Guaranteeing the network lifetime in wireless sensor networks: A MAC layer approach

Transcription

1 Computer Communications 3 (27) Guaranteeing the network lifetime in wireless sensor networks: A MAC layer approach Yongsub Nam a, Taekyoung Kwon b, *, Hojin Lee b, Hakyung Jung b, Yanghee Choi b a Information Networking Institute, Carnegie Mellon University, Pittsburgh, PA, USA b School of Computer Science and Engineering, Seoul National University, Seoul, Republic of Korea Available online 21 June 27 Abstract Energy is the most crucial but scarce resource in wireless sensor networks (WSNs). A wealth of MAC protocols are proposed only to prolong the network lifetime for energy-efficiency; whereas, others are tailored to reduce end-to-end latency in addition to extending the network lifetime. Since the requirements depend on the applications of the network, it would be difficult to design a single MAC that satisfies the wide range of applications. Specifically, some WSNs are required to survive for a certain lifetime because of too expensive deployment cost (e.g., harsh or hostile environments) to replace the energy-depleted sensor nodes. Furthermore, the portion of surviving sensor nodes can be a critical factor to satisfy the quality of surveillance (QoSv) requirements. In this paper, we propose a new adaptive MAC (A-MAC) protocol that not only guarantees the pre-configured network lifetime but also reduces the end-to-end latency. Basically, each sensor node adapts its duty cycle depending on the traffic load it suffers. By doing so, the energy consumption rate of each node approaches to the ideal energy consumption rate, which enables the node to survive for the pre-configured lifetime. Also, if a node suffers relatively less load, it can reduce the sleep delay by increasing its duty cycle. Analysis and simulation results exhibit that the proposed protocol shows less delay than S-MAC [W. Ye, J. Heidemann, D. Estrin, Medium access control with coordinated adaptive sleeping for wireless sensor networks, in: IEEE/ACM Transactions on Networking, vol. 12, No. 3, June 24.] while meeting the network lifetime requirements. Ó 27 Elsevier B.V. All rights reserved. Keywords: Wireless sensor network; Medium access control (MAC); Network lifetime; Duty cycle; Adaptation 1. Introduction A wireless sensor network (WSN) is an emerging technology for a wide range of potential applications, including the environmental monitoring, medical, and target tracking systems. A WSN comprises a set of sensor nodes deployed in an area of interest. Sensor nodes are normally located in a dense and ad hoc manner, communicating each other in a multi-hop fashion in order to collect, process, and relay data. A majority of WSNs consist of battery-powered sensor nodes. In such networks, it is hard to recharge or replace * Corresponding author. Tel.: addresses: ysnam@cmu.edu (Y. Nam), tkkwon@snu.ac.kr, tk@mmlab.snu.ac.kr (T. Kwon), lumiere@mmlab.snu.ac.kr (H. Lee), hkjung@mmlab.snu.ac.kr (H. Jung), yhchoi@mmlab.snu.ac.kr (Y. Choi). the energy-depleted nodes due to the desolate or harsh environment of the target area. Therefore, how to efficiently utilize the limited amount of energy has been the primary concern in designing MAC protocols for WSNs. This energy-efficiency is directly proportional to how much network lifetime is prolonged. Prolonging the network or node lifetime has a significant effect on the networking and sensing operations of WSNs. If a sensor node runs out of its energy, the network connectivity or the quality of surveillance (QoSv) of WSNs can be hampered. As the number of depleted nodes increases, the network may be partitioned and/or the precision or quality of sensory data may fail to meet the application requirements. Specifically, a sensing hole may occur when the sensor nodes in charge of a specific area die. A sensing hole is defined as a blind area that cannot be covered, but is included in the region of the sensor network. Also, network partitioning [8] can /$ - see front matter Ó 27 Elsevier B.V. All rights reserved. doi:.16/j.comcom

