The Effect of Physical Topology on Wireless Sensor Network Lifetime

Transcription

1 1 JOURNAL OF NETWORKS, VOL., NO. 5, SEPTEMBER 7 The Effect of Physical Topology on Wireless Sensor Network Lifetime Debdhanit Yupho Planning Department Aeronautical Radio of Thailand Limited, Bangkok, Thailand dey5@pitt.edu Joseph Kabara Graduate Program in Telecommunications and Networking University of Pittsburgh, PA jkabara@pitt.edu Abstract Wireless sensor networks must measure environmental conditions, such as temperature, over extended periods and therefore require a long system lifetime. The design of long lifetime networks in turn requires efficient sensor node circuits, algorithms, and protocols. Protocols such as GSP (Gossip-based Sleep Protocol) have been shown to mitigate energy consumption in idle listening and receiving, by turning off the receiver circuit. However, previous studies of network lifetime have been based on physical topologies in which nodes were placed on a square grid or randomly distributed throughout the service area. This paper shows that the lifetime of a sensor network depends on the physical topology of the sensor nodes. The lifetime of a sensor network varies as a function of both the size of the network and the transmission range of individual nodes. Index Terms wireless sensor network, energy efficiency, routing protocol, grid topology I. INTRODUCTION A wireless sensor network (WSN) is a general term used to describe networks with widely varying characteristics. However, in almost all cases a WSN consists of a large number of densely deployed sensor nodes [1]. To extend the lifetime of an individual WSN node it has a very limited transmission range however the network as a whole must cover a large service area so that routing protocols are necessary for end-to-end communication. Although many proposed routing protocols support wireless ad hoc networks, they are not necessarily appropriate for sensor networks []. Wireless sensor networks normally have larger size, higher density, more limited power supply and computational capacity than nodes in general ad hoc networks. Additionally, sensor energy may be conserved and network lifetime extended by treating WSNs as data centric networks, where users querying for an attribute of the phenomenon, (sensed information) rather than querying an individual node [1]. However, the requirements on the network may change with the network application and protocols must still conserve energy under changing conditions. As an example, some sensor network applications require that nodes sense the environment at one sampling rate for one purpose (e.g. average rainfall) and at another rate for another purpose (potential flooding). Also, in WSNs adjacent nodes may have identical or similar data; therefore, sensor networks should be able to aggregate similar data to reduce unnecessary transmissions and save energy. Lastly, assigning unique IDs may not be suitable in sensor networks because of the data centric characteristic data does not need to be routed to a specific node. Additionally, the large number of nodes will require long IDs, creating a large overhead cost, compared to the data being transmitted. Previous research on the impact of routing protocols on network lifetime shows that GSP can reduce overhearing transmissions and receptions by allowing nodes to enter sleep states, resulting in extended sensor network lifetime [3]. However, these results have not considered the energy consumption resulting by sleeping nodes in the idle and sleep states. We developed a timebased simulation to evaluate GSP network lifetime performance by including energy consumption of nodes in idle and sleep states. The results of the new simulation are compared with previous results to show the impact of nodes continuing to consume energy even though they are in a sleep state [3]. As in the previous reports, network lifetime is defined as the time for the first node to fail [3]. However, to better describe the energy state of the network, a new metric is introduced, Average Remaining Energy (ARE), which describes the amount of energy in all nodes, at the point where the first node fails. The new simulation model is then used to evaluate the interaction between a routing protocol (GSP) and the transmission radius of nodes and their combined effect on the lifetime of the network. The new model is also used to compare network lifetime for a traditional packet routed case, or known path (KP) case and an unknown path (UKP) case, which occurs when a WSN has multiple or mobile data sinks. 7 ACADEMY PUBLISHER

