86 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 11, NO.

1, JANUARY 2012

Fast Data Collection in Tree-Based

Wireless Sensor Networks
Özlem Durmaz Incel, Amitabha Ghosh, Bhaskar Krishnamachari, and Krishnakant Chintalapudi

Abstract—We investigate the following fundamental question—how fast can information be collected from a wireless sensor network
organized as tree? To address this, we explore and evaluate a number of different techniques using realistic simulation models under
the many-to-one communication paradigm known as convergecast. We first consider time scheduling on a single frequency channel
with the aim of minimizing the number of time slots required (schedule length) to complete a convergecast. Next, we combine
scheduling with transmission power control to mitigate the effects of interference, and show that while power control helps in reducing
the schedule length under a single frequency, scheduling transmissions using multiple frequencies is more efficient. We give lower
bounds on the schedule length when interference is completely eliminated, and propose algorithms that achieve these bounds. We
also evaluate the performance of various channel assignment methods and find empirically that for moderate size networks of about
100 nodes, the use of multifrequency scheduling can suffice to eliminate most of the interference. Then, the data collection rate no
longer remains limited by interference but by the topology of the routing tree. To this end, we construct degree-constrained spanning
trees and capacitated minimal spanning trees, and show significant improvement in scheduling performance over different deployment
densities. Lastly, we evaluate the impact of different interference and channel models on the schedule length.

Index Terms—Convergecast, TDMA scheduling, multiple channels, power control, routing trees.

1 INTRODUCTION

C ONVERGECAST, namely, the collection of data from a set of

sensors toward a common sink over a tree-based
routing topology, is a fundamental operation in wireless
2) raw-data convergecast where packets are individually
relayed toward the sink. Aggregated convergecast is applic-
able when a strong spatial correlation exists in the data, or
sensor networks (WSNs) [1]. In many applications, it is the goal is to collect summarized information such as the
crucial to provide a guarantee on the delivery time as well as maximum sensor reading. Raw-data convergecast, on the
increase the rate of such data collection. For instance, in other hand, is applicable when every sensor reading is
safety and mission-critical applications where sensor nodes equally important, or the correlation is minimal. We study
are deployed to detect oil/gas leak or structural damage, the aggregated convergecast in the context of continuous data
actuators and controllers need to receive data from all the collection, and raw-data convergecast for one-shot data
sensors within a specific deadline [2], failure of which might collection. These two types correspond to two extreme cases
lead to unpredictable and catastrophic events. This falls of data collection. In an earlier work [4], the problem of
under the category of one-shot data collection. On the other applying different aggregation factors, i.e., data compression
hand, applications such as permafrost monitoring [3] require factors, was studied, and the latency of data collection was
periodic and fast data delivery over long periods of time, shown to be within the performance bounds of the two
which falls under the category of continuous data collection. extreme cases of no data compression (raw-data converge-
In this paper, we consider such applications and focus on cast) and full data compression (aggregated convergecast).
the following fundamental question: “How fast can data be For periodic traffic, it is well known that contention-free
streamed from a set of sensors to a sink over a tree-based topology?” medium access control (MAC) protocols such as Time
We study two types of data collection: 1) aggregated Division Multiple Access (TDMA) are better fit for fast data
convergecast where packets are aggregated at each hop, and collection, since they can eliminate collisions and retrans-
missions and provide guarantee on the completion time as
opposed to contention-based protocols [1]. However, the
. Ö.D. Incel is with the Department of Computer Engineering, Bogazici problem of constructing conflict-free (interference-free)
University, Istanbul 34342, Turkey.
E-mail: [email protected].
TDMA schedules even under the simple graph-based
. A. Ghosh is with the Department of Electrical Engineering, Princeton interference model has been proved to be NP-complete. In
University, F-310, Engineering Quad, Olden Street, Princeton, NJ 08544. this work, we consider a TDMA framework and design
E-mail: [email protected]. polynomial-time heuristics to minimize the schedule length
. B. Krishnamachari is with the Ming Hsieh Department of Electrical
Engineering, University of Southern California, 3740 McClintock Avenue,
for both types of convergecast. We also find lower bounds
EEB 300, Los Angeles, CA 90089. E-mail: [email protected]. on the achievable schedule lengths and compare the
. K. Chintalapudi is with Microsoft Research, Bengaluru, India. performance of our heuristics with these bounds.
E-mail: [email protected]. We start by identifying the primary limiting factors of
Manuscript received 5 Nov. 2009; revised 14 Oct. 2010; accepted 15 Dec. fast data collection, which are: 1) interference in the wireless
2010; published online 2 Feb. 2011. medium, 2) half-duplex transceivers on the sensor nodes, and
For information on obtaining reprints of this article, please send e-mail to:
[email protected], and reference IEEECS Log Number TMC-2009-11-0482. 3) topology of the network. Then, we explore a number of
Digital Object Identifier no. 10.1109/TMC.2011.22. different techniques that provide a hierarchy of successive
1536-1233/12/$31.00 ß 2012 IEEE Published by the IEEE CS, CASS, ComSoc, IES, & SPS
INCEL ET AL.: FAST DATA COLLECTION IN TREE-BASED WIRELESS SENSOR NETWORKS 87

improvements, the simplest among which is an interfer- of aggregated convergecast. In this work, we
ence-aware, minimum-length TDMA scheduling that en- evaluate the behavior of an optimal power control
ables spatial reuse. To achieve further improvement, we algorithm [6] under realistic settings considering the
combine transmission power control with scheduling, and limited discrete power levels available in today’s
use multiple frequency channels to enable more concurrent radios. We find that for moderate size networks of
transmissions. We show that once multiple frequencies are 100 nodes, power control can reduce the schedule
employed along with spatial-reuse TDMA, the data collec- length by 15-20 percent.
tion rate often no longer remains limited by interference but . Evaluation of channel assignment methods. Using
by the topology of the network. Thus, in the final step, we extensive simulations, we show that scheduling
transmissions on different frequency channels is
construct network topologies with specific properties that
more effective in mitigating interference as com-
help in further enhancing the rate. Our primary conclusion
pared to transmission power control. We evaluate
is that combining these different techniques can provide an
the performance of three different channel assign-
order of magnitude improvement for aggregated converge- ment methods: 1) Joint Frequency Time Slot Scheduling
cast, and a factor of 2 improvement for raw-data con- (JFTSS), 2) Receiver-Based Channel Assignment [7], and
vergecast, compared to single-channel TDMA scheduling 3) Tree-Based Multichannel Protocol (TMCP) [8]. These
on minimum-hop routing trees. methods consider the channel assignment problem
Although the techniques of transmission power control at different levels: the link level, node level, or
and multichannel scheduling have been well studied for cluster level. We show that for aggregated conver-
eliminating interference in general wireless networks, their gecast, TMCP performs better than JFTSS and RBCA
performances for bounding the completion of data collec- on minimum-hop routing trees, while performs
tion in WSNs have not been explored in detail in the worse on degree-constrained trees. For raw-data
previous studies. The fundamental novelty of our approach convergecast, RBCA and JFTSS perform better than
lies in the extensive exploration of the efficiency of TMCP, since the latter suffers from interference
transmission power control and multichannel communica- inside the branches due to concurrent transmissions
tion on achieving fast convergecast operations in WSNs. on the same channel.
Besides, we evaluate the impact of routing trees on fast data . Impact of routing trees. We investigate the effect of
collection and to the best of our knowledge, this has not network topology on the schedule length, and show
been the topic of previous studies. As we will discuss in that for aggregated convergecast, the performance
Section 2, some of the existing work had the objective of can be improved by up to 10 times on degree-
minimizing the completion time of convergecasts. How- constrained trees using multiple frequencies as
ever, none of the previous work discussed the effect of compared to that on minimum-hop trees using a
multichannel scheduling together with the comparisons single frequency. For raw-data convergecast, multi-
of different channel assignment techniques and the impact channel scheduling on capacitated minimal span-
of routing trees and none considered the problems of ning trees (CMSTs) can reduce the schedule length
aggregated and raw convergecast, which represent two by 50 percent.
extreme cases of data collection, together. . Impact of channel models and interference. Under
As the new concepts in this paper, we introduce the setting of multiple frequencies, one simplifying
polynomial-time heuristics for TDMA scheduling for both assumption often made is that the frequencies are
types of data collection, i.e., Algorithms 1 and 2, and prove orthogonal to each other. We evaluate this assump-
that they do achieve the lower bound of data collection time tion and show that the schedules generated may
once interference is eliminated. Besides, we elaborate on the not always eliminate interference, thus causing
performance of our previous work, a receiver-based considerable packet losses. We also evaluate and
channel assignment (RBCA) method, and compare its compare the two most commonly used interference
efficiency with other channel assignment methods and models: 1) the graph-based protocol model, and
introduce heuristics for constructing optimal routing trees 2) the Signal-to-Interference-plus-Noise Ratio
to further enhance data collection rate. The following lists (SINR)-based physical model.
our key findings and contributions: The rest of the paper is organized as follows: in Section 2,
. Bounds on convergecast scheduling. We show that we discuss related works. In Section 3, we describe the
if all interfering links are eliminated, the schedule problem formulation and state our assumptions. In Section
length for aggregated convergecast is lower 4, we analyze the lower bounds on the schedule length for
bounded by the maximum node degree in the aggregated and raw convergecast, and propose algorithms
routing tree, and for raw-data convergecast by that achieve the corresponding bounds. In Section 5, we
maxð2nk  1; NÞ, where nk is the maximum number focus on power control and multichannel scheduling as
of nodes on any branch in the tree, and N is the mechanisms to eliminate interference. Section 6 explains the
number of source nodes. We then introduce optimal impact of routing topologies, and Section 7 presents
time slot assignment schemes under this scenario detailed evaluation results. Finally, we draw our conclu-
which achieve these lower bounds. sions in Section 8.
. Evaluation of power control under realistic setting.
It was shown recently [5] that under the idealized
setting of unlimited power and continuous range,
2 RELATED WORK
transmission power control can provide an un- Fast data collection with the goal to minimize the schedule
bounded improvement in the asymptotic capacity length for aggregated convergecast has been studied by us
88 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 11, NO. 1, JANUARY 2012

