Joint Aggregation and MAC Design to Prolong

Sensor Network Lifetime

Zi Li, Yang Peng, Daji Qiao, and Wensheng Zhang
Iowa State University, Ames, IA, USA
Email: {zili, yangpeng, daji, wzhang}@iastate.edu

AbstractThis paper proposes JAM, a Joint Aggregation and

MAC design, to improve the sensor network lifetime under
the end-to-end delay constraint. The key idea is to adjust
both network traffic (via data aggregation) and communication
overhead (via duty-cycled MAC) in a holistic manner at each
individual node as well as between neighbors. As a result,
JAM extends the sensor network lifetime more efficiently and
effectively than the state-of-the-art solutions while guaranteeing
the desired delay bound and achieving a lower level of average
nodal power consumption. JAM is a lightweight and distributed
solution with limited control information exchanged between
neighbors only, which makes it deployable in practical sensor
networks. Extensive ns-2 simulation and TinyOS experiment
results are used to demonstrate the effectiveness of JAM in
prolonging the network lifetime.

I. I NTRODUCTION
A. Motivation
Extending the lifetime of sensor networks is critically
important when the networks are deployed for long-term
monitoring applications. Besides duty cycling, balancing nodal
lifetime is another major approach for network lifetime extension, because the network lifetime is often defined as the
minimal nodal lifetime among all nodes in the network [1]
[3]. Following this approach, a variety of duty-cycled MAC
protocols [4][6] and data aggregation schemes [7], [8] have
been proposed recently.
Duty-cycled sensor networks rely on MAC protocols to
establish rendezvous between sender and receiver nodes. The
incurred MAC-layer communication overhead is distributed
between sender and receiver in different manners with different MAC protocols. To balance nodal lifetime, the authors
of [4][6] proposed to adapt the distribution of communication
overhead among neighbors according to their nodal lifetime,
i.e., longer-lifetime nodes shall absorb more communication
overhead than their shorter-lifetime neighbors.
Through eliminating inherent redundancy in raw sensory
data, in-network data aggregation [9], [10] can effectively
reduce network traffic. For the sake of aggregation, a node
needs to hold data (received or self-generated) for a while.
Clearly, a node can suppress more data traffic with a longer
holding time. However, holding data introduces extra delay,
and the value of sensory data could be greatly depreciated
c
978-1-4799-1270-4/13/$31.00 2013
IEEE

if the data is delivered to the sink with a delay longer

than an application-specified delay bound. To balance nodal
lifetime while maintaining a required end-to-end delay bound,
the authors of [7], [8] proposed to adapt the distribution of
holding time among nodes along the same route according to
their nodal lifetime; i.e., shorter-lifetime nodes shall hold data
longer to reduce its outgoing data traffic while longer-lifetime
nodes shall hold data for a shorter time so that the end-to-end
delay requirement is still guaranteed.
As the aforementioned MAC and data aggregation schemes
improve the network lifetime from two distinct perspectives,
it is naturally an attractive idea to explore the possibility of a
joint design of MAC and data aggregation to further improve
the network lifetime. This, however, is a challenging task due
to the following reasons. Firstly, most of the MAC protocols
were designed without explicitly considering the end-to-end
delay requirement. As a result, they may yield uncontrollable end-to-end delay under certain practical scenarios; this
is unacceptable for data aggregation applications that often
require a stringent end-to-end delay bound. Secondly, even if
there exists a MAC protocol that can provide a certain delay
guarantee, when it works with a data aggregation scheme,
it is non-trivial to decide how to divide the allowed endto-end delay into two parts to serve as the delay constraints
respectively for the MAC and data aggregation protocols, such
that the maximum lifetime improvement can be accomplished.
Therefore, it is important to develop an innovative approach
to integrate MAC and data aggregation protocols together via
a joint adjustment of their protocol parameters.
B. Related Work
Research has been conducted on designing duty-cycled
MAC protocols and energy-efficient or lifetime-extending data
aggregation schemes. However, there is no research on jointly
adjusting data aggregation and MAC behaviors to extend the
sensor network lifetime.
Numerous MAC protocols [4][6], [11] have been proposed
to extend network lifetime through balancing nodal lifetime.
Particularly, SEESAW [4] balances the energy consumption
between sender and receiver through adapting the data retry
interval at the sender side and the channel checking period at
the receiver side. ZeroCal [6] targets at improving the fairness

of energy utilization in duty-cycled sensor networks by dynamically tuning the nodal wakeup interval. GDSIC [5] decides the
individual nodal wakeup interval through solving distributed
convex optimization problems. Though the network lifetime
can be prolonged by these schemes, they do not consider the
end-to-end delay bound. pTunes [11] is a recently-proposed
centralized solution, which formulates a multi-objective optimization problem, where prolonging network lifetime and
guaranteeing the end-to-end delay can be solved together.
Most of existing data aggregation schemes [12][16] have
the objective of minimizing the total network energy consumption instead of extending network lifetime, or do not consider
the end-to-end delay requirement. The problem of balancing
nodal lifetime under a delay constraint has been studied in [7],
[8]. Particularly, Becchetti et al. [8] investigated the problem
of energy-efficient data aggregation within a delay bound,
and proposed two distributed schemes to balance energy
consumption among sensor nodes. LBA [7] is a recentlyproposed lifetime-balancing aggregation protocol. Through
dynamically adjusting the aggregation holding time among
neighbors to balance their nodal lifetime, LBA provides a low
cost, asynchronous, and delay-constrained data aggregation
scheme for duty-cycled sensor networks.
C. Contributions
In this paper, we propose a novel holistic approach, called
JAM (Joint Aggregation and MAC), to jointly adjust MAC
and data aggregation behaviors to extend the sensor network
lifetime. The key idea of JAM is to coordinate the aggregation
and MAC behaviors at each individual node as well as between
neighbors, with the target of extending the minimal nodal
lifetime in the neighborhood. As such coordination occurs in
all neighborhoods, the network lifetime, i.e., the minimal nodal
lifetime among all nodes in the network, may be improved. As
JAM reduces both network traffic (via data aggregation) and
communication overhead (via duty-cycled MAC), it prolongs
the network lifetime more efficiently and effectively than
previous works that only use one of the two techniques. The
contributions of this work are summarized as follows.
To the best of our knowledge, JAM is the first design
on integrating and jointly configuring MAC and data
aggregation protocols to extend the lifetime of dutycycled sensor networks.
JAM is a distributed and lightweight solution. It works
through limited control information exchanged locally
between neighbors.
JAM has been implemented and evaluated on a sensor
network testbed, and results show that it can achieve
significant improvement on network lifetime compared
to the state-of-the-art solutions.
D. Organization
The rest of the paper is organized as follows. Section II
presents the system models and problem statement. Details

