Energy-Latency Tradeoffs for Data Gathering in

Wireless Sensor Networks

Yang Yu, Bhaskar Krishnamachari, and Viktor K. Prasanna
Department of Electrical Engineering
University of Southern California
Los Angeles, CA 90089-2562
{yangyu, bkrishna, prasanna}@usc.edu

Abstract— We study the problem of scheduling packet trans- is due to data gathering, it is crucial to design energy-efficient
missions for data gathering in wireless sensor networks. The focus communication strategies in implementing such an operation.
is to explore the energy-latency tradeoffs in wireless communi- One useful approach for energy-efficient communication is
cation using techniques such as modulation scaling. The data
aggregation tree – a multiple-source single-sink communication to explore the energy-latency tradeoffs. An important obser-
paradigm – is employed for abstracting the packet flow. We vation in [4] is that in many channel coding schemes, the
consider a real-time scenario where the data gathering must transmission energy can be significantly reduced by lowering
be performed within a specified latency constraint. We present transmission power and increasing the duration of transmis-
algorithms to minimize the overall energy dissipation of the sion. Techniques such as modulation scaling [5] have been
sensor nodes in the aggregation tree subject to the latency
constraint. For the off-line problem, we propose (a) a numerical proposed for implementing such tradeoffs.
algorithm for the optimal solution, and (b) a pseudo-polynomial In this paper, we explore the above tradeoffs in the context
time approximation algorithm based on dynamic programming. of data gathering in WSNs, subject to application level perfor-
We also discuss techniques for handling interference among mance constraints. We consider a real time scenario where the
the sensor nodes. Simulations have been conducted for both raw data gathered from the source nodes must be aggregated
long-range communication and short-range communication. The
simulation results show that compared with the classic shut- and transmitted to the sink within a specified latency con-
down technique, between 20% to 90% energy savings can be straint. Our technique is applicable to any given aggregation
achieved by our techniques, under different settings of several function. The objective function is to minimize the overall
key system parameters. We also develop an on-line distributed energy dissipation of the sensor nodes in the aggregation tree
protocol that relies only on the local information available at each subject to the latency constraint. Compared with [4], [6], we
sensor node within the aggregation tree. Simulation results show
that between 15% to 90% energy conservation can be achieved use a more general and accurate energy model for abstracting
by the on-line protocol. The adaptability of the protocol with the energy characteristics for packet transmission in WSNs.
respect to variations in the packet size and latency constraint is Specifically, the transmission energy does not monotonically
also demonstrated through several run-time scenarios. decrease as the transmission time increases – the transmission
Index terms – System design, Mathematical optimization energy may increase when the transmission time exceeds some
threshold value [7]. We refer to the above general model as
the non-monotonic energy model.
I. I NTRODUCTION
For the off-line version of the problem, we present (a) a
In many applications of wireless sensor networks numerical algorithm for the optimal solution, and (b) a pseudo-
(WSNs) [1], data gathering is a critical operation needed polynomial time approximation algorithm based on dynamic
for extracting useful information from the operating environ- programming. We also discuss techniques for handling in-
ment. Recent studies [2], [3] show that data aggregation is terference. Simulations were conducted for both long-range
particularly useful in eliminating the data redundancy and communication (with radius around 32 m) and short-range
reducing the communication load. Typical communication communication (with radius around 7 m). The simulation
patterns in data aggregation involve multiple data sources and results from the scenarios we studied show that compared with
one data sink (or recipient). Thus, the corresponding packet the classic technique that transmits the packets at the highest
flow resembles a reverse-multicast structure, which is called speed and shut down the radio afterwards, between 20% to
the data aggregation tree. 90% energy savings can be achieved by our techniques, under
Energy-efficiency is a key concern in WSNs. The large different settings of several key system parameters. We also
number of sensor nodes involved in such networks and the develop an on-line distributed protocol that needs only local
need to operate over a long period of time require careful information of the aggregation tree. Simulation results show
management of the energy resources. In addition, wireless that between 15% to 90% energy conservation can be achieved
communication is a major source of power consumption. by the on-line protocol. The adaptability of the protocol is also
Since a significant portion of the communication in WSNs demonstrated through several run-time scenarios.
This work is supported by NSF under grant IIS-0330445 and by an ITR Related work: The most relevant works include [4]–[6], [8],
grant under award number 0325875. [9]. The problem of minimizing the energy dissipation for

0-7803-8356-7/04/$20.00 (C) 2004 IEEE IEEE INFOCOM 2004

 transmitting a set of packets over a single-hop link subject We consider the scenario that the data gathering must
to a specified latency constraint is discussed in [4]. An be completed within a specified latency constraint, which is
extension of the problem that considers a single transmitter necessary for real-time monitoring or mission-critical appli-
and multiple receivers is investigated in [6]. In [5], an on-line cations. To enforce the latency constraint requires the use
policy for adjusting modulation level is proposed for single- of time-synchronization schemes such as [11]. The recently
hop communication. In [8], modulation scaling is integrated proposed epoch-based scheme (refer to [10]) instantiates the
into the Weighted Fair Queuing (WFQ) scheduling policy. latency constraint by the length of each epoch. Prior work,
In [9], the problem of balancing the energy dissipation along however, has not considered the possibility of using packet-
a multi-hop communication path is studied. scheduling techniques that trade latency for energy in such a
To the best of our knowledge, this is the first paper that scenario. This is the focus of our work.
addresses packet scheduling in a general tree structure. The
challenges of our problem are multi-fold. Firstly, the energy
B. Non-Monotonic Energy Model
functions can vary for different links. It is therefore required
to develop general optimization techniques instead of explicit We model the transmission energy using the example of
solutions. Secondly, the latency constraint for data gathering modulation scaling [5] based on Quadrature Ampitude Modu-
in real applications is typically given by considering the lation (QAM) scheme [12]. Note that the algorithms presented
aggregation tree as a whole. It is difficult to directly apply in this paper are extendible to other modulation schemes as
the techniques in [4] and [6], as they require explicit latency well as other techniques that provide energy-latency tradeoffs,
constraints over each link. Lastly, we consider non-monotonic such as code scaling [13]. Consider a packet of s bits to
energy functions, which has not been previously addressed. be transmitted between two sensor nodes. Assuming that the
Paper Organization: We discuss the background of our work symbol rate, R, is fixed, the transmission time, τ , can be
in Section II. The packet transmission problem is defined in calculated as [5]:
Section III. Off-line algorithms for the problem are presented s
τ= , (2)
in Section IV. In Section V, a distributed on-line protocol is b·R
described. Simulation results are shown in Section VI. Finally, where b is the modulation level of the sender in terms of
concluding remarks are made in Section VII. the constellation size (number of bits per symbol). The corre-
sponding transmission energy can be modeled as the sum of
II. BACKGROUND
output energy and electronics energy. Though the transmission
A. Data Aggregation Paradigm energy essentially depends on the setting of b, we would like
We abstract the underlying structure of the network as a data to characterize it as a function of τ [5], denoted as w(τ ), to
aggregation tree. This is essentially a tree that aggregates and illustrate the key energy-latency tradeoffs in this paper.
gathers information from multiple sources enroute to the sink. s