2 Y. Nam et al. / Computer Communications 3 (27) happen when all the intermediate nodes between two or more partitions die. When network partitioning occurs, there would not exist any path between partitions, therefore, the network cannot work appropriately. In addition, according to [9], QoSv relies primarily on the number of sensor nodes alive. Therefore, making all the nodes survive for a certain amount of time is worth tackling. However, most of the existing MAC protocols seek to prolong the lifetime of each sensor node without considering uneven or unbalanced distribution of energy consumption among sensor nodes. This can bring about early termination of the network, even though most of sensor nodes have enough remaining energy. Thus, we propose a new adaptive MAC (A-MAC) protocol that guarantees the network lifetime by making all the nodes survive for a pre-configured time. We assume the required lifetime is already known. In order to guarantee the network lifetime configured a priori, each node tries to make its energy consumption rate approximate to the ideal energy consumption rate. This can be achieved by adapting the duty cycle. To the best of our knowledge, there is no previous work, especially at MAC layer, that guarantees the network lifetime in WSNs. In addition, A- MAC exploits the surplus energy of relatively energy-rich nodes in order to reduce end-to-end latency. The contributions of this paper are as follows. Automatic and adaptive duty cycle configuration: determining the duty cycle of a sensor node requires an intensive testing of energy consumption of the sensor device and an estimation of the traffic load. If the operator configures the duty cycle too high, the network may not survive for the expected lifetime. On the contrary, A-MAC does not need manual configuration of the duty cycle. By measuring the energy consumption rate in run-time, each node adapts its duty cycle automatically. Guarantee of pre-configured network lifetime: many applications will require the guarantee of the pre-configured network lifetime. By adapting the duty cycles of sensor nodes in run-time, A-MAC ensures the required network lifetime. Latency reduction: the listen/sleep operation of the sensor nodes may introduce high latency. If a network should survive for a required lifetime, the duty cycle of each sensor node is configured to that of the bottleneck node that suffers from the most traffic load. In that case, latency becomes larger than the minimum that the network can provide. A-MAC reduces latency by adapting the duty cycle depending on the traffic load of each sensor node. Therefore, the node with less traffic load can provide less delay, which reduces the overall delay. Implicit load balancing: in the existing MAC protocols for WSNs, the routing path will not change once it has been set up until a relay node dies. This causes unbalanced distribution of the traffic load. In A-MAC, the node with less traffic load increases its duty cycle, indicating that it can relay the message with a less delay. If the routing protocol on top of A-MAC supports the dynamic route change, the traffic load will be distributed approximately fairly among sensor nodes. The rest of this paper is organized as follows. We present a survey of existing MAC protocols in terms of the network lifetime and the end-to-end delay in Section 2. In Section 3, discussions on the network lifetime and latency are presented, followed by the proposed A-MAC protocol in Section 4. In Sections 5 and 6, brief analysis and simulation results are given. Finally, concluding remarks are given in Section Related work Compared to processing and sensing, communication has been the most power-consuming operation in WSNs. Due to diverse sensor networking environments (e.g., node density, transmission range), carrier sense multiple access (CSMA)-based MAC protocols become the norm in WSNs. However, CSMA MAC has also been the major source of energy waste, such as idle listening, overhearing, control overhead, and packet collision. Among them, idle listening is the main cause of energy waste. To mitigate energy consumption by idle listening, most of the currently proposed MAC protocols employ periodic listen/sleep scheduling. With this scheduling, each node listens to see if any other node wants to transmit data to it in the listening period. If so, it stays awake during the sleep period; otherwise, it sleeps, turning off the radio and setting a timer to awake itself later. The listen/sleep schedule is a good approach to save energy in WSNs where the traffic load is light. Fig. 1 illustrates an example of a periodic listen/sleep cycle. Here, a cycle of a listen and a sleep period is called a superframe or a frame [2], and the ratio of the listen period to the superframe is termed a duty cycle. For example, a node with 25% duty cycle has the sleep period three times as long as the listen period. However, this listen/sleep scheduling delays the data delivery. Overall, there is a trade-off between energy-saving and end-to-end latency. The existing MAC protocols fall into two categories depending on their objectives: (i) prolonging the network lifetime only, (ii) reducing end-to-end latency and prolonging the network lifetime MAC protocols for prolonging the network lifetime Following protocols aim at prolonging the network lifetime. With the limited amount of energy, these protocols Fig. 1. Example of a listen/sleep cycle.

