Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
Sensor nodes, charged with battery power and capable of wireless communications,
are distributed around to survey the environments and to send the needed data to the base
station. The limited energy resources of sensor nodes are consumed in computation and
especially in communication. To reduce power consumption for the sensor network, this
paper presents a new data gathering scheme based on the hypercube topology. The hy-
percube-based data gathering scheme collects data from all sensor nodes to the base sta-
tion through the communication tree of the constructed hypercube. It shortens commu-
nication delay by parallel transmission and replaces dead nodes through reconfiguration.
Data gathering by the distributed hypercube and binary tree is also provided.
Keywords: wireless sensor networks, data gathering, base stations, hypercubes, binary
trees, centralized and distributed approaches, simulation and performance evaluation
1. INTRODUCTION
concurrent functioning. Balancing node loads is also important. When a sensor node un-
dertaking extreme workloads dies prematurely, its neighbor nodes must consume more
energy to make up for the broken link. Besides, a good data gathering scheme should
work under all possible environments to increase system scalability.
Based on the observation, this paper presents a hypercube-based new data gathering
scheme for sensor networks. The new scheme collects data from sensor nodes to the BS
through the communication tree of a hypercube. It shortens communication delay by par-
allel transmission and replaces dead nodes through reconfiguration. Simulation results
show the new scheme outperforms other schemes in terms of transmission delay, energy
loads, network lifetime and system scalability. To complete our investigation, data gath-
ering by the distributed hypercube and binary tree is also provided.
2. BACKGROUND STUDY
The proposed data gathering scheme adopts the first order radio model in [3, 4]. In
this model, a radio dissipates Eelec = 50nJ/bit to run the transmitter or receiver circuitry
and εamp = 100pJ/bit/m2 for the transmitter amplifier. For a k-bit message and a distance d,
the needed transmission energy is
tion of their labels differs in one bit position k. Such two nodes are called dimension-k
neighbors of each other and the link between them is their dimension-k link. The advan-
tages of the hypercube topology are as follows.
• It takes only m = log2 (N) delay time to collect data from a network of N nodes.
• At each turn, each node transmits data only once.
• Nodes are selected as heads with the same probability, ensuring more balanced energy
(load) distribution.
Our hypercube-based data gathering scheme forms the total system into a log2N-
cube, N being the number of nodes. A node will be the root by turns at each round. After
data are transmitted to the root through the communication tree embedded in the hyper-
cube, the root will transmit the gathered data to the BS. The hypercube is constructed by
the bottom-up approach: Every two separate nodes are formed into a 1-cube, two 1-cubes
into a 2-cube and two 2-cubes into a 3-cube until all nodes are included in the con-
structed cube. Assume that all nodes transmit their locations to the BS after deployment
and the BS takes the following steps to build links between two nodes:
All 1-cubes are thus constructed, and the connected node pairs are dimension-0
neighbors of each other. Constructing dimension-1 links involves a similar process ex-
cept that the square of distance between two nodes is now computed to form two 1-cubes
into a 2-cube − the needed transmission energy grows according to the square of the
transmission distance. Two values (sums of squares) will result from two possible con-
nections: Take the connection with the smaller value as the dimension-1 link.
The following notations are needed for further discussion.
Constructing 1-dimensional cubes: To form 0-cubes into 1-cubes, first put all nodes in
C0 and all links between any two nodes in E0. Each link takes the square of its length as
the weight, to reflect the required transmission energy. Repeatedly delete the maximum
weight of links until a node, say a, with d(a) = 1 is found. Remove node a and the node
connecting to it, say b. Combine the two 0-cubes into a 1-cube and put it in C1. The link
between a and b is marked as the dimension-0 link. Delete all other links relevant to a
and b. If two or more node pairs have the degree of 1, choose the node pair with the least
weight. Repeat the process until C0 is null (i.e., the construction of 1-cubes is completed).
Fig. 2. The four possible connections to combine two 2-cubes into a 3-cube.
Constructing a 3-dimensional cube: There are 2i ways to form two i-cubes into an (i +
1)-cube. Fig. 2 shows the four (22) possible connections to combine two 2-cubes into a
3-cube. Each of the 2-cubes (abdc and efgh) needs four more links to form the 3-cube.
Put the selected connection between abdc and efgh (i.e., the connection with the mini-
mum weight among the four connections) to E2.
Two special situations: Lines 8-12 and 13-16 in the above algorithm deal with two spe-
cial situations.