2 JOURNAL OF NETWORKS, VOL., NO. 5, SEPTEMBER 7 II. BACKGROUND Routing protocols for sensor networks may be characterized as either cluster-based or flat [] [8]. Cluster-based routing schemes divide the network into groups of nodes and utilize a sleep mode to save energy and prolong the network lifetime. Flat routing schemes achieve energy efficiency indirectly by reducing the routing overhead. However, all cases tradeoff other performance metrics, such as low delay and high throughput for energy savings. Cluster-based routing protocols such as LEACH [], and TEEN [], organize nodes into groups with one node from each group selected to be a cluster-head. A cluster-head receives data packets from its members, aggregates them and forwards them to a data sink. In some cluster-based routing protocols, a cluster-head assigns Time Division Multiple Access (TDMA) slots to its members to create a communication and sleep schedule. Based on the number of nodes, the cluster-head creates a TDMA schedule. Nodes send data during their allocated transmission time and sleep otherwise. The TDMA schedule conserves energy by reducing idle listening, but requires extra energy for the overhead needed for timeslot assignment and synchronization. The cluster-based routing protocols can arrange the sleep mode of each node to conserve energy so that only the nodes with data to send are awake. However, this requires additional control messages and therefore energy, for cluster organization and node synchronization. Flat routing schemes, typically implement either flooding, forwarding, or data-centric based routing. In flooding, every node repeats the data once by broadcasting. Flooding does not incur the energy cost for topology maintenance and route discovery algorithms. However, it has several deficiencies, e.g. implosion, overlap, and resource blindness [1]. Gossiping protocols overcome limitations with flooding schemes by utilizing local information to forward messages. However, unlike the traditional routing protocols, forwarding schemes do not maintain end-to-end routes. Instead, intermediate nodes maintain only neighbor information. In a gossiping protocol, a node only forwards data to one randomly chosen neighbor and does not maintain any routing information. Optimal Forwarding employs a cost field to decide whether to forward data [8]. A cost field is the minimum cost from a node to the sink on the optimal path. To establish a single cost field, the sink broadcasts an ADV (advertisement) message. One restriction is that a single sink node must be the destination for all data. Data centric routing protocols avoid implosion and overlap problems through data aggregation [1]. There are two types of data-centric based routing: either the sink broadcasts the attribute for data, e.g. Directed Diffusion [7], or the sensor nodes broadcast an advertisement for the available data and wait for a request, e.g. Sensor Protocols for Information via Negotiation (SPIN) [5]. Gossiping protocols address the overhearing problem by forwarding data to some of the neighbor nodes, some of the time. However, the nature of wireless channels (broadcast) delivers a message to multiple neighbors. Gossiping reduces the number of times a message is retransmitted by requiring some of the nodes to discard the message instead of forwarding it. A node decides whether or not to forward the message with probability p, the gossip probability. Given a sufficiently large network and a gossip probability p greater than certain threshold, almost all the nodes in the network will receive the message [9]. The Gossip-based Sleep Protocol (GSP) reduces energy consumption further, by replacing the message discard operation with an operation that places some nodes into a sleep mode for a specified period of time [1], [11]. At the beginning of each gossip period, nodes either sleep with probability p, or remain awake with probability (1- p), and all sleeping nodes wake up at the end of each period to repeat the process. Therefore, with a probability p, almost all nodes in the network receive a message. If each sensor in the network enters a sleep state with probability p = ( 1 p ), almost