in [7] and [9], and also by others in [5], [10], and [11]. In [7], However, TreeMAC operates on a single channel and
we experimentally investigated the impact of transmission achieves 1=3 of the maximum throughput similar to the
power control and multiple frequency channels on the bounds presented by Gandham et al. [1] since the sink can
schedule length, while the theoretical aspects were dis- receive every three time slots.
cussed in [9], where we proposed constant factor and The problem of minimizing the schedule length for raw-
logarithmic approximation algorithms on geometric net- data convergecast on single channel is shown to be NP-
works (disk graphs). Raw-data convergecast has been complete on general graphs by Choi et al. [13]. Maximizing
studied in [1], [12], [13], and [14], where a distributed time the throughput of convergecast by finding a shortest-length,
slot assignment scheme is proposed by Gandham et al. [1] conflict-free schedule is studied by Lai et al. [14], where a
to minimize the TDMA schedule length for a single channel. greedy graph coloring strategy assigns time slots to the
The problem of joint scheduling and transmission power senders and prevents interference. They also discussed the
control is studied by Moscibroda [5] for constant and impact of routing trees on the schedule length and proposed a
uniform traffic demands. Our present work is different routing scheme called disjoint strips to transmit data over
from the above in that we evaluate transmission power different shortest paths. However, since the sink remains as
control under realistic settings and compute lower bounds the bottleneck, sending data over different paths does not
on the schedule length for tree networks with algorithms to reduce the schedule length. As we will show in this paper, the
achieve these bounds. We also compare the efficiency of improvement due to the routing structure comes from using
different channel assignment methods and interference capacitated minimal spanning trees for raw-data converge-
models, and propose schemes for constructing specific cast, where the number of nodes in a subtree is no more than
routing tree topologies that enhance the data collection rate half the total number of nodes in the remaining subtrees.
for both aggregated and raw-data convergecast. The use of multiple frequencies has been studied
The use of orthogonal codes to eliminate interference has extensively in both cellular and ad hoc networks; however,
been studied by Annamalai et al. [10], where nodes are in the domain of WSN, there exist a few studies that utilize
assigned time slots from the bottom of the tree to the top such multiple channels [8], [17], [18]. To this end, we evaluate the
that a parent node does not transmit before it receives all the efficiency of three particular schemes that treat the channel
packets from its children. This problem and the one assignment at different levels.
addressed by Chen et al. [11] are for one-shot raw-data
convergecast. In this work, since we construct degree-
constrained routing topologies to enhance the data collection
3 MODELING AND PROBLEM FORMULATION
rate, it may not always lead to schedules that have low We model the multihop WSN as a graph G ¼ ðV ; EÞ, where
latency, because the number of hops in a tree goes up as its V is the set of nodes, and E ¼ fði; jÞ j i; j 2 V g is the set of
degree goes down. Therefore, if minimizing latency is also a edges representing the wireless links. A designated node
requirement, then further optimization, such as constructing s 2 V denotes the sink. The euclidean distance between two
bounded-degree, bounded-diameter trees, is needed. A nodes i and j is denoted by dij . All the nodes except s are
study along this line with the objective to minimize the sources, which generate packets and transmit them over a
maximum latency is presented by Pan and Tseng [15], where routing tree to s. We denote the spanning tree on G rooted
they assign a beacon period to each node in a Zigbee network at s by T ¼ ðV ; ET Þ, where ET  E represents the tree
during which it can receive data from all its children. edges. Each node is assumed to be equipped with a single
For raw-data convergecast, Song et al. [12] presented a half-duplex transceiver, which prevents it from sending and
time-optimal, energy-efficient packet scheduling algorithm receiving packets simultaneously. We consider a TDMA
with periodic traffic from all the nodes to the sink. Once protocol where time is divided into slots, and consecutive
interference is eliminated, their algorithm achieves the slots are grouped into equal-sized nonoverlapping frames.
bound that we present here; however, they briefly mention We use two types of interference models for our
a 3-coloring channel assignment scheme, and it is not clear evaluation: the graph-based protocol model and the SINR-
whether the channels are frequencies, codes, or any other based physical model. In the protocol model, we assume
method to eliminate interference. Moreover, they assume a that the interference range of a node is equal to its
simple interference model where each node has a circular transmission range, i.e., two links cannot be scheduled
transmission range and cumulative interference from simultaneously if the receiver of at least one link is within
concurrent multiple senders is avoided. Different from the range of the transmitter of the other link. In the physical
their work, we consider multiple frequencies and evaluate model, the successful reception of a packet from i to j
the performance of three different channel assignment depends on the ratio between the received signal strength at
methods together with evaluating the effects of transmis- j and the cumulative interference caused by all other
sion power control using realistic interference and channel concurrently transmitting nodes and the ambient noise
models, i.e., physical interference model and overlapping level. Thus, a packet is received successfully at j if the
channels and considering the impact of routing topologies. signal-to-interference-plus-noise ratio, SINRij , is greater
Song et al. [12] extended their work and proposed a TDMA- than a certain threshold , i.e.,
based MAC protocol for high-data-rate WSNs in [16].
TreeMAC considers the differences in load at different Pi  gij
levels of a routing tree and assigns time slots according to SINRij ¼ P ; ð1Þ
k6¼i Pk  gkj þ N
the depth, i.e., the hop count, of the nodes on the routing
tree, such that nodes closer to the sink are assigned more where Pi is the transmitted signal power at node i, N is the
slots than their children in order to mitigate congestion. ambient noise level, and gij is the propagation attenuation
INCEL ET AL.: FAST DATA COLLECTION IN TREE-BASED WIRELESS SENSOR NETWORKS 89

(link gain) between i and j. We use a simple distance operating frequency, channel scheduling can be coupled
dependent path-loss model to calculate the link gains as with time slot scheduling as it is the case with the JFTSS
gij ¼ d
ij , where the path-loss exponent  is a constant
algorithm (Section 5.2.1) or first channels are assigned and
between 2 and 6, whose exact value depends on external then time slot scheduling continues as in the case of RBCA
conditions of the medium (humidity, obstacles, etc.), as well explained in Section 5.2.3. However, the TMCP algorithm
as the sender-receiver distance. We assume that the level of (Section 5.2.2) considers tree construction and channel
interference is static and does not change over time. For assignment jointly and then does the scheduling of time slots.
simplicity and ease of illustration, we use the protocol
model in all the figures. 4 TDMA SCHEDULING OF CONVERGECASTS
We study aggregated convergecast in the context of
periodic data collection where each source node generates a In this section, we first focus on periodic aggregated
packet at the beginning of every frame, and raw-data convergecast and then on one-shot raw-data convergecast.
convegecast for one-shot data collection where each node Our objective is to calculate the minimum achievable
has only one packet to send. We assume that the size of each schedule lengths using an interference-aware TDMA pro-
packet is constant. Our goal is to deliver these packets to the tocol. We first consider the case where the nodes commu-
sink over the routing tree as fast as possible. More nicate on the same channel using a constant transmission
specifically, we aim to schedule the edges ET of T using a power, and then discuss improvements using transmission
minimum number of time slots while respecting the power control and multiple frequencies in the next section.
following two constraints: 4.1 Periodic Aggregated Convergecast
. Adjacency constraint. Two edges ði; jÞ 2 ET and In this section, we consider the scheduling problem where
ðk; lÞ 2 ET cannot be scheduled in the same time packets are aggregated. Data aggregation is a commonly
slot if used technique in WSN that can eliminate redundancy and
T they are adjacent to each other, i.e., if minimize the number of transmissions, thus saving energy
fi; jg fk; lg 6¼ . This constraint is due to the half-
duplex transceiver on each node which prevents it and improving network lifetime [19]. Aggregation can be
performed in many ways such as by suppressing duplicate
from simultaneous transmission and reception.
messages; using data compression and packet merging
. Interfering constraint. The interfering constraint de-
techniques; or taking advantage of the correlation in the
pends on the choice of the interference model. In the
sensor readings.
protocol model, two edges ði; jÞ 2 ET and ðk; lÞ 2 ET
We consider continuous monitoring applications where
cannot be scheduled simultaneously if they are at 2-
perfect aggregation is possible, i.e., each node is capable of
hop distance of each other. In the physical model, an aggregating all the packets received from its children as
edge ði; jÞ 2 ET cannot be scheduled if the SINR at well as that generated by itself into a single packet before
receiver j is not greater than the threshold . transmitting to its parent. The size of aggregated data
Since we consider data collection to be periodic in transmitted by each node is constant and does not depend
aggregated convergecast, each of the edges in ET is on the size of the raw sensor readings. Typical examples of
scheduled only once within each frame, and this schedule such aggregation functions are MIN, MAX, MEDIAN,
is repeated over multiple frames. Thus, a pipeline is COUNT, AVERAGE, etc.
established after a certain frame, and then onward the sink In Figs. 1a and 1b, we illustrate the notion of pipelining
continues to receive aggregated packets from all the source in aggregated convergecast and that of a schedule length on
nodes once per frame. We explain further details about the a network of six source nodes. The solid lines represent tree
pipelining in the next section. On the other hand, in one-shot edges, and the dotted lines represent interfering links. The
data collection for raw-data convergecast, the edges in ET numbers beside the links represent the time slots at which
may be scheduled multiple times and no pipelining takes the links are scheduled to transmit, and the numbers inside
place. We use the terms link scheduling and node scheduling the circles denote node ids. The entries in the table list the
interchangeably as they are equivalent in our case. Note that nodes from which packets are received by their correspond-
the two other scenarios, which we do not consider in this ing receivers in each time slot. We note that at the end of
paper due to space constraints, are one-shot aggregated frame 1, the sink does not have packets from nodes 5 and 6;
convergecast and periodic raw-data convergecast. however, as the schedule is repeated, it receives aggregated
The key difference in terms of scheduling between packets from 2, 5, and 6 in slot 2 of the next frame. Similarly,
periodic and one-shot data collection is that a node in the the sink also receives aggregated packets from nodes 1 and
periodic case does not have to wait for data from its children 4 starting from slot 1 of frame 2. The entries f1; 4g and
before being scheduled. This is because a link is scheduled f2; 5; 6g in the table represent single packets comprising
only once within each frame and each node generates a packet aggregated data from nodes 1 and 4, and from nodes 2, 5,
in the beginning of every frame, so a pipelining is eventually and 6, respectively. Thus, a pipeline is established from
established. However, in the case of one-shot data collection, frame 2, and the sink continues to receive aggregated
a node needs to wait for data from its children before being packets from all the nodes once every six time slots. Thus,
scheduled, which we refer to as the causality constraint. the minimum schedule length is 6.
To summarize the steps in our design, we start with tree
construction and then continue with interference-aware 4.1.1 Lower Bound on Schedule Length
scheduling. If the nodes can control their transmission We first consider aggregated convergecast when all the
power, scheduling phase is coupled with a transmission interfering links are eliminated by using transmission power
power control algorithm. If the nodes can change their control or multiple frequencies. Although the problem of
90 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 11, NO. 1, JANUARY 2012

