Traffic Based Clustering in Wireless Sensor Network Vijay Kr. Chaurasiya1, S. Rahul Kumar2, Shekhar Verma3, G. C. Nandi4 Indian Institute of Information Technology Allahabad, India [email protected], [email protected] Abstract- To increase the lifetime and scalability of a wireless sensor network (WSN) it is necessary to have control over topology of the network. Dynamic clustering is one way to achieve the above defined objective. In this paper, we propose a hierarchal clustering approach to adapt the size of clusters and number of hierarchal level in accordance with the traffic patterns and density of sensor nodes deployed in a given area of interest. As all the traffic is directed to the base station, the traffic load on the cluster heads in the vicinity of a base station is far more as compared to distant cluster heads. This may result in draining out of proximity cluster heads sooner than other cluster heads. The network would fail even though most of the distant nodes have sufficient energy. A bottleneck is created near the base station. The proposed algorithm of hierarchical clustering with variable cluster size based on its distance from the base station endeavors to balance the intercluster traffic so that energy depletion in cluster-heads is uniformly distributed. Variable cluster size is also able to balance the intra-cluster traffic. Simulation results show that load balancing through hierarchical variable size approach is able to enhance the network lifetime.

I.

INTRODUCTION

WSNs are used in many applications such as military installations, environmental monitoring, disaster control, traffic control, security installations and so on. Most of these applications demands efficient organization of network topology for data collection, data aggregation and load balancing to increase the network lifetime. The network must be autonomous and self organizing. There are many constraints associated with WSN such as the nodes are light weighted and small in size, it should be of low cost and less energy consuming. These constraints demand an efficient and optimal clustering protocol in WSN for its operations. In order to facilitate low power consumption, fault tolerance, scalability, WSNs should be clustered hierarchically and aggregated data should to be routed energy efficiently with minimum latency. There are two types of networks. a) Reactive and b) Proactive. Our approach can be applied to both types of networks. In our approach, clusters are organized according to the traffic patterns so that the inter-cluster and intra-cluster traffics are balanced. The size of cluster is changing according to its distance from the base station. The remainder of this paper is organized as follows. Section 2 deals with related work in the area. In Section 3 the problem addressed in this paper is described whereas in section 4 our approach is explained and verified. Section 5 covers simulation details and results. II. RELATED WORK A clustering protocol in WSN must be energy efficient and must be able increases network life time. It should ensure that all the nodes in the network remain alive as long as possible. Various approaches have been described to

ensure that battery power could be used as efficiently as possible in all the phases of network operation (i.e. Clustering, Routing, Information Processing and Forwarding). In [3] author classified different types of clustering protocols for WSN their advantage and drawbacks. The protocol presented here are concerned on how to increase WSN lifetime and to make efficient use of critical resources located at sensor nodes by creating intelligent clustering schemes. In [10] authors have proposed that to minimize the drying out of proximity nodes, battery capacity of nodes should be inversely proportional to their distance from base station. It means a node which is closest to BS should be equipped with greatest battery capacity. In [4] authors raised the issue of doughnut effect in wireless sensor network and tried to solve that problem. Doughnut effect arises when nodes which are nearer to BS have additional responsibilities of forwarding data of distant nodes. Because of this, nodes which are in nearer to BS run out of battery very soon and create bottleneck in the communication. They solved this problem by increasing the density of nodes as we move towards BS so that more number of nodes can take part in forwarding the data of distant nodes and load is being equally distributed among all nodes and bottleneck never occurs because of dead nodes that were encountered previously. But here we are using a homogenous system where nodes are randomly deployed and all nodes have equal capability so we need optimized the cluster size so that the load can be shared by different nodes in the network. In [19] authors have been developed a scenario to increase the coverage time of the network or the life time of the network. Their aim was to provide optimal power allocation in a Rayleigh fading channel model. They have demonstrated a strategy for interaction between routing and clustering by considering impacts of intra-cluster and inter-cluster traffic on each cluster head. Their main two mechanisms for achieving balanced power consumption are:a) Routing aware optimal cluster planning b) Clustering aware optimal random delay One of the major flaws of this routing aware optimal clustering system is that it cannot have many numbers of clusters which can form independent sets i.e. non overlapping clusters. There is a limit to what could be a maximum independent sets for different level from base station and as for the first level we can’t have more than 5 independent sets [25].Our work is also based on routing aware clustering but we have solved this problem. III. PROBLEM STATEMENT Consider that we have a set of sensors dispersed in an area which need continuous monitoring. Here we are