of the JAM design and implementation are described in Section III. Section IV shows the performance evaluation results
obtained from both ns-2 simulations and TinyOS testbed
experiments. Finally, Section 5 concludes the paper.
II. P ROBLEM D ESCRIPTION
A. System Models
We consider a sensor network deployed for monitoring applications where in-network data aggregation is allowed. Each
sensor node generates and reports sensory data periodically,
and all nodes form a data collection tree rooted at the sink.
Protocols like CTP (Collection Tree Protocol) [17] could be
used to build and maintain the data collection tree.
1) Aggregation Model: The total aggregation model [18]
is adopted, which allows an arbitrary number of data packets
generated and/or received at the same node to be suppressed
into a single data packet. Such a model is useful in many
sensor network applications, for example, when users are more
interested in the maximum, minimum, average, or percentile
statistics of sensory data, rather than the raw data themselves.
With this model, a source node may not send out a sensory
data packet immediately after it is generated. Instead, the
node may wait for a certain period of time (called the selfaggregation delay (SAD)), and aggregates all data generated
during the period to reduce the amount of data traffic to its
parent node. Similarly, a forwarding node may not forward
a data packet immediately after reception; it may wait for
another period of time (called the forwarding-aggregation
delay (FAD)) and aggregate all packets received during the
period. Generally, the longer time a node waits for aggregation,
the more data traffic can be suppressed; at the same time, data
delivery latency is increased.
2) MAC Model: To conserve energy without time synchronization overhead, it is desired to employ an asynchronous and
duty-cycled MAC protocol for long-term monitoring applications. The design principle of our proposed scheme does not
require a particular MAC protocol. Instead, our design works
compatibly with any asynchronous duty-cycled MAC protocol,
as long as the protocol allows a nodes duty cycle to be
adjusted dynamically. To simplify the presentation, however,
we assume each node runs an RI-MAC [19] like protocol as
shown in Figure 1.
In RI-MAC, each node wakes up every Tr interval to interact
with potential senders. Upon wakeup, it sends out a beacon and
then checks the channel activity for time. If a data packet is
received within , it replies with an ACK; otherwise, it goes
back to sleep. On the other hand, if a node has a data packet
to send, it remains awake and waits idly for the receivers
beacon to start data transmission. In the worst-case scenario,
the sender has to stay awake for Tr time before rendezvous
with the receiver, which incurs a transmission delay of Tr.
Different from the original RI-MAC protocol that has a fixed
Tr, we assume Tr is dynamically tunable. As can be observed,
a larger Tr reduces the receivers channel polling frequency

Data Arrival

Objective:
max
min

Beacon

Data

ACK

Beacon

Data

ACK

Radio on

j{i}C(i)

its computation will be explained later in Section III-C.

C(i) is the set of is child nodes.
Subject to:
SAD(i), FAD(i), Tr(i) > 0;
End-to-End Delay Requirement:

Beacon

Beacon

Sender

Receiver

L(j), where L(j) is js nodal lifetime and

SAD(i) +
Tr

Fig. 1.

Radio on

Radio off

and hence its energy consumption; on the other hand, it

increases the senders energy consumption on idly waiting for
the beacon, and meanwhile increases the transmission delay
over the sender-to-receiver link.
3) Delay Model: Three types of delays are involved along
a source-to-sink path: (i) one SAD at the source node, (ii)
multiple FADs at the forwarding nodes, and (iii) multiple
transmission delays over the links. With the aggregation and
MAC models described above, to guarantee that the maximum
end-to-end packet delivery delay from any source node i to
the sink node is bounded by an application-specific parameter
D, we need to ensure that1
X

(Tr(m) + FAD(m)) 6 D,

(Tr(m) + FAD(m)) = D.

(2)

mSi

An RI-MAC like protocol but with a tunable Tr parameter.

SAD(i) +

(1)

mSi

where Si is the path from node is parent to sink. As larger

aggregation delays would allow more traffic to be aggregated
to save more energy, the delay bound shall be fully utilized
to maximize network lifetime, meaning that the equality shall
hold true in Inequality (1) in practical schemes.
B. Problem Statement
To effectively prolong the sensor network lifetime under
the end-to-end packet delivery delay constraint, it is critical
to have a holistic approach to adjust data aggregation and
MAC behaviors of all sensor nodes. Ideally, all sensor nodes
shall work together to maximize the minimum nodal lifetime
in the entire network. Unfortunately, this global objective is
impossible to accomplish in a realistic sensor network, as it
requires each node to know the residual energy levels and data
generation rates of all other nodes, and the topology of the
network, which are highly dynamic and often unpredictable
by nature. Instead, we study the following localized problem
for each sensor node i in the network:
1 In practice, data or ACK packets may get lost due to collision, interference,
or deteriorated channel condition. As a result, the sensor node may need to
retransmit multiple times before the data packet can be delivered successfully.
This issue has been dealt with in JAM (as shown in Section III-F2) by
replacing Tr with ETXTr in Equations (1) and (2), where ETX is the expected
number of transmission attempts to deliver a data packet successfully over one
hop. To simplify the presentation, we set ETX = 1 during the explanation of
the JAM design in the next section, while the practical Equation (14) (given
in Section III-F2) is used in the actual JAM implementation.

