Energy efficient cross layer load balancing in tactical multigateway wireless sensor networks

A tactical wireless sensor network (WSN) is a distributed network that facilitates wireless information gathering within a region of interest. A challenge in the deployment of WSNs is the limited battery power of each sensor node. This has a significant impact on the service life of the network. In order to improve the lifespan of the network, load balancing techniques using efficient routing mechanisms must be employed such that traffic is distributed between sensor nodes and gateway(s). In this paper, we study load balancing from a cross-layer point of view, specifically considering energy efficiency. We investigate the impact of deploying single and multiple gateways on the following established energy aware load balancing routing techniques: direct routing, minimum transmission energy, low energy adaptive cluster head routing, and zone clustering. Based on the node die out statistics observed with these algorithms, we develop a novel, energy efficient zone clustering algorithm called EZone. Via extensive simulations using MATLAB, we analyze the effectiveness of these algorithms on network performance for single and multiple gateway scenarios and show that the EZone algorithm maximizes network lifetime and service area coverage.


I. INTRODUCTION
A wireless sensor network (WSN) is a group of sensor nodes that are geographically distributed to provide data gathering and monitoring of tasks and events. These WSNs can be used in a variety of applications such as atmospheric monitoring, human detection and video surveillance.
As a result of their ubiquitous inclusion in society, WSNs are finding increased applicability to the Department of Defense (DoD) in areas specific to surveillance and reconnaissance. A WSN can be used to remotely monitor a battlespace making the presence of a warfighter unnecessary thereby increasing the safety to forces. In addition, a WSN can be used to remotely monitor deployed systems and trigger alerts at a command-and-control site when certain events occur.
Each sensor node in the WSN must have the ability to simultaneously serve as a sensing device and a wireless communication device that can exchange information with nearby nodes [1]. It is critical that information from every node is communicated to a desired destination outside the network. A typical WSN and its associated supporting infrastructure is shown in Fig. 1. Individual sensor nodes capture information using the battery energy they are deployed with. Nodes then utilize their peers (if necessary) to pass this information wirelessly to a gateway node and then through supporting infrastructure to a command-and-control site for further processing. The focus of this paper is the deployment of tactical WSNs. Tactical WSNs, as used by the DoD, are remotely deployed in potentially hostile areas with gateway nodes located on the outskirts of these areas. The network must operate reliably and maximize sensor network coverage for the maximum amount of time in the absence of human contact. A key challenge in the deployment of tactical WSNs is the limited battery power of each sensor node. This has a significant impact on the service life of the network. In order to improve the lifespan of the network, load balancing techniques using efficient routing mechanisms to achieve energy efficiency must be employed such that traffic is distributed between sensor nodes and gateway(s).
Modern day networks abstract all the processes that take place between any two nodes and represent them in the form of layers. The general network layering construct contains the following five layers: physical, medium access control (MAC), network, transport, and application layers. Generally, layers of one node only rely on information from the layer immediately above or below it.
In this paper, we exploit the opportunity to explore a crosslayer solution for the load balancing problem. A cross-layering method does not restrict a layer from utilizing information only from the layer directly above or below it.
Specifically, for load balancing and energy efficiency, we allow the network layer access to the physical layer for battery parameters and distance between nodes in performance of energy-efficient routing strategies. This allows us to: 1) create routing paths that conserve transmission power, and 2) favor those nodes with higher residual energy to perform high energy consumption tasks. In addition, we allow the network layer to access the application layer to perform data aggregation. Performing data aggregation reduces the size of network data packets, which reduces the energy required to transmit each packet through the network.