3 2534 Y. Nam et al. / Computer Communications 3 (27) sacrifice traditional performance metrics, such as throughput and end-to-end latency to improve the network lifetime. S-MAC [2] is a representative contention-based protocol using periodic wakeup/sleep schedules, as shown in Fig. 2. By employing the periodic listen/sleep cycle, S-MAC seeks to prolong the network lifetime since energy waste due to idle listening is reduced. Additionally, S-MAC includes mechanisms to further reduce energy waste: collision, overhearing, and control overhead. All nodes are free to choose the listen/sleep schedules and periodically broadcast SYNC packets to their immediate neighbors. Nodes that have not adopted their own schedules follow the received schedule included in the SYNC packet. Accordingly, nodes with the same listen/sleep schedules form a virtual cluster, waking up and entering into the sleep state simultaneously. In addition, collision and overhearing problems are alleviated by RTS/CTS and NAV mechanisms similar to those of the IEEE [1]. However, due to the periodic sleeping, if a downstream node does not receive the RTS message in the listen period while forwarding data, an upstream node should defer forwarding data until the next listen period starts. That is, the amount of deferred time is accumulated at each hop, which is the main drawback of S-MAC. Overall, S-MAC achieves the prolonged node lifetime at the cost of increased end-toend latency. One way to mitigate this increased latency is to employ adaptive listen. The basic idea of adaptive listen is to let the node that overhears its neighbor s transmissions wakeup for a short period of time at the end of the transmission since the node may have to relay that data packet. Even though S-MAC significantly increases energy-efficiency, it does not take into account the uneven distribution of energy consumption among sensor nodes, not to mention the requirement of a pre-configured network lifetime. B-MAC [3] is also designed to minimize idle listening energy consumption. B-MAC contains a small core of media access functionality, such as clear channel assessment (CCA) and packet backoffs for channel arbitration, link layer acknowledgements for reliability, and low power listening (LPL) for low power communication. In B-MAC, each node wakes up periodically to check for channel activity. Specifically, each time the node wakes up, it turns on the radio and checks for activity. If activity is detected, the node powers up and stays awake for the time required to receive an incoming packet. If no packet is received, a timeout forces the node back to sleep. To ensure that all packets are heard by the nodes, packets are sent with a preamble whose transmission/reception time is longer than the Fig. 2. S-MAC and T-MAC schedules. check interval. B-MAC defines eight wakeup periods (referred to as the check interval), and each check interval corresponds to one of B-MAC s eight listening modes. B- MAC therefore defines eight different preamble lengths referred to as transmit modes. For a given network configuration (each node s neighbor node size and the ideal sampling rate), B-MAC s parameters such as check interval can be calculated and hence its lifetime can be estimated. B- MAC assumes identical operation (i.e., sampling rate) of all the sensor nodes. Therefore its lifetime analysis can be applied to only special applications of WSNs. PMAC [4] dynamically adapts the schedule of sensor nodes depending on the traffic load to conserve energy. Based on the traffic of its own and its neighbors, a node generates a pattern which indicates its tentative schedule. Then, by exchanging the pattern in the periodic Pattern Exchange Time Frame, each node can be aware of the schedules of its neighbors. Since the nodes can adjust its sleep period according to the traffic, PMAC can save more energy than S-MAC under light loads. However, the node suffering more load than others will die much earlier compared to S-MAC. Consequently, more unbalanced energy distribution among nodes than S-MAC is expected, which can reduce the network lifetime MAC protocols for improving end-to-end latency Following MAC protocols endeavor to reduce end-toend latency while trying to achieve energy-efficiency. To improve the serious latency problem of periodic listen/sleep scheduling, these protocols may sacrifice energy-efficiency to a certain extent, compared to the MAC protocols in the previous section. T-MAC [5] operates similarly to S-MAC; thus, it also adopts a periodic listen/sleep cycle. In T-MAC, however, the active period is not fixed but dynamically ending when no activation event, such as reception of data packet or RTS/CTS, occurs