Such a topological structure is common to data-centric routing w(τ ) = [C · (2 τ ·R − 1) + F ] · τ · R , (3)

schemes for sensor networks such as Directed Diffusion [2], where C is determined by the quality of transmission (in terms
[3]. While there may be transients during the route creation of Bit Error Rate) and the noise power, and F is a device-
phase, we assume that this tree, once formed, lasts for a dependent parameter that determines the power consumption
reasonable period of time and provides the routing substrate of the electronic circuitry of the sender. Further, the output
over which aggregation can take place during data gathering. power, Po , and the electronics power, Pe , can be modeled as
Specific techniques have been previously proposed for com- follows [5]:
puting aggregates on such trees [10]. For our analysis we
make the following abstraction: each sensor node in the tree Po = C · R · (2b − 1) and (4)
aggregates the information from all its children or by local Pe = F ·R. (5)
sensing so that it results in a reduced size packet that is
dependent on the subtree rooted at the sensor node and an We consider the radio modules from [7], [14]. Typically, for
aggregation factor, k ∈ [0, 1]. For simplicity, we assume that short-range communication with R = 1 Mbaud, the electronics
each source node generates a data packet with the same size, power of the radio is approximately 10 mW, while the output
s. Let d denote the number of source nodes in the subtree power is approximately 1 mW (at 4-QAM). From equations (4)
rooted at a sensor node, and s
denote the amount of output and (5), it can be derived that C ≈ 3 × 10−10 and F = 10−8 .
data after aggregation. Intuitively, low correlations among data Further, we consider a d2 power loss model, where d is the
make s
close to ds, while high correlations make s
close communication radius. Assuming that it takes 10 pJ/bit/m2 by
to s. We use the above aggregation factor, k, to indicate the the amplifier to transmit one bit at an acceptable quality
√ [15],
degree of correlations among data with k = 1 meaning the we infer that the desired communication radius is 50 ≈ 7
2
highest correlations and k = 0, the lowest. Based on the above m (from 2×101 6mW 2
bit/sec = 10 pJ/bit/m × d ). In our study, we
intuition, we abstract the relationship between s
and s using consider one more case of communication √ in WSNs – long-
equation (1). It can be verified that s
= s when k = 1 and range communication with radius at 1000 ≈ 32 m and the
s
= ds when k = 0. output power at 20 mW for 4-QAM (C ≈ 7 × 10−9 ).
ds Figure 1 plots the energy functions with b ∈ [2, 8] for
s
= (1) the long and short range communication based on the above
dk − k + 1

0-7803-8356-7/04/$20.00 (C) 2004 IEEE IEEE INFOCOM 2004

 analysis. In practice, b is typically set to positive integers edges from any leaf node, Vi , i ∈ {1, . . . , M }, to Vn , denoted
(indicated by circles in the figure), resulting in discrete values as pi . We use the notation Vj ∈ pi to signify that node Vj is
of τ . It can be observed that the the transmission energy for an intermediate node of path pi .
the short-range communication eventually increases after the We assume that sensor nodes are completely shutdown when
transmission time exceeds 300 nSec. Intuitively, it is more there is no packet to transmit or receive. Mechanisms such as
beneficial to explore the energy-latency tradeoffs for the long- signaling channel [1] can be used for synchronization between
range communication. However, we demonstrate in Section VI sensor nodes before any packet transmission. However, the
that up to 60% energy savings can still be achieved by our modeling of power assumption of such mechanisms is beyond
algorithms for the short-range communication. the scope of this paper.
−7
x 10
B. Problem Definition
b=8 long−range
short−range Let Γ denote the latency constraint. For ease of analysis, it
2
is assumed that raw data is available at source nodes at time
0. Further, we assume that the energy functions for all links
1.5 in the target aggregation tree follow the model described in
energy (J)

Section II.
A schedule of packet transmission is defined as a vector
1
τ = {τi : i = 1, . . . , n − 1}, where τi is the time duration for
b=6
packet transmission over link (i, j). Since a sensor node can
0.5 transmit its packet only after receiving all input packets from
b=4
b=2 its children, the start time of each transmission is implicitly
determined by τ . The transmission latency

of a path, pi , is
0
1 2 3 4 5 denoted as Li and calculated as Li = j:Vj ∈pi τj . A schedule
transmission time duration (Sec) −7
x 10 is feasible if for each pi ∈ T , we have Li ≤ Γ.
Our goal is to improve the energy-efficiency of the system.
Fig. 1. Energy-latency tradeoffs for transmitting one bit data
Various objective functions can be developed for interpreting
energy-efficiency. For ease of analysis, the objective function
III. PACKET T RANSMISSION P ROBLEM OVER DATA defined in this paper is the overall energy dissipation of the
AGGREGATION T REES sensor nodes in the aggregation tree.
A. Data Aggregation Tree Let wi (τ ) denote the energy function of sensor node Vi ,
with mi denoting the value of τ ∈ (0, T ] when wi (·) is
Let T = (V, E) denote the data aggregation tree, where V
minimized. Moreover, by assuming a first order energy model,
denotes the set of n sensor nodes, {Vi : i = 1, . . . , n}, and
the reception energy can be modeled by doubling the value
E denotes the set of directed communication links between
of F in equation (3). Thus, we state the packet transmission
the sensor nodes. Let M denote the number of leaf nodes in
problem (called PTP) as follows:
the tree. Without loss of generality, we assume that the sensor
Given:
nodes are indexed in the topological order with V1 , . . . , VM
a. a data aggregation tree T consisting of n sensor nodes,
denoting the M leaf nodes and Vn denoting the sink node.
b. energy functions for each link (i, j) ∈ E, wi (τ ), and
Every link in E is represented as a tuple (i, j), meaning that
c. the latency constraint, Γ;
a packet, denoted as Pi , needs to be transmitted from Vi to
find a schedule of packet transmission, τ , so as to minimize
Vj . Let si denote the size of Pi .
Raw data is generated by a set of source nodes from V 
n−1

(not necessarily leaf nodes). Data aggregation is performed f (τ ) = wi (τi ) (6)
by any non-sink and non-leaf node (called an internal node i=1

hereafter). We assume that aggregation is performed only subject to

after all input information is available – either received from 
children, or generated by local sensing. The aggregated data ∀pi in T, Li = τj ≤ Γ . (7)
j:Vj ∈pi
is then transmitted to the parent node. Although we use the
expression in equation (1) as a typical aggregation function, The above formulation differs from the problem defined
please note that our technique is not limited to this function in [6] in two key aspects. (1) We employ a tree structure packet
alone. The only requirement is that we can derive the value of flow where the latency constraint is imposed on each path of
si ’s based on the functions. Thus, even different functions can the tree. (2) The non-monotonic energy model in Section II-
be used to specify the aggregation at different sensor nodes. B indicates an upper-bound on the transmission time of each
The time and energy costs for generating raw data at source packet, i.e., to optimize PTP, we should have τi ≤ mi , for each
nodes or aggregating data at internal nodes are considered to i = 1, . . . , n − 1. The consequences of the above differences
be negligible. are discussed in Section IV-A.
Let Ti denote the subtree rooted at any node, Vi , with Tn = We note that the above model assumes no MAC layer
T . A path in T is defined as a series of alternate nodes and interference, which can be realized by multi-packet reception

