and compare the recent
activity level of neighbor transceivers. Neighbors with a
relatively higher radio duty cycle are penalized through
higher routing cost, which diverts traffic away from busier
nodes to other nodes to achieve more uniform energy
consumption. The radio duty cycle promotes load balan-
cing to the extent allowed by the routing graph. As long
as the routing graph provides a node with more than one
alternative for a routing parent, the radio duty cycle
information contributes favorably to load balancing.
As an example, consider Fig. 3 again. The routing tree in
Fig. 3a puts most of the forwarding burden on node A.Asa
result, A depletes its battery resources quicker than node B.
In order to shift its forwarding load, A declares its high
duty cycle to its neighbors, causing neighbors to increase
A’s routing cost. This in turn causes most of A’s current
childrentochooseanotherparent whenever possible
(Fig. 3c). Having diverted most of its forwarding load to
node B, node A begins listening with a longer check
interval to reduce its listening power consumption.
3.3.3 Role
The nonuniform state of sensor nodes also stems from the
roles that the nodes may take during deployment. The
number of descendants is a routing aspect and it represents
the present state of the node. The duty cycle is a physical
and MAC layer aspect that represents the past state of the
node. The node’s role provides information on application
functionality and represents its projected state into the
future. In the context of our energy optimization study, the
node’s role can play a part in determining the node’s
optimal listening mode and neighbors’ routing costs. The
rationale is to incorporate additional knowledge about the
expected power profile of a node’s role in optimal local
decisions.
In order to integrate role into network decisions, nodes
should share their role status with thei r neighbors.
Sharing role information among nodes involves minimal
overhead communication. For example, in a network
where nodes may take one of two roles, each node can
include its role status within its routing update messages
through a single bit. As a result, nodes learn the roles of
all their one-hop neighbors and store this information in
their local neighbor table.
As in the case of radio duty cycles, including node roles
into ALPL’s state representation promotes local load
balancing of network traffic. Each node can consider and
compare the recent sensing activity level of neighboring
6 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 6, NO. 8, AUGUST 2007

nodes. Thus, a node can divert traffic away from nodes
with high sensing activi ty, which contribute s to load
balancin g. In homogeneous networks where all nodes
behave the same, the role variable does not contribute to
load balancing. However, in event or query-driven net-
works, certain nodes may assume more active sensing roles
during the deployment. In these cases, the power con-
sumption for sensing becomes more significant and it
should be considered for load balancing considerations.
Consider the network in Fig. 4. While nodes are
periodically sending their data with period T , much of
the forwarding load is focused at node L. At some point
during the deployment, an event E occurs at the lower
periphery of the network. Nodes K and L detect event E,
switch their role to tracker, and begin sensing and reporting
data at T=2. The neighbors of K and L learn that these two
nodes are now tracker nodes and thereby increase the
routing cost of these two nodes. Fig. 4b shows the resulting
topology. Nodes K and H choose a new parent, node J,in
order to avoid using the tracker node L as a forwarder. The
topology in Fig. 4b puts most of the forwarding load on
node J. The energy consumption resulting from the high
forwarding load of node J is balanced by the energy
consumption for more frequent sensing and data reporting
at node L.
3.4 Cost Function
Our primary platform for sensor network development is
TinyOS [35], developed at UC Berkeley. Within TinyOS, the
standard routing protocol is called MintRoute [36]. Mint-
Route is a proactive routing protocol in which nodes send
periodic routing messages to declare their local states to
their 1-hop neighbors.
In all proactive routing protocols and particularly in
MintRoute, each node periodically selects its routing parent.
A node N first selects the highest contender M in its
neighbor table with the least routing cost at the current
time. Next, N compares the cost of M with the cost of the
current routing parent RP . N chooses M as its new parent
only if
C
mint
ðMÞþ<C
mint
ðRPÞ; ð1Þ
where C
mint
ðN
j
Þ is the routing cost of node N
j
and  is the
switching threshold that ensures that a node switches its
routing parent only when there is an appreciable benefit in
doing so.
The original cost metric in MintRoute is the Expected
Transmission Count (ETX), which assigns the link weight as
the product of the reciprocal of the forward and backward
link qualities. Mintroute then uses a shortest path algorithm
for determining the best route to a destination. Nodes that
employ the ETX routing metric snoop on the neighborhood
links to keep track of the forward and reverse delivery ratio
on each link. The ETX metric in MintRoute implicity
minimizes radio power consumption. Since our framework
aims at making routing decisions dependent on the node
energy states, we introduce additional cost metrics that are
direct functions of the aspects that constitute node state.
3.4.1 Duty Cycle Awareness
The number of descendants aspect is considered for setting
local listening modes rather than setting neighbor costs, so
this aspect does not require an explicit routing cost metric.
For the duty cycle aspect, we introduce the cost metric
CðradioÞ, which denotes how busy a particular neighbor’s
radio is relative to other neighboring nodes. Highly loaded
nodes should have a higher routing cost. Therefore, we
express CðradioÞ for a neighbor N
j
as
CðradioÞ¼

j

P
k
i¼0

i
=k
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P
k
i¼0

2
i
=k 
P
k
i¼0

i
=k

2
r
; ð2Þ
where k is the number of neighbors in the node’s local
neighbor table and 
i
is the radio duty cycle of neighbor N
i
.
Equation (2) compares the radio duty cycle of neighbor N
j
to the average duty cycle in the neighborhood during the
recent time window and normalizes the difference. The
normalized difference determines the extent of statistical
deviation of the radio activity of N
j
among all the nodes in
the neighborhood.
3.4.2 Role Awareness
The duty cycle is a sufficient indicator of energy profiles in
cases where the nodes all have the same sensing and
processing patterns throughout the lifetime of the network.
In many event-driven or demand-driven sensor networks,
some nodes change their application behavior, such as the
sensing frequency or buffering strategy, based on events in
the network. Such cases require the inclusion of information
on the current roles of nodes in routing decisions.
The actual cost dependance on the node’s role is highly
application-specific. In this paper, we consider an event-
driven network that specifies that nodes can have one of
two sensing frequencies, f
1
or f
2
, where f
2
is double f
1
.A
JURDAK ET AL.: ADAPTIVE LOW POWER LISTENING FOR WIRELESS SENSOR NETWORKS 7
Fig. 4. An event-driven tracking sensor network.

node uses the following equation to evaluate the sensing
cost of a neighbor N
j
:
CðsensingÞ¼


j

P
k
i¼0

i
=k
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P
k
i¼0

2
i
=k
P
k
i¼0

i
=k

2
q

j
¼ 1
0 
j
¼ 0;
8
>
<
>
:
ð3Þ
where k is the number of neighbors in the node’s local table.