Output:
is MAC protocol parameter: Tr(i);
is data aggregation parameters: SAD(i) and FAD(i).
The goal of this problem is to maximize the minimal nodal
lifetime in is neighborhood. As such procedure occurs in
all neighborhoods, the minimal nodal lifetime in the entire
network, i.e., the network lifetime, may be improved gradually.
III. T HE P ROPOSED JAM S CHEME
In this section, we propose a protocol called JAM to address
the problem defined above. In JAM, coordination only occurs
between a sensor node and its child nodes, through exchanging
lightweight control information as well as adjusting their
aggregation and MAC behaviors together in a collaborative
manner. Figure 2 gives an overview of the JAM scheme. To
ease the presentation, we use the topology shown in Figure 3
as an example to explain the design details of JAM.
JAM consists of four modules: Aggregator, OREM (Output
Rate Estimation Module), LEM (Lifetime Estimation Module),
and ICCM (Intra-node Cross-layer Collaboration Module). In

Data to
Parent

Send

Data Generator

Receive

Data from
Children

Aggregator
Children
Info
Control
Info

Parent
Info
ACK from
Parent

OREM

ICCM

LEM

JAM
Kernel

Control
Info

Receive

Data Path

Fig. 2.

Fig. 3.

ACK to
Children

Send

Control Path

JAM Overview.

Topology used to describe JAM details.

general, when node i receives a data packet from its child

nodes or an ACK from its parent, it extracts the control
information embedded in the packet and feeds them into JAM
kernel. Here, the control information needed by JAM kernel
includes five items from each child node j of i: e(j), (j),
SAD(j), the delay introduced at node
P j: D(j) = FAD(j) +
Tr(j), and js own input data rate kC(j) (k) where (k)
is the output data rate of a child node k of j; and two items
from parent node p of i: Tr(p) and (p). We will explain the
meaning of in Section III-E. With these information, JAM
kernel decides how node i shall adjust its MAC behavior (i.e.,
Tr) and aggregation behavior (i.e., SAD and FAD) to improve
the minimal nodal lifetime in its neighborhood. The decision
is then piggybacked into the ACK packet to its child nodes.
Upon being notified, each child node adjusts its MAC and
aggregation behaviors.
A. Aggregator Module
The aggregator module controls the nodal aggregation behavior. Specifically, it works as follows at node i:
1
Case I: P
6 FAD(i) 6 SAD(i). In this case,
jC(i) (j)
node i has a data packet arrival rate no lower than 1/FAD(i)
and 1/SAD(i). A timer is fired every FAD(i) interval. When it
fires, all data packets received and self-generated since its last
firing are aggregated into a single packet, and then forwarded
to the parent node. This way, we ensure that packets received
from is child nodes are held for no longer than FAD(i) time.
Also, since the self-generated data packets are aggregated and
forwarded every FAD(i) time, which is always smaller than
SAD(i) in JAM, their delay requirement is guaranteed as
well. We will prove FAD(i) 6 SAD(i) for any node i in
Section III-B.
1
1
Case II: P
> FAD(i) and
6 SAD(i). In
(j)
(i)
jC(i)
this case, data packets from child nodes arrive at a rate lower
than 1/FAD(i) while the node itself generates data packets at
a rate no lower than 1/SAD(i). Data received from children
and self-generated data are treated differently as follows:
Whenever a data packet is received from a child node,
the packet is aggregated immediately with all the selfgenerated data packets that have not yet been aggregated, and then forwarded to the parent node. Hence, the
forwarding-aggregation delay is zero.
A timer is fired every SAD(i) interval. When it fires,
all the self-generated data packets that have not yet been
aggregated are suppressed into a single packet and then
forwarded to the parent node.
1
1
Case III: P
> FAD(i) and
> SAD(i).
(j)
(i)
jC(i)
In this case, data packets from child nodes arrive at a rate
lower than 1/FAD(i) while the node itself generates data
packets at a rate lower than 1/SAD(i). As the data packet
arrival/generation rates are low, every data packet is simply

forwarded to the parent node immediately upon its reception

or generation.
B. Output Rate Estimation Module (OREM)
This model is part of the JAM kernel (shown in Figure 4)
and is used to estimate the nodal output data rate. Before
explaining how it works in detail, we first prove FAD(i) 6
SAD(i) for any node i. According to the end-to-end delay
requirement in Equation (2), node i shall satisfy:
X
SAD(i) +
(Tr(m) + FAD(m)) = D.
mSi

Similarly, for every child node j of i, we have:

X
SAD(j) + (FAD(i) + Tr(i)) +
(Tr(m) + FAD(m)) = D.
mSi

Hence, it follows that:

FAD(i) 6 D

(Tr(m) + FAD(m)) = SAD(i).