during a certain interval. As shown in Fig. 2, the length of the active period changes depending on occurrence of the activation events. T-MAC is inspired by the burst traffic pattern of WSNs. Thus, T-MAC delivers the bursts in consecutive listen intervals. In this way, T- MAC reduces energy consumption due to idle listening by transmitting all messages in bursts, and sleeping between bursts. As a consequence, the average end-to-end delay can be decreased. However, T-MAC does not take the network lifetime into account; it only focuses on reducing idle listening and thus the average end-to-end delay. D-MAC [6] is also based on a periodic listen/sleep model. In order to reduce end-to-end latency, D-MAC focuses on the traffic patterns of WSNs, especially the data gathering pattern from sensor nodes to a sink. In many WSN applications, the messages are normally destined to the sink node only. D-MAC forms a data gathering tree, the tree structure of the delivery paths from multiple sources to sink. D-MAC also introduces a staggered active/sleep schedule, where the activity schedule of inter-

4 Y. Nam et al. / Computer Communications 3 (27) mediate nodes on the multi-hop path are staggered (actually, pipelined) to wakeup sequentially like a chain reaction; i.e., the sleep schedule of a node is calculated based on its depth on the data gathering tree. An example of staggered scheduling is given in Fig. 3; in this case, node 1 is the leaf node and node 3 is closest to the sink node. Thus, it reduces sleep latency by pipelining or staggering the sleep/wakeup schedule. In addition, D-MAC adjusts the duty cycles adaptively depending on the traffic load in the network. As a result, D-MAC is able to decrease end-toend latency significantly. It however does not take the network lifetime into consideration. Also, since data transmission is based on the data gathering tree, it cannot adapt to the change in the topology. Furthermore, it only supports unidirectional communication pattern. ASAP [7] is a contention-based protocol, exploiting to maximize the network lifetime while satisfying the latency requirements. ASAP uses a hash function to decide its listen/sleep schedule, i.e., hashing the node identifier (e.g., MAC address) to determine its listen slot in a frame. During a neighbor discovery interval, each node wakes up and broadcasts its identifier. Since each node keeps track of its neighbor s identifiers, it is aware of its neighbor s listen slot. A node which has data to send to its neighbor node wakes up at the neighbor s listen slot and transmits it. The listen/sleep schedule changes in run-time depending on the node density and its residual energy. ASAP substantially reduces end-to-end latency by introducing staggered scheduling, which is illustrated in Fig. 3. S-MAC forms a virtual cluster and every node within the same virtual cluster wakes up at the same time, thus relaying at each hop takes the duration of a single superframe. However, ASAP does not form any clusters and the listen/sleep schedule is determined arbitrarily by the hash function, which leads to an almost uniform distribution of listen/sleep schedules among nodes. Therefore, ASAP can notably reduce the end-to-end delay in the multi-hop case. However, broadcasting data requires repeated transmissions at the listen slot of each neighbor node, which causes much more control traffic. 3. Discussions 3.1. Network lifetime Fig. 3. Example of staggered scheduling. There is no single MAC protocol as a panacea for WSNs; the requirements may vary depending on the target applications. We therefore need to explore the network lifetime issue to design a more comprehensive solution. The meaning of the network lifetime may change depending on and the application requirements. The network lifetime can be defined as one of the following. Time to network partitioning: network partitioning means that a network is divided into two or more partitions since the intermediate nodes between these partitions run out of energy. When network partitioning occurs, there would not exist any path between two partitions, and consequently, the network does not work properly. Time to a sensing hole s occurrence: a sensing hole means a blind point where the sensor network cannot sense. If every node in charge of an area dies, a sensing hole takes place. Accordingly, any sensing event in that area cannot be delivered to the sink. Thus, time to a sensing hole s occurrence is equivalent to the network lifetime in the monitoring applications. Time to first node s death: some WSN applications might be sensitive to a single node s exhaustion of energy, possibly due to the sparse deployment of sensor nodes. In such