0-7803-8356-7/04/$20.00 (C) 2004 IEEE IEEE INFOCOM 2004

 Begin
1. Set k ← 0 // initialize iteration counter
2. For (i, n) ∈ E, set τik ← min{Γ, mi } // initialize transmission time for links to the sink
3. For (i, j) ∈ E such that j = n, set τik ← 0 // initialize transmission time for other links
4. Set f lag ← 0 // flag to keep track of convergence in the iterations
5. While f lag = 0
6. k ← k + 1 // increment the iteration counter by 1
7. For each Vi with i from n − 1 downto M+1 // perform local optimization for each internal node
8. ({τjk }(j,i)∈E , τik ) ← best({τjk−1 }(j,i)∈E , τik−1 ) // move right the start time of transmission from Vi
9. For (i, n) ∈ E
10. Set τik ← min{mi , Γ − (maxVi ∈pj {Lj } − τik )} //increase the transmission time for links to the sink
11. if τ = τ k−1 , f lag ← 1
k
// check convergence
End

Fig. 2. Pseudo code for EMR-Algo

(MPR) techniques [16]. We further elaborate this issue in extended MoveRight algorithm (EMR-Algo) is shown in Fig-
Section IV-C. ure 2. In the figure, τik denotes the value of τi after the k-
th iteration. Initially, we set the starting time for all packet
IV. O FF - LINE A LGORITHMS FOR PTP transmission to zero – the transmission time for all the links
to the sink is set to min{Γ, mi }, while the transmission time
In this section, we consider an off-line version of PTP for the rest links is set to 0 (Steps 2 and 3). The main idea
(called OPTP) by assuming that the structure of the aggre- is to iteratively increase (move right) the starting times of
gation tree and the energy functions for all sensor nodes packet transmissions, so that each move locally optimizes our
are known a priori. We first describe an extension of the objective function. Finally, this iterative local optimization
MoveRight algorithm [6] to get an optimal solution for OPTP. leads to a globally optimal solution.
A faster dynamic programming based approximation algorithm The best(·) function returns the transmission durations for
is then presented. Techniques for handling interference are also node Vi and its children, such that Lemma 1 holds for Vi with
discussed. respect to the invariant that τik ≤ mi . Since the value of τik
must lie within (0, τik−1 ], the best(·) function can be easily
A. A Numerical Optimization Algorithm implemented using binary search. Step 10 is important as it
moves right the complete time of transmissions on links to the
Since we must have τi ≤ mi in an optimal solution to sink. This movement stops when the latency constraint of all
OPTP, the latency of a path does not necessarily equal Γ. We paths is reached.
show the following necessary and sufficient condition for the The proposed EMR-Algo is distinguished from the
optimality of the OPTP problem. MoveRight algorithm in two key respects. (recall the differ-
Lemma 1: A schedule, τ∗ , is optimal for OPTP iff ences between our problem and the one defined in [6]). (1) The
1) for any node Vi with τi∗ < mi , the length of at least one best(·) function respects Lemma 1 regarding the optimality of
path that contains Vi equals Γ; and OPTP in a tree structure. (2) The transmission time for any
2) for any internal node, Vi , we have Vi ∈ V is bounded by mi , enforced by lines 2, 8 and 10.
 The correctness of EMR-Algo can be proved by exploring
ẇi (τi∗ ) = ẇj (τj∗ ) . (8) the convexity property of the energy functions. Let τ∗ =
(j,i)∈E
{τ1∗ , . . . , τn−1
∗
} be the optimal schedule. Let s∗i = 0, for
Corollary 1: Consider an optimal schedule, τ∗ , for OPTP; i = 1, . . . , M ; and s∗i = max(j,i)∈E (s∗j + τj∗ ), for i =
the following hold: M +1, . . . , n−1. As previously stated, {τ1k , . . . , τn−1
k
} indicate
1) If τi∗ = mi for some Vi ∈ V , we have τj∗ = mj for all the transmission time of nodes V1 , . . . , Vn−1 after the k-th
sensor nodes in Ti . pass of EMR-Algo. Let ski = 0, for i = 1, . . . , M , and
2) If τi∗ < mi for some Vi ∈ V , we have τj∗ < mj for all ski = max(j,i)∈E (skj + τjk ), for i = M, . . . , n − 1. We have:
ancestors of Vi . Theorem 1: Let ski and s∗i , i = 1, . . . , n − 1 be as defined
Due to space limitations, the proof of Lemma 1 and Corol- above. Then
lary 1 is omitted in this paper. Details of the proof can be
1) ski ≤ sk+1
i ;
found in [17].
2) ski ≤ s∗i ; and
In this section, we extend the MoveRight algorithm from [6]
to solve OPTP in a general-structured aggregation tree with 3) s∞i = si .
∗

non-monotonic energy functions. The pseudo code for the The proof of Theorem 1 is presented in Appendix I.

0-7803-8356-7/04/$20.00 (C) 2004 IEEE IEEE INFOCOM 2004

 The convergence speed of EMR-Algo depends on the struc-
ture of the aggregation tree and the exact form of the energy Γ g(V1, Γ) g(Vn, Γ)
functions. It is therefore difficult to give a theoretical upper
bound on the number of iterations. In Section VI, we show the
running time of EMR-Algo for simulated problems. However,
by approximating wi (τ ) with a set of interpolated discrete jε g(V1, jε) g(Vi, jε)
values, we develop a pseudo-polynomial time approximation
algorithm based on dynamic programming. We present the
approximation algorithm in Section IV-B.

g(V1, 3ε)
B. A Dynamic Programming Based Approximation Algorithm
2ε g(V1, 2ε)
For ease of analysis, we assume that for each sensor node, D ε g(V1, ε) g(V2, ε) g(Vi, ε) g(Vn, ε)
discrete values are evenly distributed over [0, Γ] in the domain
of τ . Let ε be the difference between two adjacent values. That V1 V2 Vi Vn
Γ
is ε = D . Hereafter, D is called the approximation accuracy. Fig. 3. The g(·) table computed by DP-Algo
Higher value of D leads to a more accurate approximation
of the energy function. By changing D, we can explore the
tradeoffs between the quality of the solution and the time cost τi that can be achieved by an available modulation level. We
of the algorithm. call the above method the rounding procedure.
Let g(Vi , t) denote the minimal overall energy dissipation
of a subtree Ti rooted at Vi within latency constraint t. The C. Handling Interference
original OPTP problem can be expressed as g(Vn , Γ). It is The definition of OPTP implicitly assumes that there is
clear that for any sensor node Vi , g(Vi , t) can be computed as no interference among the sensor nodes. Such an assumption
the sum of (a) the energy dissipation for packet transmission can be realized by using MAC layer scheduling or multi-
by the children of Vi , and (b) the energy dissipated by packet reception (MPR) through spatial, time, frequency, or
transmitting packets within the subtrees rooted at each child of code diversity [16]. However, use of such techniques may
Vi . Additionally, the packet transmission time from any child increase the hardware cost of the sensor nodes. In case the
of Vi can take εt values, namely ε, 2ε, . . . , t. Therefore, we above techniques are not available, one possible way for han-
have the following recursive representation of g(Vi , t): dling interference is to intentionally set the latency constraint
 imposed on OPTP to be less than the actual constraint. The
wi (t), for 1 ≤ i ≤ M preserved laxity can then be used for accommodating the back-


  t
ε off time of the sensor nodes when collision occurs.
g(Vi , t) = (min{wk (jε) + g(Vk , t − jε)}), (9) A more systematic way is to carefully schedule the transmis-