i
is a Boolean variable that takes the value of 1 if N
i
is
sensing data at frequency f
2
(tracker node) and takes the
value of 0 if N
i
is sensing data at frequency f
1
. Equation (3)
compares the sensing activity of neighbor N
j
to the average
sensing activity in the neighborhood during the recent time
window and normalizes the difference. The normalized
difference determines the extent of statistical deviation of
the sensing activity of N
j
among all the nodes in the
neighborhood.
3.4.3 Overall Cost Metric
The overall routing cost should strike a balance between the
need for correct and timely data delivery on one hand and
uniform energy cost on the other. Furthermore, CðradioÞ
and CðsensingÞ should play a role in determining the
routing parent only if the node M is at the same or at a
lower level than the current routing parent RP.
Therefore, the new overall cost of a neighbor includes the
original MintRoute routing metric, as well as the power cost
according to the following equation:
CðMÞ¼
C
mint
ðMÞþCðradioÞþC ðsensing Þ HðMÞHðRP Þ
C
mint
ðMÞ HðMÞ >HðRPÞ;

ð4Þ
where  and  are constants representing the weights of
CðradioÞ and CðsensingÞ, respectively. HðNÞ indicates the
hop count of node N from the base station. With the new
cost definition, neighbors with higher power consumption
(due either to radio duty cycle or to increased sensing
activity) have a higher routing cost, and they are less likely
to be chosen as forwarders.
3.4.4 Network Cost
The network cost in our framework is the sum of the costs
of individual nodes [7]:
NC ¼
X
N
CðNÞ: ð5Þ
Since optimizing NC is not feasible, each node selects its
routing parent as the node with the least routing cost CðNÞ.
3.5 Routing Issues
Enabling each node to set its own listening mode adapts
MAC protocol behavior to the node’s state. The MAC
protocol adaptation should be accompanied with routing
protocol adaptation in order to maximize performance
gains.
The concept of adaptive listening modes raises the
possibility that some nodes may not hear the packets sent
by their neighbors because of mismatched preamble lengths
and check intervals. For example, if a node A sends a packet
with a short preamble to its parent B, one of node A’s
neighbors C that is listening infrequently may miss node
A’s packet. This situation does not affect data delivery,
since it is only necessary for A’s parent B to hear the packet.
Missing a routing update packet is more detrimental, since
routing packets hold important information on neighbor-
hood routing state changes.
We implement modifications to MintRoute to address
missed routing update packets. A central issue in designing
ALPL is to ensure that asymmetric listening modes do not
affect maintaining an up-to-date neighborhood view at each
node. Achieving this goal requires that nodes always hear
the routing update packets of their neighbors. Thus, ALPL
specifies that nodes always send their routing update
packets with the longest preamble, so that a neighbor in
any listening mode can hear the update packets. Secondly,
because MintRoute determines link quality by snooping on
forward and backward links, we modify MintRoute so that
nodes only snoop on periodic routing updates instead of
data packets to determine the link quality to their
neighbors. Monitoring route update packets for determin-
ing link quality ensures that asymmetric listening modes at
neighboring nodes have no detrimental effect on link
quality, because all routing update packets are sent with
the longest preamble.
The current implementation of ALPL piggybacks state
information on routing update messages, which closely
couples the period of the state information updates with
routing updates. We view this coupling as favorable, since
routing protocols that send frequent updates target highly
dynamic networks, where frequent state information up-
dates are also required. In our current implementation of
ALPL, the default routing update period is set to half of the
data sampling period.
The longer preambles used for routing update messages
represent the main cause of energy overhead of ALPL. For
applications with data sampling periods in the order of
minutes, the control overhead is relatively low as routing
update messages with long preambles are sent infre-
quently. The increase in control overhead is counter-
balanced by the reduction of idle listening power, since it
is the idle listening power consumption that dominates the
overall power consumption at a node in most monitoring
applications.
Because state information is sent infrequently, nodes
may transiently have stale state information about the
listening mode of their parents due to missed route update
packets. If the parent is still alive and within range, the stale
information persists until the time of the next route update
message (which is two data sampling periods in our current
implementation). During that time, the mismatch between
the node’s preamble length and the parent’s listening mode
affects data delivery only if the parent has switched to a less
frequent listening mode. In this case, the child node does
not receive acknowledgements for its data packet and it
realizes that the parent node has not received the packet. If
the child no de sends the data packet aga in with no
acknowledgement, it tries to send the packet with the
8 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 6, NO. 8, AUGUST 2007