Fig. 1. Aggregated convergecast and pipelining: (a) Schedule length of 6 in the presence of interfering links. (b) Node ids from which (aggregated)
packets are received by their corresponding parents in each time slot over different frames. (c) Schedule length of 3 using BFS-
TIMESLOTASSIGNMENT when all the interfering links are eliminated.

minimizing the schedule length is NP-complete on general To illustrate, we show the same network of Fig. 1a in Fig. 1c
graphs, we show in the following that once interference is with all the interfering links removed, and so the network is
eliminated, the problem reduces to 1 on a tree, and can be scheduled in three time slots.
solved in polynomial time. To this end, we first give a lower Although BFS-TIMESLOTASSIGNMENT may not be an
bound on the schedule length, and then propose a time slot approximation to ideal scheduling under the physical
assignment scheme that achieves the bound. interference model, it is a heuristic that can achieve the
Lemma 1. If all the interfering links are eliminated, the schedule lower bound if all the interfering links are eliminated.
length for aggregated convergecast is lower bounded by ðT Þ, Therefore, together with a method to eliminate interference,
where ðT Þ is the maximum node degree in the routing tree T . the algorithm can optimally schedule the network.
Proof. If all the interfering links are eliminated, the Theorem 1. If all the interfering links are eliminated, the
scheduling problem reduces to 1 on a tree. Now since schedule length for aggregated convergecast achieved by BFS-
each of the tree edges needs to be scheduled only once TIMESLOTASSIGNMENT is the minimum, i.e., ðT Þ.
within each frame, it is equivalent to edge coloring on a Proof. The proof is by induction on i. Let T i ¼ ðV i ; ETi Þ
graph, which needs number of colors at least equal to the denote the subtree of T in the ith iteration constructed in
maximum node degree. u
t the BFS order, where ETi comprises all the edges that are
assigned a slot, and V i comprises the set of nodes on which
Once all the interfering links are eliminated, concurrency the edges in ETi are incident. Note that, jETi j ¼ i, because at
is still limited by the adjacency constraint due to the half- every iteration, exactly one edge is assigned a slot. For
duplex transceivers, which prevents a parent from trans- i ¼ 1, clearly the number of slots used is 1, equal to ðT 1 Þ.
mitting when it is already receiving from its children, or Now, assume that the number of slots NðiÞ needed to
when its parent is transmitting. schedule the edges in T i is ðT i Þ. In the ði þ 1Þst
iteration, after assigning a slot to the next edge in BFS
4.1.2 Assignment of Time Slots
order, the number of slots needed in T iþ1 can either
Given the lower bound ðT Þ on the schedule length in the remain the same as before, or increase by 1. Thus,
absence of interfering links, we now present a time slot
assignment scheme in Algorithm 1, called BFS-TIMESLO- Nði þ 1Þ ¼ maxfNðiÞ; NðiÞ þ 1g: ð2Þ
TASSIGNMENT, that achieves this bound.
If it remains the same, Nði þ 1Þ is still the maximum
Algorithm 1. BFS-TIMESLOTASSIGNMENT degree of T iþ1 at end of ði þ 1Þst iteration. Otherwise, if it
1. Input: T ¼ ðV ; ET Þ increases by 1, the new edge must be incident on a node
2. while ET 6¼  do v , common to both T i and T iþ1 , such that the number of
3. e next edge from ET in BFS order incident edges on v that were already assigned a time
4. Assign minimum time slot t to edge e respecting slot at the end of ith iteration was ðT i Þ. This is so
adjacency and interfering constraints because in the BFS traversal, all the edges incident on a
5. ET ET n feg node are assigned a slot first before moving on to the
6. end while next node, and because the slot assigned to the new edge
is the minimum possible that is different from all that
In each iteration of BFS-TIMESLOTASSIGNMENT (lines 2- already assigned to the edges incident on v until the ith
6), an edge e is chosen in the Breadth First Search (BFS) iteration. Thus, at the end of ði þ 1Þst iteration, the
order starting from any node, and is assigned the minimum number of slots used NðiÞ þ 1 is equal to the number of
time slot that is different from all its adjacent edges assigned edges incident on v which, in turn, equals
respecting interfering constraints. Note that, since we ðT iþ1 Þ. This proves the inductive step. Therefore, it
evaluate the performance of this algorithm also for the case holds at every iteration of the algorithm until the end
when the interfering links are present, we check for the when i ¼ jV j  2, yielding a schedule length equal to the
corresponding constraint in line 4; however, when inter- maximum degree ðT Þ ¼ ðT jV j1 Þ. Now, since assign-
ference is eliminated this check is redundant. The algorithm ing different time slots to the adjacent edges of T is
runs in OðjET j2 Þ time and minimizes the schedule length equivalent to edge coloring T , which requires at least
when there are no interfering links, as proved in Theorem 1. ðT Þ colors, the schedule length is minimum. u
t
INCEL ET AL.: FAST DATA COLLECTION IN TREE-BASED WIRELESS SENSOR NETWORKS 91

which requires each node to maintain a buffer that stores at

most two packets and minimizes the schedule length. In the
following, we describe a time slot assignment scheme that
computes a schedule of length exactly equal to the lower
bound when interference is eliminated and does not require
Fig. 2. Raw-data convergecast: largest top-subtree with nk nodes. to store more than one packet in buffers at any time.

