Self-Organization Techniques for Large Multi-hop Networks: Energy-Efficiency and Robustness Sonia Fahmy Center for Wireless Systems and Applications (CWSA) Department of Computer Sciences Purdue

1
Self-Organization Techniques for Large Multi-hop
Networks Energy-Efficiency and RobustnessSonia
FahmyCenter for Wireless Systems and
Applications (CWSA)Department of Computer
SciencesPurdue Universityfahmy_at_cs.purdue.eduhtt
p//www.cs.purdue.edu/homes/fahmy/
2
Scenario

• Application-specific networks, e.g., sensor
  networks used for monitoring and surveillance
• Nodes typically
• Densely deployed
• Unattended during their lifecycle
• Limited in processing, memory, and communication
  capabilities
• Constrained in battery lifetime

3
Application Goals
Scalability, data and state aggregation,
robustness, and prolonged network lifetimeHow?

• Deploy redundant nodes
• Apply topology management to distribute energy
  consumption

4
In this talk, we give an overview of

• Distributed node clustering (HEED) utilizing
  multiple communication power levels
• Data aggregation using HEED in TinyOS
• Robust node clustering in harsh environments
• Node synchronization

5
System Model

• A set of n nodes are uniformly and independently
  dispersed in a field

Network

• ad-hoc, unattended, and does not necessarily have
  any infrastructure

• quasi-stationary, equally significant, location
  un-aware, and each possesses a number of
  transmission power levels

Nodes
Application

• Source-driven or data-driven

6
Node Clustering Goals

• Design a new clustering approach that has the
  following properties
• Completely distributed
• Terminates in O(1) iterations
• Has low message/processing overhead
• Generates high energy, well-distributed cluster
  heads
• Can accommodate other goals, such as balanced or
  dense clusters

7
Node Clustering Approach

• Hybrid, Energy-Efficient, Distributed (HEED)
  clustering is distributed
• Every node only uses information from its 1-hop
  neighbors
• HEED is hybrid
• Node elects to become a cluster head based on two
  parameters primary (residual energy) and
  secondary (communication cost)
• HEED is energy-efficient
• Elects cluster heads rich in residual energy
• Reclustering distributes energy consumption

8
HEED Algorithm

• Discover neighbors within cluster range
• Compute the initial cluster head probability
  CHprob f(Er/Emax)

• Initialization

• Clustering

• If v received some cluster head messages, choose
  one head with min cost
• If v does not have a cluster head, elect to
  become a cluster head with CHprob .
• CHprob min(CHprob 2, 1)
• Repeat until CHprob reaches 1

• If cluster head is found, join its cluster
• Otherwise, elect to be cluster head

• Finalization

9
HEED Example
Discover neighbors
(0.4,3)
(0.6,2)
a10
(0.1,4)
a13
c2
a11
(0.2,2)
(0.2,5)
a7
Compute CHprob and cost
a8
(0.5,3)
a12
(0.2,3)
c3
(0.2,3)
a9
(0.8,4)
(0.1,4)
Elect to become cluster head
(0.1,2)
c1
a5
a6
(0.9,4)
a14
(0.5,4)
a2
c4
Resolve ties
a4
(0.6,4)
(0.3,2)
a3
(0.7,5)
(0.2,3)
Select your cluster head
(0.3,2)
a1
10
HEED Properties

• Completely distributed
• Clustering terminates in O(1) iterations
• Message overhead O(1) per node
• Processing overhead O(n) per node
• Cluster heads are well distributed
• Produces a connected multi-hop cluster head graph
  (structure) asymptotically almost surely if Rt
  6Rc

11
Performance
Cluster head energy
Termination
Clustering overhead
Network Lifetime
12
Application Data Aggregation

• An observer collects readings from all sensors in
  the network
• A routing tree is constructed on the overlay of
  cluster heads
• Cluster heads act as aggregation points
• Consider simple aggregates, such as MAX, AVG, etc

observer
13
iHEED

• Integrates TinyOS multi-hop routing with node
  clustering to construct a clustered aggregation
  tree
• Has a soft-state view of residual energy
• Uses a Credit-Point (CREP) system to compute the
  dissipated energy for different sensor operations
• Decrements the available points periodically
  according to the node duty cycle and the
  transmission power level used

14
iHEED Design
15
iHEED Evaluation
Mica2 nodes
barrier
observer
Mica2Dot nodes
16
Robust Clustering

• In hostile environments, such as military fields
  or volcanic areas, sensor nodes may fail
  unexpectedly
• Thus, higher degree of connectivity is desirable

17
Robust Clustering REED

• Goal
• construct k fault-tolerant network
• How?
• Every node should have k distinct cluster heads
• Every cluster head should be able to communicate
  with at least k other cluster heads

REED approach Construct k distinct
(node-disjoint) cluster head overlays
18
REED Example
(1) Discover neighbors within cluster range
(0.3)
(0.35)
c2
(0.1)
c1
c1
(0.2)
(0.4)
c2
(0.25)
(2) Compute CHprob and cost
c2
c2
(0.2)
c1
(0.2)
(0.8)
(0.1)
(3) Elect your cluster head for overlay c1 then c2
(0.1)
c1
(0.9)
(0.4)
c2
c2
c1
(0.6)
(0.3)
c2
(4) Elect to become cluster head if one is not
found
(0.2)
(0.3)
(0.3)
19
Node Synchronization
TW Time Window

• Some applications require that data falls within
  certain time windows to be aggregated
• Synchronization is crucial for sleep-wakeup
  cycles to save energy
• Services, such as TDMA scheduling, requires at
  least regional node synchronization

TW for sensor 1
TW for sensor 2
TW for sensor 3
TW for sensor 4
1 min
Observer
20
Node Synchronization

• Within a cluster, a cluster head can synchronize
  its members in constant time using
  receiver-receiver low-level synchronization
  (e.g., RBS)
• For data-driven applications, reactive path
  synchronization can be used
• For source-driven applications, another overlay
  of cluster heads can synchronize the routing
  cluster head overlay in constant time

21
Summary

• Node clustering can significantly prolong network
  lifetime
• We have proposed HEED for clustering nodes
• HEED is integrated with data aggregation trees
  and implemented in TinyOS
• An extension of HEED is proposed for constructing
  k fault-tolerant networks in harsh environments
• High-level node synchronization can be carried
  out on a clustered network for efficient data
  aggregation and duty-cycle scheduling