longest preamble in order to aggressively reach the parent
node [13], [20]. If there is still no acknowledgment after the
third attempt, then the child node gives up and looks for
another routing parent within its routing table.
Two factors determine the stability of the new cost
metric: 1) the switching threshold  and 2) the time-
averaging of the cost components. Suppose that a node
has two potential routing parents and that the overall
routing cost of the two parents is close in value. The node
does not switch its current parent until the other candidate’s
routing cost is less than the current parent’s cost by at least
, whose value can be tuned for different scenarios. Another
issue is route flaps. If CðpowerÞ and CðroleÞ are computed
for a short time window into the past, then one might
expect a node to keep switching between the two potential
parents. Once it switches to a new parent, the overall cost of
the other original parent drops and the node switches back
to the original parent during the next period. However,
setting a sufficiently large time window into the past,
during which CðpowerÞ and CðroleÞ are computed, ensures
that the overall cost metric reflects the aggregated energy
consumption of a node so far and avoids dependence of
routing decisions on transient changes. The optimal value
of  and the optimal length of the time window over which
the values are averaged are dependent on how dynamic the
target network scenario is, and they remain open issues for
further research.
4QUALITATIVE ANALYSIS
This section presents the analytical basis for using greedy
local decisions to reduce global network power consump-
tion. We assume that the sensor nodes collect sensor data
and transmit the data in a packet once in every period T .
The following equation governs power consumption E at a
sensor node [8]:
E ¼ E
t
þ E
r
þ E
d
þ E
listen
þ E
sleep
; ð6Þ
where E
t
is the power spent on transmissions during time
T, E
r
is the power for packet reception during time T , E
d
is
the power required to collect sensor values, E
listen
is the
power consumed for checking the channel for activity, and
E
sleep
is the power consumed while the node is asleep. The
quantitative expressions for each energy component are
given in [8]. We limit the discussion here to the qualitative
aspects that are relevant to ALPL.
In monitoring and data gathering applications, the
sampling period T is typically in the order of minutes.
Therefore, each node collects sensor data, transmits packets,
and receives packets once every few minutes. On the other
hand, nodes wake up to monitor the channel for activity
much more frequently, for instance, once every several
milliseconds. Thus, idle listening on the channel has a
profound effect on the overall power consumption, so
reducing idle listening yields significant power savings.
4.1 Topology
The aim of the number of descendants metric is to minimize
the overall power consumption in the network through
local optimization decisions. Each node can optimize its
own power consumption E locally by selecting its own
optimal liste ning mode while maintain ing correct and
timely data delivery. Minimizing E on a per-node basis
reduces the overall network power consumption and builds
on the following observations about (6):
1. E
d
and E
sleep
are not significant factors in determin-
ing optimal listening mode in a homogeneous
monitoring network. E
d
is equal for all nodes. E
sleep
is at least an order of magnitude smaller than the
other terms in (6), so it has a negligible effect on E.
2. E
t
and E
r
depend on the node’s position in the
logical topology. If a node is a leaf in the routing tree,
it has fewer packets to forward.
3. The listening mode at node N determines E
listen
.It
also determines the preamble length for packets that
are received at N. Consequently, the listening mode
at N affects E
r
at N and E
t
at N’s children.
For example, a more frequent listening mode at N
increases E
listen
and decreases E
r
at N. E
listen
increases
because N wakes up more frequently to check for channel
activity, and E
r
decreases because the frequent listening
enables N’s neighbors to send their packets to N with
shorter preambles, thereby reducing packet reception time
at N.
Similarly, the listening mode of N affects E
t
at the
children c
i
of N. A more frequent listening mode at N
enables c
i
to send its packets with shorter preambles, thus
reducing E
t
i
. Through similar reasoning, less frequent
listening at N forces c
i
to send packets with long preamble
and consume more power for packet transmissions.
These dependencies further support the need for setting
per-node listening modes. In practice, each node locally
computes E
t
, E
d
, and E
sleep
and then selects the listening
mode that provides the combination of E
listen
and E
r
that
yields the lowest power consumption E .
Our cross-layer optimization framework builds on the
need to reduce idle listening and to balance power
consumption. The framework most basic state representa-
tion only considers number of descendants in the routing
tree, which yield s energy savin gs at nod es with f ew
descendants. We now consider a case study analysis to
illustrate the benefits, scalability, and shortcomings of this
most basic form of ALPL.
4.2 Case Study
To analyze the trade-offs involved in ALPL and to compare
the power consumption of ALPL and BMAC, we consider a
case study of a static 127-node network with a binary tree
topology. Although the topology of an actual sensor
network can be both irregular and transient according to
environmental conditions as well as location, this case study
serves the purpose of validating the analytical basis and the
scalability of ALPL. In the next section, we perform
experiments with actual sensor nodes to further validate
our model in a more dynamic and realistic scenario. Our
analysis for this network assumes that the sensor nodes
sense the environment and send their data to the base
station once every 3 minutes.
In their evaluation of BMAC, Polastre et al. use the
following method for assigning a listening mode that favors
the busiest node to improve network lifetime:
JURDAK ET AL.: ADAPTIVE LOW POWER LISTENING FOR WIRELESS SENSOR NETWORKS 9

. compute the expected number of descendants of the
busiest node in the network and
. set the network-wide listening mode to favor the
busiest node.
Since the busiest node in the network has the largest
forwarding load, this method for setting the listening mode
typically means that all the nodes in the network use the
higher duty cycle setting which suits the busy node. This
gives rise to one of the motivations of ALPL: to enable
nodes that are not as busy to choose their appropriate duty
cycle in a decentralized manner.
Table 1 compares the check interval between network-
wide listening modes (plain BMAC) and ALPL at each level
in the 127 node binary tree network. In plain BMAC, the
check interval is set to 10 ms for all the nodes in order to
accommodate the forwarding load of the busiest node. In
ALPL, the busiest node (the node at level 1) selects the same
listening mode as BMAC nodes, and the remaining nodes
select their own listening modes based on their topology
position. The following example illustrates the listening
mode selection process in ALPL.
We consider how a level 2 node in the 127-node binary
tree selects its optimal check interval in ALPL. A level 2
node sends 63 data packets each update period, where one
packet is locally generated and the rest are forwarded
packets from its descendants. The level 2 node also uses a
preamble length of 28 bytes in order to match its parent’s
check interval of 10 ms. The resulting transmission power
E
t
in ALPL for a level 2 node is 0.5321 mW. Using the fact
that it receives 62 data packets from its descendants each
update period, the level 2 node then computes its reception
power E
r
for each of the eight possible check intervals. It
also computes E
listen
and E
sleep
for each check interval.
Finally, the level 2 node selects a check interval of 20 ms,
with a reception power E
r
of 0.8617 mW and a listening
power E
listen
of 0.865 mW, as the optimal check interval that
yields the minimum overall power consumption E of
2.5 mW at the level 2 node.
All other nodes in ALPL use the same process to select
the optimal listening mode. As mentioned earlier, the level 1
node selects a check interval of 10 ms as in the BMAC case
because it is the busiest node in the network. Nodes at
level 2 and at higher levels choose less frequent listening
modes because they have a smaller forwarding load than
the level 1 node.
We note here that, as the network size grows, more
nodes at lower levels of the tree converge to the generic
BMAC behavior. For example, if we double the network
size of this case study to 256 nodes, nodes at level 2 choose
their optimal check interval as 10 ms. Doubling the network
size further causes level 3 nodes to select the highest duty
cycle, and so on.
Fig. 5 compares the sources of energy consumption for
the two cases. E
d
is omitted from Fig. 5a because it does not
vary with listening mode, and E
sleep
is omitted because it
has a negligible impact on overall energy consumption. The
plots for the listening power consumption in Fig. 5 illustrate
the energy savings of choosing per-node check intervals
(See Table 1). We observe that ALPL saves more on idle
listening power for nodes at higher levels in the tree
because these nodes have fewer packets to forward and
they can listen to the channel less often. For the level 1 node,
the idle listening power consumption is the same for a
network-wide listening mode and ALPL because these
nodes must listen often enough to accommodate their high
forwarding load.
ALPL saves on idle listening power consumption at the
cost of increased transmission and reception power, which
is an inherent trade-off of the underlying BMAC protocol
[8]. The increase in reception and transmission power stems
from the use of longer check intervals, which requires long
preambles for packets. Note that the level 1 node has the
same transmission and reception power in both cases
because this node receives and sends packets with the
same preamble.
Fig. 6 compares the overall power consumption for
BMAC and ALPL on the basis of node levels. For ALPL,
power consumption follows a similar trend as the idle
listening power consumption in Fig. 5. For plain BMAC, the
overall power consumption follows the trend of E
r
, mainly
because E
listen
is the same for all nodes. Nodes at higher
layers exhibit more power savings with ALPL because their
low forwarding load enables them to sleep more often.
The lifetime of the network is constrained by the power
consumption of level 1 nodes [16], [38]. At first glance,
Fig. 6 seems to indicate that the basic form of ALPL does
not extend the lifetime of the network despite significant
energy savings at all but one level in the tree. However,
10 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 6, NO. 8, AUGUST 2007
TABLE 1
A Comparison of the Listening Modes
at Each Level in the Network
Fig. 5. A comparison of BMAC and ALPL for listening, transmission, and
reception power consumption in a sensor network with a binary tree
topology.

