1
Directed Budget-Based Clustering for WSN
Leonidas Tzevelekas, Ioannis Stavrakakis
Department of Informatics & Telecommunications
National and Kapodistrian University of Athens
LOCAN 2006
2
Large-scale Wireless Sensor Networks
Sensor motes characteristics Cheap, tiny, embedded devices Used in orders of thousands (or millions) Resulting in extremely high network densities Also: extreme energy constraints of individual nodes Also: short-range wireless connectivity only possible
Wireless Sensor Network characteristics Autonomous operation (no human interventions possible) Should be self-organizing, self-healing, self-* Overall: may exhibit arbitrarily sophisticated behavior Highly distributed networking environment (new methods of
network organization and operation)
3
Hierarchical network self-organization
Network self-organization in clustersEnhances sensor node coordinationNetwork managementIn-network processing of sensed data
Clusters with fixed-size number of sensorsReduced routing protocol overheadAccommodating specific service requirements
Distributed/ decentralized cluster formationRadically decentralized algorithms required (Rapid, Persistent)Low message complexity -> energy efficiency
Our contribution– A strictly localized algorithm for fixed-size cluster formation in
large-scale Wireless Sensor Networks
4
Distributed cluster formation: Budget-based Clustering
Distributed cluster formation methodology (adopted in our work)
Main idea: Growing cluster’s nodes do even budget distributions of tokens among their first-hop neighbors
Algorithm description (formal)– An initiator node is chosen randomly in the network (among
unclustered nodes)– Initiator assigned a budget of B tokens of which it accounts
one for himself and distributes B-1 evenly among its neighbors– Subsequent nodes receiving a budget do the same until the
budget is exhausted or no more growth is possible
5
Rapid, Persistent examplesB
2
A
CD
E F
G
2
2
2
1
B=112
Rapid cluster
2A
B
CD
EF
G
2
2
6
5
B=11 2
H
I
J1 1
2
Persistent cluster
B.B. Persistent algorithmPersistent algorithm• Recursive elaboration of RapidRecursive elaboration of Rapid• EvenEven distribution of budget to neighbors distribution of budget to neighbors
except from the parent nodeexcept from the parent node• Persistent re-distribution of budget shortfall Persistent re-distribution of budget shortfall
(if any)(if any)• Good clustering performance network-wide Good clustering performance network-wide
at cost of higher message complexityat cost of higher message complexity
A.A. Rapid algorithmRapid algorithm• Fast, one-way budget Fast, one-way budget distribution processdistribution process• Even Even distribution of budget distribution of budget among neighbors except from among neighbors except from the parent nodethe parent node• No accounting for wasted No accounting for wasted tokenstokens• Poor clustering performance Poor clustering performance network-widenetwork-wide
6
Initiators selection process
1. Randomized initiators methodology– Nodes run count-down timers with exponentially distributed
initial values– Nodes become initiators when their timer fires– Bounded probability that multiple initiators are concurrently
active in the same neighborhood
2. Sequential approximation for initiator picking (useful for computer simulations)
– Next initiator picked only after currently growing cluster completes
– Associated cluster is allowed to fully grow– Only one initiator active network-wide at each time instant– Identical with optimistic randomized timers methodology
7
Rapid, Persistent major drawback “Blindness” in budget distribution process
No awareness of neighbor’s clustering status at each distributing node
Even budget distribution always among ALL physical neighbors
Tokens directed to “bad” neighbors => token waste
Tokens are frequently wasted/ returned (Rapid/ Persistent)
– Resulting in bad clustering performance: Very low average clustersizes (Rapid) High number of budget shortfall redistributions (Persistent)
8
Proposals for improved clustering performance
1. Fighting inter-cluster token distribution contentions– Nodes already clustered under previous inititiator => receive NO
tokens from growing cluster Tokens should be directed away from clustered nodes Eliminate inter-cluster token distribution contentions (sequential
initiators) Significantly reduce inter-cluster token distribution contentions
(randomized initiators)
2. Fighting intra-cluster token distribution contentions– A growing cluster’s tokens should not contend for common
unclustered nodes Significantly reduces token distribution contentions for a single
growing cluster
9
Directed Budget-Based Clustering (DBB)
Assumption for radically distributed networks – Periodic HELLO messages to set-up/ maintain local physical
network topology=> Same HELLO messages to set-up/ maintain local clustered
network topology DBB algorithm’s specific characteristics
Utilize HELLO messages to convey additional clustering status information of nodes