(3)

mSi

P
With the four inputs: FAD(i), SAD(i), jC(i) (j), and
(i), OREM estimates the output data rate of node i according
to the three cases described in Section III-A as follows:

(i) =

FAD(i)

jC(i) (j) +

:
1
SAD(i)

jC(i) (j) + (i)

:
:

1
P

jC(i) (j)

jC(i) (j)

6 FAD(i),
> FAD(i)

1
and (i)
6 SAD(i),
1
P
> FAD(i)
jC(i) (j)
1
> SAD(i).
and (i)

(4)

Note that a nodes input data rate is simply the sum of the
output data rates from all of its children.
C. Lifetime Estimation Module (LEM)
The LEM module is another module in the JAM kernel and
is used to estimate the nodal lifetime. Its input consists of
e(i), P
Tr(i), Tr(p) (i.e., Tr of node is parent node p), (i),
and jC(i) (j). Its output is the estimated nodal lifetime
L(i), which is computed as follows:
L(i) =

e(i)
,
c(i)

(5)

where e(i) is the residual energy and c(i) is the energy

consumption rate:
c(i) =

Tr(p)
+
2

(i)P +

P +
(j)P.
Tr(i)

(6)

jC(i)

Here, P is the power consumption rate when a nodes radio is

on. To send a data packet, node i waits for Tr(p)
time on av2
erage for its parent node p to wake up and then
spends

 time
Tr(p)
for the transmission. Hence, it consumes
2 + (i)P
power on average for data transmissions. The second term
in Equation (6) represents the average amount of power
consumed to monitor channel for time every Tr(i) interval,
while the third term is the average amount of power consumed
for data receptions. As radio is the most energy-consuming

component, we ignore other energy consumptions such as

sensing and computation, which could be easily plugged into
Equation (6).

FAD (i) = argFAD(i) max{L (i)},

Tr (i) = D(i) FAD (i).

(12)
(13)

E. JAM Kernel
D. Intra-node Cross-layer Collaboration Module (ICCM)

As the core of the JAM scheme, the JAM kernel of node i is

triggered to execute periodically every W time, or on demand
whenever its parent node p changes Tr(p) or FAD(p) and
consequently the delay D(p). We use (p) to denote the
change in D(p). The JAM kernel is executed to decide:
how the extra delay (p) introduced at parent node p shall
be absorbed by node i (in the amount of (p) via
updating D(i) to D (i) = D(i) ((p) )) and its
child nodes (in the amount of via updating D(j) to
D (j) = D(j) ); and

how node i shall split D (i) into Tr (i) and FAD (i),
so that the minimal nodal lifetime within node is neighborhood can be increased. Detailed working procedure of the JAM
kernel is illustrated in Figure 4 and explained below, while the
effect of W will be discussed in Section III-F3.
The JAM kernel of node i takes the following inputs: Tr(p)
and
P (p) from parent node p, and (j), SAD(j), D(j), and
kC(j) (k) from each child node j. Based on these inputs,
it checks the candidate values of from an ordered sequence
= h0, , , 2, 2, i one by one till the first feasible
value has been found to increase the minimal nodal lifetime
within node is neighborhood, or till all values in have
jC(i) (j)
been exhausted. Here, is a system parameter.
k
j To ensure
D(i)(p)

1
.
D
(i)
>
0,
the
smallest
element
in

is
set
to

Case II: FAD(i) > P

. In this case, the output

(j)

jC(i)
Similarly,
to ensure
j
k D (j) > 0, the largest element in is set
1
minjC(i) D(j)
data rate is (i) =
according to Equation (4). In to
. For each candidate value of , the JAM

FAD(i)
order to achieve optimal nodal lifetime in this case, we need kernel executes the following:

Iteratively, FAD (i) takes a value from range [0, D (i)]

to minimize the following term according to Equation (6):


with a small step . Here, is a system parameter. For
Tr(p)
1

+
+
,
(8)
each FAD (i) value, Tr (i) = D (i) FAD (i).
2
FAD(i)
Tr(i)

Based on FAD (i), Tr (i), and the inputs from each

which we denote as f (FAD(i)). To ease the presentation, we
child node j, the ICCM module is called to compute
use to denote Tr(p)
the maximum lifetime that node j can achieve, denoted
2 + . Plugging in Tr(i) = D(i) FAD(i),
we can rewrite f (FAD(i)) as:
as L (j), and the corresponding FAD (j), according to
Equations (11) and (12), respectively.

f (FAD(i)) =
+
.
(9)

FAD (j) is fed into the OREM module to compute the

FAD(i)
D(i) FAD(i)
output data rate (j) of node j. With (j) from all
By solving f (FAD(i)) = 0, we have:
child nodes, the output data rate (i) of node i is also
!

1
computed using the OREM module.

L i  FAD(i) > P
=
(j)

Then, the LEM module is called to estimate the nodal

jC(i)

 
lifetime L (i). If min{L (i), L (j), j Ci } is larger


D(i)
D(i)
1

P
q
q

L i  FAD(i) = 1+ : when jC(i) (j) 6 1+ 6 D(i),

than the highest L that has been found so far, L is

n
 
o

1
updated, and the corresponding FAD (i) and Tr (i) values
max L (i | FAD(i) = D(i)) , L i  FAD(i) = P

jC(i) (j)

are recorded as FAD (i) and Tr (i).

: otherwise.

When FAD (i) reaches the boundary condition, i.e.,

(10)

FAD (i) > D (i), if the best achievable L improves the

To summarize, for a given D(i), L (i) and the correspondminimal nodal lifetime within node is neighborhood, i.e.,
ing optimal FAD(i) and Tr(i) can be computed by:
L > min{L(i), L(j), j Ci }, FAD (i) and Tr (i)

L (i) = max{Equation (7), Equation (10)},

(11)
are output, (i) = (p) is appended to ACK packets

According to Equation (2), SAD value at the source node

can be computed once the FAD and Tr values of its ancestor
nodes have been determined. As both MAC and aggregation
behaviors affect nodal lifetime, their behaviors should be
coordinated. Specifically, as each forwarding node i introduces
a delay of D(i) = FAD(i) + Tr(i), the objective of the ICCM
module in the JAM kernel is to determine proper FAD(i) and
Tr(i) values for a given D(i), so that nodal lifetime L(i) is
maximized (denote as L (i)). To achieve this goal, there are
two cases to consider:
1
Case I: FAD(i) < P
. Depending on the relation
jC(i) (j)
between SAD(i) and (i), (i) equates the second or third
case of Equation (4). In these two cases, (i) is not affected
by FAD(i) and hence nodal lifetime is only affected by Tr(i)
according to Equations (5) and (6). As the energy consumption
rate (c(i)) in Equation (6) is a decreasing function of Tr(i),
optimal nodal lifetime is achieved when D(i) is fully allocated
to Tr(i) and hence FAD(i) = 0:
!