1. Lines 8-12: In finding the dimension-(i + 1) neighbor for each i-cube, if there remains
only one i-cube in Ci, increase the dimension number of the cube’s links by 1, let each
node be its own dimension-0 neighbor, and put it into Ci+1. (Note that links of higher
dimensions are less likely to be used. In such a pseudo-complete hypercube, each node
may have more than one label and yet transmit only once at each stage.)
2. Lines 13-16: When a cube c in Ci with d(c) = 0, i.e., c has no related connections with
other cubes in Ci, re-compute the links between c and the other cubes in Ci. For exam-
ple, when three cubes connect with one another like a triangle, deleting any one con-
nection will isolate one of the cubes.
1160 PO-JEN CHUANG, BO-YI LI AND TUN-HAO CHAO
• Fig. 3 (a): There are 6 nodes in the system. Delete the longer links until there remain
only 6 links.
• Fig. 3 (b): Delete bc in E0, choose nodes a and b, and then delete links relevant to node
a, leaving node c with d(c) = 0.
• Fig. 3 (c): Put all links relevant to c and the other nodes in E0.
• Fig. 3 (d): Dimension-0 links are thus established.
• Fig. 3 (e): Connect the two 1-cubes ab and cd into a 2-cube (abcd). C1 now has a
1-cube ef.
• Fig. 3 (f): Increase the dimension number of links for e and f, let e and f each be their
own dimension-0 neighbors. Put eeff to C2.
Deciding the dimension-k neighbor of a node: The above algorithm does not decide
the dimension-k neighbor of a node by the shortest distance because when choosing the
shortest link on one side, we may be forced to choose a longer link on the other side,
producing uneven load distribution for the nodes. Fig. 4 gives an example case. In the
case, the dimension-0 link between nodes b and c is decided by the shortest distance,
which leads to the construction of the dimension-0 link between nodes a and d. Fig. 4
finally forms a 2-cube similar to the one in Fig. 1 (f) except that in Fig. 1 (f) the longest
connection is a dimension-1 link.
The hypercube transmission scheme: After the hypercube is constructed, data can be
transmitted through the embedded communication tree. The transmission scheme [5, 6]
for a 4-cube is illustrated in Fig. 5, in which the node at the start of an arrow line/at the
arrow head is a sender/receiver node. To start transmission, a head node is first chosen by
turns. Each of the other nodes compares its own label with the head’s label from LSB to
HYPERCUBE-BASED DATA GATHERING 1161
MSB: Among the difference bits, the smallest bit number (say k) indicates the stage at
which the node needs to transmit data to its dimension-k neighbor. For example, if the
head node is 0000, node 0101 needs to transmit data at stage 0 and the receiver will be its
dimension-0 neighbor node 0100. Similarly, if the head node is 1001, node 0101 must
transmit data at stage 2 to node 0001. Transmission completes at stage log2 N, N = the
number of nodes. (As the probability for a link to be used is 1/2k, k being the dimension
of the link, we will find the lowest dimension link first when constructing the cube.)
In the proposed scheme, the BS dominates the construction of a hypercube and the
hypercube needs to be reconstructed when a node is dead. To trim down the complexity
of re-constructing a cube, we set a node replacement policy. Each node in the hypercube
will broadcast, at each round, a packet to all neighbors to notify it is alive. If a node fails
to receive such a packet from a specific neighbor node j for a certain period of time, it
will inform other nodes that j is dead. Each of j’s neighbor nodes then broadcasts a mes-
sage including its own location and the connection bit number k between itself and node j
to the network. Upon receiving the information, each node n computes C(n) − the in-
creased average cost per round when n attains the label of the dead node j − and get
d (i, n) 2
C ( n) = ∑ 2k
.
i∈N
N is the set of neighbors of node j; d(i, n) is the distance between nodes i and n. The
cost is computed by the square of distance d(i, n) times the probability of the link to be
used in a round (1/2k). After each neighbor node of j broadcasts its computed value to the
network, the node with the minimum value will take the label of the dead node.
1162 PO-JEN CHUANG, BO-YI LI AND TUN-HAO CHAO
The proposed hypercube transmission scheme builds a cube by the bottom-up ap-
proach conducted in a centralized way by the BS. For completeness, we also investigate
the feasibility of gathering data through a hypercube or a binary tree which is constructed
by a different approach − the top-down approach − in a distributed way.
The algorithm shows that, at each step, every selected node will request a nearest
unselected node to be its upstream. If an unselected node receives only one such request,
it will take the sender as its downstream and mark itself “selected”; if receiving multiple
requests, it will take the nearest sender as its downstream.
