Clustering in Mobile Ad-Hoc Networks Ovidiu Valentin, DRUGAN - - PowerPoint PPT Presentation

Clustering in Mobile Ad-Hoc Networks

Ovidiu Valentin, DRUGAN Department of Informatics, University of Oslo, Norway

Outline

• Clustering in MANETs
• Routing Protocol Clustering in MANETs

– Issues for clustering in routing – Clustering approaches for routing

• Dynamic clustering in the overlay

– Communication non-intrusive clustering – Evaluation

• Conclusions & References

2

Motivation

• Application Scenario for Mobile Ad-Hoc Network (MANET): Rescue operations

and emergency interventions

– Properties:

• Network without a fixed infrastructure and topological structure that allows mobile nodes to

create a temporary communication network

– Information sources:

• Mobile devices, wireless sensors, stationary devices, Internet, …

– Important information to be shared:

• Medical records, layout of buildings, installations, dangerous goods, collected evidence, …

– Cooperation is necessary …

3

Clustering

• Definition: division of the network into different virtual groups, based on

rules in order to discriminate the nodes allocated to different sub-networks

• Goal: achieve scalability in presence of large networks and high mobility
• Information sources: routing and higher level

4

Properties: Geographically allocated Balance resource use Service localization

Nodes roles in a cluster

• Roles of nodes in a cluster

– Cluster-Head: local coordinator of a cluster – Cluster-Member: ordinary node – Cluster-Gateway: node with inter- cluster links, forwards information between clusters – Cluster-guests: a node associated to a cluster

5

Cluster-Head Cluster Cluster-Member Cluster-Gateway

Graphs

• A network is an undirected graph

– G(V,E)  Graph G with a set V of nodes (vertices) and a set E of links (edges)

• Graphs specific measures

– Node degree: number of edges incident to the node – Paths in the graph

• Diameter: length of the longest path in the graph
• Shortest path: between 2 nodes in the network

– Centrality measures

• Closeness: measures how many steps is required to access

every other node from a given node

• Betweenness: number of shortest paths going through a node
• r an link

6

2 5 4 6 7 8 3 9 1

   

8 ) 5 , ( 45 . ) 5 , ( 9 , 6 , 3 , 4 , 5 )) 9 , 5 ( , ( 5 ) ( 2 2, 3, 4, 2, 2, 3, 2, 2, 2, degree(G)      G Betw G Clos G SP G Diam

Outline

• Clustering in MANETs
• Routing Protocol Clustering in MANETs

– Issues for clustering in routing – Clustering approaches for routing

• Dynamic clustering in the overlay

– Communication non-intrusive clustering – Evaluation

• Conclusions & References

7

Routing and Communication

• Routing: Nodes perform route discovery

and maintenance

– Flat: works fine for small networks but might not work in large MANETs

• Proactive: messages communication
• verhead
• Reactive: high overhead just from route

discovery

– Hierarchical: may work fine for large networks

• Localized route search and information

dissemination

• Communication flows: follow hierarchical

structures (i.e., social and organizational)

8

) (

2

n O

Routing Protocol Clustering in MANET

• Clustering Goals

– Achieve communication scalability for a large number of nodes and high mobility – Spatial reuse and coordination of resources

• Increase system capacity
• Reduce retransmissions and collisions
• Balance the use of resources in the network

– Virtual communication backbone

• Inter-cluster communication can be restricted to cluster-heads and cluster-gateways

– Local changes

• Update and maintain cluster information only locally
• Minimize information stored and propagated in the network

9

Advantages and Disadvantages

• Advantages

– Reusability: spatial reuse of resources at nodes – Simplification: of addressing – Stability and Localization: smaller and potentially mode stabile sub-network structures

• Disadvantages

– Explicit control messaging: clustering related information exchange – Ripple effect: rebuild of cluster structure in case of network structure changes – Stationary period: collect and exchange information for cluster formation – Computation rounds: number of rounds to complete the cluster election – Communication complexity: amount of control messages exchanged – No common solution

10

Classification

• DS-based clustering

– Route maintenance actions to the nodes from the dominating set

• Mobility-aware clustering

– Cluster based on the mobility behavior of the mobile nodes

• Energy-efficient clustering

– Consider the energy available at the nodes

• Load-balancing clustering

– Limit the number of nodes in a cluster in order to distribute the workload.