1


L i  FAD(i) < P
= L(i | FAD(i) = 0). (7)

Node
is info

= <0, + , - , +2 , -2 , ., - (D(i) (p))/

(i), e(i), SAD(i),

FAD(i), D(i)

k = 0, L = 0
D(i) = D(i) ( (p) )

D(i), FAD(i)

Child
js info

j C(i)>

(j), e(j), SAD(j),

D(j), k C(j)(k)

Pick first element of

as

FAD(i) = k
Tr(i) = D(i) - FAD(i)
SAD(j) = SAD(j) (FAD(i) - FAD(i)), j C(i)

k=k+1

FAD(i)
SAD(i),
(i)

, , + min{D(j)}/

SAD(j)

Tr(i)

SAD(j)

OREM

(j), e(j),
k C(j)(k)

ICCM
D(j)

'(i)

Tr(i)
e(i)

FAD*(j)

LEM

*(j)

Tr(p)

D(j)

D(j) = D(j)

OREM

(j),
k C(j)(k)

Tr(p)

L(i)

(p)

L*(j)

min{L(i), L*(j), j C(i)} > L ?

L = min{L(i), L*(j), j C(i)}

FAD(i) = FAD(i)
Tr(i) = Tr(i)

Parent
ps info

(p)
N

Parent
ps info

(k+1) > D(i) ?

Y

L > min{L(i), L(j), j C(i)}?

Is

Empty?

FAD (i), Tr (i), (i) = (p)

FAD(i), Tr(i), (i) = 0

Fig. 4. JAM kernel of node i. With parent, children, and self information, node i decides a proper value so that L > max min{L(i), L(j), j Ci }.
The figure only shows the interactions between ICCM, OREM, and LEM modules for child node j of i. In the actual JAM implementation, these interactions
are executed for every child node of i.

to all child nodes, and search completes; otherwise, the

next candidate value will be tested.
Note that, if all the candidate values of in have been
checked but none of them improves the minimal nodal lifetime within node is neighborhood, FAD(i) and Tr(i) remain
unchanged and (i) = 0.
In our implementation of the JAM scheme, we set = 0.5 s
and = 0.1 s. This means that, when the average per-hop
delay is 10 s, there is a total of 40 candidate values for ,
and 100 candidate values for FAD . Therefore, in the worstcase scenario, the JAM kernel needs to iterate 4000 times to
complete the execution, which is acceptable in practice. To
further reduce the complexity, we adopt a two-step heuristic
as follows. We first use a larger value (i.e., 1 s) to conduct
the initial search; when a feasible FAD has been found, we
use a smaller value (i.e., 0.1 s) to refine the search around
it. With this simple heuristic, the search space for FAD is
reduced to 30 and the total number of iterations in the worstcase scenario is reduced to 1200 in the above example.
F. Other Design Considerations

1) JAM Initialization: After the data collection tree has

been established (i.e., the routing table of each node becomes
stable), each node i needs to decide its initial FAD(i) and
Tr(i) values. In JAM, all nodes start with a default Tr value
and the rest of the end-to-end delay bound is evenly distributed
to nodes along the source-to-sink path. For this purpose, each
node i periodically measures the most updated (i) accumulative delays from its parent node to the sink (D(i s)),
i.e., the second term in the end-to-end delay requirement in
Equation (2), and (ii) hop count to its farthest descendant
(h(i)).
We assume that the sinks radio is always on and its wakeup
interval is Tr(s) = 0 s. As the sink is the end of data delivering
paths, FAD(s) = 0 s. D(s s) = 0 s is then broadcast to all
of its children. Upon receiving D(i s), each child node j
acts as follows:

If node j is not a leaf node, it sets D(j) = DD(is)

h(j)+1 ,
FAD(j) = D(j) Tr(j), SAD(j) = D D(i s),
and sends D(j s) = D(i s) + D(j) to all of its
children. In the example shown in Figure 5, node 1 sets

D(1) = 15/3 = 5 s, FAD(1) = 4 s, SAD(1) = 15 s, and

sends D(1 s) = 5 s to nodes 2 and 3.
If node j is a leaf node, it simply sets SAD(j) = D
D(i s). In the example shown in Figure 5, leaf node
4 sets SAD(4) = 15 10 = 5 s.
3

SAD(3) = 10s

Tr(s) = 0s
FAD(s) = 0s
D(ss) = 0s

Tr(1) = 1s
FAD(1) = 4s
SAD(1) = 15s
D(1s) = 5s

2
Tr(2) = 1s
FAD(2) = 4s
SAD(1) = 10s
D(2s) = 10s

G. JAM Implementation
1) Software Component: We have implemented JAM in
TinyOS 2.1.0. As shown in Figure 6, the shaded parts illustrate
the core components of JAM in the software architecture: (i)
the aggregation component that sits between application and
routing layers, and (ii) the MAC component that is designed
based on RI-MAC [19]. CTP [17] is adopted as the routing
layer protocol to set up the data collection tree.

Application

SAD(4) = 5s

Receive

Send

FwdAggM

SelfAggM

Fig. 5. JAM initialization example where the default Tr is 1 s and D is 15 s.

2) Handling of Packet Loss: In practice, data or ACK

packets may get lost due to collision, interference, or deteriorated channel condition, and the sensor node may need
to retransmit multiple times before the data packet can be
delivered successfully. This issue has been dealt with in
JAM by replacing Tr with ETX Tr in the end-to-end delay
requirement (i.e., Equation (2) in Section II-B), where ETX
is the expected number of transmission attempts to deliver a
data packet successfully over one hop:
SAD(i) +

(ETX(mc , m)Tr(m) + FAD(m)) = D.