recall that sensor networks have a dynamic topology.
Varying interference conditions and moving objects typi-
cally cause changes in the topology, causing nodes to
choose new parents. For instance, a level 2 node may
choose the base station as its parent at some point,
effectively assuming the role of a level 1 node. Similarly,
nodes may choose a parent at a lower level at some point in
order to avoid low quality links.
The expanded state representations in our framework
yield further energy savings at the most loaded nodes,
which contributes to prolonging network lifetime. By
having more detailed information about neighbors’ states,
nodes can more expressively select neighbors’ routing costs.
As a result, forwarding traffic moves away from loaded
nodes and energy consumption becomes more balanced.
The next two subsections describe the interaction of duty
cycle and role on power consumption in our framework.
4.3 Duty Cycle
The energy consumption in sensor networks is nonuniform
among the nodes due to communication asymmetries.
First, data always flows to one or a few sinks. The nodes
that are one hop away from a data sink are called critical
nodes [16]. Critical nodes have a larger forwarding burden
and consume more energy than nodes further away from
the sink [38]. These factors indicate that nodes can use up
battery resources at different rates. ALPL can contribute to
balancing the power consumption rates among network
nodes by manipulating d ata forwarding patterns and
listening modes. Of course, the degree of achievable load
balancing depends on the node density and the layout of
the nodes.
In terms of (6), highly loaded critical nodes have larger
E
t
and E
r
. Because ALPL sets the listening mode according
topology information, a loaded critical node will also
choose to listen frequently to the channel, so it has a high
E
listen
. As a result, the energy consumption E and the duty
cycle at a loaded critical node are higher than that of its
neighbors. Through our framework, the loaded critical node
informs its neighbors of its busy state. As a result, the
loaded node’s children increase its routing cost and choose
new parents to forward their packets, thereby reducing its
E
r
and E
t
at the loaded critical node. The loaded critical
node also switches to a listening mode with a longer check
interval to reduce E
listen
. The reductions in E
r
, E
t
, and
E
listen
yield a significant decrease in E at the critical node.
4.4 Role
In applications where n odes may dynamically select
different roles, the sensing power E
d
may become a
significant player in determining the node’s power con-
sumption relative to its neighbors. For instance, suppose the
occurrence of a particular event causes some nodes in the
network to double their sampling frequency in order to
better track the event. The increased sampling frequency at
these nodes increases the sensing power consumption E
d
and it also increases E
t
because the node sends packets
more frequently to report the more frequent sensed data.
The increase in E
t
falls within the radio duty cycle value so
it does not cause any distortion for energy balancing. The
more appreciable increase in E
d
has no explicit effect on the
radio duty cycle at the node, but it can cause energy
imbalances in the network. Including the power consump-
tion E
d
due to the node’s role remedies this problem and
maintains a balance in overall power c onsumption E
among the nodes.
5DEPLOYMENT RESULTS
In this section, we investigate the impact of local state-
driven optimizations on a testbed of sensor nodes deployed
in our laboratory. The sensor nodes in our experiments
consist of 14 mica2 motes from Crossbow [39]. Our
implementation of ALPL is in NesC [40], a component-
oriented variant of C customized for networked embedded
systems and built into TinyOS.
The nodes are placed at random positions in the
laboratory and the base station is placed near one of the
walls of the room. We reduce the transmit power of nodes
to limit their radio range, enabling multihop communica-
tion. The aim of the experiments is twofold: 1) to assess the
effect of state-driven optimizations on the global network
power consumption and 2) to evaluate the local node power
consumption and energy balancing benefits for state-driven
optimizations.
We adopt the method suggested by [8] for computing
power consumption. The underlying BMAC design in-
cludes several radio states, including active and sleep states.
The average current draw for each radio state is fixed. The
method employs a counter that keeps track of the time that
the radio spends in each power state. Having the current
draw and time spent in each state, each node can
continually compute its aggregated power consumption
so far. The power consumption results shown for our
deployments represent the aggregated power consumption
for the entire duration of the deployment, including all
overhead communication for maintaining routing graphs.
To normalize the results, we assign the data point with the
highest aggregated power consumption a value of one, and
JURDAK ET AL.: ADAPTIVE LOW POWER LISTENING FOR WIRELESS SENSOR NETWORKS 11
Fig. 6. A comparison of BMAC and ALPL for overall power consumption
in a sensor network with a binary tree topology.

we assign the remaining values for that comp arison
according to that data point.
5.1 Time-Driven Sensor Network
The first experiment set considers a time-driven monitoring
sensor network with a single data sink. We conduct three
experiments for the same physical network topology. In the
first experiment, we initially determine the listening mode
for the busiest node in the network, and we assign that
listening mode in BMAC to all the nodes. In the second
experiment, each node runs ALPL and sets its listening
mode according to its number of descendants in the routing
tree. The third experiment expands the view of local state to
include the duty cycle, so nodes run ALPL and select
listening modes according to the expanded local state. We
refer to this case as Energy Aware ALPL (EA-ALPL).
Each experiment lasts for 43 hours. In all experiments,
nodes run the Surge application that is available with the
standard distribution of TinyOS. In Surge, nodes sample the
sensors and send the data once every minute. Thus, each
node sends 2,580 data packets during each experiment. The
routing update period for the network-wide listening mode
experiment is 120 seconds. For both ALPL and EA-ALPL
experiments, the routing update period is 90 seconds, to
allow more adaptive link qualities based on routing update
messages. The radio duty cycle weight  is set to 2 for this
experiment set.
5.1.1 Delay and Pack et Delivery
We begin by examining the effect of applying ALPL on the
packet delivery time to the base station. Fig. 7 compares the
interpacket delays arriving at the base station in each of the
three experiments. For the majority of packets in all
experiments, interpacket delivery time follows a uniform
distribution centered at 60 seconds. The deviation of a few
seconds for a small number of packets is attributed to clock
skews in the motes. It is worth noting the local peaks at
0 seconds and at constant multiples of the update period.
An interpacket delivery time close to 0 seconds indicates a
packet retransmission. The number of packet retransmis-
sions is about the same for the three experiments. This
result confirms that ALPL does not increase the number of
retransmissions over BMAC. The other local peaks repre-
sent missed packets. An interpacket arrival time of
120 seconds from a particular node to the base station
indicates a single missed packet from that node. Through
the same logic, an interpacket arrival time of AT seconds
indicates ðAT  60Þ=60 consecutive missed packets. The
number of missed packets for the three experiments are of
the same order, indicating that ALPL does not cause
additional packet losses over BMAC. In sum, the results in
Fig. 7 have shown that the introduction of ALPL does not
affect delay or packet delivery in the network when
compared to BMAC.
5.1.2 Overall Power Consumption
The average data yield for BMAC, ALPL, and EA-ALPL is
about 98.5 percent. Because the data yield is the same with
or without state-driven optimizations, we focus our analysis
here on power consumption issues.
Fig. 8a plots the average global network energy con-
sumption for BMAC, ALPL, and EA-ALPL. Both ALPL and
EA-ALPL reduce global energy consumption by about
35 percent on average during the deployment. The
reduction in global energy consumption stems from the
optimal local decisions at each node. In particular, the
12 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 6, NO. 8, AUGUST 2007
Fig. 7. A comparison of the distribution of interpacket delays arriving at
the base station.
Fig. 8. (a) Global network power consumption during deployment. (b) Average check intervals at each node. The error bars indicate the 99 percent
confidence interval.