• Combined-metrics clustering

– Considers multiple metrics

• Low-maintenance clustering

– Perform clustering for upper-layers and reduce the maintenance cost

11

DS-Clustering

• Idea: Dominating Set (DS): in a graph G =(V, E) is a subset D
• f V such that every node not in D is joined to at least one

member of D by an edge from E

– Agglomerative methods: each node assumes at the beginning a cluster-head role and connected cluster can be merged

12

• Example1: Connected DS

▪ A node announces in the set of connected nodes ▪ Inspects its neighborhood for complete inclusion into D, if true it removes itself from D ▪ Moving nodes send beacons at periodic intervals to inform the CDS about movement

DS-Clustering (2)

• Summary

– Clusters:

• 1-hop non-overlapping clusters

– Communication complexity in case of mobility:

• |V| moving nodes  cost O(2|V|) (i.e., two messages for each cluster related status

claim)

– Ripple effect

• Recomputed the entire DS on local re-election and global re-clustering

13

• Example2: Weak CDS

▪ DS includes dominating and non-dominating (i.e., connect 2 dominating nodes) ▪ Favors the nodes with high degree (i.e., nodes with many links) for inclusion in WCDS ▪ Merges the coverage zones of the nodes in DS until the entire network is covered

Mobility-aware clustering

• Idea: cluster nodes with similar moving patterns are clustered

together.

14

 Example1: MOBIC

• Nodes disseminate their mobility

information (speed and direction)

• Cluster-head:

▪ The node with the lowest relative mobility in a neighborhood is elected ▪ Cluster-Heads encounter: timers and lowest id cluster policy s m s s s

n n n

] 7 , 4 [ , ,

4 2 1

 s m s s s

n n n

] 1 , [ , ,

6 5 3

 s m s s s

n n n

] 3 , 2 [ , ,

9 8 7



2 5 4 6 7 8 3 9 1

1

C

3

C

2

C

Mobility-aware clustering (2)

• Summary:

– MOBIC: 1-hop, high communication complexity (absolute and relative speed is distributed in the neighborhood of a node) – DDCA: multi-hop, larger clusters, overlapping clusters

15

 Example2: DDCA

• (α,t)-every mobile node in a cluster has a path to every other

node that will be available for some time period for a time period t with a probability ≥α

▪ Independent of the hop count between nodes

• Cluster-Member:

▪ Bidirectional path to the Cluster-Head which satisfy the clusters (α,t) ▪ Favor the highest availability path cluster

2 5 4 6 7 8 3 9 1

 

60 , 5 .

2

  t C 

 

60 , 5 .

 

120 , 75 .

 

120 , 75 .

 

180 , 8 .

 

180 , 8 .

 

150 , 75 .

1

  t C 

Energy-efficient clustering

• Idea: balance energy consumption on nodes by moving the cluster-heads
• Example1: IDLBC

– Limit the time a node can be cluster-head based on time counters

• The counter is decremented while a node is cluster head
• The cluster head relinquish its cluster-head role when counter is 0 and a new cluster-head the

node with higher counter

• Example2: Energy based DS

– Limits the size of the DS by removing the nodes with low residual energy than direct neighbor nodes in DS

• Summary

– Active clustering schemes with stationary assumption – Affected by ripple effect – High communication complexity

16

Load-balancing clustering

• Idea: limit the minimum and maximum number of clusters in a cluster
• Example1: AMC

– Cluster-Members and Cluster-Heads: Periodic broadcast of clustering information – Cluster-Gateways: Periodic exchange own cluster info with neighbor clusters – Tries to maintain for each cluster

• 17

U C L

i 

 U C C C C C L

j i j i i

    if merge and U C C C C C U C

i i i i i i

   

' ' ' ' ' '

, L such that and into splits 2 5 4 6 7 8 3 9 1

3

C

1

C

2 5 4 6 7 8 3 9 1

1

C

3

C

2

C

2 5 4 6 7 8 3 9 1

1

C

3

C

2

C

5 3   U L

Load-balancing clustering (2)

• Example2: DLBC

– Optimal number of nodes for each cluster head – Increase the stability: variation interval around the optimal number of nodes

• Summary

– Multi-hop clusters – AMC localizes the ripple effect, but DLBC is affected by it – The communication complexity is high

18

2 5 4 6 7 8 3 9 1

2

C

1

C

2 5 4 6 7 8 3 9 1

2

C

1

C

1 1 3 1 1 1 4 1 2 2 5 4 6 7 8 3 9 1

2

C

1

C

1 1 3 1 1 1 4 1 2

3

C

1 5

var 

 C Copt

Combined-metrics clustering

• Idea: use multiple metrics to elect a cluster-head
• Example: On-demand WCA

– Parameters: degree-difference (difference node degree with the optimal number of cluster-members), distance to neighbor nodes, average moving speed and cluster- head serving time – Cluster-head: local area minimum for the combined weighted factor, where the sum