considering a large scale WSN such as sensor network deployed for forest fire detection or environmental monitoring. We have assumed that sensor network has following properties:a) All sensor nodes are similar i.e. having same intial batery power and having identical capabilities with all other sensor nodes. b) Links are symetrical. c) Sensors are stationary. d)nodes are unattended. e) Nodes have fixed number of trasmission power levels We have to keep in mind that clustering should be uniform throughout the area and the cluster heads are uniformly distributed to have proper load balancing. Here the nodes A,BC,D represents cluster heads of different cluster. Here data from any cluster head get forwarded to the adjecent cluster head and finally delivered to BS. For example data from node A will follow a path A B C D BS. Here, more load will be on node D because data is accumulated on node D from various CHs present in the network.

Fig. 1 Simple WSN Topology where A,B,C,D,P are cluster heads and BS is base station

This will result in dying out of node D and hence failure of the whole system. This bottleneck will reduce overall life time of the network. Therefore, we have to construct a network topology based on the node density in different regions of the overall area covered by the network. This will avoid bottleneck problem and may result in long life of the network by proper load balancing. IV. PROPOSED APPROACH We have proposed a traffic based routing aware clustering approach. The main motive for developing this approach is to increase the life time of the network as a whole. To increase network life time we have to utilize battery capacity of each node as efficiently as possible. To get greater life time we have to decide on the topology of the network formation i.e. level of clustering, size of clusters at different levels based on traffic and density of the sensor nodes in a specific region of deployment. We have seen in cluster based sensor network that most of the energy is being consumed by the nodes which act as a cluster head. So it’s very important to save the cluster head’s energy. The responsibility of the cluster head near the base station increases as it has to relay the data of all previous cluster heads means its failure will result in failure of the system as a whole. In this scenario all other

nodes becomes worthless as they will not be able to communicate with the base station. The following are the parameters where energy consumption at any node takes place:1. Sensing 2. Computation 3. Communication (Reception and Transmission). The consumption of energy in reception and sensing is independent from transmission distance. Energy consumption in computation is negligibly small compared to energy dissipation in other process such as sensing, transmission and reception of a bit. [27] Suppose energy required to transmit (to unit distance) and to receive a bit is being represented by E. Energy required to receive a bit received is represented by ER and Energy required to transmit (to unit distance) a bit to distance d is represented by ET. So ER and ET can be defined as :ER =E*b ---------(1) ET =E*b*dα ---------(2) Where b is no. bits transmitted or received or sensed α is path loss exponent which varies from 2 to 6[26]. For ideal case we can assume α= 2 and the value of one unit of energy E range from 10 nJ/bit to 100 nJ/bit [27,8,30]. By observing the formulae we can conclude that transmission energy is more than receiving or sensing energy and is also dependent on the distance between the nodes. So to reduce the consumption of energy in both transmission and reception we must minimize the data received or transmitted i.e. energy consumption directly depends on amount of data transferred. To reduce data transferred we can aggregate total data at each cluster head fused the data and forward it to adjacent cluster heads. To transfer data from all cluster heads to the base station, every cluster head need to be alive. To achieve that and to avoid doughnut effect [4] we have to balance inter-cluster and intra-cluster traffic. Intra-cluster traffic: Intra-cluster traffic can be varied if we change the size of the cluster accordingly the number of nodes in a cluster. So a change in a cluster size and the change in number of nodes will result in change in traffic load on a cluster head of that cluster. If we increase the cluster size we can accommodate more number of nodes. Therefore load on cluster head increases and it may lead to depletion of the residual energy of the cluster head more rapidly and vice versa. Inter-cluster traffic: Inter-cluster traffic means the data transfer between adjacent cluster heads on its way to the base station i.e. relaying data to the adjacent cluster head on its way to reach the base station. The farthest clusterhead from the base station have no work of forwarding any data but the cluster head following it has to forward its own traffic and the traffic it received from its nearest hops. By so the cluster head which is nearest to BS will be dealing with maximum data traffic. Also larger the size of the cluster the cluster head requires more energy to directly transmit or relay data to its adjacent cluster head on its

path to the base station and smaller the size of the cluster , lesser will the amount of energy spent in transmitting it to the next hop cluster head or to the base station. So as to optimize these two traffics we can vary the cluster size according to its distance from the base station. One of the solutions is that we can increase the number of hierarchical levels to cover the same area as covered by lesser number of hierarchical levels of cluster. This will result in smaller size of the all clusters and thus the nearer cluster head have to deal with a lesser load due to reduced intra-cluster communication as its accommodating lesser number of nodes in one cluster and also it has to now transmit to a comparatively lesser distance to its next hop (which may be adjacent cluster head or base station). The other solution can be varying the cluster size as its distance from base station varies i.e. smaller cluster size will be nearer to the base station and larger as we move away from central base station. This topology will provide the good solution to balance inter-cluster and intra-cluster traffic and also the total number of clusters and cluster heads is not increased as in previous case. We have simulated both the proposed clustering protocols and examined result of both the simulations.