applications, the network would not work as appropriately as expected when even a single node dies. In such cases, time to first node s death degrades the QoSv or incurs the sensing hole problem. In a densely deployed network, this definition can be extended to the time until a certain percentage of the sensor nodes die. Most of the existing MAC protocols for WSNs focus on improving the lifetime of each node without considering unbalanced energy consumption among nodes. With these approaches, prolonging the node s lifetime does not always result in the longer network lifetime. A node suffering more traffic may die far earlier than others, resulting in the early termination of the network, regardless of the remaining energy levels of other nodes. Therefore, it is of great significance to take into account not only the node lifetime but also the entire network lifetime in designing a MAC protocol for WSNs End-to-end latency Some applications of the WSNs, e.g., surveillance, habitat monitoring, or target tracking systems, are likely to be delay-sensitive. The data gathering tree in [6] or staggered scheduling in [7] can be good approaches to tackle the latency problem. However, these schemes do not consider the unbalanced energy consumption among the nodes. For example, a sensing hole will occur or even the network can be partitioned as more nodes run out of energy. Therefore, the network lifetime mentioned above should also be considered in addressing the latency problem. Overall, there is a trade-off between minimizing latency and prolonging the network lifetime. Therefore, both objectives should be taken into account in designing a

5 2536 Y. Nam et al. / Computer Communications 3 (27) MAC protocol for lifetime-critical WSNs. To this end, the ultimate goal is to distribute the load equally among the nodes. There are two possible solutions. One is to employ the topology control schemes, such as [15] and [16]. However, topology control will incur the additional overhead, such as a number of control messages. Furthermore, it is more susceptible to the sensing hole problem. The other is to configure the periodic listen/sleep schedule at the MAC layer like S-MAC. If we set the schedule appropriately, we can achieve both objectives. However, there are inherent asymmetries in WSNs; e.g., traffic is forwarded toward the sink and the occurrence rate of events can be non-uniform. Hence, the rationale behind our approach seeks to balance the energy consumption among nodes by adjusting the duty cycles dynamically and independently. 4. A-MAC: an approach for guaranteeing the network lifetime Fig. 4. Example of an ideal energy consumption rate. We consider the surveillance or target tracking systems, which are likely to be deployed in a harsh environment and to operate in an unmanned manner. Under such a circumstance, it is economically infeasible to replace or recharge batteries of energy-depleted sensor nodes. Also, when we do not receive sensory data from the specific nodes, we cannot tell whether it is due to the occurrence of the sensing hole or no event to be reported has happened. This may be a critical problem particularly in surveillance systems because some important events are not conveyed to the sink and yet the operator cannot figure out whether the sensor node in charge is dead or not. Making each node survive for longer than a required network lifetime can be a good solution since it prevents the network from partitioning or suffering the sensing hole problem during the network s lifetime. Here, we assume there is an application requirement for the network lifetime, which is referred to as the pre-configured network lifetime. In many applications, because of the restriction of the environments, guaranteeing the network lifetime is necessary to provide the WSN service. Fig. 4 shows an ideal energy consumption rate, b/a, where a and b denote the pre-configured network lifetime and the amount of initially supplied energy, respectively. If a node consumes energy at this rate, this node will die right after the end of the pre-configured network lifetime. Then, the lifetime of the network is predictable, so that the operators can determine when to re-deploy new sensor nodes. However, the traffic load is different among nodes and varies over time. So, fixing the same energy consumption rate for each node is not possible. Therefore, we choose to adjust the duty cycles of sensor nodes dynamically, which can make a sensor node consume its energy approximately at the ideal consumption rate. Another advantage of our design is that surplus energy that would remain after the pre-configured network lifetime can be exploited to reduce latency. That is, if