all the nodes remaining awake to receive the message, subject to channel fading [9]. Thus, a percentage (p) of the nodes may be in a sleep mode, without losing network connectivity. Unlike other protocols using TDMA-based sleep mode [], [], GSP is extremely simple and requires no information, even from immediate neighbors. The gossip sleep probability (p) depends on the network density, which however must be configured before the deployment of the network [1]. III. PERFORMANCE EVALUATION To compare these results with previous work, networks employ square grid topologies with a single sink node. All calculations assume the period of time to transmit one bit of data, i.e. bit-time. Also, the traffic load remains constant with or without GSP, i.e. the number of bits generated by the sensors in a bit time are the same. Although the actual application may generate bursty traffic, this assumption will not change the results because the incurred extra energy consumption in GSP comes from the amount of traffic, not the fashion of the traffic [1]. A. Simulation Model to Determine Gossip Sleep Probability (p) Simulations were used to determine the highest gossip sleep probability (p) that results in all awake nodes receiving a message. To study the change of average path length for different network sizes, three square grid topologies were created with a single sink node in the center, 1x1, x, and 3x3 node networks with 11, 1, and 91 total nodes respectively, i.e. 1,, and 9 sensors and a sink. In these experiments the sink has an unlimited energy source. The case where a network has all awake nodes in non-gsp, i.e., (p = ) is compared to the case of (1 - p) % awake nodes in GSP. Fig. 1 presents the flowcharts to determine the average path length in hops and average number of awake nodes that will not receive the message (Disconnected, DC nodes). The dropping of the curves in average path length (Fig. ) and increase in the ratio of DC nodes after p = 7 ACADEMY PUBLISHER

3 1 JOURNAL OF NETWORKS, VOL., NO. 5, SEPTEMBER 7.3 (Fig. 3) indicate a loss of network connectivity. Figs and 3 indicate that a.3 gossip sleep probability, i.e. p =.3, is the highest value resulting in a connected network. remaining in the network, a high average number of gossip periods indicates longer network lifetime. Simulation results are based on 5 simulation runs. Average Remaining Energy (ARE) represents the energy efficiency in the way that network can continuously use the energy remaining when the network is able to refigure itself, or the network considers lifetime as the multiple depleted nodes. In analysis, the simulation used the highest gossip sleep probability (p) that maintains network connectivity, to test the GSP performance a) b) Figure 1. Flowcharts to determine a) average path length in hops and b) average number of DC nodes. Figure 3. Probability of sleeping nodes vs. ratio of nodes disconnected from the sink. TABLE I. ENERGY CONSUMPTION MODEL Transmit Receive Sleep Initial energy stored.8 μjoules / bit.3 μjoules / bit ~ Joule 1 Joules Figure. Probability of sleeping nodes vs. average path length for connected nodes with 95% c.i. B. Simulation Model to Determine Network Lifetime WSN lifetime is usually assumed to be the most critical network constraint, since sensor networks have limited energy stores. One definition of network lifetime for a sensor network is the shortest lifetime of any participating node in the network [1]. In some applications, any sensor node may be responsible for performing critical functions. One dead node may create a loss of required system information. Network lifetime analysis determines the gossip period (G p ) and energy Table I summarizes the TinyOS Mica mote s measured energy consumption model for transmitting and receiving [3]. The frame size is 1 bytes and the data rate is 19. kbps. A node may initially store 1 joule [1] or up to 5 Joules of energy as in [3]. However, in this, analysis used 1 Joules as a convenient initial energy stored on each node. Although there are various communication schemes among source nodes and sink nodes, in our analysis, GSP performed operated using two schemes: Known path (KP) and Unknown path (UKP) schemes. In the Known path (KP) scheme, the simulation employed a minimum cost forwarding mechanism [7]. At the beginning of each period, p% of nodes perform sleeping mode, and the simulation does not include them in the network topology. A sink broadcasts a message to setup the paths. The path setup process consumes energy at the beginning of each gossip period (G p ). A KP scheme is useful when network has low mobility and GSP has very long gossip periods. An Unknown path (UKP) scheme may reflect a more typical use for GSP. In the UKP case, the network contains multiple sinks or sinks that enter and leave the 7 ACADEMY PUBLISHER