A. Motivations and Contributions
Various traditional load balancing algorithms exist in the literature [1], [2], [3]. In this paper we show that these traditional algorithms have a negative impact on the service life of a tactical WSN because their design does not take into account energy efficient strategies required to extend the network lifetime. Our aim in this paper is to provide realistic insights on how to incorporate an energy efficient load balancing strategy into a routing algorithm to maximize network service life.
Load balancing must be considered to achieve optimal performance. We consider optimal performance of the WSN to be when all nodes function for the maximum amount of time. We control the topology of live nodes as the concentration of dead nodes increases such that we achieve uniform service coverage throughout the area of interest through the lifetime of the WSN. We consider and obtain performance improvements by tactically including an additional gateway node in our simulations. We model one specific node-gateway arrangement and extend signal processing methodologies to model WSN die out parameters as random variables. This allows us to generate thousands of data points and draw our conclusions on an expansive subset of data instead of just one trial.
Our contributions in this paper can be summarized as follows: • Survey and identify existing load balancing routing algorithms for WSNs. We also identify performance improvements of adding an additional gateway to these algorithms. The routing algorithms studied are: 1) Direct Routing, 2) Minimum Transmission Energy (MTE), 3) Low Energy Adaptive Clustering Hierarchy (LEACH), and 4) Zone routing. • Develop a novel energy efficient WSN routing algorithm that uses a cross-layer approach and identify performance improvements compared to algorithms that do not consider energy efficiency. We refer to this algorithm as EZone. • Show detailed energy statistics for a specific nodegateway(s) arrangement and how this affects the continuous service coverage throughout the sensor field. • Model network die out statistics as random variables to better characterize the distribution of the algorithm results over thousands of trials. This technique allows us to better substantiate the performance of classical routing algorithms and our novel energy efficient routing algorithm (EZone). The remainder of this paper is organized as follows. In Section II we discuss the WSN protocol stack implementation that will facilitate our cross-layer load balancing strategies. In Section III we describe the various traditional algorithms that we simulated as well as our novel EZone algorithm that is developed paper. We provide our simulations and analysis of the results in Section IV. We conclude the paper in Section V.

II. WSN PROTOCOL STACK IMPLEMENTATION
As a result of our literature search on load balancing at each layer, we implemented the following models into each layer of the protocol stack. We build these models for both single and multigateway implementations of each routing algorithm analyzed.
Physical layer: Our WSN is comprised of 100 uniformly distributed sensor nodes located in a 50 m x 50 m grid and the gateway (gateways for the multigateway simulation) is placed 100 m away from the grid. We utilize a first order power amplifier and sensor model [2], [4], [5]. This model assigns an energy cost-per-bit to collect, transmit and receive information. It considers direct path and multi-path wireless signal propagation theory to identify the amount of information required to transmit one bit of information a certain distance between nodes while guaranteeing adequate signal-to-noise ratio at the receiving node. We utilize the first order radio energy model, which is common throughout the literature [6], [1], [2], [7]. This model relates the energy expended to send and receive an L-bit message over a distance d when considering direct path and multi-path propagation. We use energy symbols and parameters introduced in [5] to produce Table I. These parameters are employed in our physical layer simulations. All our simulations assume that each node is within wireless transmission of the gateway which also means that each node is within communication range of any other node in the WSN.