number of descendants variable is the main contributor for
the reductions in energy consumption, as the next subsec-
tion reveals in more detail. This reduction also confirms the
benefits of greed y local decisions in reducing overall
network power consumption.
The global energy consumption for ALPL and EA-ALPL
experiments is almost the same with a slight difference of
1 percent. The main distinction between ALPL and EA-
ALPL is in the distribution of the power consumption
among network nodes rather than in the aggregated power
consumption. The remainder of this subsection explores the
local power consumption at individual nodes, which
uncovers the load balancing benefits of EA-ALPL.
5.1.3 Local Power Consumption
We now turn our attention to local power consumption at
each node. We collect node statistics by piggybacking node
status information into data packets. Each arriving data
packet for a node provides one data point regarding node
status. Recall that during each 43 hour experiment, each
node sends 2,580 data packets, yielding 2,580 data points
for each node.
Fig. 8b shows the average check interval of each node in
the BMAC, ALPL, and EA-ALPL deployment experiments,
with the error bars indicating the 99 percent confidence
interval. In plain BMAC, all nodes use a check interval of
20 ms. The average check interval for ALPL and EA-ALPL
ranges between 20 ms and 200 ms throughout the
deployment. In both ALPL and EA-ALPL, the average
check intervals of all nodes are longer than in BMAC, so all
nodes save on idle listening power consumption.
EA-ALPL yields a more balanced spread of check
intervals among nodes than ALPL or BMAC. The nodes
with the lowest average check interval in ALPL, such as
nodes 1 and 8, have a higher average check interval with
EA-ALPL. Similarly, some of the nodes with the highest
average check intervals in ALPL have lower check intervals
in EA-ALPL. The balancing out of average check intervals
yields more balanced energy consumption among network
nodes, as Fig. 9a reveals.
The narrow 99 percent confidence interval for all nodes
in the ALPL and EA-ALPL experiments indicates the
stability of this approach in the long-term experiments.
We also note that busier nodes with the smaller average
chec k i nterval had slightly less stable check intervals
compared to other nodes du e to occasional routing
oscillations. The stability of check intervals for busy nodes
is of the same order for both ALPL and EA-ALPL. Since the
basic version of ALPL does not introduce any routing cost
modifications, we attribute these transient oscillations to
link state changes in the dynamic network topology.
Fig. 9a shows the effect of average check intervals on
total local power consumption at each node for the duration
of each experiment. Each point in Fig. 9a relates the
aggregated power consumption of a single node during
an experiment to the node’s average check interval. In plain
BMAC, all nodes consume almost the same power as the
busiest node because all nodes use the same listening mode.
In ALPL, nodes with an average check interval close to
200 ms achieve energy savings of more than 50 percent
compared to the busiest node. EA-ALPL yields a more
balanced traffic load, as it reduces power consumption at
the most active nodes by about 16 percent at the cost of
small increases in power consumption at less active nodes.
This trade-off is favorable because network lifetime
depends on the most active critical nodes.
Through a similar representation, Fig. 9b plots the
aggregate local power consumption of each node based
on the node’s average number of descendants during the
deployment. For both ALPL and BMAC, the range and
distribution of the number of descendants is the same
because both methods use the same routing metrics. This
reinforces the claim in Section 3.3.1. EA-ALPL incorporates
radio activity into routing decisions to balance the forward-
ing load, so the most loaded node in EA-ALPL has an
average of two descendants in comparison with an average
of 2.5 in ALPL and BMAC.
Power consumption for the BMAC case is correlated
with the number of descendants, but the variation is
limited. The overall trend for both ALPL experiments is
that it yields more energy savings for nodes with fewer
descendants because these nodes can use longer check
intervals. As the number of descendants increases, the
power savings of using ALPL are reduced because the
JURDAK ET AL.: ADAPTIVE LOW POWER LISTENING FOR WIRELESS SENSOR NETWORKS 13
Fig. 9. (a) Local power consumption as a function of average check interval. (b) Local power consumption as a function of the average number of
descendants.