Fig 2 :-Selecting cluster size for proper load balancing

V. SIMULATION AND FINDING We have simulated the above approach in a simulator called OMNET++ . We have seen till now that data is relayed to the base station through next-hop cluster heads but the cluster head receiving the relayed data is not always free to just relay the data as it also got some of its own data to be processed so the relayed data enters into a queue till the existing data is processed and this simple queuing model of wireless sensor network is assumed to be M/M/1 queuing model. The function of the member nodes is approximated by the generator block which has the parameter called inter-arrival rate i.e. this parameter represents the number of nodes a particular cluster is containing as its only the number of member nodes which is of importance for assuming a sensor network. Sensor nodes doesn’t do any function on its own rather than transmiting the sensed data to the cluster head. So the cluster head is approximated by a FIFO queue block which maintains the data received from different hops and from its own member nodes to which it belongs. For simulation we have used an area of 100m*100m and 1000 sensor nodes. Suppose we need 5 hierarchical levels

to cover the whole area and distance between each level is r then the farthest level will be at the distance 5r from the base station. We have assumed 4 cases of topology with different hierarchal levels. Here we are proposing a horizontal hierarchy and not the vertical hierarchy as clear by figure 2(showing horizontal hierarchy). So we have simulated a hierarchal topology with different levels with equal distance between different levels, also we have assumed that the nodes (cluster head) will die out after performing 500 units of job. 1 unit of job = 1 bit transfer (reception + transmission) of data We assume that energy spent at each node is consumed on reception and transmission only as energy spent on processing is negligible [27]. It has also been assumed that:For reception of one bit, energy spent= 0.5 unit of energy. For transmission of one bit, energy spent = 1 unit of energy for unit distance, i.e. if each node(cluster head) spends 1.5 unit of energy per job (each bit of data transfer) then it will perform total of 500 units of jobs before dying out. The density of nodes in a cluster in our simulation environment is represented by a parameter known as interarrival time [25], which is defined by time gap between adjacent receptions by the cluster head from its member nodes. We have assumed a fixed number of nodes in each cluster (in this 5 level scenario) the inter-arrival time gap between consecutive reception of data from member nodes to the cluster head = 10 s A total number job to be generated at each cluster is taken as 15 units of job. – (a) CASE 1:- Hierarchal topology with 3 levels with same cluster size: We have simulated it for the same area of 100*100 m2 and 1000 sensor nodes, if we change the no. of hierarchical levels to 3 for covering same area then the distance between different levels will be r’ and value of r’ will be as follows: r’= 5r/3 i.e. 1.66r –(3) where, r is the radius of cluster in 5 level hierarchical clustering for the same area. Inter-arrival time will depend on area of the cluster as area of the cluster is increased so the rate of arrival of data at cluster head will also increases as number of nodes in a cluster will increases. Inter-arrival time is proportional to 1/area i.e. Inter-arrival time α 1/r2 Factor (r2) = (1.66r)2 = 2.75 Inter-arrival time = 10*1/(1.66)2 units= 3.62 s –(4) Where: 10 is inter arrival time in 5 level hierarchical clustering for the same area. From the simulation we got following results:So from the graph the average life-time of cluster heads of the cluster at a particular level is: Life time of the node at 1st level= 34.7 s Average life time of the nodes at 2nd level =45.6 s Average life-time of the nodes at 3rd level= does not reach dies out for condition (a).

Jobs done (in units)

Analysis of simulation: Here the nearest Node is dying out in 35.2 s so the system lifetime is 35.2 s as the system will fail if the nearest node stops working. The life-time is improved compared to case lets simulate the scenario of same level of clustering but varying the cluster size according to the distance of cluster head from the base station.

Time taken (in seconds) Fig 3: Simulation result for hierarchal topology with 3 levels with equal cluster-size

Here the nearest Node is dying out after 34.7 s so the system lifetime is 34.7 s as the system will fail if the nearest node will stop working.