4.2.2 Assignment of Time Slots

4.2 One-Shot Raw-Data Convergecast
We now describe a time slot assignment scheme in
In this section, we consider one-shot data collection where
Algorithm 2, called LOCAL-TIMESLOTASSIGNMENT, which
every sensor reading is equally important, and so aggrega-
is run locally by each node at every time slot. The key idea is
tion may not be desirable or even possible. Thus, each of the
to: 1) schedule transmissions in parallel along multiple
packets has to be individually scheduled at each hop en route
branches of the tree, and 2) keep the sink busy in receiving
to the sink. As before, we focus on minimizing the schedule
packets for as many time slots as possible. Because the sink
length. Unlike in the case of periodic aggregated converge-
can receive from the root of at most one top-subtree in any
cast where a pipelining takes place and each of the tree edges
time slot, we need to decide which top-subtree should be
is scheduled only once within each frame, here the edges
made active. We assume that the sink is aware of the
could be scheduled multiple times and there is no pipelining.
number of nodes in each top-subtree. Each source node
The problem of minimizing the scheduling length for
maintains a buffer and its associated state, which can
raw-data convergecast is proved to be NP-complete even
be either full or empty depending on whether it contains a
under the protocol interference model by a reduction from
packet or not. Our algorithm does not require any of the
the well known Partition Problem [13]. Before getting into the
nodes to store more than one packet in their buffer at any
details, we first define the following terms: a branch is
time. We initialize all the buffers as full, and assume that
defined as a subtree containing the sink as an endpoint; a
the sink’s buffer is always full for the ease of explanation.
top-subtree is defined as a subtree that has a child of the sink
as its root. For instance, in Fig. 3, the branches are fs; 1; 4g, Algorithm 2. LOCAL-TIMESLOTASSIGNMENT
fs; 2; 5; 6g, and fs; 3; 7g, while the top-subtrees are f1; 4g, 1. node.buffer ¼ full
f2; 5; 6g, and f3; 7g. 2. if {node is sink} then
3. Among the eligible top-subtrees, choose the one with
4.2.1 Lower Bound on Schedule Length the largest number of total (remaining) packets, say
As mentioned in Section 4.1.1, if all the interfering links are top-subtree i
eliminated using multiple frequencies, the only limiting 4. Schedule link ðrootðiÞ; sÞ respecting interfering
factor in minimizing the schedule length is the half-duplex constraint
transceivers. In the following, we give a lower bound on the 5. else
schedule length under this scenario. 6. if {node.buffer == empty} then
Lemma 2. If all the interfering links are eliminated, the schedule 7. Choose a random child c of node whose buffer
length for one-shot raw-data convergecast is lower bounded by is full
maxð2nk  1; NÞ, where nk is the maximum number of nodes 8. Schedule link ðc; nodeÞ respecting interfering
in any top-subtree of the routing tree, and N is the number of constraint
sources in the network. 9. c.buffer = empty
Proof. Let ni denote the number of nodes in top-subtree i. 10. node.buffer = full
Order the top-subtrees in nonincreasing order of their 11. end if
sizes: nk  nk1      n1 . Consider the routing tree 12. end if
shown in Fig. 2. Since the nodes cannot receive multiple The first block of the algorithm in lines 2-4 gives the
packets simultaneously, N is a trivial lower bound to scheduling rules between the sink and the roots of the top-
receive all the packets. Next, consider the largest top- subtrees. We define a top-subtree to be eligible if its root has at
subtree k, the root of which has to transmit nk packets to least one packet to transmit. For a given time slot, we
the sink, and the children of this root have to forward schedule the root of an eligible top-subtree which has the
nk  1 packets in total. Because of the half-duplex largest number of total (remaining) packets. If none of the top-
transceivers, time slots assigned to the root of this top- subtrees are eligible, the sink does not receive any packet
subtree must be distinct from all those assigned to its during that time slot.
children. Thus, in total, we need at least nk þ ðnk  1Þ ¼ Inside each top-subtree, nodes are scheduled according
2nk  1 distinct time slots. u
t to the rules in lines 5-12. We define a subtree to be active if
there are still packets left in the subtree (excluding its root)
We note that this bound of maxð2nk  1; NÞ, which to be relayed. If a node’s buffer is empty and the subtree
applies only when all the interfering links are removed, is rooted at this node is active, we schedule one of its children
smaller than the lower bound of 3N for general networks at random whose buffer is not empty. Our algorithm
and that of maxð3nk  3; NÞ for tree networks, as computed guarantees (as proved in Lemma 3) that in an active subtree,
by Gandham et al. [1] for the 2-hop interference model. They there will always be at least one child whose buffer is not
proposed a time slot assignment scheme for tree networks, empty, and so whenever a node empties its buffer, it will
92 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 11, NO. 1, JANUARY 2012

. Node i is full, while pðiÞ and cðiÞ are both empty.

. Nodes i and pðiÞ are both full, while cðiÞ is empty.
For the first case, we need to show that both pðiÞ and
cðiÞ (since now they are empty) have a child and a parent
whose buffers are full. Clearly, pðiÞ has a child with a full
buffer because i is now full. Similarly, pðiÞ also has a
parent with a full buffer because a transmission took
place from pðiÞ to its parent at t ¼ k þ 1. For the latter,
Fig. 3. Raw-data convergecast using algorithm LOCAL-TIMESLOTAS- cðiÞ has a parent with a full buffer because transmission
SIGNMENT: (a) Schedule length of 7 when all the interfering links are took place from cðiÞ to i at t ¼ k þ 1. If the child of cðiÞ,
removed. (b) Schedule length of 10 when the interfering links are
i.e., gcðiÞ, was empty at t ¼ k, then gcðiÞ also had a child
present.
with a full buffer because the lemma was true at t ¼ k.
Therefore, at t ¼ k þ 1, the child of gcðiÞ transmits and
receive a packet in the next time slot, thus emptying buffers
fills up its parent’s buffer. Otherwise, if gcðiÞ was full at
from the bottom of the subtree to the top.
t ¼ k, then it also remains full at t ¼ k þ 1 because it
We run through an example shown in Fig. 3a to explain
cannot transmit to its parent cðiÞ, which was full at t.
the algorithm. In the first time slot, since the eligible top-
For the second case, cðiÞ transmitted and pðiÞ did not.
subtree containing the largest number of remaining packets
For this to happen, gpðiÞ was full at t ¼ k and either
is f2; 5; 6g, we schedule the link ð2; sÞ, and the sink receives a
empties or remains full at t ¼ k þ 1. If it empties, gpðiÞ
packet from node 2 in slot 1. In the second time slot, the
has a parent with a full buffer because it transmitted at
eligible top-subtrees are f1; 4g and f3; 7g, both of which have
t ¼ k þ 1, and also has a child with a full buffer because
two remaining packets. We choose one of them at random,
pðiÞ did not transmit. If it remains full, at t ¼ k þ 1 nodes
say f1; 4g, and schedule the link ð1; sÞ. Also, in the same time
i, pðiÞ, and gpðiÞ are full, cðiÞ is empty, and gcðiÞ is full as
slot since node 2’s buffer is empty, it chooses one of its
we showed in the first case. So, the lemma holds for
children at random, say node 5, and schedule the link ð5; 2Þ.
t ¼ k þ 1, and the proof follows. u
t
In the third time slot, the eligible top-subtrees are f2; 5; 6g
and f3; 7g, both of which have two remaining packets. We Theorem 2. If all the interfering links are eliminated, the
choose the first one at random and schedule the link ð2; sÞ, schedule length for raw-data convergecast achieved by
and so the sink receives node 5’s packet (relayed by node 2). algorithm LOCAL-TIMESLOTASSIGNMENT is the minimum,
We also schedule the link ð4; 1Þ in the third time slot because i.e., maxð2nk  1; NÞ.
node 1’s buffer is empty at this point. This process continues Proof. Let ni be the number of nodes in top-subtree i.
until all the packets are delivered to the sink, yielding an Order the top-subtrees in nonincreasing order of their
P
assignment that requires seven time slots. Note that, in this sizes: nk  nk1      n1 . Suppose nk > k1 i¼1 i ; then
n
example, 2nk  1 ¼ 5, and so maxð2nk  1; NÞ ¼ 7. In Fig. 3b, maxð2nk  1; NÞ ¼ 2nk  1. From Lemma 1, we know
we show an assignment when all the interfering links are that it takes at least 2nk  1 slots to schedule all the
present, yielding a schedule length of 10. packets originated in top-subtree k. Out of these, the sink
In the following, we prove that the algorithm requires can use at most nk  1 slots to receive packets from the
exactly maxð2nk  1; N) slots when all the interfering links other top-subtrees, which P have a total of at most nk  1
are eliminated. Before giving the details of the proof, we packets. Also, when nk > i¼1 k1
ni , the root of the largest
first highlight the two key insights of the algorithm: 1) the top-subtree k gets scheduled once in every two time
sink is kept busy in receiving packets for as many time slots slots. Therefore, the schedule length is at most 2nk  1.
as possible, and 2) a node’s buffer is not empty for two or P
Now suppose nk  k1 i¼1 ni ; then maxð2nk  1; NÞ ¼
more consecutive time slots so long as the subtree rooted at N. We need to show that there always exists an eligible
this node is active. The first one is evident from the top-subtree to complement for the largest one when it is
scheduling rule between the sink and the top-subtrees. We not eligible. In this case, the sink will receive packets in
prove the second insight in the following lemma: every slot, because otherwise it remains P idle during some
Lemma 3. In an active subtree, a node with an empty buffer time slots and the first condition nk > k1 i¼1 ni will be met.
always has a child and a parent whose buffers are full. Thus, we will P prove that the algorithm keeps the
inequality nk  k1 i¼1 ni as an invariant.
Proof. We prove it by induction on time slot t. The parent
In any given time slot t, the algorithm schedules an
and grandparent of node i are denoted by pðiÞ and gpðiÞ;
eligible top-subtree that has the largest number of
similarly, a child and a grandchild of i are denoted by
remaining packets. At slot t þ 1, therefore, we have
cðiÞ and gcðiÞ, respectively. Slightly abusing notation, we
nk ¼ nk  1, and the following three cases might arise:
also use these symbols to denote the state of the buffers
on the respective nodes. . Top-subtree k still has the largest number of
At t ¼ 1, the lemma is trivially true because all the remaining packets with nk  nk1      n1 .
buffers are full. Suppose the lemma holds for t ¼ k, i.e., Then, the root of k is again chosen to transmit
every node whose buffer is empty has a child and a at t þ 1, and the inequality still holds as nk  1 
parent whose buffers are full. At t ¼ k þ 1, each node P k1
i¼1 ni .
with an empty buffer schedules one of its children whose . Top-subtree k and at least another one, say j, have
buffer is full. The following two situations can occur: an equal number of remaining packets. Then, the
INCEL ET AL.: FAST DATA COLLECTION IN TREE-BASED WIRELESS SENSOR NETWORKS 93