• f weights is 1
• Summary:

– High communication complexity – High overhead – Longer frozen periods – Ripple effect on re-clustering

19

Low-maintenance clustering

• Idea: increase the tolerance to topology changes

– Reduce Re-affiliation and Re-clustering  lower the communication overhead

• Re-affiliation: change the affiliation cluster for a node
• Re-clustering: change the structure of a cluster
• Mechanisms:

– Cluster head election: Lowest ID or Highest Connectivity

• Periodically check in a node’s neighborhood
• Nodes which satisfy the condition in a neighborhood is elected as cluster head
• Example1: Least Cluster Change

– A cluster-head has the lowest ID in a neighborhood – In range cluster-heads the one with the lowest id gives up

• Example2: 3-hop Between Adjacent Cluster-Heads

– Role of Cluster-Guest which allows a higher stability for the clusters – Require a stationary period

20

Low-maintenance clustering (2)

• Example3: Passive Clustering

– Nodes states:

• Initial, Cluster-Head, Gateway, and Ordinary
• Timers to reset the states to “Initial”

– Initial  Cluster-head: a node that has something to send

• Piggybacking the cluster-head claim

– Initial  Ordinary: node receiving one cluster-head claim – Initial  Gateway: node receiving multiple cluster-head claim

• The number of gateways in an area is controlled (a constant based on the difference between no
• f cluster-heads and gateways in an area)

21

2 5 4 6 7 8 3 9 1

Low-maintenance clustering (3)

• Summary

– 1-hop clusters – Motion frozen period – Neighborhood Lowest ID or Highest Degree – Non-constant number of rounds – Time complexity is equal to the number of clusters – Nodes are wiling to renounce their Cluster-Head position

• PC – clustering when there is data to send

22

Outline

• Clustering in MANETs
• Routing Protocol Clustering in MANETs

– Issues for clustering in routing – Clustering approaches for routing

• Dynamic clustering in the overlay

– Communication non-intrusive clustering – Evaluation

• Conclusions & References

23

Overlay Clustering in MANET

• Goal

– Achieve service scalability and improved information dissemination in the network – Service placement

• Reduce the distance between providers and consumers
• Balance load on service providers

– Adaptation to application needs

• Allow applications define own clustering objectives
• Provide concomitantly different clusterings for different objectives

24

Resources & Service Placement – Example

25

Resources & Service Placement – Example

26

Resources & Service Placement – Example

27

Communication non-intrusive clustering

• Question: Where to place services in the network?
• Issue: Minimize the distance to resources in order to balance

the use of resources

• Requirements:

– Management overhead independence  zero dedicated cluster management – Position independence

• Solution: clustering with routing table information

28

Usability

29

• Replication of data

– 3 Replicas – 4 Writers – 4 Readers

Usability (2)

• Replica placement

– Influences the accessibility and availability of the data – Well placed replicas  Reduced bandwidth consumption

• 10% nodes with replica  Close to optimal traffic for replica maintenance
• Random placement increases the traffic by 380 KB/s while clustering increases the

traffic by 216 KB/s considering an ideal placement of replica

30

5 10 15 20 25

1 21 41 61

Link

Bandwidth usage (KB/s) 1 replica (n1) 1 replica (n24) 38 replicas 5 replicas

Bandwidth usage (S04, 40 readers, 40 writers)

• 3
• 2
• 1

1 2 3 4 5

2000 4000 6000 8000 10000 12000 14000 Time (s) Bandwidth usage (MB/s)

Re-clustering every 60 s Potential traffic if the change was not made Potential savings (+) and costs (-)

Usability (3)

• Replica reconfiguration

– Needs to replicate the data to the new node – Adding and removing data replica in the network can cost more in terms of transmitted data

• Delaying the reconfiguration can help …

31

Point 10500: 1.207 MB/s Point 4400: -2.072 MB/s

Temporary services

• Solution: Temporary Clustering with dynamic clustering

methods (i.e., consider the dynamic in the network)

– Clustering which adapts to the current network layout

• Adaptive number of clusters
• Unconstrained number of nodes in a cluster

– Temporary service positioning:

• Number of data replica and services,
• Data and service placement,
• Network partitioning
• Problem: No methods to handle dynamics issues of

MANETs

32

A non-intrusive information source

• Information Source:

– Routing table in the routing protocol

• Information type: Topology information vs. Position of nodes

– Advantages

• Updated view of the network,
• Location independent

– Disadvantages

• Sensitive to existence of communication
• Sensitive to mobility and communication patterns
• Partial topology of the network

– Issues to investigate

• Accuracy
• Consistency

33

Topology Data Consistency

• Issue: Topology consistency (i.e., nodes may have different topology information)
• Question: How consistent is the topology information at the different nodes?
• Solution: Compare topology information on all the node in the network

– Compute the Hamming Distance between topologies (count the differences between the adjacency matrixes of different nodes)

34

1 1 1 1 1 1 1 , 1 1 1 1                                HD

          1 1 1 1           1 1 1 1 1 1

1 2 3 1 2 3

Topology Data Consistency (2)

• Issue: Topology consistency
• Result: Similar topologies, if node are connected.

35

Ground truth

Non-intrusive Clustering

• Method 1: Physical Position of Nodes

– Clustering based on the position of nodes

• Method 2: Community Detection

– Separate the regions with dense network connections, and sparse connections

• utside the groups

– Clustering based on the network topology in the route table – Divide or agglomerate to detect the groups of nodes in the network with dense network connections, and sparse connections outside the groups – Types:

• Modularity based method
• Random walk method
• Potts based method
• Evaluation

– Cluster head placement: Cluster head election based on centrality measures – Measurements: quality, stability, similarity, consistency, and significance

36

Non-intrusive Clustering (2)

• PAM: Clustering based on the

position of nodes

– Map a distance matrix of

• bjects into k number of

clusters – Finds k nodes which have the smallest distance to the nodes around them

37

2 5 4 6 7 8 3 9 1

1

C

2

C

Non-intrusive Clustering (3)

• Community Detection:

Modularity based method

– NG [Newman and Grivan 2004]: recursively finds and deletes the links with high weight in the network

38

2 5 4 6 7 8 3 9 1

1

C

2

C

Non-intrusive Clustering (4)

39

2 5 4 6 7 8 3 9 1

1

C

2

C

• Community Detection: Random

walk method

– vD [van Dongen 2008] simulates flow diffusion in a graph by random walks, a dense region in a graph will easily trap a random walker

Non-intrusive Clustering (5)

40

2 5 4 6 7 8 3 9 1

1

C

2

C

• Community Detection: Clustering

based on the network topology in the route table

– RB [Reichard and Bornholdt 2006] where community membership of a node is determined by its neighborhood (i.e., number of neighbors and neighbors’ membership)

Clustering Evaluation: Quality

• Question: Is the clustering valid?
• Measure: Silhouette index

– How well is a node clustered considering its distance to the center and of the center of the closest cluster

41

   

 

  

  

j i h i j i h i i

C p d C p d C p d C p d p s , , , max , ,  

 

j N i i j

N p s S

j







1

Silhouette Node Index: Silhouette Cluster Index:

k S GS

k j j







1

Silhouette Network Index: 2 5 4 6 7 8 3 9 1

1

C

2

C

       

72 . 2 72 . 71 . 71 . 5 5 . 6 . 6 . 1 75 . 5 , 2 1 1 , 2 max 1 2 1 , 2 ,

1 5 1 5 2 5

               GS S p s C p d C p d

Clustering Evaluation: Quality (2)

• Question: Is the clustering valid?
• Measure: Silhouette index

42

   

 

  

  

j i h i j i h i i

C p d C p d C p d C p d p s , , , max , ,  

Silhouette Node Index: Silhouette Cluster Index:

 

j N i i j

N p s S

j







1

Clustering Evaluation: Quality (3)

• Question: Is the clustering valid?
• Measure: Silhouette index
• Results: The created clusters are good.

43

k S GS

k j j







1

Silhouette Network Index:

Clustering Evaluation: Stability

• Question:

– Is the clustering stabile?

• Measure:

– Stability quantifies the changes of the clustering with respect to the new network structure

• Measure the cluster-head time

– Delay the clustering

• Results: Delayed clustering can improve the stability of the clusters.