a sensor node has more energy than to survive for the pre-configured lifetime, end-to-end latency can be reduced by increasing its duty cycle. Our approach is also applicable to heterogeneous networks [17], where the sensor nodes have different initial energy levels or different types of missions, e.g., sensing versus actuating. In such networks, nodes are likely to consume energy at different rates, even though the events occur with the same probabilities. By adjusting the duty cycle dynamically, A-MAC is capable of adapting to such networks Protocol overview We primarily focus on how to ensure the configured lifetime determined a priori. To this end, we adopt a strategy where each node consumes energy approximately constantly over time. Since we also aim at reducing end-toend latency while maintaining the pre-configured lifetime, we employ a duty cycle adaptation mechanism. Our proposed protocol is based on the S-MAC s periodic listen/ sleep schedule; however, each sensor node dynamically adjusts its duty cycle depending on its energy consumption rate; while in S-MAC, the duty cycle is fixed for each node once it is configured. Thus, in A-MAC, the node with relatively higher remaining energy wakes up more frequently and relays more traffic. This way, energy consumption among all the nodes will be almost evenly distributed over time. In designing A-MAC protocol, we assume the network is densely deployed and the sensing events occur in a low frequency. As mentioned earlier, A-MAC basically employs a periodic listen/sleep mechanism. An example of the duty cycles of three nodes is illustrated in Fig. 5. As shown in Fig. 6, a superframe comprises a listen and a sleep periods, and the listen period is composed of SYNC and RTS/CTS time slots. The SYNC and RTS/CTS slots are used to exchange the listen/sleep schedule and to indicate a data message to deliver, respectively, as in S-MAC. The length of the listen period is fixed in A-MAC, while the length of the sleep period varies depending on duty cycle adaptation. During the

6 Y. Nam et al. / Computer Communications 3 (27) listen period, the SYNC information and RTS/CTS packets are exchanged. When the RTS/CTS messages are successfully exchanged, both the sender and the receiver should remain active at the sleep period and send/receive data Self-organization A node (a synchronizer) initially determines its own listen/sleep schedule and periodically broadcasts it in the SYNC message. The SYNC message basically has three fields: the source address, the next wake-up time and the listen/sleep schedule (duty cycle). The source address is the address of the node sending the SYNC message. The next wake-up time and the sleep schedule fields are announced to inform when the node will be active again and how often the listen period will be, respectively. If a neighbor node hears a schedule from the node, it adopts the received schedule as its own and re-broadcasts it, like S-MAC. In this phase, nodes do not enter into sleep state for the quick propagation of the SYNC message. After this self-organization process is finished, the network starts its operation Network operation In the network operation phase, nodes dynamically change their duty cycles depending on the remaining energy and the remaining time until the required network lifetime expires. The detailed mechanism of duty cycle adaptation will be given later. By receiving the SYNC packets of the neighbors, each node keeps track of all of the one-hop neighbors schedules. Each node wakes up, exchanges RTS/CTS and sends data at the next-hop s listen period when a packet should be transmitted to that node. Since each node dynamically changes its own schedule depending on its energy consumption rate, a node should maintain every one-hop neighbor s schedule. Note that the listen/sleep schedule is changed on a much longer time scale than the duration of the superframe. Fig. 7 illustrates a data transmission from the source node (A) to the sink (E). The percentage value next to the node ID indicates the duty cycle. Each node without data to send wakes up only at the scheduled listen period. On the other hand, if a node has a packet to send, it also wakes up at the nexthop s listen period, and after RTS/CTS exchange, it transmits data to the next-hop node. In this manner, data is relayed to the sink. A node broadcasts its schedule when the neighbor of the lowest duty cycle wakes up. Fig. 5. Example of duty cycle distribution. In A-MAC, too frequent broadcast of the SYNC message can lower the network performance. If two or more nodes send the SYNC messages simultaneously, they may collide with each other. In such a case, the changes in the duty cycle cannot