root of j is chosen, P and the inequality still holds higher than what is required by the threshold value, it
because nj  1  k1 i¼1 ni  1 (since nj ¼ nk  1). should decrease its power and if it is below the threshold, it
. Top-subtree k does not have the largest number of should increase its transmission power, within the available
remaining packets, implying that there were other range of power levels on the radio. If all the nodes meet
top-subtrees with an equal number of packets left the interfering constraint, the algorithm proceeds with the
as k in slot t. Then, the root of a new largest top- schedule calculation for the next time slot. On the other
subtree j P is chosen, and the inequality holds since hand, if the maximum number of iterations is reached and
nj  1  k1 i¼1 ni  1 (since nj ¼ nk ). there are nodes which cannot meet the interfering con-
Thus, the algorithm keeps the inequality as an straint, the algorithm excludes the link with minimum
invariant, and there always exists a top-subtree that can SINR from the schedule and restarts the iterations with the
be alternatelyPscheduled with the largest top-subtree. new subset of nodes. The power control phase is repeated
k1 until an admissible transmission scenario is found.
When nk ¼ 1, i¼1 ni  1 ¼ 1, which means that there are
two packets left at two different top-subtrees that can be
5.2 Multichannel Scheduling
scheduled in alternate slots. Since this inequality holds for
all the N steps, the sink always finds a top-subtree to Multichannel communication is an efficient method to
receive packets from, and therefore it takes N slots. eliminate interference by enabling concurrent transmissions
Moreover, Lemma 1 implies that a top-subtree becomes over different frequencies [21]. Although typical WSN radios
eligible after a transmission because its root is filled up in operate on a limited bandwidth, their operating frequencies
the next slot. Therefore, the theorem follows. u
t can be adjusted, thus allowing more concurrent transmis-
sions and faster data delivery. Here, we consider fixed-
bandwidth channels, which are typical of WSN radios, as
5 IMPACT OF INTERFERENCE opposed to the possibility of improving link bandwidth by
So far, we have focused on computing spatial-reuse TDMA consolidating frequencies. In this section, we explain three
schedules where transmissions take place on the same channel assignment methods that consider the problem at
frequency at a constant transmission power. In this section, different levels allowing us to study their pros and cons for
we focus on different methods to mitigate the effects of both types of convergecast. These methods consider the
interference on the schedule length. First, we discuss the channel assignment problem at different levels: the link level
benefits of using transmission power control and explain (JFTSS), node level (RBCA), or cluster level (TMCP).
the basics of a possible algorithm. Then, we discuss the
advantages of using multiple channels by considering three 5.2.1 Joint Frequency Time Slot Scheduling
different channel assignment schemes. JFTSS offers a greedy joint solution for constructing a
maximal schedule, such that a schedule is said to be
5.1 Transmission Power Control maximal if it meets the adjacency and interfering constraints,
In wireless networks, excessive interference can be elimi- and no more links can be scheduled for concurrent
nated by using transmission power control [6], [20], i.e., by transmissions on any time slot and channel without
transmitting signals with just enough power instead of violating the constraints. Approximation bounds on JFTSS
maximum power. To this end, we evaluate the impact of for single-channel systems and its comparison with multi-
transmission power control on fast data collection using channel systems are discussed in [22] and [23], respectively.
discrete power levels, as opposed to a continuous range JFTSS schedules a network starting from the link that has
where an unbounded improvement in the asymptotic the highest number of packets (load) to be transmitted.
capacity can be achieved by using a nonlinear power When the link loads are equal, such as in aggregated
assignment [5]. We first explain the basics of one particular convergecast, the most constrained link is considered first,
algorithm that we use in our evaluations in Section 7. i.e., the link for which the number of other links violating
The algorithm proposed by ElBatt and Ephremides [6] is the interfering and adjacency constraints when scheduled
a cross layer method for joint scheduling and power control simultaneously is the maximum. The algorithm starts with
and it is an optimal distributed algorithm to improve the an empty schedule and first sorts the links according to the
throughput capacity of wireless networks. The goal is to loads or constraints. The most loaded or constrained link in
find a TDMA schedule that can support as many transmis- the first available slot-channel pair is scheduled first and
sions as possible in every time slot. It has two phases: 1) added to the schedule. All the links that have an adjacency
scheduling and 2) power control that are executed at every constraint with the scheduled link are excluded from the list
time slot. First, the scheduling phase searches for a valid of the links to be scheduled at a given slot. The links that do
transmission schedule, i.e., largest subset of nodes, where no not have an interfering constraint with the scheduled link
node is to transmit and receive simultaneously, or to can be scheduled in the same slot and channel whereas the
receive from multiple nodes simultaneously. Then, in the links that have an interfering constraint should be sched-
given valid schedule, the power control phase iteratively uled on different channels, if possible. The algorithm
searches for an admissible schedule with power levels chosen continues to schedule the links according to the most
to satisfy all the interfering constraints. In each iteration, the loaded (or most constrained) metric. When no more links
scheduler adjusts the power levels depending on the can be scheduled for a given slot, the scheduler continues
current RSSI at the receiver and the SINR threshold with scheduling in the next slot. Fig. 4a shows the same tree

according to the iterative rule: Pnew ¼ SINR  Pcurrent . Accord- given in Fig. 1a which is scheduled according to JFTSS
ing to this rule, if a node transmits with a power level where aggregated data are collected. JFTSS starts with link
94 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 11, NO. 1, JANUARY 2012

are on frequency F 1. RBCA starts with the most interfered

parent, node 2 in this example, and assigns F 2. Then, it
continues to assign F 3 to node 3 as the second most
interfered parent. Since all interfering parents are assigned
different frequencies, sink can receive on F 1.

6 IMPACT OF ROUTING TREES

Fig. 4. Scheduling with multichannels for aggregated convergecast:
Besides transmission power control and multiple channels,
(a) Schedule generated with JFTSS. (b) Schedule generated with the network topology and the degree of connectivity also
TMCP. (c) Schedule generated with RBCA. affect the scheduling performance. In this section, we
describe schemes to construct topologies with specific
(2; sink) on frequency 1 and then schedules link (4,1) next properties that help to reduce the schedule length.
on the first slot on frequency 2. Then, links (5,2) on
6.1 Aggregated Data Collection
frequency 1 and (1; sink) on frequency 2 are scheduled on
the second slot and links (6,2) on frequency 1 and (3; sink) We first construct balanced trees and compare their
on frequency 2 are scheduled on the last slot. performance with unbalanced trees. We observe that in both
An advantage of JFTSS is that it is easy to incorporate the cases, the sink often creates a high-degree bottleneck. To
physical interference model; however, it is hard to have a overcome this, we then propose a heuristic, as described in
distributed solution since the interference relationship Algorithm 3, by modifying Dijkstra’s shortest path algorithm
between all the links must be known. to construct degree-constrained trees. Note that constructing
such a degree-constrained tree is NP-hard. Each source node
5.2.2 Tree-Based Multichannel Protocol i in our heuristic keeps track of the number of its children,
TMCP is a greedy, tree-based multichannel protocol for data CðiÞ, which is initialized to 0, and a hop count to the sink,
collection applications [8]. It partitions the network into HCðiÞ, which is initialized to 1. The algorithm starts with
multiple subtrees and minimizes the intratree interference by the sink node, and adds a node i0 62 T at every iteration to the
assigning different channels to the nodes residing on tree such that HCði0 Þ is minimized. It stops when jT j ¼ jV j,
different branches starting from the top to the bottom of or when no more nodes can be added to the tree because the
the tree. Fig. 4b shows the same tree given in Fig. 1a which is neighbors of all these new nodes have reached the limit on
scheduled according to TMCP for aggregated data collec- their maximum degree. Consequently, in this latter situation,
tion. Here, the nodes on the leftmost branch are assigned the heuristic might not always generate a spanning tree. In
frequency F 1, second branch is assigned frequency F 2, and our evaluation presented in Section 7.3, we consider only
the last branch is assigned frequency F 3 and after the those instances of the topologies where spanning trees with
channel assignments, time slots are assigned to the nodes the specified degree constraint are produced.
with the BFS-TimeSlotAssignment algorithm. The advantage Algorithm 3. DEGREE-CONSTRAINED TREES
of TMCP is that it is designed to support convergecast traffic 1. Input: GðV ; EÞ, s, max degree
and does not require channel switching. However, conten- 2. T fsg
tion inside the branches is not resolved since all the nodes 3. for all i 2 V do
on the same branch communicate on the same channel. 4. CðiÞ 0; HCðiÞ 1
5.2.3 Receiver-Based Channel Assignment 5. end for
6. HCðsÞ 0
In our previous work [7], we proposed a channel assign-
7. while jT j 6¼ jV j do
ment method called RBCA where we statically assigned the
channels to the receivers (parents) so as to remove as many 8. Choose i0 62 T such that:
interfering links as possible. In RBCA, the children of a 9. (a) ði; i0 Þ 2 E, for some i 2 T with CðiÞ <
common parent transmit on the same channel. Every node max degree  1
in the tree, therefore, operates on at most two channels, thus 10. (b) HCði0 Þ is minimized
avoiding pairwise, per-packet channel negotiation over- 11. T T [ fi0 g
heads. The algorithm initially assigns the same channel to 12. HCði0 Þ ¼ HCðiÞ þ 1
all the receivers. Then, for each receiver, it creates a set of 13. CðiÞ CðiÞ þ 1
interfering parents based on SINR thresholds and itera- 14. if 8i 2 V , CðiÞ ¼ max degree then
tively assigns the next available channel starting from the 15. break
most interfered parent (the parent with the highest number 16. end if
of interfering links). However, due to adjacent channel 17. end while
overlaps, SINR values at the receivers may not always To illustrate the gains of degree-constrained trees,
be high enough to tolerate interference, in which case the consider the case when all the N nodes are in range of
channels are assigned according to the ability of the each other and that of the sink. If the nodes select their
transceivers to reject interference. We proved approxima- parents according to minimum hop without a degree
tion factors for RBCA when used with greedy scheduling in constraint, then all of them will select the sink, and this
[9]. Fig. 4c shows the same tree given in Fig. 1a scheduled will give a schedule length of N. However, if we limit the
with RBCA for aggregated convergecast. Initially, all nodes number of children per node to 2, then this will result in
INCEL ET AL.: FAST DATA COLLECTION IN TREE-BASED WIRELESS SENSOR NETWORKS 95