JAMKernelM
NeighborMgmtM
Receive

Send

Routing
Send

Receive

MACControllerM
ReceiverM

SenderM

(14)

MACSchedulerM

mSi

RoutingInfo

Here, m is the child node of m along the path from i to the

sink. Similarly, Equations (6) and (10) have also been updated
to include the ETX information. Note that measurement of
ETX is readily available in many routing protocols such as
CTP [17], and thus not an extra overhead.
3) Adaptive Adjustment Interval W : The JAM kernel is
triggered to execute periodically every W interval, which
poses a tradeoff. A small interval, i.e., frequent adjustment,
makes JAM more responsive to the changes in the network and
allows it to approach a balanced nodal lifetime distribution in
the network sooner, but uses more computational resources.
A large interval, on the other hand, uses less computational
resources but may let nodes stay in suboptimal states for a
longer time.
We adopt an adaptive approach to adjust W dynamically
(between Wmin and Wmax ) to the network condition. Specifically, if the current state is already the best that can be found,
i.e., no in improves the minimal nodal lifetime within
the neighborhood as described in Section III-E, W is doubled
till reaching Wmax . Otherwise, W is reset to Wmin . This way,
when the nodal lifetime distribution is heterogeneous or the
network configuration changes, JAM allows the network to reconverge quickly to the balanced nodal lifetime distribution;
otherwise, it decreases the adjustment frequency exponentially
to save the control overhead in the long term. In the JAM
implementation, we set Wmin and Wmax conservatively to 1
minute and 16 minutes, respectively.

Receive

PowerControl

Send

RadioCore
Fig. 6.

Implementation of JAM in TinyOS.

When data packets from a child node j arrives at node

is MAC layer, the piggybacked control information will be
extracted, passed to, and processed at the JAMKernelM module, which implements the JAM kernel. The NeighborMgmtM
module manages all senders information while the FwdAggM
and SelfAggM modules maintain the aggregation timers. After
deciding the new Tr(i) and FAD(i) values, the JAMKernelM
module notifies the MACControllerM module to adopt the
latest wakeup interval, and the FwdAggM and SelfAggM
modules to adjust the aggregation timers. JAMKernelM also
informs MACControllerM of (i) and piggybacks it in the
ACKs to child nodes.
2) Hardware Component: Among all the control information piggybacked in data packets, nodal residual energy
is an important piece. As TelosB motes do not provide an
interface to measure nodal residual energy, we have designed
and fabricated a TelosB power meter kit as shown in Figure 7.
This kit measures the nodal power consumption rate, based on
which we can calculate the total energy consumed so far. The
nodal residual energy is then the difference between the battery
energy capacity [20] and the consumed energy.

200

AVG
LBA
JAM

150

(h)

150

(h)

200

AVG
LBA
JAM

100
50

100
50

0
20

30

40

50

20

e2e delay requirement (s)

(a) Network lifetime.

Fig. 7.

TelosB power meter kit used in JAM.

AVG
LBA
JAM

(mW)

(mW)

10

15
10
5
0

0
20

30

40

20

50

30

40

50

e2e delay requirement (s)

e2e delay requirement (s)

(c) Average nodal power consump- (d) Average nodal power consumption.
tion.
1

0.8

0.8

CDF

0.6
0.4

20s
30s
40s
50s

0
0

10

20

30

40

50

0.6
0.4

20s
30s
40s
50s

0.2
0
60

e2e delay (s)

In the simulation, nodes are randomly deployed in a 500 m

500 m area and the sink is located at the center of the
area. The evaluation results are averaged over 30 different
random topologies. We vary the end-to-end delay requirement,
initial nodal energy distribution, and the network density in
the simulation. The default initial nodal energy is 4500 Joules.
The maximal communication range is 70 meters and the power
consumption is 69 mW when radio is on. All nodes are sources
and the data generation rate is a random value between 0.1
and 1 packet per second. The packet size is 128 bytes. In both
simulations and testbed experiments, = 0.5 s, = 0.1 s,
Wmin = 1 minute, and Wmax = 16 minutes.
1) Performance under Different End-to-End Delay Requirements: Figure 8 compares the performances of all the evaluated schemes with the end-to-end delay requirement varying
between 20 s and 50 s. Left column of Figure 8 shows
evaluation results when all nodes start with 4500 J energy
while right column shows results when the initial nodal energy
is a random value between 4500 (1 40%) J.
Different from LBA which does not improve the network
lifetime much when the initial nodal energy distribution is
homogeneous as shown in Figure 8(a) (similar finding has also
been observed in [7]), JAM consistently improves the network
lifetime in both homogeneous and heterogeneous initial energy
settings under all end-to-end delay requirements. Specifically,
when all nodes start with the same amount of energy, compared with LBA, JAM improves the network lifetime by 158%
and 59% when the end-to-end delay requirement is 20 s and
50 s, respectively. When nodes start with different energy, the
improvement ratio is 165% and 50% when the requirement is
20 s and 50 s, respectively.

AVG
LBA
JAM

20

15

0.2

A. NS-2 Simulations

50

(b) Network lifetime.

CDF

NS-2 based simulations and TinyOS testbed experiments

have been conducted to evaluate the JAM performance in
terms of network lifetime, average nodal power consumption,
and end-to-end delivery delay. We compare JAM with a
naive scheme called AVG which simply sets FAD and Tr
values according to Section III-F1 without runtime adjustment,
and LBA [7] which is a state-of-the-art lifetime-balancing
aggregation protocol under delay constraint.

40

25

25
20

IV. P ERFORMANCE E VALUATION

30

e2e delay requirement (s)

(e) End-to-end delay in JAM.

10

20

30

40

50

60

e2e delay (s)

(f) End-to-end delay in JAM.