be informed to neighbors. To alleviate this problem, each node normally broadcasts the SYNC message once in every n superframes of the minimum duty cycle neighbor; we choose for n in our experiments. Also, a node randomly chooses when to send the SYNC message among n superframes. By doing so, the SYNC message storm problem can be mitigated. In addition, when a node wants to adjust its duty cycle, it can only double it or cut it in half. This is to make the sensor nodes more tolerant of the SYNC message loss. Also with the help of superframe synchronization which will be given in the next section, any two nodes are active simultaneously in the beginning of the superframe of the lower duty cycle node. Thus, when a node does not hear its neighbor s new schedule, its data transmission fails only when the neighbor decreases its duty cycle; however, the next try will succeed Synchronization Fig. 6. A-MAC superframe structure. Fig. 7. Illustration of data transmission. In addition to the fields in the original SYNC message, we need three more fields: the address of the synchronizer, the timestamp of the transmission time, and the next starting time of the minimum duty cycle schedule. The synchronizer refers to the originator of the schedule. The timestamp of SYNC transmission time is used to correct clock skews among the nodes. The next starting time of the minimum duty cycle schedule is used to synchronize the starting time of each node s superframe. When a node changes its duty cycle, it starts a new superframe in the beginning of the next superframe of the minimum duty cycle. Without superframe synchronization, two nodes may not communicate with each other because if the starting times of the two schedules are differ-

7 2538 Y. Nam et al. / Computer Communications 3 (27) ent, they cannot receive the SYNC messages from each other. Note that a node s schedule is the same as that of the synchronizer only in the self-organization phase; it will change depending on the energy consumption rate in the network operation phase. On the other hand, if the synchronizers of two nodes are equal, they will awake simultaneously at least once during the period of the minimum duty cycle Duty cycle adaptation Regarding the duty cycle adaptation, A-MAC takes into account the remaining energy level of each node and the remaining time until the required network lifetime. Intuitively, the node with less energy should sleep more in order to balance the energy consumption. Let T conf and T elap denote the pre-configured network lifetime and the elapsed time, respectively. Then, the ratio of T elap to T conf represents the ideal energy consumption rate because if a node consumes energy with that rate the node will die exactly at the end of the required lifetime. However, there will be non-uniform energy consumption among nodes due to unbalanced distribution of the traffic load. Let d in Eq. (1) denote the difference between the ideal and the currently calculated energy consumption rates. d ¼ T elap E cons : ð1þ T conf E init Here, E cons and E init denote the dissipated energy of a node so far and the initial energy, respectively. Thus, the second term of the right-hand side in Eq. (1) indicates the energy consumption rate of the node so far. If d is greater than zero, that means the node has more remaining energy than it should have to last for the required lifetime. On the other hand, if d is smaller than zero, the node cannot survive until the pre-configured lifetime with the current energy consumption rate. Thus, the node doubles its duty cycle if d is greater than zero; otherwise, it decreases its duty cycle by half. The reason for exponentially increasing/decreasing adaptation is to ease the superframe synchronization among nodes. To prohibit too frequent change of the duty cycle, we adopt two thresholds. That is, the duty cycle is doubled if d is greater than the upper threshold and is cut in half if d is smaller than the lower threshold. If d is in-between, the current duty cycle is maintained. This way, the energy consumption rate would approach the ideal energy consumption rate with hysteresis mitigated. Another significant advantage of A-MAC is that it does not require to configure the initial duty cycle carefully. Here, we assume that there is a minimum duty cycle and it can make the node survive for a pre-configured network lifetime, no matter how much the traffic load is Cross-layer approach Since A-MAC adapts its duty cycle independently of other nodes, we suggest a routing protocol running on top of A-MAC choose a path with more remaining energy rather than a path with the lowest hop count. Existing routing protocols for WSNs, e.g., [12 14], consider