MAC layer:
We implement a Time-Division Multiple Access (TDMA) scheme that assigns each node in the WSN a timeslot during each round. The node transmits information to the gateway during the timeslot. We simulate the MAC layer simply through the performance of transmission rounds. Each simulation begins at round zero and ends at some maximum number of rounds (or when the last node dies). During each round, each node in the WSN sends an L-bit packet from the application layer to the gateway. We are not concerned with how the TDMA assignment takes place, just that during each round, each node transmits its packets to the gateway.
Network layer: There are a variety of routing algorithms applied to the network layer in the literature, some of which are described in [8], [9], [10], [11], [12]. We implement several traditional and established routing algorithms observed in the literature. We also develop and implement our own energy efficient routing algorithm (EZone). The routing algorithms we implement are as follows: • Direct: All nodes transmit their data message directly to the nearest gateway during each transmission round. • MTE: All nodes transmit their data to the nearest gateway using a shortest-path MTE route that is calculated using Dijkstra's shortest path routing algorithm [1]. The link cost parameter input into Dijkstra's algorithm is the distance squared between nodes along the path. • LEACH: A popular network clustering algorithms specific to WSNs [1], [2]. • Zone: A network clustering algorithm that partitions the network into zones and hierarchically determines node to gateway routes using a cluster head (CH) assigned in each zone [3] • EZone: Energy efficient zone routing is the novel routing algorithm developed and implemented in this paper. EZone implements Zone routing as described in [3], except EZone elects the node with the most energy at the beginning of each round to be the CH. EZone will be further discussed in Section III. Transport layer: Our transport layer implements a strategy similar to the modern day internet's use of the User Datagram Protocol (UDP). UDP is a connectionless oriented protocol in which the source node gets no feedback that its messages reached the destination. This is applicable for WSNs as it provides a mechanism to prevent feedback transmissions that would unnecessarily deplete WSN energy levels.
Application layer: Our application layer implements two strategies: 1) use of a traffic generator, and 2) use of a data aggregation technique. The traffic generator of each node generates a 2000 bit data message during each round for transmission to the gateway. Data aggregation is used only for the clustering algorithms and the CH is the only node that can perform data aggregation. The CH receives all the messages from nodes in the cluster. It then includes its own sensor's message, compresses all the messages into one 2000bit message, and transmits the compressed message to the gateway at the end of each round.

III. ROUTING ALGORITHMS FOR LOAD BALANCING: TRADITIONAL VS EZONE
In this section, we describe the five routing algorithms that were simulated: Direct, MTE, LEACH, Zone, and EZone. We simulate these algorithms in single and multigateway scenarios. Our interest is to investigate the result of load balancing techniques in single and multigateway WSNs by employing load balancing techniques at each layer while focusing on the impact of varying the network layer routing algorithms on the WSN.