Transmission for a distributed hypercube: Each node will pass the data received from
its upstream nodes. After gathering data from all upstream nodes, the head will transmit
HYPERCUBE-BASED DATA GATHERING 1163
the fused data to the BS. To gather data by such a distributed hypercube approach takes a
total of log2 N steps. Whenever a node dies, the topology needs re-construction.
To gather data by the distributed hypercube approach, each node needs to have one
immediate upstream node to conduct tree transmission. If there are no sensor nodes in the
neighboring area of a node, it needs to take farther upstreams to assist data gathering –
consuming extra energy. For improvement, we can use the distributed binary tree which
is more flexible on the number of allowed immediate upstream nodes to gather data. (The
flexibility may save a significant amount of energy in the data gathering process.) The
distributed binary tree approach allows each node to have at most two immediate up-
stream nodes. To construct a binary tree, the BS first selects the nearest node as the root.
At each round data are collected through the tree to the root which then transmits the
fused data to the BS. The topology must be re-constructed when a node dies (possibly the
root − because the dead nodes will always appear in the neighboring area of the BS).
Constructing the binary tree: The BS first picks up the nearest node as the root. The
root then takes two closest nodes to be its children and asks them to find their own chil-
dren. The offspring of all nodes are decided in this way. To avoid choosing a child of
long distance, we set a maximum distance constraint D. When the distance between two
nodes exceeds D, one can not be the child of the other. If a node is chosen as child by
more than one node, it will become the child of the nearest node. A node which fails to
find any child or receives connection tables from all of its children must return an inte-
grated table (including its connection status) to the parent. After receiving tables from all
children, the root will broadcast a query to check if any node is left out of the tree. If any,
the root will pass the information along the tree and those nodes with less than two chil-
dren now get the chance to find another child with a larger D. When all nodes are in-
cluded, tree construction is completed. Repeatedly enlarging the distance constraint to
build a complete tree is acceptable because a binary tree needs re-construction only when
a node dies. Fig. 7 gives an example of constructing the binary tree.
In Fig. 7, node d is first taken as the root, with D as the initial distance constraint,
for constructing the binary tree (a). Node d then selects the two nearest nodes, i.e., nodes
c and g, as its upstreams (b). The two upstream nodes move on to have their own up-
streams: Node c takes node e; node g takes nodes e and h. As node e is closer to node g,
it becomes the upstream of node g (c). Node f becomes the upstream of node h in the
same way, and tree construction stops here (d). Seeing that nodes a and b are not in-
cluded in the tree, the root passes the command along to re-find children with a larger D.
Node a is thus taken as the upstream of node c (e) and node b becomes the upstream of
node a (f).
After the root is decided, all nodes will execute the following algorithm to construct
the binary tree (the status of a node is specified by a variable S).
The algorithm does not immediately set S to 1 following “send a request to become
the child of the node” (line 6), but instead “notify my upstream to change S to 1 and the
HYPERCUBE-BASED DATA GATHERING 1165
current D” later in case 1. The reason can be traced in the following example in which
node A (already with a child) requests node B to be its another child.
Deciding the transmission stages: After all nodes are included in the binary tree, the BS
computes the stage by the following algorithm for each node to send data in each round
and passes the schedule up the tree. Nodes then send data based on the schedule in after
rounds. The transmission stage starts from stage 0. A node will transmit data to its parent
only after receiving and fusing data from all of its children. If two children of a parent
conduct data transmission at the same stage, one has to delay to the next stage. The fol-
lowing are parameters used in the algorithm to decide the transmission stages.
A node in WAIT will be assigned a stage number and removed from WAIT when
no children are in the WAIT and no siblings carry the same stage number to be assigned
to it. The arrangement allows a parent to get a transmission stage after its children and
sibling nodes to have different stage numbers.
5. PERFORMANCE EVALUATION
Table 1 lists the required transmission stages (the data gathering delay time) at each
round for various schemes (n = the number of nodes). As shown here, the hypercube-
based scheme takes less stages than PEGASIS and LEACH to gather data.
1166 PO-JEN CHUANG, BO-YI LI AND TUN-HAO CHAO
The simulation model: Conducted using NS-2, our simulation adopts the same radio
model and network topology as PEGASIS [4]. The system has 100 nodes, initially
charged with 0.25J energy and randomly scattered to a 100m * 100m network. The BS is
located at (50, 250). Each node has a packet (2,000 bits) to send to the BS at each round.