Fig. 8. Performance comparison under different end-to-end delay requirements. Number of nodes in the network is 60, and initial Tr is 1.5 s. Figures
on the left column are results when initial nodal energy is 4500 J while figures
on the right are results when initial nodal energy is a random value between
4500 (1 40%) J.

JAM yields more significant network lifetime improvement

when the end-to-end delay requirement is more stringent
because smaller delay bound means more limited FAD values
for each node along source-to-sink paths, which may result
in insufficient data suppression and hence heavier network
traffic. As MAC behaviors in AVG and LBA are not jointly
adjusted with aggregation behaviors, the communication overhead could be expensive. Hence, the heavy traffic could soon
deplete the nodal energy and constrain the network lifetime. In
comparison, JAM yields a much lower average nodal energy
consumption due to its joint MAC and aggregation design, as
shown in Figures 8(c) and 8(d).
Figures 8(e) and 8(f) plot the CDF (Cumulative Distribution Function) of the end-to-end delay for the JAM scheme.
Results show that all the end-to-end delay requirements are
well-satisfied. Similar to RI-MAC, JAM drops a data packet
after a certain number (4 in our implementation) of failed
transmission attempts. Therefore, when the end-to-end delay
requirement is relatively small, i.e., 20 s, the heavy traffic
deteriorates channel contentions and about 4% of data packets
are dropped, while the data delivery ratio approaches 100%
when the end-to-end delay requirement increases to 50 s. Due
to space limitation, we omit the CDF results of the end-to-

200

100

AVG
LBA
JAM

12
10

(mW)

150

(h)

14

AVG
LBA
JAM

200

14

AVG
LBA
JAM

AVG
LBA
JAM

12
10

(mW)

150

(h)

end delay for other evaluation scenarios, where JAM exhibits a

similar performance as the ones shown in Figures 8(e) and 8(f).
2) Performance under Different Initial Energy: We now
evaluate the impact of the initial nodal energy heterogeneity
level (denoted as ). It is defined as follows. With a heterogeneity level of , the initial nodal energy is a random value
between 4500 (1 ) Joules. Results are shown in Figure 9.

100

8
6
4

50

2
0

0
40

60

80

100

40

60

node number

80

100

node number

(a) Network lifetime.

(b) Average nodal power consumption.

Fig. 10. Performance comparison under different network densities. Endto-end delay requirement is 40 s, initial Tr is 1.5 s, and initial nodal energy
level is a random value between 4500 (1 40%) J.

8
6
4

50

2
0

0
0

0.2

0.4

(a) Network lifetime.

0.6

0.2

0.4

0.6

10

12
4

19

26

11

(b) Average nodal power consumption.

1
0

13

Fig. 9. Performance comparison under different initial nodal energy heterogeneity levels. Number of nodes in the network is 60, end-to-end delay
requirement is 40 s, and initial Tr is 1.5 s.

28

17

15

16
24

14

21
18

29

With AVG, as Tr and FAD are not adjusted at runtime, network lifetime is bounded by the initial minimal-lifetime node.
As the initial energy distribution becomes more and more
heterogeneous (i.e., increases), network lifetime achieved by
AVG drops quickly. On the other hand, both LBA and JAM redistribute the end-to-end delay so that energy bottleneck nodes
could be saved by other high-energy nodes along the same
route and hence yield a significantly longer network lifetime.
As a larger demands more re-distribution efforts, the network
lifetime drops as well with LBA and JAM. However, as all
the nodes along the route absorb this effect collaboratively,
the network lifetime drops less quickly. For example, network
lifetime drops 21% with JAM when increases from 0% to
60%, in comparison to 46% with AVG. Finally, thanks to the
joint MAC and aggregation design, JAM yields a consistently
70% longer network lifetime than LBA under different
values.
3) Performance under Different Densities: As shown in
Figure 10, JAM always yields a significantly longer network
lifetime than other schemes, regardless of the network density.
It is interesting to see (in Figure 10(b)) that the average
nodal power consumption decreases as the network density
increases. This is because, with a higher network density and
consequently a higher node degree on average, a source node
may reach the sink via a shorter path. This means less nodes
are involved in the path and the overall energy consumption
is thus reduced.
B. Testbed Experiments
We set up a testbed network of 32 TelosB motes, forming a
fixed tree topology shown in Figure 11. All nodes are sources,
and the data generation interval is uniformly distributed between 0.8 and 1.2 s. The default Tr value is 1.5 s, and the
end-to-end delay requirement varies between 20 s and 40 s.

20
30

31

Fig. 11.

25

23

22
27

Network topology in testbed experiments.

In order to complete the experiments within a reasonable

amount of time, we study how fast a node consumes a small
designated amount of energy, and evaluate its nodal lifetime
as the time period during which the designated amount of
energy is consumed. We run two sets of experiments. In the
first set, the initial available energy distribution is uniform and
all nodes have 450 Joules; results are shown in the left column
of Figure 12. In the second set, the initial available energy at
an individual node is a random value between 250 and 450
Joules; results are shown in the right column of Figure 12.
Overall, experiment results confirm our observations in the
simulation study.
To illustrate how JAM adaptively tunes MAC and aggregation operational parameters to balance nodal lifetime within
the neighborhood, we plot in Figure 13 the changing traces
of the operational parameters of the nodes along the path
24 21 16. We observe that during time period [0, 0.2] h,
as shown in Figure 13(a), node 21 has a lower nodal lifetime
than nodes 16 and 24. To improve node 21s lifetime, (i) node
16, as node 21s parent, decreases its Tr to save node 21s
energy cost on idly waiting during transmissions; (ii) node 21
increases its FAD to reduce the amount of outgoing traffic; and
(iii) node 24 decreases its Tr to ensure the delay requirement
is satisfied. As a result, their nodal lifetimes are balanced
gradually. During time period [0.2, 0.5] h, as all three nodes
have reached a similar level of nodal lifetime, their operational
parameters are relatively stabilized.