4 JOURNAL OF NETWORKS, VOL., NO. 5, SEPTEMBER 7 17 network. Nodes transmit the packets in broadcast fashion to neighbors within their transmission ranges without the knowledge of the neighbor nodes locations. Then the awake/active neighbors or intermediate nodes will relay these packets to the sink. This process stops when the timer expires to form a new topology at the beginning of each gossip period. Sleeping nodes in each period do not participate in any activities. The simulation restricted nodes to relay the individual packet only once. When a duplicate packet arrives, a node will receive that packet and discard it. An UKP scheme may be useful when the physical topology is changing quickly, or multiple sinks are part of the network. non-gsp (p = ) as a function of different network sizes. The plot illustrates the increase in network lifetime when using GSP. However, as the network size increases, the percentage of improvement in network lifetime decreases. Fig. demonstrates the Average Remaining Energy (ARE) in the network after the simulation found the first completely depleted node, averaged over 5 runs. The UKP scheme is most likely to consume more energy than the KP scheme since it employs a universal broadcast mechanism. As a result, the UKP scheme has less ARE in the network. By comparing GSP to non-gsp, little change occurs between the KP schemes. On the other hand, the UKP scheme using GSP with.3 gossip sleep probability shows a significant amount of ARE comparing to non-gsp. This is because sleeping nodes reduced a significant number of overhearing transmissions and receptions through the network. Figure. Average number of gossip periods (G p ) until network termination vs. Network size (nodes). Figure. Average remaining energy (ARE) vs. network sizes (nodes). Figure 5. Change in network lifetime for both Known and Unknown path schemes when using GSP with p =.3 compared to p = (non- GSP). Fig. displays the average number of gossip periods (G p ) for different network sizes. The UKP scheme offers fewer average number of gossip periods than the KP scheme since the UKP scheme has a broadcast characteristic. GSP with p =.3 provides higher average number of gossip periods on both KP and UKP schemes. As a result, they offer a longer network lifetime than non- GSP. Fig. 5 shows the changes in network lifetime in percentage on both KP and UKP schemes when using GSP with gossip sleep probability p =.3 compared to Figure 7. Average energy consumed per gossip period vs. network size. Fig. 7 shows the average energy consumed per gossip period. The largest energy consumption occurs for the UKP scheme in Non-GSP network. When networks increase in size, the average energy consumed per gossip period increases. Figs. 8 and 9 demonstrate the ARE for the KP scheme in a 1 node grid network with and.3 gossip sleep 7 ACADEMY PUBLISHER

5 18 JOURNAL OF NETWORKS, VOL., NO. 5, SEPTEMBER 7 probabilities respectively. X and s represent the coordinated nodes locations on a grid topology. By placing a sink at the center of the grid, the energy usage increases toward the sink or center of the grid. This is because of data-centric characteristic in wireless sensor networks. However, by using GSP with p =.3, the shape of the plot is more symmetric (see Fig. 9). The symmetry in shape indicates balanced energy usage throughout the network. As a result, GSP with p =.3 provides longer network lifetime as shown in Fig Figure 8. A 3-D surface plot on ARE for the Known path scheme in 1 nodes grid network with p = Figure 9. A 3-D surface plot on ARE for the Known path scheme in 1 nodes grid network with p = Figure 1. A 3-D surface plot on ARE for the Known path scheme in 9 nodes grid network with p =. Figs. 1 and 11 show three dimensional plots of ARE for the KP scheme in 9 nodes grid network with and.3 gossip sleep probabilities respectively. Figs. 1 and 13 show ARE for the UKP scheme in 1 node grid networks