The Binary tree approach sets a 10m basic distance constraint between any two
nodes. When the distance constraint fails to include all nodes in the tree, the approach
enlarges the constraint by 10m at each tree construction attempt − until the tree is com-
pleted. As the BS is located far from the sensor nodes, LEACH is set to have only one
head at each round to save energy. In PEGASIS and Hypercube, the BS performs the
task of constructing topologies. The cost for constructing topologies is negligible and is
therefore left out. All schemes will reconstruct their topologies when a node dies.
1200
1000 Hypercube
800 PEGASIS
PEGBin
Round
600 LEACH
400 Direct
Binary tree
200
DH
0
100 90 80 70 60 50 40 30 20 10 0
Number of Living Nodes
100 87.76
76.49
80
Average Delay
60
40
20 14.87
6.79 6.84 6.48
0
LEACH PEGASIS PEGBin Hypercube DH Binary tree
Protocol
Fig. 10. The average delay of total inter-node transmissions per round.
nodes as the PEGASIS. By contrast, the Binary tree approach maintains the most living
nodes − partly because it is more flexible with the number of upstream nodes and partly
because it selects the node that uses the least energy to transmit data to the BS.
• the average delay of total inter-node transmissions per round before all nodes turn
dead (Fig. 10)
The DH, PEGBin and Hypercube are shown to yield the lowest average delay. The
Binary tree yields more than twice the delay because it takes double time of log2 n.
• the average energy dissipation for inter-node transmissions per round (Fig. 11)
The Binary tree, PEGASIS and Hypercube consume less energy per round. PEGBin
performs not as well because it optimizes only the first transmission step of each round.
(For clearer illustration, the result excludes the large amount of energy consumed by the
head node when sending the gathered data to the BS.)
Fig. 11. The average energy dissipation for inter-node transmissions per round.
5 4.58
Average Energy*Delay
4
3
1.99
2
800
700 re-construction
600 reconfiguration
500
Round
400
300
200
100
0
100 90 80 70 60 50 40 30 20 10 0
Number of Living Nodes
Fig. 13. The round vs. the number of living nodes for the Hypercube.
• the rounds vs. the number of living nodes for the Hypercube (Fig. 13)
The result is collected for the Hypercube under reconstruction or reconfiguration
when a node dies. “Reconfiguration” indicates the reconfiguring process to replace the
dead node rather than re-construct the hypercube itself. Reconfiguration, whose per-
formance slightly trails behind that of reconstruction, actually saves a lot of work for the
BS. When it is not desirable for the BS to reconstruct a faulty hypercube, system recon-
figuration can be an alternative.
HYPERCUBE-BASED DATA GATHERING 1169
6. CONCLUSIONS
To save the limited energy resources of sensor nodes consumed during data trans-
mission, this paper presents a new data gathering scheme for the sensor network. Based
on hypercube transmission, the new scheme concurrently collects data from all sensor
nodes to the base station through the communication tree of the constructed hypercube.
Such a hypercube transmission scheme is desirable because it can shorten communica-
tion delay by parallel transmission and replace a dead node through topology reconfigu-
ration. Experimental evaluation shows that our hypercube-based data gathering scheme
outperforms previous schemes in terms of transmission delay, energy load distribution,
system scalability and system lifetime. To complete our investigation, data gathering by
the distributed hypercube and binary tree is also provided and discussed.
REFERENCES
1. M. Dong, K. Yung, and W. Kaiser, “Low power signal processing architectures for
network microsensors,” in Proceedings of International Symposium on Low Power
Electronics and Design, 1997, pp. 173-177.
2. I. F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci, “A survey on sensor
networks,” IEEE Communications Magazine, Vol. 40, 2002, pp. 102-114.
3. W. Heinzelman, A. Chandrakasan, and H. Balakrishnan, “Energy-efficient communi-
cation protocol for wireless microsensor networks,” in Proceedings of the 33rd Ha-
waii International Conference on System Sciences, 2000, pp. 3005-3014.
4. S. Lindsey, C. Raghavendra, and K. Sivalingam, “Data gathering algorithms in sensor
networks using energy metrics,” IEEE Transactions on Parallel and Distributed Sys-
tems, Vol. 13, 2002, pp. 924-935.
5. Y. R. Leu and S. Y. Kuo, “A fault-tolerant tree communication scheme for hypercube
systems,” IEEE Transactions on Computers, Vol. 45, 1996, pp. 641-650.
6. P. J. Chuang, S. Y. Chen, and J. T. Chen, “Constructing fault-tolerant communication
trees in hypercubes,” Journal of Information Science and Engineering, Vol. 20, 2004,
pp. 39-55.