two subtrees rooted at the sink, and if there are enough

frequencies to eliminate interference, the network can be
scheduled using only two time slots, thus achieving a factor
of N=2 reduction in the schedule length.

6.2 Raw-Data Collection

As emphasized in [13], routing trees that allow more parallel
transmissions do not necessarily result in small schedule
lengths. For instance, the schedule length is N for a network
Fig. 5. Balanced tree construction: Node 4 is attached to b2 based on the
connected as a star topology, whereas it is ð2N  1Þ for a line search sets; load on both b1 and b2 is 5.
topology once interference is eliminated. Theorem 1
suggests that the routing tree should be constructed such the largest cardinality for further processing, breaking ties
that all the branches have a balanced number of nodes and based on the smallest id. We define the growth set of a node
the constraint nk < ðN þ 1Þ=2 holds. In this section, we
as the set of neighbors (potential children) that are not yet
construct such routing trees.
connected to the tree and have larger hop counts.
A balanced tree satisfying the above constraint is a
Rule 3. Once a node is chosen based on the growth sets
variant of a capacitated minimal spanning tree [24]. The CMST
according to Rule 2, we construct search sets (SS) to decide
problem, which is known to be NP-complete, is to
determine a minimum-hop spanning tree in a vertex which potential branch the node should be added to. A
weighted graph such that the weight of every subtree search set is thus branch specific and includes the nodes
linked to the root does not exceed a prescribed capacity. In that are not yet connected to the tree and are neighbors of a
our case, the weight of each link is 1, and the prescribed node that are at a higher hop count. In particular, if the
capacity is ðN þ 1Þ=2. Here, we propose a heuristic, as chosen node has access to branch b, and has a neighbor that
described in Algorithm 4, based on the greedy scheme can connect to only branch b if b is selected, then this
presented by Dai and Han [25], which solves a variant of the neighbor and its potential children are included in the
CMST problem by searching for routing trees with an equal search set for b. However, if the neighbor has access to at
number of nodes on each branch. We augment their scheme least one other branch even after b is selected, then it is not
with a new set of rules and grow the tree hop by hop included in the search set.
outward from the sink. We assume that the nodes know The search sets guarantee that the choices for the nodes
their minimum-hop counts to sink. at longer hops to join a particular branch are not limited by
the decision of the joining node. This balances out the
Algorithm 4. CAPACITATED-MINIMALSPANNINGTREE
number of nodes on different branches and prevents one to
1. Input: GðV ; EÞ, s grow faster than others. Once the search sets are con-
2. Initialize: structed, we choose the branch for which the sum of its load
3. B roots of top subtrees // the branches (W ) and the size of the search set is minimum.
4. T fsg [ B To illustrate the merit of search sets, consider the
5. 8i 2 V , GSðiÞ unconnected neighbors of i situation shown in Fig. 5. Dotted lines represent potential
at further hops communication links and solid lines represent already
6. 8b 2 B, W ðbÞ 1 included tree edges. At this point, node 4 is being
7. h 2 processed, and the loads on branches b1 and b2 are 2 and
8. while h 6¼ max hop count do 4, respectively, where bi denote the branch rooted at node i.
9. Nh unconnected nodes at hop distance h The search set SSð4; b1Þ is f8; 9; 10g, because the neighbor
10. Connect nodes Nh0 that have a single potential node 8 has access to only b1 if b1 is selected by node 4.
S
parent: T T Nh0 However, the search set SSð4; b2Þ is empty, because the
11. Update Nh Nh n Nh0 neighbor node 8 has access to another branch b1 (via node
12. Sort Nh in non-increasing order of jGSj 3). Therefore, the sum of the load and the size of the search
13. for all i 2 Nh do set for b1 is 5, and that for b2 is 4. So, we attach node 4 to b2,
14. for all b 2 B to which i can connect do and in the next step attach node 8 to b1. This balances out
15. Construct SSði; bÞ the number of nodes over the two branches.
16. end for
17. Connect i to b for which W ðbÞ þ jSSði; bÞj is 7 EVALUATION
minimum
In this section, we evaluate the impact of transmission
18. Update GSðiÞ and W ðbÞ
S S power control, multiple channels, and routing trees on the
19. T T fig SSði; bÞ
scheduling performance for both aggregated and raw-data
20. end for
convergecast.
21. h hþ1 We deploy nodes randomly in a region whose dimen-
22. end while sions are varied between 20  20 m2 and 300  300 m2 to
Rule 1. Nodes with single potential parents are con- simulate different levels of density. The number of nodes is
nected first. kept fixed at 100. For different parameters, we average each
Rule 2. For nodes with multiple potential parents, we point over 1,000 runs. We use an exponential path-loss
first construct their growth sets (GS) and choose the one with model for signal propagation with the path-loss exponent 
96 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 11, NO. 1, JANUARY 2012

Fig. 6. Scheduling on minimum-hop trees with and without power Fig. 7. Scheduling on minimum-hop trees with multiple channels:
control: (a) Aggregated convergecast. (b) Raw-data convergecast. (a) Aggregated convergecast. (b) Raw-data convergecast.

varying between 3 and 4, which is typical for indoor 7.1.2 Raw-Data Convergecast
environments. We also use the physical interference model For raw-data convergecast, we observe in Fig. 6b that the
and simulate the behavior of CC2420 radios that are used schedule length increases as the network gets sparser on
on Telosb and TmoteSky motes and are capable of minimum-hop trees. This is counterintuitive because in
operating on 16 different frequencies. The transmission sparse networks, the reuse of slots should be higher which
power can be adjusted between 24 and 0 dBm over eight would reduce the schedule length. However, as the network
different levels, and the SINR threshold is set to  ¼ 3 dB.1 gets sparser, the number of nodes that can directly reach the
We first evaluate the schedule length for single-channel
sink decreases and packets have to be relayed over more
TDMA, and then its improvement using transmission
hops. Thus, more packets need to be scheduled than in a
power control, multiple channels, and routing trees.
single hop. We see that the number of packets to be scheduled
7.1 Impact of Transmission Power Control increases faster than the reuse ratio. In the densest setting
We investigate two cases: 1) when nodes transmit at where all the nodes can directly reach the sink, the schedule
maximum power, and 2) when nodes adjust their length is 99, which is equal to the number of sources.
transmission power according to the algorithm described With power control, we observe a reduction in the
in Section 5.1. In both cases, nodes communicate on the schedule length in Fig. 6b as some of the interfering links
same channel and use minimum-hop routing trees. In the are eliminated, thus increasing slot reusability. When
first case, time slots are assigned according to BFS-  ¼ 3:0, most of the interference can be eliminated by
TIMESLOTASSIGNMENT for aggregated data, and accord- power control, and beyond which the structure of the
ing to LOCAL-TIMESLOTASSIGNMENT for raw data. In the routing tree, especially the number of nodes nk on the
second case, we follow the scheduling rules in [6]. largest branch with ð2nk  1Þ > N, becomes the bottleneck.
However, for   3:5, power control cannot always elim-
7.1.1 Aggregated Convergecast inate interference as networks get sparser and nodes tend to
Fig. 6a shows the variation of schedule length with density transmit at their maximum power.
for different values of  on minimum-hop trees. We
observe that the schedule length decreases as the deploy- 7.2 Impact of Multichannel Scheduling
ment gets sparser. This happens because at low densities, In this section, we analyze the performance of the channel
the interference is less, and so more concurrent transmis- assignment methods discussed in Section 5.2. We use
sions can take place. In the densest deployment (L ¼ 20) CC2420 radios that have 16 channels in the 2.4 GHz range,
when all the nodes are within the range of each other, the with adjacent channels overlapping according to the
sink is the only parent, and the network is scheduled in rejection and blocking values given in the data sheet.
99 time slots regardless of power control. However, in We assume that the nodes transmit at maximum power and
sparser scenarios, using power control, the network can be use minimum-hop trees. In TMCP and RBCA, time slots are
scheduled with fewer time slots as the level of interference assigned according to BFS-TIMESLOTASSIGNMENT for
goes down. We achieve a 10-20 percent reduction in
aggregated convergecast and LOCAL-TIMESLOTASSIGN-
schedule length for the best case.
MENT for raw-data convergecast. The path-loss exponent
We also observe that power control is more effective in
is taken as 3.5.
reducing the schedule length for denser deployments than
in sparser ones where the results tend to be similar. This is 7.2.1 Aggregated Convergecast
due to the discrete power levels and limited power range.
Comparing the plots in Figs. 7a and 6a, we observe that the
Moreover, due to the 95 dBm threshold for the transcei-
channel assignment methods achieve schedule lengths that
vers to be able to decode a signal successfully, further
are shorter than those achieved by power control. While it’s
power reduction is limited.
true that power control helps in reducing the effects of
1. Due to variation in signal strength, a fading margin can be included interference, this gain is limited due to the discrete levels and
such that some of the packets can still be captured if the RSSI is slightly limited range of transmission power (e.g., CC2420 has eight
lower than the threshold. Such a model [26] can easily be incorporated in
our experiments, in which case retransmissions of lost packets should also different power levels between 0 and 24 dBm). In sparse
be considered in calculating the schedule length. deployments, nodes cannot reduce their transmit power
INCEL ET AL.: FAST DATA COLLECTION IN TREE-BASED WIRELESS SENSOR NETWORKS 97