A. Traditional Routing Algorithms: Direct, MTE, LEACH and Zone
Direct transmission to the gateway involves each node sending a packet to the gateway directly without using any other nodes along the way. During each round, the Euclidean distance is calculated between the node and the gateway. The distance along with the transmit amplifier parameters given in Table I is used to determine the propagation mechanism [6]. The node's energy is decremented in proportion to the required energy for packet transmission to the gateway. For the multigateway scenario, the node chooses the gateway that requires the smaller transmit energy (i.e., the closer gateway) and transmits the packet to that gateway.
To simulate the MTE algorithm, a route from every node to the gateway must be generated. We desire to minimize propagation distance to the gateway in order to produce a route that minimizes the overall sensor energy depletion rate. We utilize propagation distance as our link cost parameter to input into the MTE algorithm. We use Dijkstra's algorithm to generate our MTE routes. In MTE routing, the node closest to the gateway is always chosen to be included in the route. This node is known as the hot node. Since the hot node is the relay point between the gateway and all traffic from other nodes, it is overwhelmed with traffic during each round and dies quickly. Another hot node is then immediately chosen. This hot node concept in MTE routing causes nodes that are closest to the gateway to die out first. This results in the node closest to the gateway dying out very quickly and the least preferred node (least preferred position for routing) dying out last because no other node utilizes it in the calculation of the preferred route. Our MTE algorithm does not employ any data aggregation strategy since at each round every node is assumed to transmit its message in a TDMA scheme where there is only one message passing from source to gateway through the network at a time.
We now move into describing clustering techniques. Our motivation for employing a clustering technique is aimed at rotating the energy intensive role of the node that performs the long-range wireless transmission to the gateway as well as providing the opportunity to perform data aggregation. In clustering, a cluster head (CH) is chosen from a group of nodes to serve as an intermediate relay between a group of nodes and the gateway(s). Utilizing a clustering mechanism, we rotate the role of the CH to minimize the probability that any node is a hot node in an effort to balance the energy depletion of all nodes and take into consideration the topology of the network as subsequent nodes die out.
The LEACH algorithm is a well-known clustering algorithm developed specifically for WSNs. LEACH routing elects a CH and nodes associate with the CH according to the LEACH algorithm [2]. Each node picks a random number between zero and one. Each node also computes a threshold number (T n ), which is a number between zero and one and is proportional to the current round. The probability for any node to serve as a CH is denoted as p. If a node has been a CH in the last 1 p rounds, it is excluded from being a CH during the round. Otherwise, if the temporary random number is less than T n , the node is elected as a CH during the round. The desired probability for a node to be chosen as a CH is an input to the algorithm and must be specified. The original author of LEACH performed analysis to determine the optimum value for p to be 0.5 [1]. Nodes that were not elected as CHs during each round then associate in clusters with the nearest CH.
Each node then transmits its data message to its CH. The CH collects all the messages of its nodes and retransmits them collectively to the gateway. This process repeats during subsequent rounds until all nodes have died.
Zone clustering appears less frequently in the literature as compared to LEACH. However, for a tactical network, it may be a preferred routing algorithm because the user can specify how zones are characterized for the network. The general methods that we use for our zone routing algorithm are based on techniques described in [3]. In [3], the authors utilize a sensor field comprised of homogenous zones. A sensor in each zone has a probability p of becoming a CH during each round. The probability p is determined relative to the number of nodes in the zone: p = 1 (number of nodes in zone) . Their methodology is useful for a tactical network in that the objective of zoning in a WSN is to ensure that CHs are uniformly selected through the network. Our zone clustering algorithm divides the sensor field into z equal zones. Equal zones were chosen because the distribution of nodes is uniform in the field. Equal zones span along the Cartesian x-axis to create z vertical rectangular zones. We use five zones in our simulations. Five zones were chosen to provide a comparison with the LEACH algorithm. Recall that in the LEACH algorithm, the probability of any node being chosen a CH is p=0.05. Thus, in a 100 node network, we would on average have five CHs. To ensure there are five CHs for our zone clustering algorithm, we must have five zones and each zone is only allowed to have one CH.
Once all nodes are assigned to a zone, we begin the simulation at round one and complete the simulation at the maximum desired round. In each round, the set of live nodes for each zone is identified, and the CH is chosen based on a random assignment from this set. Each node in the zone then transmits its L-bit packet to the zone's CH and its energy is decremented according to our radio energy model. The CH for the zone then aggregates all the messages from the nodes in the zone and transmits the aggregated message to the gateway. The process is repeated for each zone at each round.

B. EZone: Zone Clustering with Energy Efficient Cluster Head Selection
The zone clustering case described in Section III-A chooses the CH for each zone randomly. A clustering algorithm that partitions nodes into specific zones is an energy saving technique when compared to the LEACH algorithm because there is a lower maximum distance that any node must transmit to reach its CH. Our implementation of the zones guarantees a nearby CH in the zone as compared to that of LEACH. In LEACH the nearest CH may be on the other side of the network since the criteria for a node to be elected as a CH may have only been met randomly on the other side of the field.
There are significant differences in energy distribution of the nodes in the network. The differences in energy levels across the WSN caused some nodes to die out earlier and some nodes to die out later. Therefore, we modify the CH election criteria in the following way: in any given round, if the highest energy node is chosen to be the CH, individual node energy depletion rates are minimized with the battery levels in any zone depleting at a uniform rate.
To accomplish this strategy, we modify our zone routing algorithm to revise the process of electing the CH in each zone at each round. Instead of randomly choosing the CH from the live nodes in the zone, we choose the CH that has the maximum energy level in the zone. Based on this election criteria, nodes that are in a more preferred location (a location that minimizes energy depletion rate such as locations closer to the gateway) are chosen to be the CH for the zone more than those in a less preferred location (a location farther away from the gateway).
In practice, electing the highest energy node to be the CH during each round in each zone requires additional processing by the gateway to perform CH election. Our simulations perform this aspect automatically with the assumption that it is normally performed by the gateway.