 j=1

(k,i)∈E sion of sensor nodes that can potentially interfere (or simply
otherwise. interfere) with each other. The goal is to ensure that the
The above representation is suitable for a dynamic program- corresponding time periods for a group of interfering sensor
ming based algorithm (DP-Algo for short). DP-Algo can be nodes do not overlap with each other. Intuitively, children of
viewed as a procedure to build a table of size D×n (Figure 3). a sensor node are interfering – they cannot send packets to
The i-th column from the left side corresponds to sensor node the parent at the same time. In the following, we describe
Vi , while the j-th row from bottom-up corresponds to jε. After a modified DP-Algo under the hypothesis that any group of
the execution of DP-Algo, the cell crossed by the j-th row and interfering sensor nodes are children of the same node. Such
the i-th column shall contain the value of g(Vi , jε). a hypothesis can be supported by carefully reconstructing the
To build the table, we start from the bottom left cell aggregation tree (refer to Appendix II for details).
that contains g(V1 , ε) = w1 (ε). The table is then completed To solve OPTP with the above interference restriction is
column by column, from left to right. To calculate the value actually non-trivial, as for any sensor node, the order of packet
of g(Vi , jε) for i > M , we need to compare, for each child of transmission from its children matters – the child that transmits
Vi , j different values by varying the packet transmission time earlier has a larger latency constraint over the subtree rooted
of the child. Therefore, the time cost for building up the table at the child. Our basic idea is to divide the latency constraint
is O(( Γε )2 (|V | + |E|), which is pseudo-polynomial due to the over any subtree Ti (rooted at Vi ) into two consecutive parts.
factor Γ2 . We schedule the packet transmission in the subtrees rooted at
A Special Case for Modulation Scaling: In practice, the each child of Vi with respect to the first part of the latency
modulation levels are typically set to positive even integers. constraint. The packets transmitted to Vi from its children are
Based on equation (2), it can be verified that the τi ’s resulted then scheduled in the second part. Hence, the order of packet
from different modulation levels are not evenly distributed transmission in the second part has no effects on the packet
among [0, Γ]. Thus, DP-Algo cannot be directly applied. scheduling in the first part. The optimal division of the latency
However, one practical method is to, for each i, set τi obtained constraint over Ti can be found using dynamic programming
by EMR-Algo or DP-Algo to the largest time duration below with the following recursive representation of g(Vi , t):

0-7803-8356-7/04/$20.00 (C) 2004 IEEE IEEE INFOCOM 2004

 for sequentializing the packet transmissions in each group

w (t), for 1 ≤ i ≤ M of interfering nodes can be accommodated by intentionally
 ti

 ε  reducing the latency constraint. When MPR technique is used,
g(Vi , t) = min{z(i, jε) + g(Vk , t − jε)}, (10) however, there is no impact on the latency constraint.

 j=1

 (k,i)∈E In the following, we first describe the local data structure
otherwise. maintained at each sensor node. A distributed adaptation
The function z(i, jε) returns a schedule for the packets policy for minimizing the energy dissipation is then proposed.
from children of Vi within time duration jε so that run- Local Data Structure: Each sensor node, Vi , maintains a
time contentions can be avoided. Obviously, if j is less than simple local data structure (r, τi , τd ). The flag r equals one
the number of children of Vi , no feasible solution exists. if Vi is the node with the highest positive energy gradient in
Otherwise, we use the following greedy heuristic. Initially, the the subtree rooted at Vi , Ti , and zero otherwise. Field τi is
transmission time from all children of Vi to Vi are set to ε. the time cost for transmitting the packet from Vi to its parent,
Let the energy gradient of a sensor node be the energy gain while τd records the time cost for data gathering within subtree
that can be obtained by increasing the current transmission Ti (excluding τi ).
time of the node by ε. Note that the energy gradient can The local data structure is maintained as follows. Every leaf
be negative due to the non-monotonic energy functions. We node attaches its energy gradient to the outgoing packet. Once
then increase the transmission time of the child with the a sensor node, Vi , receives packets from all its children, the
maximal positive energy gradient by ε. The above operation node compares the energy gradients attached to each packet
is repeated until the sum of the transmission time of all and the energy gradient of its own. The value of r at Vi is
children reaches jε, or no more energy savings can be achieved then set accordingly. If Vi is not the sink, the largest energy
by increase the transmission time (i.e., the gradients of all gradient from the above comparison is attached to the packet
children are negative). We call the modified DP-Algo as the sent to the parent of Vi . The above procedure continues till all
DP-IA algorithm. It can be verified that the time complexity the sensor nodes have the correct value of r. Fields τi and τd
of DP-IA is also O(( Γε )2 (|V | + |E|). can be easily maintained based on the above assumptions.
Adaptation Policy: The sink node periodically disseminates
a feedback packet to its children that contains the value of its
V. D ISTRIBUTED O N - LINE P ROTOCOL
local τd and the difference between Γ and τd , denoted as δ.
The algorithms presented in Section IV all assume a com- Note that if Directed Diffusion [2] is used for maintaining the
plete knowledge of the aggregation tree. However, the discrete aggregation tree, the feedback packet can be easily embedded
approximation of the energy function motivates an on-line into the interest packet sent by the sink.
distributed protocol that relies on local information of the Once a sensor node, Vi , receives the feedback packet, it
aggregation tree only. To facilitate the on-line scheduling, we checks its local data and performs one of the following actions.
make the following assumptions: To distinguish from the field τd in Vi ’s local data, let τd
denote
1) Some local unique neighbor identification mechanisms the field τd in the feedback packet.
are available at each sensor node for identifying the parent 1) If δ < 0, the transmission time for packet from Vi is
and children. decreased by a factor of β, where β is a user-specified
2) Every sensor node Vi can derive the time cost for data parameter. The feedback packet is then forwarded to all
gathering within subtree Ti . of Vi ’s children.
3) Every sensor node is able to measure its contemporary 2) If r = 1 and δ ≥ ε, the transmission time of Vi ’s
power consumption, and hence its energy gradient – outgoing packet is increased by ε. The local data structure
the energy gain that can be obtained by increasing the at Vi is updated accordingly; and the feedback packet is
transmission time of the node by ε. suppressed.
4) Interference among sensor nodes is minimized by using 3) Otherwise, the feedback packet is updated by setting δ =
either MPR techniques or MAC layer scheduling. δ + (τd
− τi − τd ) and τd
= τd . The updated packet is
The local identifier in assumption 1 is commonly imple- then forwarded to all children of Vi .
mented in protocols such as Directed Diffusion [2]. The The rationale behind the above adaptation policy is that
parent and children information is set up after constructing the when the latency constraint is violated, all the sensor nodes
aggregation tree. Assumption 2 can be fulfilled by attaching send out packets in an increased speed. If Vi is the node with
a time stamp to each packet from the leaf nodes (we shall the largest positive energy gradient in Ti and the latency laxity
be assuming that time synchronization schemes, such as [11], allows, the second action is performed to reduce the energy
are available). In assumption 3, the power consumption and dissipation of Vi . Otherwise, the latency laxity is accumulated
energy gradient of a sensor node can be determined using and the sensor nodes in Ti are recursively examined.
the system parameters provided by the hardware vendors Discussion: During each dissemination of the feedback packet,
and the operating configuration of the system, such as the the proposed on-line protocol increases the transmission time
modulation level. Assumption 4 only applies to each group for at most one sensor node per path. Such an increment
of interfering nodes, which are children of the same sensor is guaranteed not to violate the latency constraint for each
node from the hypothesis in Section IV-C. If the interference path. Therefore, the on-line protocol converges when the
is handled by MAC layer scheduling, the incurred time cost transmission latency of all paths reach the latency constraint,