Table 1

Distance between each level Units of energy spent for a job Inter-arrival time No. of jobs possible before dry-out

1st level

2nd level

3rd level

4th level

0.5r

0.75r

1r

2.25r

0.75

1.0625

1.5

5.5

40

17.8

10

2

1000

705

500

135

Time taken (in seconds) Fig 4: Simulation result for hierarchal topology with 4 levels with equal cluster-size

So from the graph the average life-time of cluster heads of the cluster at a particular level is: Life time of the node at 1st level= 35.2 s Average lifetime of nodes at 2nd level=45.3 s Average life time of nodes at 3rd level= never dies out for condition (a) Average life-time of the nodes at 4th level = does not reach dies out for condition (a)

Jobs done (in units)

Jobs done (in units)

CASE 2:- Hierarchal topology with 4 levels with same cluster size: Now if we change the no. of levels to 4 but covering same area then the distance between different levels will be r’ = 5r/4 i.e. 1.25r. We got following results from the simulation:-

CASE 3: Hierarchal topology with 4 levels with varying cluster size: In this case we vary the distance between each level i.e. least size at first level and larger size for higher levels i.e. 0.5r, 1r, 1.5r and 2r for first, second, third and fourth level respectively. From equation (2), (3) and (4) and by assuming whatever we assumed for base case and by calculating the corresponding relative values using the same calculation we used for case 1 and case 2 we get the following table:

Time taken (in seconds) Fig 5: Simulation result for hierarchal topology with 4 levels with varying cluster-size

The average life-time of cluster heads of the cluster at a particular level is:

Life time of the node at 1st level= 58.6 s Average lifetime of nodes at 2nd level=65.2 Average life time of nodes at 3rd level= 98.6 s Average life-time of the nodes at 4th level = does not reach dies out for condition (a) Analysis of simulation: Here the nearest Node is dying out in 58.6 s so the system lifetime is 58.6 s as the system will fail if the nearest node stops working. So by far, this is the best topology for having a system with large coverage time or good life time. Findings: It has been seen that lifetime of the sensor network increase with the increase in number of hierarchical levels. But whatever number of hierarchal level he may be using, if he uses varying cluster sizes in accordance with the distance from base station the lifetime of the system will improve considerably. VI. CONCLUSIONS In this paper we have devised an approach for clustering for a homogenous system having deployed randomly without prior node position information . The main aim of our work was to find a traffic based topology for the formation of clusters and to counter the problem of doughnut effect and to have a load balance between intercluster and intra-cluster traffics. We have changed the cluster size and number of level of clusters accordingly based on traffic pattern and density of nodes deployed. Our simulation in OMNET++ shows that as we change the size of clusters in accordance with it’s distance from the base station imroves the life of the system drastically. This approach has thus proved to be superior than the existing static approaches (i.e. fixed number of level or equal cluster size for each level ) as the total lifetime of the system is increased with improved load balancing. Future work we will be to increase the number of nodes and number of level of hierarichy to a much higher number and to test weather our approach is that much efficient if we increase the number of nodes in ever changing environment. REFERENCES [1]

J.Kulik, W. Heinzelman, and H. Balakrishnan, Negotiation-based protocols for disseminating information in wireless sensor networks, Wireless Networks, 8, 169–185, March–May 2002.

[2]

Younis and S. Fahmy, “Distributed Clustering in Ad Hoc Sensor Networks: A Hybrid, Energy-Efficient Approach,” Proc. IEEE INFOCOM, Hong Kong, Mar. 2004; an extended version appeared in IEEE Trans. Mobile Comp.,vol. 3, no. 4, Oct.–Dec. 2004, pp. 366–79.

[3]

K. Sohrabi, J. Gao, V. Ailawadhi, G.J. Pottie, “Protocols for self organization of a wireless sensor Network”, IEEE Personal Communications, October 2000, pp. 16–27.

[4]

Doughnut Effect in Wireless Sensor Network and its Solution Padmanabh, K. Roy, R SECON '06. 2006 3rd Annual IEEE Communications Society.

[5]

Wei Ye, John Heidemann, Deborah Estrin, “An Energy-Efficient MAC Protocol for Wireless Sensor Networks”, IEEE Infocom 2001.

[6]

S. S. Raghuwanshi, A. Mishra, “An Approach Towards Improved Energy efficiency in Wireless Sensor Nets”, IEEE RAWCON 2003.

[7]

S. Lindsey and C. Raghavendra, PEGASIS: power-efficient gathering in sensor information systems ,Int. Conf. Communication Protocols, 149–155, 2001.