average check interval gets closer to the case of network-
wide listening modes. EA-ALPL conforms to the trend of
larger power savings for nodes with fewer descendants.
However, EA-ALPL yields higher energy savings than
ALPL at the nodes with the highest number of descendants,
mainly by shifting the forwarding load away from busier
nodes when possible. By reducing the power consumption
of the busiest node, the inclusion of radio duty cycle
prolongs network lifetime.
5.2 Event-Driven Sensor Network
In this section, we consider an event-driven sensor network
application to investigate the benefits of using role
information to alter network behavior. Fig. 10a shows the
topology of the test-bed network. The application models a
target tracking scenario where the nodes react to the
appearance of the target by providing data more frequently.
The nodes in the network collect and send their sensor data
periodically every 60 seconds by default. We designate one
of the nodes, node 3, as the target node. Nodes that detect
the target node’s presence, nodes 1 and 11 in Fig. 10a, begin
sampling their sensors and sending the data at 30 second
intervals. Nodes that do not detect the presence of the target
node continue sampling their sensors at 60 second intervals.
We conduct four experiments for the same physical
network topology. The first three experiments use BMAC,
ALPL, and EA-ALPL as in the time-driven case above. The
fourth experiment considers the node’s number of descen-
dants, duty cycle, and role as the local states and adapts
network behavior accordingly. We refer to this case as Role
and Energy Aware ALPL (REA-ALPL). The weights of both
the radio duty cycle and sensing cost metrics,  and , are
set to 2 for this scenario.
5.2.1 Global Power Consumption
The average data yield for BMAC, ALPL, EA-ALPL, and
REA-ALPL remains at 98.5 percent. As in the time-driven
case, data delivery is not affected by the introduction of
ALPL.
Fig. 10b plots t he average global network power
consumption for BMAC, ALPL, EA-ALPL, and REA-ALPL.
All three state representations in ALPL yield almost the
same overall power consumption, since they all include the
number of descendants variable, which is the main
contributor to reduction of global power consumption.
ALPL reduces the overall network power consumption by
21 percent over the BMAC case for this event-driven
network. The reductions in global power consumption are
lower than the time-driven network case because the
sensing power consumption constitutes a larger portion of
overall power consumption in the event-driven case.
5.2.2 Local Power Consumption
We now turn our attention to the local power consumption
at each node. Before presenting the local power consump-
tion results, we note that the goal of this analysis is to
illustrate how the addition of the role aspect contributes to
load balancing, leading to a longer network lifetime. Ideally,
all nodes would deplete their batteries at about the same
time. In this experiment, the goal is to explore the extent to
which nodes can recognize the increased sensing activity of
the tracker nodes 1 and 11 in order to favor these nodes in
routing decisions.
We first examine the average check interval for each
node during the deployment. Fig. 11a plots the average
check interval of each node for the duration of each of the
three ALPL experiments. We omit the constant BMAC
check interval from the figure for presentation clarity. In the
ALPL experiment, the tracker node with an ID of 1 is the
busiest node with a check interval of 43 ms. By including
radio duty cycle information, EA-ALPL shifts some of
node 1’s forwarding load to node 9, enabling a slightly
longer average check i nterval at node 1. REA-ALPL
includes information on sensing activity, which shifts most
of node 1’s forwarding load to nodes 8 and 9. As a result,
we observe an average check interval of 183 ms at node 1 in
the REA-ALPL, while the check intervals of nodes 8 and 9
drop to 51 and 75 ms, respectively. The expanded state
representation of REA-ALPL did not cause any added
instability in the network, evidenced by the similar width of
error bars for the three experiments.
Fig. 11b examines the topological distinctions of each
experiment by plotting the average number of descendants
for each node in the ALPL experiments. Fig. 11b omits the
results for t he BMAC experiment results because the
number of descendants is the same as ALPL. For the case
of ALPL, the top three forwarders have a number of
descendants that ranges between 2.62 and 1.1. EA-ALPL
considers radio duty cycle as the main reason for energy
imbalance, which narrows the gap in the average number of
descendants of the three busiest nodes to a range between
1.875 and 1.32. REA-ALPL provides a more expressive node
energy profile by considering the potential for increased
sensing frequency at tracker nodes. REA-ALPL recognizes
that node 1 already has a higher power burden for tracking
the target node and reporting the data at double the
14 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 6, NO. 8, AUGUST 2007
Fig. 10. (a) Physical network topology for the event-driven experiments. (b) Global network power consumption during deployment.

frequency of other nodes, so it relieves node 1 from most of
its forwarding burden. As a result, REA-ALPL forces nodes
to avoid node 1 as a routing parent, which shifts most of the
forwarding to nodes 8 and 9. The number of descendants at
nodes 8 and 9 experiences higher instability than at other
nodes in the RE A-ALPL experiment, but the absolute
instability in the number of descendants for all nodes
appears to be small and transient.
Fig. 12a shows the effect of shifting the descendants
away from node 1. The labeled arrows in Fig. 12a indicate
the point corresponding to node 1 for the four cases. The
power consumption for node 1 in the cases of BMAC,
ALPL, and EA-ALPL is about the same although EA-ALPL
reduces the number of descendants by about 25 percent.
This effect is explained by the dominant effect of sensing
power consumption of node 1 over its listening power
consumption. REA-ALPL reduces the power consumption
of node 1 by about 25 percent. By shifting the descendants
of node 1 to nodes 8 and 9, REA-ALPL enables node 1 to
have a longer lifetime than node 8 and keep on reporting
the target node even after node 8 dies. REA-ALPL thus
provides the most balanced energy consumption of the four
experiments through an extended cross-layer state repre-
sentation that considers all the causes of power consump-
tion imbalance.
Shifting descendants away from node 1 reduces both the
reception and transmission power consumption at node 1,
but it also reduces the idle listening power consumption at
node 1 as indicated by Fig. 12b. In contrast, the three
experiments that do not consider node role fail in reducing
node 1’s listening power consumption. Because REA-ALPL
diverts almost all forwarding traffic away from node 1, it
enables node 1 to have an average check interval of about
183 ms. This increase in check interval significantly lowers
the listening power consumption at node 1 and reduces
overall power consumption at this node. The t raffic
diverted from node 1 to node 8 causes node 8 to have
the shortest average check interval among all three state
representations. As mentioned before, the traffic shift
enables power consumption balancing in the network. As
a result, the plot for REA-ALPL exhibits the l owest
variation in the power consumption of the four most
loaded nodes for all experiments.
Through its consideration of a more comprehensive
definition of node state that includes role information, REA-
ALPL has yielded better load balancing in the event-driven
tracking network experiments. In particular, it has shifted
the traffic load away from the nodes with high sensing
activity. This traffic shift has reduced the power consump-
tion of the busiest node in the network by 23 percent,
resulting in significant improvement in the longevity of the
critical tracker nodes.
6DISCUSSION
This paper has proposed a framework for greedy cross-
layer local optimizations in sensor networks that reduces
JURDAK ET AL.: ADAPTIVE LOW POWER LISTENING FOR WIRELESS SENSOR NETWORKS 15
Fig. 11. (a) Average check intervals at each node. (b) Average number of descendants. The error bars show the 99 percent confidence intervals.
Fig. 12. (a) Power versus average number of descendants. (b) Power versus average check interval.

