[ppt]
TRANSCRIPT
Efficient Clustering for Improving Network
Performance in Wireless Sensor Networks
Bracha HodJoint work with:
Tal Anker, Danny Bickson and Danny Dolev
The Hebrew University of Jerusalem
Outline
Introduction Related work Motivation and main contribution Belief Propagation (BP) Clustering using BP Simulation results Summary
Introduction Cluster-based network is divided
into subsets Each group of nodes contains a
single leader (cluster head) and several ordinary nodes
Introduction Clustering main objectives
Minimize the total transmission power aggregated over all nodes in the selected path
Balance the load to prolong the network lifetime
Clustering advantages Increase network scalability Support data aggregation Reduce energy consumption
Clustering challenges Optimal cluster selection is a hard problem Cluster maintenance is essential
Related Work Many research efforts
LEACH - Low Energy Adaptive Clustering Hierarchy (Heinzelman et al, 2002)
HEED - Hybrid, Energy-Efficient, Distributed clustering (Younis et al, 2004)
VCA - Voting-based Clustering Algorithm (Qin et al, 2005)
EEUC - Energy-Efficient Unequal Clustering (Li et al, 2005)
Motivation
Network performance is important Retransmission and dropped
packets may waste energy Since the network is usually dense
and several nodes are redundant, network lifetime should be measured by the time that the system is available for providing services
Main Contribution We propose a novel approach based on
Belief Propagation (BP) Considers both local properties of a node
and joint characteristics of a group of nodes Utilizes better the available information Incurs small constant overhead
Resulting in Better network performance Balanced power consumption among the
nodes Our scalable and practical implementation
of BP can be used for other inference goals
Belief Propagation (BP) BP is an iterative algorithm for
computing maximum or marginal posterior probabilities by a local message passing
BP is associated with rapid convergence, accurate results and good performance in asynchronous environment
When performed on trees, BP converges to the correct values in a finite number of iterations
The Min-Sum Algorithm (MS) The goal is to minimize the overall cost
in the network, based on the local cost functions and the constraints between the nodes
Each node transmits to its neighbors a message with its local and joint costs. Each neighbor updates its own belief accordingly and transmits the new belief
Gradually the information is propagated through the network until the nodes converge to a common belief
Efficient Implementation The Min-Sum variation of BP
requires simple operations and works well with integer values Saves floating-point calculation
Broadcast messages instead of the traditional unicast messages in BP Preserves communication resources
The routing tree is used as the message-passing tree No special maintenance and overhead
Example Network Routing tree
which used also as the message-passing treeA
B
E
C D5
76
4
4
8
100
80
85 83
151
A
B
E
C D
Example Round 1: Messages transmitted by all the
nodes A B C D E
Processing by node A: A: A->A+B->A+C->A+D->A = 80+7+6+8 =
101 B: A->B+B->B+C->B+D->D = 7+85+5+83 =
180 If B is selected to be the cluster head, D
selects itself …
A
B
E
C D5
7 6
4
4
8100
80
85 83
151
E 151C 4B 65535D 65535
D 83A 8C 4
C 100A 6B 5D 4E 4
B 85A 7C 5E 65535
A 80B 7C 6D 8
Example
E 155C 104B 65535D 65535
D 91A 88C 10
C 110A 237B 163D 163E 235
B 92A 87C 11E 65535
A 101B 180C 115D 180
Round 2: Messages transmitted by all the nodes
A B C D E
Processing by node A: A->A: 80 B->A: 87 – 80 = 7 C->A: 237 – 80 = 157
The message from E is propagated to A D->A: 88 – 80 = 8 Total cost: 80 + 7 + 157 + 8 = 252
A
B
E
C D5
7 6
4
4
8100
80
85 83
151
Example
E 235C 110B 65535D 65535
D 180A 101C 115
C 119A 252B 331D 331E 252
B 180A 101C 115E 65535
A 252B 331C 119D 331
E 252C 119B 65535D 65535
D 331A 252C 119
C 119A 252B 331D 331E 252
B 331A 252C 119E 65535
A 252B 331C 119D 331
Round 3: Messages transmitted by all the nodes
A B C D E
After round 3, all the nodes converge to a common belief – node C should be the cluster head A B C D E
A
B
E
C D5
7 6
4
4
8100
80
85 83
151
Clustering using MS Two events trigger a clustering process
A regular node does not have a cluster head
Periodically by each cluster head to balance the power
Message passing properties 1-hop vicinity for localized and distributed
process Complete asynchronous operation Number of rounds is bounded and
determines a-priory to avoid impact of the environment
Clustering using MS Every message contains
Self cost of being a cluster head Cost of connecting to other cluster
heads Final state decision (on last round)
Cost metric Self cost is based on the expected
energy consumption in a period and the residual battery power The expected energy consumption
considers degree and distance from the base station
Joint cost is based on link quality and the residual battery power
Simulation Model Simulation in TOSSIM, TinyOS simulator 250 nodes including a single base station Link Estimation and Parent Selection
routing protocol Shortest path metric combined with link quality
Surge application for data aggregation Power information of Berkeley Mica2 mote Variable power levels for cluster heads and
regular nodes
HEED Cluster heads are selected with a
probability based on their residual energy
When there are no cluster heads’ announcements, a node selects itself with the probability it has or alternatively doubles its probability for the next round
Local and efficient method which achieves very good results on simulations
Performance Evaluation Data collection time
Clustering using BP achieves more than 40% higher throughput than HEED
Performance Evaluation Data collection rate during the
network lifetime
Clustering with BP achieves better routing, deployment and stability
Performance Evaluation BP has better deployment and
network stability than HEED
Average hop count Re-clustering processes
Performance Evaluation Clustering
Overhead
BP suffers from more overhead because larger size of messages
Network Lifetime
HEED has very small advantage because using BP more packets are transmitted
Summary We present a new framework for
clustering based on BP This approach is fully distributed,
localized, asynchronous, robust and scalable
Utilization of all available information and not only subset of parameters yields better results and better network performance
Future work Comparing the BP algorithm with the
theoretical optimal clustering algorithm
Thank You!
Appendix – Notations of BP - set of possible states of node i - local distribution function of
node i - joint function of two
connected nodes i and j - unicast message from
node i to node j at round t about the state that node j should be
- broadcast message from node i to its direct neighbors N(i)
MS Formulation Message passing
Message update rule
Belief calculation
Efficient implementation in sensors by broadcast messages and integer calculations only