[8]

RAPPAPORT, T. 1996. Ad Hoc Mobile Wireless Networks: Protocols and Systems. Prentice Hall, Englewood Cliffs, NJ..

[9]

S. Banerjee and S. Khuller. A Clustering Scheme for Hierarchical Control in Multi-hop Wireless Networks. IEEE INFOCOM 2001, Anchorage, Alaska, April 2001.

[10] Mihail L. Sichitiu, Rudra Dutta: Benefits of Multiple Battery Levels for the Lifetime of Large Wireless Sensor Networks. NETWORKING 2005:1440-1444. [11] Liudong Xing and Akhilesh Shrestha “QoS Reliability of Hierarchical Clustered Wireless Sensor Networks” 2006 IEEE [12] F. Garcia, J. Solano, I. Stojmenovic, “Connectivity based k-hop clustering in wireless networks”, Telecommunication Systems, vol. 22, no. 1-4, pp. 205-220, 2003. [13] Heinzelman W.B., “Application Specific Protocol Architectures for Wireless Networks”, PhD dissertation, Massachusetts Institute of Technology, June 2000 [14] V. Kawadia and P.R. Kumar, “Power Control and Clustering in Ad Hoc Networks,” Proc. IEEE INFOCOM, Apr. 2003 [15] Raghavendra C.S., Sivalingham K.M., Znatti T.F., “Wireless Sensor Networks”, Kluwer Academic Publishers, 2004. [16] D. Culler, D. E. M. Srivastava, “Overview of Sensor Network”, IEEE Computer Magazine, vol. 37, no. 8, pp. 41-49, August 2004. [17] Al-Karaki J.N., Kamal A.E., “Routing Techniques in Wireless Sensor Networks: A Survey”, Department of Electrical and Computer Engineering, Iowa State University, December 2004 [18] Manjeshwar A., Agrawal D.P., “TEEN: A Routing Protocol for Enhanced Efficiency in Wireless Sensor Networks”, Proceedings of the 15th International Parallel and Distributed Processing Symposium, San Francisco, 2001. [19] Power balanced coverage-time optimization for clustered wireless sensor network. Tao Shu, Marwan Krunz, Sarma Vrudhula MobiHoc’05, May 25–27, 2005, Urbana-Champaign, Illinois, USA. [20] MELO,D. AND LIU,M. 2002. Analysis of energy consumption and lifetime of heterogeneous wireless sensor networks. In Proceedings of the IEEE Globecom 2002 [21] G. Gupta and M. Younis, “Load-balanced clustering of wireless sensor networks, ” in Proc. IEEE International Conference on Communications (ICC’03), vol. 3, pp. 1848-1852, May 2003 [22] HEINZELMAN,W. B., CHANDRAKASAN, A. P., AND BALAKRISHNAN,H. 2000. Energy-efficient communication protocol for wireless sensor networks. In Proceedings of the Hawaii International Conference on System Sciences (HICSS). 1–10 [23] Jin-Mook kim, Hyeon-Kyu joo, Seong-Sik on, Woo-Hyun Ahn 'Hwang-Bin Ryou “An Efficient Clustering Scheme through Estimate in Centralized Hierarchical Routing Protocol” 2006 International Conference on Hybrid Information Technology [24] HEINZELMAN,W. B., CHANDRAKASAN, A. P., AND BALAKRISHNAN,H. 2000. Energy-efficient communication protocol for wireless sensor networks. In Proceedings of the Hawaii International Conference on System Sciences (HICSS). 1–10 [25] “Connected Dominating Sets in Wireless Networks with Different Transmission Ranges” My T. Thai, Feng Wang , Dan Liu, Shiwei Zhu, Ding-Zhu Du IEEE TRANSACTIONS ON MOBILE COMPUTING, 2007 [26] WSN18-1: k-Connectivity in Shadowing Wireless Multi-Hop Networks Zhang, Lili; Soong, Boon-Hee Global

Telecommunications Conference, 2006. GLOBECOM apos;06. IEEE [27] “Collaborative Broadcasting and Compression in Cluster-Based Wireless Sensor Networks” Anh Tuan Hoang and Mehul Motani ACM Transactions on Sensor Networks, August 2007. [28] Vardhan, S.; Wilczynski, M.; Portie, G.J.; Kaiser, W.J.; “Wireless integrated network sensors (WINS): distributed in situ sensing for mission and flight systems” Aerospace Conference Proceedings, 2000 IEEE, Volume: 7 , 18-25 March 2000, 459 -463 vol.7