overall power consumption in the network and promotes
load balancing among nodes to customize network beha-
vior to an application’s performance requirements. The
framework enables nodes to use their local and neighbor-
hood state information to determine their behavior at the
MAC layer and at the network layer. At the network layer, a
flexible cost function enables nodes to customize routing
cost metrics according to an application’s performance
requirements. At the MAC layer, the nodes set their check
intervals in BMAC according to their local state. Bringing
together the optimizations at the network and MAC layers,
ALPL is ensures seamless adaptation to local node state.
We have validated the framework through two sets of
experiments on a testbed of sensor nodes. The first set of
experiments represented a typical time-driven monitoring
sensor network with a single data sink. The second set of
experiments was an event-driven network that models a
target-tracking application.
The experiments have explored three representations for
local node state. The analysis and results show that
including more information from across the network stack
into the local state representation better reflects nodes’
energy profiles and enables more informed adaptations of
network behavior. In the time-driven case, the state
representat ion combining d uty cycle and descendants
yields the most balanced power consumption distribution.
In the event-driven case, the sta te representation that
includes role, duty cycle, and descendants yields the most
balanced spread of power consumption for the event-
driven case.
The adaptive and flexible nature of our framework
supports the dynamic nature of sensor networks and can
exploit local state information about the present, the past,
and the predicted future state of sensor nodes to reduce
power consumption. We have studied how adapting
listening modes to the node’s current logical topology
position (which represents the node’s present state), duty
cycle (which represents the node’s past state), and role
(which represents the node’s predicted future state) can
reduce the global network power consumption and the
local power consumption at each node.
The use of a proactive routing strategy fo r sensor
networks requires careful tuning of the routing update
period for the network application in order to balance route
adaptivity and energy efficiency. In long term monitoring
applications, sending route update message with a period
of the same order of data message periods ensures that the
communication overhead of proactive routing is small.
One concern of using greedy approaches is that local
decisions may not be globally optimal. For instance, a well-
known case for greedy approaches is when the cheapest
next hop does not represent the best path to the destination.
In our framework’s implementation, the sh ortest path
algorithm in MintRoute uses a modified cost metric
including link quality and power to route packets to the
next hop. Because each node learns its hop count to the base
station by a adding a single hop to its parent hop count,
nodes are guaranteed to have an accurate view of their hop
count. In contrast, nodes learn the link quality and energy
cost metrics for their direct neighbors only, so nodes may
not adapt instantaneously to abrupt changes for these
metrics elsewhere in the network. However, we have
observed through our deployments that the nodes’ energy
cost metrics change gradually rather than abruptly. As for
link qualities, transient changes occur due to movement of
objects around the nodes, but the high data yield rate in our
deployment has shown that these changes have a minimal
effect on data delivery or energy savings.
Our framework adopts BMAC for the MAC layer
protocols and introduces ALPL to interface the MAC layer
with a proactive strategy at the routing layer. For any
proactive routing protocol, ALPL does not significantly
increase communication, processing, or storage complexity.
The results in Section 5 have shown that the communication
overhead of a few bits of state information in periodic
routing messages is a small price to pay for the energy
savings of ALPL. In terms of processing complexity, cost
metric calculations in ALPL involve simple arithmetic
operations and routing decisions involve a few simple
conditional statements. Finally, the state information stored
at each node is a function of node density and not network
size, which adds scalability to the greedy approach.
Because of the reduced transmission power in our
deployment and the limited space in the laboratory, each
node had an average of about four neighbors, yielding a
relatively high network density. The high node density in
our experiments raised the degree of contention among
nodes. Coupled with the framework’s locally optimal
decisions, the experiments have confirmed that the frame-
work is scalable to large or dense networks despite the
limited size of our network test-bed.
Reduction of global network power consumption
through local decisions is an approach that is widely
applicable to many sensor network applications and quality
of service requirements. Our framework is independent of
the underlying routing protocol or MAC protocol. Instead,
it can build on other underlying mechanisms with a
modular design to optimize the behavior of any sensor
network application.
ACKNOWLEDGMENTS
This work was partially supported by US National Science
Foundation grant No. CCF-0347902. The work of Raja
Jurdak has been partially supported by IRCSET.
REFERENCES
[1] B. Karp and H.T. Kung, “GPSR: Greedy Perimeter Stateless
Routing for Wireless Networks,” Proc. MobiCom, 2000.
[2] P. Skraba, H. Aghajan, and A. Bahai, “Cross-Layer Optimization
for High Density Sensor Networks: Distributed Passive Routing
Decisions,” Proc. Int’l Conf. Ad-Hoc, Mobile, and Wireless Networks
(ADHOC-NOW), pp. 266-279, 2004.
[3] R.L. Cruz and A.V. Santhanam, “Optimal Routing, Link Schedul-
ing and Power Control in Multi-Hop Wireless Networks,” Proc.
INFOCOM, 2003.
[4] T. ElBatt and A. Ephremides, “Joint Scheduling and Power
Control for Wireless Ad Hoc Networks,” IEEE Trans. Wireless
Comm., vol. 1, pp. 74-85, 2004.
[5] U.C. Kozat, I. Koutsopoulos, and L. Tassiulas, “A Framework for
Cross-Layer Design of Energy-Efficient Communication with QoS
Provisioning in Multi-Hop Wireless Networks,” Proc. INFOCOM,
2004.
[6] R. Jurdak, “Modeling and Optimization of Ad Hoc and Sensor
Networks,” PhD dissertation, Bren School of Information and
Computer Science, Univ. of California Irvine, Sept. 2005.
[7] P. Baldi, L. De Nardis, and M.G. Di Benedetto, “Modelling and
Optimization of UWB Communication Networks through a
16 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 6, NO. 8, AUGUST 2007