with and.3 gossip sleep probabilities respectively. One observation is that heavy energy consumption is toward the center area of the grid. Moreover, nodes at the edges of the grid deplete energy faster than in KP scheme. However, the UKP scheme with GSP results in increased ARE Figure 11. A 3-D surface plot on ARE for the Known path scheme in 9 nodes grid network with p = Figure 1. A 3-D surface plot on ARE for the Unknown path scheme in 1 nodes grid network with p = Figure 13. A 3-D surface plot on ARE for the Unknown path scheme in 1 nodes grid network with p = Figure 1. A 3-D surface plot on ARE for the Unknown path scheme in 9 nodes grid network with p =. 7 ACADEMY PUBLISHER

6 JOURNAL OF NETWORKS, VOL., NO. 5, SEPTEMBER Figure. A 3-D surface plot on ARE for the Unknown path scheme in 9 nodes grid network with p =.3. Figs. 1 and plot ARE for the UKP scheme in 9 node grid networks with and.3 gossip sleep probabilities. GSP with p =.3 for a large network as in 9 nodes grid has increased ARE compared to non-gsp. In Fig., GSP results in greater improvement of ARE at every point of the grid in a larger network. IV. PERFORMANCE EVALUATION FOR D Network connectivity can also be achieved by increasing transmission power/radius, which in turn may allow for a higher gossip sleep probability (p) and additional energy conservation. By increasing transmission power, a node will use higher energy when transmitting and relaying packets. However, allowing more sleeping nodes with increased transmission power can improve overall energy efficiency. Since GSP was originally built upon the UKP scheme, which may reflect a more typical use for GSP. This section discusses GSP energy efficiency when increasing transmission/power to d performing only for the UKP scheme. A.. Revised Energy Model TABLE II. CROSSBOW TINYOS MICA MOTE MEASURED ENERGY CONSUMPTION IN WATTS 5 dbm dbm - dbm Transmit 8.33 mw 59.3 mw 5.3 mw Receive 5.35 mw.1 mw 5.3 mw Sleep 17.3 mw 1.9 mw 1.9 mw TABLE III. CROSSBOW TINYOS MICA MOTE MEASURED ENERGY CONSUMPTION IN JOULES/BIT Transmit.8 Receive.3 Sleep.9 5 dbm dbm - dbm Tables II and III present a revised energy consumption model for TinyOS Mica mote [3]. By measuring energy consumption in Mica motes, this model shows that the energy consumption in transmission and reception should be higher than values in the classic radio model [13]. Thus, this analysis will use this energy consumption model values. Previous analysis employed 5 dbm transmission power as the power necessary to communicate over a distance of length d. However, to conduct the network lifetime experiments, this section requires two transmission power values for d and d transmission power/radius. The energy consumption model in [3] did not determine the transmission range of the TinyOS Mica mote. However, by knowing the transmission power, the free space propagation model can approximate the Mica mote s transmission range. Consider a simple case where there is a direct path between the transmitter and receiver, where d is the distance between them. Assuming the transmitter and receiver gains are equal to 1, the received power (P r ) is expressed as the following [1]. Pt λ P ( d) = (π ) d r The transmitted power is P t in mwatts. The free space loss, L free, given by λ L free = log1( ) πd c / f L free = log1( ) πd Where d is in km, and c is the free-space velocity, which is equal to 3 x 1 8 m/s. f is the frequency in MHz. L free can be expressed as the following. L db db L free = 3. + log 1 ( f ) + log 1 ( d ) Thus, the received power is described as below. P receive ( dbm ) = P ( dbm ) L ( db ) transmit Assume a free space propagation model with f = 93 MHz, -98 dbm receiver sensitivity [], the Mica mote can transmit with the range of 3.7 km and.1 km by using transmitted powers 5 dbm and dbm respectively. The transmission range of 5 dbm transmitted power is approximately double the transmission range of dbm. Thus, the next subsection employs this analysis to simulations to determine the network lifetime when increase transmission power/radius from d to d. Figs. 1 and 17 evaluate the gossip sleep probability (p) that creates a connected network. The plots suggest p =. as the gossip sleep probability for all GSP network sizes when using d transmission power/radius. free 7 ACADEMY PUBLISHER