44

Clustering Evaluation: Consistency

• Question: Are the clusters consistent in the network?
• Measure: Damerau-Levenshtein Distance between detected communities at different

nodes

– Counts the number of insertions, deletions, substitutions of single characters, and transpositions between two sets

45

2 5 4 7 8 9 3 6 1

1

C

2

C

2 5 4 7 8 9 3 6 1

' 1

C

' 2

C

   

5 , 4 , 3 , 2 , 1 , 5 , 4 , 3 , 2 , 1

' 1 1

  C C 1 3 4 5 6 6 5 1 2 3 4 5 5 4 2 1 2 3 4 4 3 3 2 1 2 3 3 2 4 3 2 1 2 2 1 5 4 3 2 1 1 5 4 3 2 1 5 4 3 2 1

Clustering Evaluation: Consistency (2)

• Question: Are the clusters consistent in the network?
• Measure: Damerau-Levenshtein Distance between detected cluster-heads at different nodes
• Results: There are differences in the detected cluster-heads at different nodes.

46

Cluster Head – Central Node Cluster Head – Marginal Node

Clustering Evaluation: Consistency (3)

• Question: Are the clusters consistent in the network?
• Measure: Damerau-Levenshtein Distance between detected communities at different nodes
• Results: Communities are similar at different nodes.

47

Community – Marginal Node Community – Central Node

Clustering Evaluation: Similarity

• Question: What is the difference between different clusterings?
• Measure: The similarity measures the variation of information between clustering over

the same network

– Variation of Information:

• Result: The NG, RB, and vD clustering techniques produce similar results.

48

2 5 4 7 8 9 3 6 1

1

C

2

C

2 5 4 7 8 9 3 6 1

' 1

C

' 2

C

NGRout vs. RBRout NGRout vs. vDRout RBRout vs. vDRout *Rout vs. PAMPos NGRout vs. NGGrTop RBRout vs. RBGrTop vDRout vs. vDGrTop NGGrTop vs. RBGrTop NGGrTop vs. vDGrTop RBGrTop vs. vDGrTop *GrTop vs. PAMPos 0 … 0.7 0 … 0.7 0 … 0.8 0.7 … 2.2 0.2 … 2.0 0.4 … 1.8 0.3 … 1.9 0 … 0.6 0 … 0.4 0 … 0.6 0.4 … 1.0

Clustering Evaluation: Significance

• Question: Is it a relevant clustering?
• Measure: The significance measures the resilience of a clustering to changes in the

structure of the graph.

– c-score: the probability of the node with the lowest internal degree in a community is in the same community in a equivalent random graph (≤ 5%)

• Result: Most of the clusters represent relevant structures in the graph.

49

2 5 4 7 8 9 3 6 1

1

C

2

C

2 5 4 7 8 9 3 6 1

1

C

2

C

Passive Clustering Discussion

• Topology information

– Does not apply for reactive routing protocols – Requires a consistent view of the topologies at the nodes

• Clustering measures

– Not for dynamic networks – Quality

• Not a general accepted metrics
• Different metrics may give contradicting conclusions

– Stability

• Does not consider the changes in the number of clusters
• Does not consider the changes in the number of nodes in the network

– Consistency

• Detected communities are more consistent than elected cluster heads

– Similarity:

• Not applicable to clustering from different nodes in the network

– Significance:

• Not all clusters represent significant structures in the network

50

Outline

• Clustering in MANETs
• Routing Protocol Clustering in MANETs

– Issues for clustering in routing – Clustering approaches for routing

• Dynamic clustering in the overlay

– Communication non-intrusive clustering – Evaluation

• Conclusions & References

51

Conclusions

• Clustering in routing

– Clustering schemes have different focus and objectives

• Cluster structure stability, reduce overhead in cluster construction and maintenance, limit energy consumption,

balance traffic load, or cluster-head balancing

• Different metrics  hard to compare

– Communication overhead and complexity

• Explicit control messages  high overhead
• PC  piggybacks messages

– Cluster diameter and Ripple effect

• Multi-hop clusters are less affected by ripple effect in re-clustering

– Localize the cluster management

• 1-hop clustering schemes usually create highly overlapping structures
• Clustering in overlay

– Dependent on the performance of the routing protocol – Dependent on the objectives of the application

52

References

• Jane Y. Yu and Peter H. J. Chong, “A survey of Clustering Schemes for Mobile Ad Hoc

Networks”, IEEE Communications Surveys, vol. 7, no. 1, 2005

• Matija Puzar, Thomas Plagemann, “Evaluation of Replica Placement Strategies for Mobile

Ad-Hoc Networks”, The 13th International Conference on Network-Based Information Systems (NBiS-2010), Takayama, Gifu, Japan

• Ovidiu V. Drugan, Thomas Plagemann, and Ellen Munthe-Kaas, “Detecting Communities in

Sparse MANETs”, Accepted for publication in IEEE/ACM Transactions on Networking, 2011

53