AVG
LBA
JAM

6
(h)

6
(h)

AVG
LBA
JAM

4
2

ACKNOWLEDGMENT

This work is supported partly by the NSF under grants CNS0831874 and EECS-1128312, and by the ONR under grant
N00014-09-1-0748.

0
20

30

40

20

e2e delay requirement (s)

30

40

e2e delay requirement (s)

(a) Network lifetime.

(b) Network lifetime.

AVG
LBA
JAM

40
30

AVG
LBA
JAM

40
(mW)

(mW)

and aggregation design may prolong the network lifetime even

further.

20
10

30
20
10

0
20

30

40

20

e2e delay requirement (s)

30

40

e2e delay requirement (s)

(c) Average nodal power consump- (d) Average nodal power consumption.
tion.

(h)

Fig. 12. Experiment results under different end-to-end delay requirements.

Figures on the left column are results when initial available energy is 450 J
while figures on the right are results when initial available energy is a random
value between 250 and 450 J.
12
10
8
6
4
2
0

node 16

node 21

0.2

node 24

0.5
Time (h)

(s)

(a) Nodal lifetime.

10
8
6
4
2
0
0

0.2

0.5

(s)

(b) Nodal FAD values.

5
4
3
2
1
0
0

0.2

0.5

(c) Nodal Tr values.

Fig. 13. Changing traces of nodal lifetime, FAD, and Tr values along the
path 24 21 16.

V. C ONCLUSIONS

AND

F UTURE W ORK

In this paper, we present a new holistic approach called JAM

to prolong the sensor network lifetime. Different from the existing works which adapt either MAC or aggregation behavior
alone, JAM integrates the advantages of both approaches and
therefore can extend the network lifetime more effectively. In
addition, JAM can also meet the end-to-end delay requirement
specified by the applications.
JAM was designed to work with static data collection trees.
In the future, we will improve the JAM design by taking
into account the routing strategy. As routing behavior also
affects the network traffic distribution, a joint MAC, routing,

R EFERENCES
[1] W. Wang, V. Srinivasan, and K. C. Chua, Using Mobile Relays to
Prolong the Lifetime of Wireless Sensor Networks, in MobiCom, 2005.
[2] J. Chang and L. Tassiulas, Energy Conserving Routing in Wireless
Ad-hoc Networks, in INFOCOM, 2000.
[3] , Maximum Lifetime Routing in Wireless Sensor Networks,
IEEE/ACM Trans. Netw., 2004.
[4] R. Braynard, A. Silberstein, and C. Ellis, Extending Network Lifetime
Using an Automatically Tuned Energy-Aware MAC Protocol, in EWSN,
2006.
[5] Z. Li, M. Li, and Y. Liu, Towards Energy-Fairness in Asynchronous
Duty-Cycling Sensor Networks, in INFOCOM, 2012.
[6] A. Meier, M. Woehrle, M. Zimmerling, and L. Thiele, ZeroCal:
Automatic MAC Protocol Calibration, in DCOSS, 2010.
[7] Z. Li, Y. Peng, D. Qiao, and W. Zhang, LBA: Lifetime Balanced Data
Aggregation in Low Duty Cycle Sensor Networks, in INFOCOM, 2012.
[8] L. Becchetti, A. Marchetti-Spaccamela, A. Vitaletti, P. Korteweg,
M. Skutella, and L. Stougie, Latency-Constrained Aggregation in
Sensor Networks, ACM Trans. Algorithms, 2009.
[9] S. R. Madden, M. J. Franklin, J. M. Hellerstein, and W. Hong, TinyDB:
An Acquisitional Query Processing System for Sensor Networks, ACM
Trans. Database Syst, 2005.
[10] R. Kumar, M. Wolenetz, B. Agarwalla, J. Shin, P. Hutto, A. Paul, and
U. Ramachandran, DFuse: a Framework for Distributed Data Fusion,
in SenSys, 2003.
[11] M. Zimmerling, F. Ferrari, L. Mottola, T. Voigt, and L. Thiele, pTunes:
Runtime Parameter Adaptation for Low-Power MAC Protocols, in
IPSN, 2012.
[12] Q. Xiang, J. Xu, X. Liu, H. Zhang, and L. Rittle, When In-Network
Processing Meets Time: Complexity and Effects of Joint Optimization
in Wireless Sensor Networks, in RTSS, 2009.
[13] Z. Ye, A. Abouzeid, and J. Ai, Optimal Policies for Distributed Data
Aggregation in Wireless Sensor Networks, in INFOCOM, 2007.
[14] J. Zhang, X. Jia, and G. Xing, Real-time Data Aggregation in
Contention-based Wireless Sensor Networks, ACM Trans. Sen. Netw.,
2010.
[15] C. Hua and T.-S. P. Yum, Optimal Routing and Data Aggregation
for Maximizing Lifetime of Wireless Sensor Networks, IEEE/ACM
Transactions on Networking, vol. 16, no. 4, pp. 892 903, 2008.
[16] L. Xiang, J. Luo, and A. V. Vasilakos, Compressed Data Aggregation
for Energy Efficient Wireless Sensor Networks, in SECON, 2011.
[17] O. Gnawali, R. Fonseca, K. Jamieson, D. Moss, and P. Levis, Collection
Tree Protocol, in SenSys, 2009.
[18] S. F. Madden, M. J. Hellerstein, and W. J. M. Hong, TAG: A Tiny
AGgregation Service for Ad-Hoc Sensor Networks, in OPERATING
SYSTEMS REVIEW, 2002, VOL 36.
[19] Y. Sun, O. Gurewitz and D. Johnson, RI-MAC: a receiver-initiated
asynchronous duty cycle MAC protocol for dynamic traffic loads in
wireless sensor networks, in SenSys, 2008.
[20] Online Link, http://www.allaboutbatteries.com/Energy-tables.html.