7 JOURNAL OF NETWORKS, VOL., NO. 5, SEPTEMBER 7 Figure 1. Probability of sleeping nodes vs. average path length for connected nodes in square grids with d transmission power/radius. determine the impact of the energy consumption when nodes turn on their transceiver in the idle/listening periods and the impact on the packet collisions. The shorter gossip period will force the network to change the topology faster than the longer one, which will be useful in application such as patients vital sign monitoring, because this application requires a sensor to frequently report a patient condition [1]. The longer gossip period can be used in the environment monitoring applications such as bridges, and airport runways monitoring since these applications require sensor to transmit data less frequently [1]. To facilitate comparison, the longer gossip period (G p = 3 seconds), simulation generates the same total amount of traffic as the short one (G p = 3 seconds). C. GSP Performance. Figure 17. Probability of sleeping nodes vs. ratio of nodes disconnected in square grids with d transmission power/radius. Figure 18: Average number of gossip periods vs. network size for the square grids with transmission power/distance d, G p = 3 and 3 seconds. B. Simulation Parameters. TABLE IV. SIMULATION PARAMETERS AND ENERGY CONSUMPTION MODEL WHEN USING TRANSMISSION POWER/RADIUS D. Data rate Packet size MAC Initial energy stored Transmit Receive Idle/Listening Sleep 19. kbps 1 bytes CSMA/CA 5 Joules Figure 19: The changes in network lifetime on the different sizes of the square grids with transmission power/radius d when using GSP with p =. compared to Non-GSP, G p = 3 and 3 seconds. Table IV shows the simulation parameters and energy consumption model used in transmission power/radius d analysis. With the increased transmission power, the more nodes may sleep in each gossip period with creating a disconnected network. A study employing two gossip periods, G p = 3 and 3 seconds, was created to Fig. 18 presents the network lifetime in term of the average number of gossip periods when using p = in the Non-GSP and p =. in the GSP network. The results show that the highest average number of gossip periods occurs for the small GSP network ( periods). Because of the high traffic load, when networks increase in size, 7 ACADEMY PUBLISHER

8 JOURNAL OF NETWORKS, VOL., NO. 5, SEPTEMBER 7 1 the average number of gossip periods reduces for both GSP and Non-GSP. Since the longer gossip period consumes more total energy in the idle/listening states, when using G p = 3 seconds with d transmission power/radius, the average number of gossip period decreases. Fig. 19 shows the change in network lifetime when using GSP d compared to Non-GSP d. Since, in Non- GSP d, nodes transmit with high transmission power without sleeping nodes in the network, energy is consumed faster than Non-GSP d. The largest change occurs for the small network with the 3 second gossip periods, where the changes tend to decrease when the networks increase in size. network shows lower energy consumption per node per gossip period for all network sizes compared to Non- GSP. Figure : Average energy consumed per gossip period vs. network size for the square grids with transmission power/radius d, G p = 3 and 3 seconds. Figure : Simulated network lifetime vs. network size for the square grids with transmission power/distance d, G p = 3 and 3 seconds. Figure 3: Packet loss ratio vs. network size for the square grids with transmission power/radius d, G p = 3 and 3 seconds. Figure 1: Average remaining energy (ARE) vs. network size for the square grids with transmission power/radius d, G p = 3 and 3 seconds. Fig. presents the simulated network lifetime when using the transmission power d in three sizes of square grids. The longest network lifetime occurs for the small GSP network with the 3 second gossip period. Fig. 1 plots ARE per node as a function of network size. The ARE improves when the network size increases. However, for a Non-GSP network, ARE decreases as the network size increases. Fig. shows the average energy consumed per gossip period for both GSP and non-gsp networks. The GSP Fig. 3 shows the packet loss ratio for both GSP and Non-GSP networks when using a d transmission power/radius. The GSP packet loss ratio drops under 1% in the 3 seconds gossip period, which improves from the ones with the transmission power/radius d. However, the GSP network has a higher packet loss