IV. SIMULATIONS AND RESULT ANALYSIS
In our simulations, sensors and gateways are all placed on a Cartesian grid with axes x and y. Our simulations and analysis involve a grid of 100 sensors such that each sensor's x and y coordinate is modeled as a uniformly distributed random variable between 0 and 50 meters (m). The single gateway scenario employs the gateway at (x, y)=(25m, -100m), while the multigateway simulations have gateways positioned at (x 1 , y 1 )=(25m, -100m) and (x 2 , y 2 )=(25m, 150m). The single WSN is graphically shown in Fig. 2. The multigateway scenario is similar except an additional gateway is placed on top of the network topology at the position (25m, 150m). The zoning arrangement shown in these figures are only used for the zone related algorithms; the other algorithms make no use of the vertical zone partitions. Gateways are displayed as solid green, and nodes are represented by a blue outline circle. We show the perimeter and field zones as solid red lines. All nodes have a starting energy of 0.5 J, and gateway(s) are assumed to have unlimited energy (they are not energy constrained).
We employ two application layer strategies, 1) a constant bit rate (CBR) generator and 2) a data aggregation application. CBR allows each node to send an L=2000 bit message to the gateway at each round. We are not concerned with the contents of each message. Instead, we only care that messages are produced so that we may observe how energy is depleted throughout the network due to the network routing algorithm in use.
Only clustering mechanisms use data aggregators in our scenarios. Data aggregation requires energy to perform the signal compression, which must be accounted for. We adopt a similar technique, used in the literature, which applies an energy cost to the data aggregator for the task of aggregating all the data during a round [1], [2], [13], [14], [7]. The node performing data aggregation is always the CH, and a data aggregation constant, EDA [15], is used to account for the energy to compress messages into one final L=2000 bit

A. Comparisons of Algorithms based on Energy Consumption
We plotted the total WSN system energy level during each transmission round (Fig. 3), the energy variance that resulted from the distribution of individual node battery levels (Fig. 4) and the number of live nodes during each round (Fig. 5). We visually observed how nodes geographically die out throughout the simulation. In each legend of Figs. 3-5, an S after an algorithm name refers to the single gateway scenario, and M refers to the multigateway scenario. The clustering algorithms dramatically outperformed the MTE and direct routing algorithms as a result of rotating and distributing the high energy role of nodes performing a long-range transmission and allowing the CHs to perform data aggregation. The single and multigateway clustering algorithms generally displayed similar energy depletion rates  Total number of alive nodes versus transmission round for all algorithm simulated. Our energy efficient zone routing algorithm provided the longest timeframe of 100 percent service area coverage that are illustrated in the linear regions of Fig. 3. The clustering algorithms minimized the energy variance of the WSN, and our energy efficient zone routing algorithm, EZone, provided an indistinguishable flat variance plot compared to other algorithms as shown in Fig. 4. EZone maximized the time when all nodes are alive with the single gateway simulation outperforming other multigateway algorithms. This is significant in that it reveals the efficiencies that can be gained by implementing an energy efficient cross-layer approach.
Our EZone algorithm outperformed all other algorithms from a topology perspective during node die out as well. While other algorithms created a pattern for die out, our energy efficient algorithm caused nodes to quickly die out immediately after the first node died. This also is significant in that we utilized a cross-layer approach to maximize 100 percent service coverage of the WSN. Node die out of other routing algorithms occurred in an unfavorable fashion. For example, in the direct case, live nodes farther from the gateway died first since their energy is depleted proportional to their distance from the gateway. As a result, areas farthest from the gateway lost service first, while areas closest to the gateway remained in service longest. In MTE routing the nodes closest to the gateway died first. The LEACH algorithm inefficiently creates clusters that cause the network to die out starting in the center of the sensor field and progressing radially outward. As a result of this die out mechanism, we lose coverage in the middle of the sensor field first. These die out mechanisms warrant the choice of our energy efficient zone routing algorithm for a tactical WSN since it preserves 100 percent network coverage the longest.
The single gateway case performed better than the multigateway cases for LEACH and zone clustering with random CH election. This demonstrates the impact of optimizing an energy efficient network layer strategy. As efficiency is gained at the network and application layer, the impact of the additional gateway is lowered when looking at the energy depletion rate (without consideration for topology of WSN die out). The clustering algorithms all demonstrated approximately similar energy depletion rates for respective single and multigateway configurations. This value was obtained in the linear region of the plots with all nodes alive and five CHs elected for each round.
The addition of another gateway was most significant in the direct and MTE algorithms as the energy variance is lowered by approximately 50 percent. This can be seen in as shown in Fig. 4. Energy variance of the zone routing algorithms were both lower than LEACH, with the single gateway scenarios performing better than LEACH in a multigateway configuration. The direct and MTE number of nodes alive do not cross in Fig. 5 as they do in [1] because the authors only used a direct path propagation model, while our research uses both direct path and multi-path propagation models. EZone offered the most time with all nodes alive; however, LEACH offered the most time with at least one node alive.
Statistics comparing the impact of an additional gateway is given in Table II. The table shows the round and distribution of nodes when 10%, 50% and 100% of nodes are dead. The statistics are graphically shown in Fig. 5. Table II shows the ratio of time the network is depleted from 100 percent to 20 percent (80 percent die out range) and the timeframe that the network provides 100 percent service coverage (the round the first node dies).