below a certain threshold because a transceiver cannot

decode signals below the sensitivity level 95 dBm). As
shown in Fig. 6a, for L > 200, the schedule lengths are similar
when nodes transmit at maximum power or when they
adjust their power levels. Moreover, in mid-sparse deploy-
ments (60  L  180, Fig. 6a), the limited range and discrete
power levels restrict the nodes to adjust their transmit
powers. On the other hand, multichannel communication
even with just two frequencies (Fig. 7a) can eliminate the
interference limitations, and beyond this, the performance
gains are limited by the connectivity structure. By transmit-
ting on different channels, interference is eliminated by the Fig. 8. (a) Bounds on the number of frequencies. (b) Percentage of
high adjacent/alternate channel rejection values of the C2420 incorrectly scheduled links.
radio, and the channels behave like orthogonal.
In Fig. 7a, sparser deployments (L > 140) with multi- With RBCA and JFTSS, the number of channels required is
channel communication show a 40 percent reduction in low for dense networks as the number of receivers is low. In
schedule length as compared to transmitting on a single particular, when L ¼ 20, all the nodes can directly connect to
channel with maximum power. However, in denser the sink, and so only one frequency is needed. As the network
deployments, multiple channels do not help much due to gets sparser, the number of receivers increases, and thus
increased connectivity, with the sink as a bottleneck in the more frequencies are required to support concurrent
densest setting. From Fig. 7a, we observe that JFTSS and transmissions. However, for L  80, the number of nodes
RBCA can optimally schedule the network using 16 that are being connected to the same parent slowly dominates
channels, i.e., they achieve the lower bound, as shown by the effect of the number of receivers, and since the network
the line “Lower Bound-MHST.” The advantage of RBCA gets very sparse, the number of channels required further
over JFTSS is that it takes into account the topological goes down as the level of interference decreases.
characteristics: a parent node receives data on the same The trends of both RBCA and JFTSS are quite similar, and
channel from its children, and does not have to switch the number of channels required is no more than the number
channels in every slot. In dense deployments, TMCP of available channels on CC2420 radios (16 channels). On the
performs better due to the different routing trees con- other hand, TMCP requires many more channels as each
structed, i.e., when L ¼ 20, RBCA and JFTSS construct a star branch is on a different channel. This is expensive for
topology, whereas TMCP constructs a 2-branch tree with deployments where a lot of nodes can directly connect to the
two channels and a 16-branch tree with 16 channels. sink, and thus are assigned different channels because they
form different branches. Thus, one needs to optimize the
7.2.2 Raw-Data Convergecast
channel usage.
In Fig. 7b, we observe that none of the methods can
eliminate interference completely with two channels; 7.2.4 Interference Models and Orthogonal Channels
however, JFTSS and RBCA can do so with six or more We now evaluate the impact of interference and channel
channels at different densities (plots are not shown due to models on the schedules, in particular, the idealistic
lack of space). We also see that TMCP needs 16 channels to assumption of the protocol model and the orthogonality
reach a performance similar to that achieved by RBCA and
assumption. We examine the feasibility of the schedules
JFTSS with only two channels. This is because in JFTSS and
based on the adjacent and alternate channel rejection values
RBCA, when a node is receiving from its children, its parent
of the transceivers and the SINR threshold.
can transmit simultaneously on a different channel, which
Fig. 8b shows the results for JFTSS in terms of the
is not possible due to intrabranch interference in TMCP.
percentage of nodes that are incorrectly scheduled (hence-
The results also verify that JFTSS and RBCA can achieve a
forth, referred to as errors). The top two lines show the
schedule length which is bounded by maxð2nk  1; NÞ,
shown as “Lower Bound-MHST,” so long as the number of errors for two and 16 channels with both the assumptions,
available channels is sufficient to eliminate interference. whereas the bottom line shows the errors only for the
Compared to the results on a single channel in sparser orthogonality assumption. We observe that the errors
scenarios, we achieve a reduction of up to 40 percent on the are much higher in sparser deployments, because although
schedule length. In very dense scenarios, the improvement the interference created by an individual sender is not high
is small because most of the nodes can directly reach the enough to jam concurrent transmissions, the cumulative
sink, and so the limiting factor becomes the half-duplex effect from multiple senders is very high, which is not
transceiver. captured in the protocol model. On the other hand, in dense
deployments, a single transmitter can jam another one
7.2.3 Required Number of Channels because of smaller internode distances and higher level of
In this section, for the different channel assignment interference. In such cases, some of the nodes might select
methods, we evaluate the required number of channels to the next available channels for concurrent transmissions;
completely eliminate interference as a function of deploy- however, interference can still be high because the channels
ment density. In our simulation results, as shown in Fig. 8a, in reality are not perfectly orthogonal. After the peak, the
we assume that the number of available channels is network gets sparser and interference reduces. We note
unlimited so as to show the upper bounds. that, our simulations corroborate previous results [27] that
98 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 11, NO. 1, JANUARY 2012

holds. In such cases, the schedule length is limited by

2nk  1. These results indicate that RBCA and JFTSS
combined with a suitable tree construction mechanism can
achieve a reduction of up to 50 percent in the schedule
length as compared to single-channel communication on
minimum-hop spanning trees.

8 CONCLUSIONS
In this paper, we studied fast convergecast in WSN where
nodes communicate using a TDMA protocol to minimize
Fig. 9. (a) Scheduling on degree-constrained minimum-hop trees. the schedule length. We addressed the fundamental
(b) Scheduling on CMST. limitations due to interference and half-duplex transceivers
on the nodes and explored techniques to overcome the
the protocol model may result in serious interference, and same. We found that while transmission power control
adjacent channel interference cannot always be ignored. helps in reducing the schedule length, multiple channels are
more effective. We also observed that node-based (RBCA)
7.3 Impact of Routing Trees and link-based (JFTSS) channel assignment schemes are
In the preceding sections, we observed that although more efficient in terms of eliminating interference as
interference can be substantially eliminated by using power compared to assigning different channels on different
control and multiple channels, connectivity of the tree still branches of the tree (TMCP).
limits the performance. In the following, we discuss the Once interference is completely eliminated, we proved
improvements with routing tress. that with half-duplex radios, the achievable schedule length
is lower bounded by the maximum degree in the routing
7.3.1 Aggregated Convergecast on Degree-Constrained tree for aggregated convergecast, and by maxð2nk  1; NÞ
Trees for raw-data convergecast. Using optimal convergecast
Fig. 9a shows the variation of schedule length with density scheduling algorithms, we showed that the lower bounds
when the maximum tree degree is 3 (in sparser scenarios, are achievable once a suitable routing scheme is used.
with a maximum degree of 2, it was not always possible to Through extensive simulations, we demonstrated up to an
construct connected topologies). The top two lines are for order of magnitude reduction in the schedule length for
nodes transmitting at maximum power, and nodes using aggregated, and a 50 percent reduction for raw-data
power control. The bottom two lines are for JFTSS and convergecast. In future, we will explore scenarios with
RBCA. When nodes transmit with maximum power, we variable amounts of data and implement and evaluate the
observe a reduction in the schedule length in dense combination of the schemes considered.
deployments as compared to non-degree-constrained trees
shown in Fig. 6a. We also notice further improvement with REFERENCES
power control in denser deployments than in sparser ones. [1] S. Gandham, Y. Zhang, and Q. Huang, “Distributed Time-Optimal
When nodes are assigned channels using RBCA, we see a Scheduling for Convergecast in Wireless Sensor Networks,”
factor of more than 2 reduction in the schedule length in Computer Networks, vol. 52, no. 3, pp. 610-629, 2008.
[2] K.K. Chintalapudi and L. Venkatraman, “On the Design of MAC
dense deployments (L < 120), as compared to that using Protocols for Low-Latency Hard Real-Time Discrete Control
RBCA on minimum-hop spanning trees. We also observe Applications over 802.15.4 Hardware,” Proc. Int’l Conf. Information
that the schedule lengths are much larger than the Processing in Sensor Networks (IPSN ’08), pp. 356-367, 2008.
maximum degree in the routing tree for dense deploy- [3] I. Talzi, A. Hasler, G. Stephan, and C. Tschudin, “PermaSense:
Investigating Permafrost with a WSN in the Swiss Alps,” Proc.
ments, as compared to those in sparse scenarios (L  120).
Workshop Embedded Networked Sensors (EmNets ’07), pp. 8-12, 2007.
Considering deployments at different densities, routing [4] S. Upadhyayula and S.K.S. Gupta, “Spanning Tree Based
over minimum-hop degree-constrained spanning trees Algorithms for Low Latency and Energy Efficient Data Aggrega-
together with RBCA achieves an order of magnitude tion Enhanced Convergecast (DAC) in Wireless Sensor Net-
works,” Ad Hoc Networks, vol. 5, no. 5, pp. 626-648, 2007.
improvement than routing over minimum-hop spanning
[5] T. Moscibroda, “The Worst-Case Capacity of Wireless Sensor
trees while transmitting at maximum power. When we use Networks,” Proc. Int’l Conf. Information Processing in Sensor
JFTSS, the schedule length is close or equal to the maximum Networks (IPSN ’07), pp. 1-10, 2007.
degree since it can handle interference using multiple [6] T. ElBatt and A. Ephremides, “Joint Scheduling and Power
Control for Wireless Ad-Hoc Networks,” Proc. IEEE INFOCOM,
channels more effectively by reusing and assigning them to
pp. 976-984, 2002.
the links instead of the receivers. [7] Ö. Durmaz Incel and B. Krishnamachari, “Enhancing the Data
Collection Rate of Tree-Based Aggregation in Wireless Sensor
7.3.2 Raw-Data Convergecast on CMST Networks,” Proc. Ann. IEEE Comm. Soc. Conf. Sensor, Mesh and
Ad Hoc Comm. and Networks (SECON ’08), pp. 569-577, 2008.
Fig. 9b shows the variation of schedule length on CMST. [8] Y. Wu, J.A. Stankovic, T. He, and S. Lin, “Realistic and Efficient
The impact of such routing trees is more prominent in Multi-Channel Communications in Wireless Sensor Networks,”
sparser networks (L  200) than routing over minimum- Proc. IEEE INFOCOM, pp. 1193-1201, 2008.
hop spanning trees (Fig. 7b). When L < 200, the length is [9] A. Ghosh, Ö. Durmaz Incel, V.A. Kumar, and B. Krishnamachari,
“Multi-Channel Scheduling Algorithms for Fast Aggregated
bounded by N. Beyond this point, it is almost always not Convergecast in Sensor Networks,” Proc. IEEE Int’l Conf. Mobile
possible to construct trees where the constraint 2nk  1 < N Adhoc and Sensor Systems (MASS ’09), pp. 363-372, 2009.
INCEL ET AL.: FAST DATA COLLECTION IN TREE-BASED WIRELESS SENSOR NETWORKS 99