0-7803-8356-7/04/$20.00 (C) 2004 IEEE IEEE INFOCOM 2004

 or for each Vi ∈ V , we have τi = mi . We assume that For ease of analysis, we assume that the size of raw data
each sensor node has q available modulation settings. Before generated by any source node equals 400 bits. In addition, the
the protocol converges, a feedback packet would reduce the aggregation factor, k, is assumed to be the same for all the
modulation setting for at least one sensor node every time it sensor nodes.
traverses the aggregation tree . Thus, the protocol converges
1
after the dissemination of at most nq feedback packets, where
n is the number of sensor nodes in the aggregation tree. 0.9
sources
Various tradeoffs can be explored in implementing the above 0.8
protocol. Ideally, the adaptation should be performed under
0.7
a stable system state. Thus, the period, α, for disseminating
the feedback packet should be large enough to accommodate 0.6

oscillation in system performance. However, a larger period 0.5

means a longer convergence process with greater energy
dissipation. There is also a tradeoff involved in selecting the 0.4

value of β. A larger value of β leads to higher transmission 0.3

speed when the latency constraint is violated. However, extra 0.2

energy dissipation is caused if the violation is not dramatic.
Intuitively, β should be related to the severity of the violation, 0.1
sink
which is indicated by the value of δ. 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Another option to handle latency violations is to repeatedly
reduce the transmission time of the sensor nodes with the Fig. 4. A example data aggregation tree based on the RS model (connectivity
smallest energy gradient till the latency constraint is satisfied. parameter ρ = 0.15, number of sources N = 20)
Such an option is more aggressive in reducing incurred energy
cost. However, it requires more complex control protocol and The energy function used in the simulation was in the form
more importantly, increases the response time in handling of equation (3). Note that τi ’s are assumed to be continuous
latency violations. variables except in the special case of modulation scaling,
when τi ’s have discrete values determined by the modulation
VI. S IMULATION R ESULTS level of the sensor nodes. We set Ri = 106 and Fi = 10−8
To conduct the simulations, a simulator was developed for all the sensor nodes, while the value of Ci was determined
using the PARSEC [18] software, which is a discrete-event by the distance from node Vi to its parent in the aggregation
simulation language. The baseline in our simulations is that tree. More specifically, we assume a d2 power loss model,
all sensor nodes send the packet at the highest speed and where d is the distance between Vi and its parent. That is
shutdown the radio afterward. This policy is proposed, for Ci = Cbase · ( dρ )2 . Based on our analysis in Section II, Cbase
example, in the PAMAS protocol [19]. For a fair comparison, was set to 7 × 10−9 for the long-range communication and
we first show data that does not consider the energy for starting 3 × 10−10 for the short-range communication.
up the radio in both the baseline and our techniques. We also To investigate the performance of our algorithms under
show data when the start-up energy is considered. various latency laxity, we use a user-specified parameter –
The purposes of the simulations are: (1) to demonstrate the normalized latency constraint to adjust the tightness of the
energy gain achieved by EMR-Algo and DP-Algo compared latency constraint, Γ. Specifically, let u ∈ (0, 1] denote the
with the baseline; (2) to evaluate the impact of several key normalized latency constraint, with a higher value of u mean-
system parameters to the performance of our algorithms; and ing a tighter latency constraint, and consequently, less laxity
(3) to show the energy saving and the adaptation capability of for exploring the energy-latency tradeoffs. We use equation
our on-line protocol in various run-time scenarios. (2) to model the time cost for packet transmission, with the
highest value of bi equal to 8. The minimal time cost for data
gathering, tmin , was calculated by assuming bi = 8 for all the
A. Simulation Setup sensor nodes. Then, we set Γ = tmin u .
A sensor network was generated by randomly scattering 200
sensors in a unit square. The sink node was put at the left-
bottom corner of the square. The number of sensor nodes that B. Performance of the Off-Line Algorithms
can communicate directly to a specific node is determined by The performance metric is defined as the percentage of
a connectivity parameter, ρ ∈ (0, 1], such that the average energy savings achieved by using our techniques, compared
number of neighbors of a sensor node is 200πρ2 . We used the with the baseline. In the simulation, the number of sources, N ,
so-called random sources (RS) model [3] for generating the was varied from 10 to 30 in increments of 10. The connectivity
location of the data sources. Specifically, N (the number of parameter, ρ, was varied from 0.1 to 0.3 in increments of
sources) sensor nodes are randomly selected to be the sources. 0.1. The aggregation factor, k, was varied from 0.3 to 1.0 in
The Greedy Incremental Tree (GIT) algorithm [3] was used increments of 0.35. The normalized latency constraint, u, was
for constructing the data aggregation tree. An example data varied from 0.1 to 1.0 in increments of 0.1. In addition, the
aggregation tree is illustrated in Figure 4 with ρ set to 0.15. approximation accuracy, D, was set to be 50 or 100.

0-7803-8356-7/04/$20.00 (C) 2004 IEEE IEEE INFOCOM 2004

 For each case of the simulation, more than 100 instances becomes large with solutions having high modulation levels,
were conducted. The presented data has a 95% confidence which is caused by a large value of u.
interval with a 5% (or better) precision. In each instance, the It can be observed that in short-range communication,
source nodes were randomly selected from the sensor network the performance of EMR-Algo and DP-Algo becomes satu-
shown in Figure 4. rated when u decreases below 0.5. This is because the non-
Performance of Our Off-Line Algorithms: Figure 5 demon- monotonic energy function limits the amount of latency laxity
strates the energy saving achieved by our off-line algorithms that can be traded for energy conservation. The little perfor-
for both long and short range communication. The investigated mance degradation of DP-Algo and MS when u approaches 0
algorithms include EMR-Algo, DP-Algo (with D = 100 or is because with a fixed D, the approximation accuracy actually
50), and the special case of DP-Algo for modulation scaling decreases when the latency constraint increases. Simulation re-
with D = 100 (denoted as MS). For MS, the available sults show that such performance degradation can be overcome
modulation levels are even numbers between 2 and 8. Note by increasing D.
that the performance for MS at u = 1.0 is not shown in the The simulation was performed on a SUN Blade1000 with a
figure. This is because in this case, to round the approximated 750 MHz SUN UltraSPARC III processor. The running time
solutions from DP-Algo leads to infeasible solutions. of EMR-Algo is between 0.1 to 2 second. The running time
of DP-Algo is around 0.003 second when D = 50 and 0.008
100 second when D = 100. Thus, the value of D can be used to
long−range
trade the performance of DP-Algo for the running time.
80 Energy Conservation vs. Various System Parameters: We
energy conservation (%)

show the results of DP-Algo with D = 100 in the following

study. Figure 6 shows the energy conservation achieved by DP-
60 Algo with respect to variations in k and u. It was observed
that while u mainly determines the energy gain, the energy
40 gain of DP-Algo decreases when k decreases for a fixed u.
short−range
This is because smaller value of k causes larger size of data
EMR−Algo packet after aggregation. Thus, the energy dissipated by links
20
DP−Algo (D=100) close to the sink node dominates the overall energy dissipation
DP−Algo (D=50) of the tree. It is however difficult to reduce the energy cost
MS (D = 100)
0 of these links since they have high likelihood to lie on the
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
normalized latency constraint longest path of the aggregation tree.