ratio compared to Non-GSP network in the d case. GSP has a high number of sleeping nodes (%), and therefore less traffic is from forwarded/relayed packets. Thus, the ratio of the total number of dropped packets divided by the total number of transmitted/relayed packets, results in a higher packet loss rate in d case. The results show a decreased packet loss ratio for only the small GSP and Non-GSP networks in the 3 seconds gossip period, which is increased when the network size increases. Fig. demonstrates the changes in percentage of the average number of gossip periods and AREs when using transmission power/radius d compared to d in the GSP network on both G p = 3 and 3 seconds in the square grids. Even though GSP d employs a higher transmission power than GSP d, with higher p GSP d shows increasing network lifetime compared to GSP d. When the network 7 ACADEMY PUBLISHER

9 JOURNAL OF NETWORKS, VOL., NO. 5, SEPTEMBER 7 size increases, the network lifetime improvement in GSP d increases. Also, the results show that the networks employing GSP d improve on the measure of ARE. The largest change in ARE occurs for the shorter gossip period (G p = 3 seconds), with ARE improvement increasing as the network size increases. Employing GSP d over GSP d in the small network (1 nodes) shows small improvements on both network lifetime and ARE Figure7: ARE for 3x3 (9 nodes) square grid Non-GSP network (p = ) with transmission power/radius d. Figure : The changes of the average number of gossip periods and AREs in the square grids when using GSP d compared to GSP d with G p = 3 and 3 seconds Figure8: ARE for 3x3 (9 nodes) square grid GSP network (p =.) with transmission power/radius d Figs. 5 and represent the plots on ARE in the 1 nodes square grid with transmission power/radius d on both Non-GSP and GSP respectively. In the Non-GSP network, all nodes rapidly deplete their energy stores. Networks employing GSP show increases in ARE from the average of 7.3 to 1.3 Joules (%). As networks increase in size, ARE increases, which is from the average of.8 to. Joules (3%) as shown in Figs. 7 and 8. Figure 5: ARE for 1x1 (1 nodes) square grid Non-GSP network (p = ) with transmission power/radius d Figure : ARE for 1x1 (1 nodes) square grid GSP network (p =.) with transmission power/radius d. V CONCLUSIONS GSP reduces energy consumption in large wireless sensor networks by reducing overhearing without an increase in complexity over gossiping protocols. The energy efficiency of GSP is measured by network lifetime. GSP network lifetime for different sizes of square grids when using transmission power d was compared with non-gsp networks. When the networks employ GSP d compared to Non-GSP d, the network lifetime is extended by 3 lifetimes and the ARE increases by 1 Joules ranged from small to large network sizes. However, the packet loss ratio in GSP is higher than Non-GSP by approximately %. The simulation calculated the packet loss ratio as the total number of packet collisions divided by the total number of transmitted/relayed packets. Thus, the GSP results show higher packet loss rate in the d case. When comparing GSP d and GSP d, the network lifetime is increased by 7 1% in the 3 seconds gossip period and 8% in the 3 seconds gossip period ranging 7 ACADEMY PUBLISHER

10 JOURNAL OF NETWORKS, VOL., NO. 5, SEPTEMBER 7 3 from small to large networks. Moreover, the ARE increases approximately 5 1% when using GSP d over GSP d. Simulation results also show a significant increase in network lifetime when using GSP not only the Known Path but also the Unknown Path schemes, on different network sizes. The largest percentage improvement in network lifetime occurs for smaller networks. However, only in the Unknown path scheme that GSP shows a significant amount of ARE comparing to non-gsp. In addition to the Unknown path scheme, when network grows larger, GSP presents a better ARE improvement. Since GSP can integrate with other routing and MAC protocols, this ARE improvement may be of benefit, e.g., cluster-based protocols to use the remaining energy in network when rotating the cluster-heads or sinks. REFERENCES [1] I. F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci, A survey on sensor networks, IEEE Communication Magazine, vol., issue 8, pp. 1-1, August. [] B. Chen, K. Jamieson, H. Balakrishnan, and R. Morris, Span: An energy-efficient coordination