[10] V. Annamalai, S.K.S. Gupta, and L. Schwiebert, “On Tree-Based Özlem Durmaz Incel received the BSc and
Convergecasting in Wireless Sensor Networks,” Proc. IEEE MSc degrees in computer engineering from the
Wireless Comm. and Networking Conf. (WCNC ’03), vol. 3, Yeditepe University, Turkey, in 2002 and 2005,
pp. 1942-1947, 2003. respectively. She received the PhD degree in
[11] X. Chen, X. Hu, and J. Zhu, “Minimum Data Aggregation Time computer science from the University of
Problem in Wireless Sensor Networks,” Proc. Int’l Conf. Mobile Ad- Twente, Netherlands, in March 2009. She is
Hoc and Sensor Networks (MSN ’05), pp. 133-142, 2005. currently a postdoctoral researcher in the
[12] W. Song, F. Yuan, and R. LaHusen, “Time-Optimum Packet Networking Laboratory (NETLAB) of Bogazici
Scheduling for Many-to-One Routing in Wireless Sensor Net- University, Turkey. Her dissertation is focused
works,” Proc. IEEE Int’l Conf. Mobile Ad-Hoc and Sensor Systems on efficient data collection in wireless sensor
(MASS ’06), pp. 81-90, 2006. networks and was entitled “Multi-Channel Wireless Sensor Networks:
[13] H. Choi, J. Wang, and E. Hughes, “Scheduling for Information Protocols, Design and Evaluation.” She was a visiting student in the
Gathering on Sensor Network,” Wireless Networks, vol. 15, pp. 127- Autonomous Networks Research Group of the University of Southern
140, 2009. California as part of her PhD studies from 2007-2008.
[14] N. Lai, C. King, and C. Lin, “On Maximizing the Throughput of
Convergecast in Wireless Sensor Networks,” Proc. Int’l Conf. Amitabha Ghosh received the BS degree in
Advances in Grid and Pervasive Computing (GPC ’08), pp. 396-408, physics from the Indian Institute of Technology
2008. Kharagpur, India, in 1996, and the ME degree
[15] M. Pan and Y. Tseng, “Quick Convergecast in ZigBee Beacon- in computer science and engineering from the
Enabled Tree-Based Wireless Sensor Networks,” Computer Comm., Indian Institute of Science, Bengaluru, in 2000.
vol. 31, no. 5, pp. 999-1011, 2008. He received the PhD degree in electrical
[16] W. Song, H. Renjie, B. Shirazi, and R. LaHusen, “TreeMAC: engineering from the University of Southern
Localized TDMA MAC Protocol for Real-Time High-Data-Rate California in August 2010. Between 2000 and
Sensor Networks,” J. Pervasive and Mobile Computing, vol. 5, no. 6, 2004, he worked for Alcatel-Lucent and
pp. 750-765, 2009. Honeywell on GSM/GPRS networks and wire-
[17] G. Zhou, C. Huang, T. Yan, T. He, J. Stankovic, and T. Abdelzaher, less ad hoc networks. He is currently a postdoctoral researcher at
“MMSN: Multi-Frequency Media Access Control for Wireless Princeton University. His research interest is in design and analysis
Sensor Networks,” Proc. IEEE INFOCOM, pp. 1-13, 2006. of algorithms for scheduling, routing, and power control for cellular
[18] Y. Kim, H. Shin, and H. Cha, “Y-MAC: An Energy-Efficient Multi- and wireless sensor networks.
Channel MAC Protocol for Dense Wireless Sensor Networks,”
Proc. Int’l Conf. Information Processing in Sensor Networks (IPSN ’08), Bhaskar Krishnamachari received the BE
pp. 53-63, Apr. 2008. degree in electrical engineering from The Cooper
[19] B. Krishnamachari, D. Estrin, and S.B. Wicker, “The Impact of Union in 1998, and the MS and PhD degrees in
Data Aggregation in Wireless Sensor Networks,” Proc. Int’l Conf. electrical engineering from Cornell University in
Distributed Computing Systems Workshops (ICDCSW ’02), pp. 575- 1999 and 2002, respectively. He is an associate
578, 2002. professor in the Ming Hsieh Electrical Engineer-
[20] J. Zander, “Performance of Optimum Transmitter Power Control ing Department at the University of Southern
in Cellular Radio Systems,” IEEE Trans. on Vehicular Technology, California Viterbi School of Engineering. His
vol. 41, no. 1, pp. 57-62, Feb. 1992. research interests are in the design and perfor-
[21] P. Kyasanur and N.H. Vaidya, “Capacity of Multi-Channel mance analysis of protocols for next-generation
Wireless Networks: Impact of Number of Channels and Inter- wireless networks. He received the US National Science Foundation
faces,” Proc. ACM MobiCom, pp. 43-57, 2005. CAREER Award in 2004.
[22] G. Sharma, R.R. Mazumdar, and N.B. Shroff, “On the Complexity
of Scheduling in Wireless Networks,” Proc. ACM MobiCom, Krishnakant Chintalapudi received the MS
pp. 227-238, 2006. degree in electrical engineering from Drexel
[23] X. Lin and S. Rasool, “A Distributed Joint Channel-Assignment, University and the PhD degree in computer
Scheduling and Routing Algorithm for Multi-Channel Ad-Hoc science from the University of Southern
Wireless Networks,” Proc. IEEE INFOCOM, pp. 1118-1126, 2007. California. He is currently working as an
[24] C.H. Papadimitriou, “The Complexity of the Capacitated Tree associate researcher at Microsoft Research,
Problem,” Networks, vol. 8, no. 3, pp. 217-230, 1978. India. His research interests include delay-
[25] H. Dai and R. Han, “A Node-Centric Load Balancing Algorithm sensitive and fault-tolerant wireless sensor
for Wireless Sensor Networks,” Proc. IEEE Conf. Global Telecomm. networks, high-data-rate wireless sensor net-
(GlobeCom ’03), pp. 548-552, 2003. work applications, hybrid sensor networks, and
[26] M. Zuniga and B. Krishnamachari, “An Analysis of Unreliability wireless control and automation.
and Asymmetry in Low-Power Wireless Links,” ACM Trans.
Sensor Networks, vol. 3, no. 2, p. 7, 2007.
[27] J. Grönkvist and A. Hansson, “Comparison between Graph-Based
and Interference-Based STDMA Scheduling,” Proc. ACM Mobi- . For more information on this or any other computing topic,
Hoc, pp. 255-258, 2001. please visit our Digital Library at www.computer.org/publications/dlib.