Fig. 5. Performance of our off-line algorithms (connectivity parameter ρ = 100

0.15, number of sources N = 20, aggregation factor k = 0.7) long−range

80
energy conservation (%)

The first thing to note is that up to 90% of energy saving is

achieved by all algorithms in the long-range communication
when u is small. Moreover, more than 20% energy saving is 60
achieved by all algorithms when u ≤ 0.9. Even when u = 1.0,
EMR-Algo and DP-Algo with D = 100 can still save more short−range
40
than 30% of the energy in both communication scenarios.
The reason for successful energy saving even when u = 1 is k=0.3
as follows. Intuitively, the latency constraint is determined by 20 k=0.65
the longest path in the aggregation tree. Thus, energy can be k=1.0
reduced for links on paths other than the longest one by trading
0
the latency laxity of the paths. On one hand, when there exists 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
only one path in the aggregation tree, no energy can be saved normalized latency constraint
when u = 1. On the other hand, when the aggregation tree Fig. 6. Energy conservation versus normalized latency constraint and
forms a star-like structure, all links, except the longest ones, aggregation factor k (connectivity parameter ρ = 0.15, approximation
can be optimized for saving energy even when u = 1. accuracy D = 100, number of sources N = 20)
The plot shows that the performance of DP-Algo improves
when D increases. When D = 100, the performance of DP- Figure 7 plots the performance of DP-Algo with respect to
Algo is quite close to the performance of EMR-Algo. Our variations in u and N . It can be seen that when u is large, the
simulation results show that the DP-Algo with D = 100 energy gain of DP-Algo increases as the number of sources
achieved at least 92% of the performance of EMR-Algo in increases. This is because larger number of sources facilitates
all the simulated instances. However, the performance of MS the optimization of links on paths other than the longest one.
quickly degrades when u increases. This is because the first Figure 8 demonstrates the performance of DP-Algo with
derivative of the energy function tends to −∞ as τ tends to respect to variations in u and ρ. It can be observed that the
0. Thus, the performance loss due to the rounding procedure energy saving of DP-Algo increases when ρ increase. This

0-7803-8356-7/04/$20.00 (C) 2004 IEEE IEEE INFOCOM 2004

 100 100
long−range long−range

80
energy conservation (%)

80

energy conservation (%)

60 60

short−range
40 short−range
40
N=10
20 N=20
N=30 20 DP−Algo
DP−IA
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0
normalized latency constraint 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
normalized latency constraint
Fig. 7. Energy conservation versus normalized latency constraint and number
of sources N (connectivity parameter ρ = 0.15, approximation accuracy Fig. 9. Performance of the DP-IA algorithm (connectivity parameter R =
D = 100, aggregation factor k = 0.7) 0.15, aggregation factor k = 0.7, number of sources N = 20)

packet size to 200 bits. Figure 10 shows the performance

is understandable since large ρ reduces the height of the
of EMR-Algo, DP-Algo (D = 100), and MS with start-up
aggregation tree (the extreme case is a star-like aggregation
energy. It can be observed that the impact of the start-up
tree formed by setting ρ = 1).
energy to the long-range communication is almost negligible.
100 This is understandable since the 1 µJ start-up energy is
long−range negligible compared with the transmitting energy of the radio.
However, the start-up energy is comparable to the transmitting
80 energy in the short-range communication. Thus, a decrease
energy conservation (%)

around 20% in energy saving is observed for the short-range

60
communication.
100
short−range long−range
40

ρ=0.1 80
energy conservation (%)

20
ρ=0.2
ρ=0.3 EMR−Algo
60 DP−Algo (D=100)
MS (D = 100)
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
normalized latency constraint 40

Fig. 8. Energy conservation versus normalized latency constraint and

connectivity parameter ρ (approximation accuracy D = 100, aggregation 20 short−range
factor k = 0.7)

Together, the above results suggest that when u is small, 0

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
DP-Algo is quite robust with respect to variations in different normalized latency constraint
system parameters, including k, ρ, and N .
Impact of Interference Avoidance: We also examined the Fig. 10. Performance of off-line algorithms with start-up energy (connectivity
performance of DP-IA for the two communication scenarios parameter ρ = 0.15, aggregation factor k = 0.7, number of sources N = 20)
with respect to variations in u (Figure 9). It can be seen that
the performance degradation due to interference avoidance is
severe when u is large. In particular, a decrease of more than C. Performance of the On-Line Protocol
10% in energy gain is observed for the long-range communica- Energy Conservation: We show the energy conservation
tion when u = 1. Thus, it is worthwhile to employ mechanisms achieved by the on-line protocol in Figure 11. The presented
for multi-packet reception or develop more efficient algorithms data is averaged over more than 150 problem instances and
for handling interference. has a 95% confidence interval with a 5% (or better) precision.
Impact of the Start-up Energy: We estimate the energy for In each instance, we generated a sensor network with 200
starting up the radio as 1 µJ [7]. In each epoch, the radio randomly dispersed sensor nodes. After randomly selecting
of each sensor node is started at most once. In addition, 20 source nodes, the data aggregation tree was then generated
to emphasize the impact of start-up energy, we change the using GIT.

0-7803-8356-7/04/$20.00 (C) 2004 IEEE IEEE INFOCOM 2004

 100
long−range −3
x 10
1 optimal
actual
80
energy conservation (%)

energy (J)
baseline
0.5
60

short−range 0
0 1 2 3 4
40
−3 time (Sec)
x 10
k=0.3 constraint

latency (Sec)
20 k=0.65 1.5
actual
k=1.0 baseline

0 1
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
normalized latency constraint
0.7
0 1 2 3 4
Fig. 11. Performance of the on-line protocol versus normalized latency time (Sec)
constraint and aggregation factor k (connectivity parameter ρ = 0.15, number
of sources N = 20)
(a) Scenario A

It can be seen that the energy conservation achieved by −3

x 10
1
the on-line protocol is quite close to the performance of the optimal
off-line algorithms shown in Figure 5. There is observable
energy (J)
actual
performance degradation when the normalized latency con- baseline
0.5
straint is set to one – around 15% less energy conservation
for long-range communication and 10% less for short-range
communication. Such a performance degradation is reasonable 0
0 1 2 3 4
due to the fact that only 4 options are available to set the trans- −3 time (Sec)
mission time for each sensor in the on-line protocol, instead x 10

of the fine granularity adjustment of the transmission time constraint

latency (Sec)

2 actual
in the off-line algorithms. Surprisingly, the on-line protocol
baseline
actually outperforms the modulation scaling case (MS) shown
in Figure 5, implying a large performance degradation of the 1

rounding technique used by MS.