Minimal overhead of 1-bit flag only (at HELLO messages) Some overhead for storing clustering status information in tables
(at nodes) Nodes update their neighbors’ clustering status information prior to
executing the algorithm’s steps Algorithm’s steps coincide with the periodicity of HELLO messages Clustering messages (tokens, ACKs) embedded into HELLO
messages
10
Directed Budget-Based Clustering (DBB)(Fighting inter-cluster token distribution contentions)
Example: Spatial evolution of clustering process for DBB
Tokens “bounce” on clustered nodes of another initiator and are directed away, thus avoiding to be wasted or returned
clustering process becomes completely transparent in localized HELLO message exchanges STRICTLY LOCALIZED CLUSTERING PROTOCOL
11
Directed Budget-Based Clustering with Random Delays (DBB-RD)
Effect: to “desynchronize” budget distributions at neighboring nodes for a single growing cluster
1. Introducing: random delay factor r as an integer amount of rounds of HELLO message exchanges to delay the current budget distribution at each node
2. Advantage: subsequent HELLO message exchanges allow for updating of the clustering status information among nodes
3. Drawback: additional delays to complete overall network decomposition
(Fighting intra-cluster token distribution contentions)
12
Network simulation scenario
Network simulation settings N=6000 nodes Square plane of size l=1000m with random x, y coordinates
for each node Individual nodes transmission range r=25m Average connectivity degree ρ=11.781 nodes Clusterbound targeted B=30/ 60 for medium/ large-sized
clusters Connectivity of graph is checked by displacing any
disconnected nodes after initial random placement Sequential picking of initiators is used (always one cluster
growing in the network) Random delay factor r є [0,α-1), where a is random delay
parameter
13
Network simulation scenario
Metrics:1. Time required or consecutive rounds of HELLO
message exchanges required till overall network decomposition
2. Average clustersize achieved over all clusters formed (optimum is the targeted bound B)
3. Average number of clusters in the network formed (optimum is I=N/B)
Analysis of results
K=5 independent runs for each set of parameter settings
Measured quantities are averaged and 0.95-confidence intervals are presented
14
Simulation results for Rapid/ Persistent
Rapid: low average clustersize compared with B (8.69 when B=30 and 11.13 when B=60)
Rapid: very fast network decomposition (3258 rounds for 6000 nodes when B=30)
Persistent: high average clustersize compared to B (21.03 when B=30 and 37.99 when B=60)
Persistent: up to six times more rounds required than Rapid (18992 rounds for 6000 nodes with B=30 compared with 3258 rounds for Rapid)
Sims verify negative effects of token waste in clustering performance of Rapid, Persistent
15
Simulation results for DBB/ DBB-RD
DBB: results indicate clustering performance improvements due to avoiding inter-cluster token waste
DBB: average clustersizes significantly higher compared with Rapid (28% higher for B=30, 26.6% higher for B=60)
DBB: Faster network decomposition than both Rapid AND Persistent
DBB-RD: results confirm the positive effect of “desynchronization” of budget distributions (fighting intra-cluster token distribution contentions)
DBB-RD: higher clustersizes than DBB for all values of α, though additional overall delay for network decomposition
16
DBB/ DBB-RD average decomposition time
a є {0, 3, 5, 10}
(a=0 is DBB algorithm)
Additional delay in network decomposition time with growing interval for random delay factor r
17
DBB/ DBB-RD average clustersizes
a є {0, 3, 5, 10}
From a=0 to a=3, relative increase of metric by 27.85%
From a=3 to a=5, relative increase of metric by 4.88%
From a=5 to a=10, relative increase of metric lower than 4%
18
THANK YOU
Any questions?
19
Hierarchical network self-organization for large scale sensor networks
Our workOur work: a : a strictly localized strictly localized protocolprotocol aiming at decomposing large aiming at decomposing large scale sensor networks into non-scale sensor networks into non-overlapping clusters of bounded sizeoverlapping clusters of bounded size
Localized protocols, algorithms for network self-organization seem to fit to special characteristics/ constraints of the WSN
20
Localized/ strictly localized protocols /algorithms
LOCALIZED protocols/ algorithmsInherently distributed algorithms utilizing local interactions among neighbor nodes to achieve a well-defined global objective overall in the networkAlready used for maintaining/updating local network topology at each node, and other things like energy efficient flooding, broadcasting, etc.Primarily enabled through the exchange of periodic HELLO messages among 1-hop neighbors of nodes
STRICTLY LOCALIZED protocols/ algorithmsInformation processed by a node is either (a) local in nature or (b) global in nature, but obtainable by querying only the node’s neighbors or itself