B. WSN Die Out Statistics and Random Variable Modeling
It must be noted that the results presented in Section IV-A only focused on the WSN sensor field arrangement shown in Fig. 2. To investigate other arrangements, we modeled the network die out parameters as random variables and obtained the distribution of network die out.
During every iteration, a new random WSN with uniform node distribution is created and run using similar parameters as before with die out parameters being appended to each random variable (RV) array. We utilized similar parameters for the number of nodes in the field, field dimensions, gateway locations, and physical and networking parameters. The only difference is that during each iteration of the algorithm, nodes are placed in different uniform locations in the grid.
All algorithms were executed for 5,000 iterations except for the MTE algorithms that were executed for 1,000 iterations. These numbers were chosen to offer a large sample size to obtain a representative distribution, yet small enough to limit total processing time. Each 5,000 iteration run required about one day of dedicated processing time on a modern Windows personal computer while the MTE algorithms required four and seven days for single and multigateway configurations, respectively. The MTE algorithms required significantly more time because of the computational complexity in calculating the MTE path for each node, each round, and each iteration using Dijkstras algorithm. This is why only 1000 iterations were run.
Results for our random variable testing are contained in Table III. A graphical bar plot of our mean value results of Table III is shown in Fig. 6. The standard deviation of network die out statistics is given in Table IV. The performance improvement of WSN clustering algorithms with data aggregation (LEACH, Zone, and EZone) and the improvement of WSN lifetime by the addition of an additional gateway compared to MTE and Direct Routing is illustrated in Fig. 6. We noted similar results in Section IV-A for the one uniform WSN arrangement tested. EZone maximized the service life when all nodes are alive by rotating the high energy CH role to the node in each zone with the most energy. LEACH provided the most time through 80 percent of network die out because of the random approach of CH election and the instability of the network to maintain a uniform number of CHs during each round after the first node dies.  In this paper we showed that network layer load balancing can be used to cause the network to die out in a tactically oriented fashion. The inclusion of an additional gateway extended WSN service life and offered improved coverage during die out as compared to the single gateway scenario. Our EZone algorithm offers the best opportunity to extend WSN service life while maintaining tactical control of the network layer in both single and multigateway configurations. It produced the least variance in energy distribution at any round and smartly balanced cluster and node loading since our zones were implemented based on knowledge of physical layer topology and anticipated application layer loading.