0.2
Adaptability to System Variations: Our simulations were 0 1 2 3 4
performed based on the aggregation tree shown in Figures 4 time (Sec)
that has 35 sensor nodes, out of which, 20 are source nodes.
We assume that modulation scaling is used by all the nodes (b) Scenario B
with the available modulation levels being even numbers
between 2 and 8. The data gathering was requested every 2 Fig. 12. Adaptability of the on-line protocol (connectivity parameter ρ =
0.15, aggregation factor k = 0.7)
mSec. For the sake of illustration, we set α = 4 mSec, and
β = 10.
Two run-time scenarios, namely A and B, were investigated the energy cost is also increased. After that, the energy cost
to demonstrate the efficiency and adaptability of our protocol. drops again as time advances.
The energy cost and latency for data gathering over 4 seconds Note that by setting β = 10, the modulation levels of
are depicted in Figure 12, where the optimal solutions are the sensor nodes were restored to the highest levels when a
obtained by using EMR-Algo. violation is detected, reflected by the high peaks in the energy
Scenario A: We fixed s at 400 bits, while setting u to 0.7, curve at time 1, 1.5 and 3 seconds.
0.5, 0.8, 0.9, 0.8, 0.6, and 0.8 at time 0, 0.5, 1, 1.5, 2, 2.5, Scenario B: We set Γ = 0.6 mSec, while setting s to 400,
and 3.5 seconds, respectively. In real life, such variations can 300, 200, 400, 500, 450, and 300 at time 0, 0.5, 1, 1.5, 2, 2.5,
be caused by for example, changed user requests. and 3.5 seconds, respectively. In real life, the change of packet
It can be observed that when u is fixed, the actual energy size may be caused by variations in gathered information, or
cost gradually decreases till it is close to the optimal, while the aggregation factor of individual sensor node. An analysis
the latency approaches the constraint. At time 1 second, u is similar to the one in scenario A can be performed.
varied from 0.5 to 0.8, which causes a violation of the latency In short, our on-line protocol is capable of saving significant
constraint. Due to the feedback mechanism, the transmission energy in the studied scenarios. It is also capable of adapting
latency dramatically decreases as the modulation settings of all the packet transmission time with respect to the changing
the sensor nodes are restored to higher levels. Consequently, system parameters.

0-7803-8356-7/04/$20.00 (C) 2004 IEEE IEEE INFOCOM 2004

 VII. C ONCLUDING R EMARKS [14] A. Wang, S.-H. Cho, C. G. Sodini, and A. P. Chandrakasan, “Energy-
efficient modulation and MAC for asymmetric microsensor systems,” in
In this paper, we have studied the problem of scheduling ISLPED, 2001.
packet transmissions over a data aggregation tree in wireless [15] W. Heinzelman, A. P. Chandrakasan, and H. Balakrishnan, “An appli-
sensor networks by exploring the energy-latency tradeoffs. For cation specific protocol architecture for wireless microsensor networks,”
IEEE Trans. on Wireless Networking, 2002.
the off-line version of the problem, we have provided (a) a [16] S. Roy and H. Y. Wang, “Performance of CDMA slotted ALOHA multi-
numerical algorithm for optimal solutions, and (b) a pseudo- ple access with multiuser detection,” in IEEE Wireless Communications
polynomial time approximation algorithm based on dynamic and Networking Conference (WCNC), vol. 2, 1999, pp. 839–843.
[17] Y. Yu, B. Krishnamachari, and V. K. Prasanna, “Exploring energy-
programming. Techniques for handling interference has also latency tradeoffs for data gathering in wireless sensor networks,” De-
been discussed. Our simulation results show that between 20% partment of Electrical Engineering, University of Southern California,
to 90% energy saving can be achieved by the algorithms. Tech. Rep. CENG-2003-05, 2003.
[18] PARSEC Project. [Online]. Available: http://pcl.cs.ucla.edu/projects/
We have investigated the performance of our algorithms with parsec
different settings of several key system parameters. We have [19] C. S. Raghavendra and S. Singh, “PAMAS – power aware multi-access
also proposed a distributed on-line protocol that relies only on protocol with signaling for ad hoc networks,” Computer Communication
Review, July 1998.
local information of each sensor node in the aggregation tree.
Our simulation results show that the energy saving achieved A PPENDIX I
by the protocol is between 15% to 90%. Also, the ability P ROOF OF T HEOREM 1
of the protocol to adapt the packet transmission time upon
We define the level of a tree as the greatest number of
variations in the system parameters has been demonstrated
edges contained by any path in the tree. We consider an OPTP
through several run-time scenarios.
problem with a two-level aggregation tree that has exactly one
We are interested in integrating the concept of adaptive
internal node with p children (see Figure 13). We call such a
fidelity computation [2] for aggregation and compression into
problem 2-Lev-OPTP. Let Vp+1 denote the internal node, with
our work. The fidelity of the computation can be characterized
Vp+2 denoting its parent and C = {V1 , . . . , Vp } denoting the
by the size of the output data, which affects the consequent
set of children. We assume that for any Vi ∈ C, a packet
transmission time and energy costs. Thus, a broader tradeoff
is ready to transmission at time si and Vp+2 must received
space could be explored.
aggregated information from Vp+1 by time t.
R EFERENCES
V1 V2 Vp
[1] D. Estrin, L. Girod, G. Pottie, and M. B. Srivastava, “Instrumenting the
world with wireless sensor networks,” in International Conference on
Acoustics, Speech and Signal Processing (ICASSP), May 2001.
[2] C. Intanagonwiwat, R. Govindan, and D. Estrin, “Directed Diffusion:
A scalable and robust communication paradigm for sensor networks,”
in ACM/IEEE International Conference on Mobile Computing and Vp+1
Networking (MobiCom), 2000.
[3] B. Krishnamachari, D. Estrin, and S. Wicker, “The impact of data
aggregation in wireless sensor networks,” in International Workshop on
Distributed Event-Based Systems, 2002.
[4] B. Prabhakar, E. Uysal-Biyikoglu, and A. E. Gamal, “Energy-efficient Vp+2
transmission over a wireless link via lazy packet scheduling,” in IEEE
InfoCom, 2001. Fig. 13. A problem instance of 2-Lev-OPTP
[5] C. Schurgers, O. Aberhorne, and M. B. Srivastava, “Modulation scaling
for energy-aware communication systems,” in ISLPED, 2001, pp. 96–99.
[6] A. E. Gamal, C. Nair, B. Prabhakar, E. Uysal-Biyikoglu, and S. Za- We first show the following lemma:
hedi, “Energy-efficient scheduling of packet transmissions over wireless Lemma 2: Let τ∗ = {τ1∗ , . . . , τp+1
∗
} denote an optimal
networks,” in IEEE InfoCom, 2002. schedule to the 2-Lev-OPTP problem as defined above, then
[7] V. Raghunathan, C. Schurgers, S. Park, and M. B. Srivastava, “Energy-
aware wireless microsensor networks,” IEEE Signal Processing Maga- the following hold:
zine, March 2002. 1) The schedule τ∗ is unique.
[8] V. Raghunathan, S. Ganeriwal, C. Schurgers, and M. B. Srivastava, 2) Let sp+1 denote the start time of packet transmission
“E2WFQ: An energy efficient fair scheduling policy for wireless sys-
tems,” in International Symposium on Low Power Electronics and from Vp+1 to Vp+2 in the optimal schedule, i.e., sp+1 =
Design (ISLPED’02), Aug. 2002, pp. 30–35. maxVi ∈C (si + τi∗ ). Then sp+1 never decreases when (a)
[9] Y. Yu and V. K. Prasanna, “Energy-balanced multi-hop packet transmis- some si ’s, Vi ∈ C, increase, holding t fixed; or (b) t
sion in wireless sensor networks,” in IEEE GlobeCom, Dec. 2003.
[10] S. R. Madden, M. J. Franklin, J. M. Hellerstein, and W. Hong, “TAG: a increases, holding si ’s fixed, for all Vi ∈ C; or (c) both
Tiny AGgregation service for ad-hoc sensor networks,,” in Symposium some si ’s and t increase.
on Operating Systems Design and Implementation (OSDI), Dec. 2002. Let D denote the set of sensor nodes that increase their
[11] J. Elson, L. Girod, and D. Estrin, “Fine-grained network time syn-
chronization using reference broadcasts,” in Symposium on Operating transmission start time in cases (a) and (c). Then, in
Systems Design and Implementation (OSDI), Dec. 2002. particular, sp+1 increases in case (a) if for any Vi ∈ D,
[12] T. Ue, S. Sampei, N. Morinaga, and K. Hamaguchi, “Symbol rate and we have sp+1 − si ≤ mi ; or in case (b) we have
modulation level-controlled adaptive modulation/TDMA/TDD system
for high-bit rate wireless data transmission,” IEEE Trans. on Vehicular t − sp+1 < mp+1 ; or in case (c) we have either of the
Technology, vol. 47, no. 4, pp. 1134–1147, Nov. 1998. previous two conditions hold.
[13] E. Armanious, D. D. Falconer, and H. Yanikomeroglu, “Adaptive mod- 3) sp+1 never increases when (a) some si ’s, Vi ∈ C,
ulation, adaptive coding, and power control for fixed cellular broadband
wireless systems,” in IEEE Wireless Communications and Networking decrease, holding t fixed; or (b) t decreases, holding si ’s
Conference (WCNC), Mar. 2003. fixed, for all Vi ∈ C; or (c) both some si ’s and t decrease.