Flexible Cost Function,” IEEE J. Selected Areas in Comm., vol. 20, no.
9, pp. 1733-1744, 2002.
[8] J. Polastre, J. Hill, and D. Culler, “Versatile Low Power Media
Access for Wireless Sensor Networks,” Proc. ACM Conf. Embedded
Networked Sensor Systems (SenSys ’04), 2004.
[9] R. Jurdak, P. Baldi, and C.V. Lopes, “Energy-Aware Adaptive
Low Power Listening for Sensor Networks,” Proc. Int’l Workshop
Networked Sensing Systems (INSS ’05), 2005.
[10] R. Jurdak, P. Baldi, and C.V. Lopes, “State-Driven Energy
Optimization in Sensor Networks,” Proc. IEEE Int’l Conf. Sensor
Networks (SENET ’05), 2005.
[11] A. Woo and D. Culler, “A Transmission Control Scheme for Media
Access in Sensor Networks,” Proc. MobiCom, 2001.
[12] D. Culler et al., “Towards a Sensor Network Archi tectu re:
Lowering the Waistline,” Proc. Workshop Hot Topics in Operating
Systems (HotOS X), 2005.
[13] J. Polastre et al., “A Unifying Link Abstraction for Wireless Sensor
Networks,” Proc. ACM Conf. Embedded Networked Sensor Systems
(SenSys ’05), Nov. 2005.
[14] K. Romer and F. Mattern, “Event-Based Systems for Detecting
Real-World States with Sensor Networks: A Critical Analysis,”
Proc. DEST Workshop at Int’l Conf. Intelligent Sensors, Sensor
Networks and Information Processing (ISSNIP), 2004.
[15] J. Liu et al., “State-Centric Programming for Sensor and Actuator
Network Systems,” IEEE Pervasive Computing Magazine, vol. 2,
no. 4, 2003.
[16] V. Mhatre et al., “A Minimum Cost Heterogeneous Sensor
Network with a Lifetime Constraint,” IEEE Trans. Mobile Comput-
ing, vol. 4, no. 1, 2005.
[17] W. Ye and J. Heidemann, “Medium Access Control in Wireless
Sensor Networks,” Wireless Sensor Networks, T. Znati,
K.M. Sivalingam, and C. Raghavendra, eds., Kluwer Academic
Publishers, 2003.
[18] IEEE 802.15.4, Wireless Medium Access Control (MAC) and
Physical Layer (PHY) Specifications for Low-Rate Wireless Personal
Area Networks (LR-WPANs), IEEE, http://standards.ieee.org/
getieee802/download/802.15.4-2003.pdf, 2003.
[19] R. Jurdak, C.V. Lopes, and P. Baldi, “A Survey, Classification, and
Comparative Analysis of Medium Access Control Protocols for
Ad Hoc Networks,” IEEE Comm. Surveys and Tutorials, vol. 6, no. 1,
2004.
[20] A. El-Hoiydi, J.D. Decotignie, and J. Hernandez, “Low Power
MAC Protocols for Infrastructure Wireless Sensor Networks,”
Proc. Fifth European Wireless Conf., Feb. 2004.
[21] Q. Jiang and D. Manivannan, “Routing Protocols for Sensor
Networks,” Proc. Consumer Comm. and Networking Conf. (CCNC
’04), 2004.
[22] C.E. Perkins and P. Bhagwat, “Highly Dynamic Destination-
Sequenced Distance-Vector Routing (DSDV) for Mobile Compu-
ters,” Proc. SIGCOMM, pp. 234-244, 1994.
[23] T. Chen and M. Gerla, “Global State Routing: A New Routing
Scheme for Ad-Hoc Wireless Networks,” Proc. Int’l Conf. Comm.
(ICC ’98), 1998.
[24] C.E. Perkins, E.M. Royer, and S.R. Das, “Ad Hoc On-Demand
Distance Vector Routing,” IETF Internet draft, work in progress,
Oct. 1999.
[25] D.B. Johnson and D.A. Maltz, “The Dynamic Source Routing
Protocol for Mobile Ad Hoc Networks,” IETF Internet draft, work
in progress, Oct. 1999.
[26] C.K. Toh, “Maximum Battery Life Routing to Support Ubiquitous
Mobile Computing in Wireless Ad Hoc Networks,” IEEE Comm.
Magazine, pp. 138-147, June 2001.
[27] R.C. Shah and J. Rabaey, “Energy Aware Routing for Low Energy
Ad Hoc Sensor Networks,” Proc. Wireless Comm., and Networking
Conf. (WCNC ’02), 2002.
[28] C. Park, K. Lahiri, and A. Raghunathan, “Battery Discharge
Characteristics of Wireless Sensor Nodes: An Experimental
Analysis,” Proc. IEEE Comm. Soc. Conf. Sensor and Ad Hoc Comm.
and Networks (SECON ’05), 2005.
[29] R. Szewczyk et al., “Habitat Monitoring with Sensor Networks,”
Comm. ACM, vol. 47, no. 6, pp. 34-40, 2004.
[30] W. Ye, J. Heidemann, and D. Estrin, “An Energy-Efficient MAC
Protocol for Wireless Sensor Networks,” Proc. INFOCOM, 2002.
[31] T. Van Dam and K. Langendoen, “An Adaptive Energy-Efficient
MAC Protocol for Wireless Sensor Networks,” Proc. Int’l Conf.
Embedded Networked Sensor Systems, pp. 171-180, 2003.
[32] W. Ye, J. Heidemann, and D. Estrin, “Medium Access Control
with Coordinated, Adaptive Sleeping for Wireless Sensor Net-
works,” IEEE Trans. Networking, Apr. 2004.
[33] M. Haenggi, “Energy-Balancing Strategies for Wireless Sensor
Networks,” Proc. IEEE Int’l Symp. Circuits and Systems (ISCAS ’03),
2003.
[34] R. Szewczyk et al., “Lessons from a Sensor Network Expedition,”
Proc. European Conf. Wireless Sensor Networks (EWSN ’04), 2004.
[35] “Tiny Operating System,” http://tinyos.net, 2005.
[36] A. Woo, T. Tong, and D. Cu ller, “Taming the Underly ing
Challenges of Reliable Multihop Routing in Sensor Networks,”
Proc. ACM Conf. Embedded Networked Sensor Systems (Sensys ’03),
2003.
[37] R. Sivakumar, B. Das, and V. Bharghavan, “The Clade Vertebrata:
Spines and Routing in Ad Hoc Networks,” Proc. IEEE Symp.
Computers and Comm. (ISCC ’98), 1998.
[38] R. Jurdak, C.V. Lopes, and P. Baldi, “Battery Lifetime Estimation
and Optimization for Underwater Sensor Networks,” Sensor
Network Operations, May 2006.
[39] Crossbow Technology Inc., “Mica 2 Motes,” http://www.xbow.
com/Products/productdetails.aspx?sid=72, 2005.
[40] D. Gay et al., “The nesC Language: A Holistic Approach to
Networked Embe dded Systems,” ACM
SIGPLAN Notices, vol. 38, no. 5, pp. 1-11,
2003.
Raja Jurdak received the PhD degree in
information and computer sciences at the Uni-
versity of California, Irvine in 2005. From 2005 to
2006, he was a postdoctoral researcher at the
University of California, Irvine. He is currently a
senior researcher in the Adaptive Information
Cluster at University College Dublin. He was a recipient of the IRCSET
Embark fellowship in 2006. Dr. Jurdak is the author of more than 21
refereed journal and conference publications, as well as the book
Wireless Ad Hoc and Sensor Networks: A Cross-Layer Design
Perspective (Springer, 2007). His research focuses on application-
driven networks, modeling of ad hoc and sensor networks, emerging
communication technologies, underwater acoustic networks, and cross-
layer design. He is a member of the IEEE and
the IEEE Computer Society.
Pierre Baldi receive d the PhD degree in
mathem atics from the California Institute of
Technology in 1986. From 1986 to 1988, he
was a postdoctoral fellow at the University of
California, San Diego. From 1988 to 1995, he
was a member of the faculty and technical staff
at the California Institute of Technology and at
the Jet Propulsion Laboratory (JPL). He was CEO of a startup company
from 1995 to 1999 and joined the University of California, Irvine (UCI) in
1999. He is currently a chancellor professor in in the School of
Information and Computer Sciences, director of the Institute for
Genomics and Bioinformatics, and member of the California Institute
for Telecommunications and Information Technology (Calit2) at UCI. He
is the recipient of a 1993 Lew Allen Award at JPL and a Laurel Wilkening
Faculty Innovation Award at UCI. Dr. Baldi is the author of more than
200 research articles and four books, including Modeling the Internet
and the Web–Probabilistic Methods and Algorithms (Wiley, 2003). His
research focuses on probabilistic modeling and statistical inference,
machine learning, bioinformatics, data mining, and communication
networks. He is a senior member of the IEEE.
Cristina Videira Lopes is an associate profes-
sor at the University of California, Irvine. She
conducts research in programming languages,
software engineering, and pervasive computing.
She is the recipient of a US National Science
Foundation CAREER Award. She is a 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.
JURDAK ET AL.: ADAPTIVE LOW POWER LISTENING FOR WIRELESS SENSOR NETWORKS 17