algorithm for topology maintenance in ad hoc wireless networks, Wireless Networks, no. 8, pp. 81-9,. [3] M. Calle and Joseph Kabara, Measuring Energy Consumption in Wireless Sensor Networks Using GSP, in Proceedings of the 17 th Annual IEEE International Symposium on Personal Indoor and Mobile Radio Communications, PIMRC, Helsinki, Finland. [] W. R. Heinzelman, A. Chandrakasan, and H. Balakrishnan, Energy-efficient communication protocol for wireless microsensor networks, in Proceedings of the 33 rd International Conference on System Sciences (HICSS ), IEEE,. [5] W. R. Heinzelman, J. Kulik, and H. Balakrishnan, Adaptive protocols for information dissemination in wireless sensor networks, in Proceedings of the 5 th ACM/IEEE Mobicom Conference, Seattle, WA, August [] A. Manjeshwar and D. P. Agrswal, TEEN: A Routing Protocol for Enhanced Efficient in Wireless Sensor Networks, in Proceedings of the th International, Parallel and Distributed Processing Symposium, IEEE, April 1, pp. 9-. [7] C. Intanagonwiwat, R. Govindan, D. Estrin, J. Heidemann, and F. Silva, Directed diffusion for wireless sensor networking, Networking, IEEE/ACM Transactions, vol. 11, issue 1, pp. -1, February 3. [8] F. Ye, A. Chen, S. Lu, and L. Zhang, A scalable solution to minimum cost forwarding in large sensor networks, in Proceedings of the 1 th International Conference of the Computer Communications and Networks, IEEE, October 1, pp [9] Z. J. Haas, J.Y. Helpern, and L. Li, Gossip-based ad hoc routing, in Proceedings of the IEEE INFOCOM,. [1] X. Hou, D. Tipper, D. Yupho, and J. Kabara, GSP: gossip-based sleep protocol for energy efficient routing in wireless sensor networks, presented at 1 th International Conference on Wireless Communications, Calgary, Alberta, Canada,. [11] X. Hou and D. Tipper, Gossip-based sleep protocol (GSP) for energy efficient routing in wireless ad hoc networks, Wireless Communications and Networking Conference, IEEE, vol. 3, pp , March. [1] H. Karl and A. Willig, Protocols and Architectures for Wireless Sensor Networks, England: John Wiley & Sons Ltd, 5. [13] M. Calle, Energy Consumption in Wireless Sensor Networks Using GSP, M.S. Thesis, Department of Information Science & Telecommunications, University of Pittsburgh, Pittsburgh, PA, USA,. [1] B. Sklar, Digital communications fundamentals and applications, Prentice Hall PTR, 1. [] Crossbow Technology Inc. Wireless Sensor Networks: MOTE VIEW 1. User s Manual, October 5, Online: MOTE-VIEW_Users_Manual_.pdf. [1] D. Yupho and J. Kabara, Continuous vs. event driven routing protocol for WSNs in healthcare environment, in Proceedings of the Workshop on Security and Privacy in Mobile Health Care (PMHCS), Innsbruck, Austria, November. Debdhanit Yupho is an engineer in the Planning Department at Aeronautical Radio of Thailand Limited. Debdhanit is a graduate of the University of Pittsburgh (Ph.D. School of Information Sciences Telecommunications and Networking, 7), University of Colorado at Boulder (M.S. Telecommunications, ) and Thammasat University, Thailand (B E.E., 1999). His research interests include energy efficient wireless sensor network design, power limited networked sensor devices and routing algorithms for wireless networks. Joseph Kabara is an Assistant Professor in the Department of Information Science and Telecommunications at the University of Pittsburgh. Prior to joining Pitt in the fall of 1997 he lectured at Vanderbilt University. Between 1988 and 1991 he was employed as an Electronics Engineer at the National Institutes of Health Division of Computer Research and Technology where he designed their networking infrastructure. Professor Kabara is a graduate of Vanderbilt University (Ph.D. E.E., 1997), Johns Hopkins University (M.S. E.E., 1991) and Marquette University (B.S. E.E., 1987). His current research interests include wireless network design to support capacity requirements, information assurance for wireless networks, efficient algorithms for controlling information flow in wireless networks, power limited networked sensor devices and the use of embedded neural networks as algorithms for managing data and networks. He is a Senior member of the IEEE. His research has been supported by the University of Pittsburgh, the state of Pennsylvania Link--Learn program, National Science Foundation, National Institute of Standards and Technology and Microsoft. 7 ACADEMY PUBLISHER