0-7803-8356-7/04/$20.00 (C) 2004 IEEE IEEE INFOCOM 2004

 Let D
denote the set of sensor nodes that decrease their we have τi∞ = mi for any Vi ∈ Tj . Based on the
transmission start time in cases (a) and (c). Then, in definition of s∞ ∗
j and sj , we obtain the contradiction that
particular, sp+1 decreases in case (a) if for any Vi ∈ D
, ∞
sj = sj .∗

we have sp+1 − si < mi ; or in case (b) we have

t − sp+1 ≤ mp+1 ; or in case (c) we have either of the A PPENDIX II
previous two conditions hold. J USTIFICATION OF THE H YPOTHESIS FOR DP-IA
Due to space limitations, the proof of Lemma 2 is omitted The DP-IA algorithm is designed based on the hypothesis
in this paper. Details of the proof can be found in [17]. that any group of interfering sensor nodes are children of the
Now we prove Theorem 1. same node. Such a hypothesis can be satisfied by carefully
Proof: constructing the aggregation tree as follows.
1) Recall that EMR-Algo works in iterations: for each We consider possible scenarios for interference between
iteration k, the algorithm determines ski by decreasing sensor nodes that are not children of the same node. First we
i from n − 1 to M + 1. Since the EMR-Algo initializes consider packet transmission along each path – the packets
si = 0 for i = 1, . . . , n − 1, it follows that s0i ≤ s1i for must be transmitted by following the topological order of
each i = 1, . . . , n−1. Suppose that i
> 1 and k
> 1
are sensor nodes along the path. Therefore, no interference could


the first time that there is a violation; that is, ski
> ski
+1 . happen between sensor nodes on the same path.
We consider the 2-level aggregation tree formed by Vi
Now we examine the case when sensor nodes from two
together with its parent, denoted as Vp , and its children, different paths interfere with each other. Let p1 and p2 denote





denoted as set C. We have skp ≤ skp +1 , and sik −1 ≤ ski , the two paths and let Va ∈ p1 and Vb ∈ p2 denote the
for each Vi ∈ C.
two interfering sensor nodes. More precisely, let Vc denote


From line 8 in EMR-Algo, skp and sik −1 ’s actually give the parent of Va and Vd denote the parent of Vb (shown


the boundaries within which EMR-Algo determines ski
. in Figure 14 (a)). Let d(Vi , Vj ) denote the physical distance
k
+1 k

Similarly, sp and si ’s give the boundaries within between sensor nodes Vi and Vj . Without loss of generality, we


which EMR-Algo determines ski
+1 . From part (2) of say that Vb interferes Va if d(Vb , Vc ) ≤ d(Vb , Vd ). Intuitively,



Lemma 2, we have ski
≤ ski
+1 . This contradicts the it implies that Vc is within the communication range of both



assumption ski
> ski
+1 and hence property (1) holds. Va and Vb . Hence, packets simultaneously transmitted from
2) It is obvious that s0i ≤ s∗i , for each i = 1, . . . , n − 1. Va and Vb may collide at Vc .
Similar to the proof for property (1), suppose that i
≥ 1
and k
≥ 1 are the first time that there is a violation; that communication
radius


is, ski
> s∗i
. V1 V1
Again, we consider the 2-level aggregation tree formed V2 V2
by Vi
together with its parent, denoted as Vp , and its


children, denoted as set C. We have skp ≤ s∗p , and



sik −1 ≤ s∗i , for each Vi ∈ C. We know that skp and Va
Vb
Va
Vc Vb

Vc
sik −1 ’s actually give the boundaries within which EMR-

Vd Vd
Algo determines ski
. Similarly, s∗p and s∗i ’s give the
boundaries within which EMR-Algo determines s∗i
. Part
(2) of Lemma 2 again leads to the contradiction that


ski
≤ s∗i
and proves property (2).
3) We prove by contradiction and hence assume that j = Vn Vn
max{i : s∞ ∗
i < si }. Let Vp denote the parent of Vj and Vg (a) before transform (b) after transform
denote the parent of Vp . We have s∞ ∗ ∞
p = sp and sg = sg .
∗
∗ ∗
Since τj is optimal, we have τj ≤ mj . We consider two Fig. 14. A example for transforming the data aggregation tree
cases:
Case (i): We suppose that τj∗ < mj . Considering the We use a simple transform procedure to re-construct the
2-level tree formed by Vp , Vg and the children of Vp , aggregation tree. The key idea is to group interfering sensor
denoted as C, we have s∞ ∗ ∞ ∗
j < sj and si ≤ si , for each
nodes to be the children of the same sensor node. Specifically,
Vi ∈ C ∧ i = j. Suppose that we run EMR-Algo for one we make Vb also a child of Vc and break the links from Vb
more pass and let ŝp denote the resulting start time for the to its ancestors until an ancestor that performs local sensing
transmission from Vp to Vg . From part (3) of Lemma 2, is reached. The aggregation tree after the transformation is
we have s∞ ∗ ∞ ∞ ∗
p − (sj − sj ) < ŝp < sp = sp , contradicting
illustrated in Figure 14 (b).
both property (1) for Vp and the definition of j. It can be shown that by applying the above transform
Case (ii): We assume that τj∗ = mj . From part (1) of procedure for each pair of interfering sensor nodes from two
Corollary 1, we have τi∗ = mi for any Vi ∈ Tj . Moreover, different paths as previously defined, we obtain an aggregation
we have s∞ ∞ ∗ ∗ ∗
p − sj > sp − sj = τj = mj . Since EMR-
tree such that any group of interfering sensor nodes are
Algo maintains the invariant that τi ≤ mi for all Vi ∈ V ,
k children of the same sensor node.
we have τj∞ = mj . Again from part (1) of Corollary 1,

0-7803-8356-7/04/$20.00 (C) 2004 IEEE IEEE INFOCOM 2004