energy consumption when they select the routing path. We believe these routing protocols will work well with A-MAC. However, they do not take into account balanced energy distribution. In this case, the potential of A-MAC may not be highlighted. Hence, we present two energy-aware routing metrics, which can take advantage of A-MAC. These routing metrics will reflect the changes in the duty cycles, finding out a new path with more remaining energy. The two possible metrics for routing are listed below. To compare the performance of these metrics with the traditional one, i.e., the minimum hop count, we use AODV [] as the base routing protocol. We modify AODV to implement these routing metrics. Although AODV exhibits a large overhead, it is expected to clearly show the difference among the routing metrics since it is based on a simple design. Note that these routing metrics can be employed in any routing protocols or data dissemination protocols for WSNs. Max-Min metric: this routing metric seeks to find out the path whose minimum duty cycle of the nodes along the path is maximum. The duty cycle of the intermediate node with the minimum duty cycle is recorded in the RREQ/RREP messages while discovering the routes. In choosing the routing path, the path with the largest minimum duty cycle value is selected. If two routes have the same minimum duty cycle values, the route which is found with less delay is selected. This metric can distribute the load in diverse paths, considering the energy level of relay nodes. However, it can take a long detour when there exists a node with the high duty cycle but far away from the normal route, which deteriorates the delay performance as well as consumes far more energy. Max-Avg metric: the route with the largest average duty cycle is selected. The number of hops and the accumulated duty cycle values are stored in the RREQ/RREP messages and the average duty cycle is calculated at each hop. The path with the largest average duty cycle value is selected. Again, the delay is used for tie-breaking. This metric is an eclectic approach because it divides total duty cycle values of the nodes on the path by the number of hops. Therefore, this metric can mitigate the problem of the above Max-Min metric. In our approach, when to trigger a new route discovery is an important issue, since the current route is still alive. Because the route discovery process usually floods the control messages, a periodic route discovery may cause an unnecessary control overhead. Thus, we introduce a sinktriggered route discovery. That is, the sink requests a new route discovery to the sender when it detects the average duty cycle value of the received packet goes below the average duty cycle at the previous route discovery by a certain threshold. This triggering is designed to work with the

8 Max-Avg metric; the same approach can be applied to the Max-Min metric. The duty cycle value of each node on the path is accumulated and recorded in the packet header along with the number of hops. The sink keeps track of this duty cycle value, and when it falls below the average value at the previous route discovery by a certain threshold, the sink sends a route discovery trigger message to the sender. In this way, the routing path can change dynamically and the traffic can be distributed among nodes. Y. Nam et al. / Computer Communications 3 (27) Latency analysis In this section, we analyze the multi-hop delays of A- MAC and S-MAC. Let D(h) the packet transfer delay over h hops. According to [2], the average of the total delay of S- MAC over h hops is given by E½DðhÞŠ ¼ h 1 T f þ t cs þ t tx ; ð2þ 2 where T f, t cs, and t tx represent the time duration of the frame (a cycle of listen and sleep), the carrier sense delay, and the transmission delay, respectively. As mentioned in [2], T f is, in general, much larger than (t cs + t tx ). Therefore, the delay over h hops is almost proportional to T f. From the definition of the duty cycle, the frame duration, T f,is inversely proportional to the duty cycle. (Recall that the listen interval is fixed in both S-MAC and A-MAC.) Then, we have E½DðhÞŠ / T f / 1 duty cycle : To figure out the energy consumption rate of each node, let us assume the target network is formed in a dense and uniform manner, as shown in Fig. 8. For simplicity, we assume the sink is located in the center, and we draw concentric circles, each of which delineates the number of hops to the sink. Let H denote the maximum possible number of hops of the network. Let q be the node density of the sensor network, i.e., the number of nodes per square of hop counts from the sink. Then the total number of nodes in the network is qh 2 p. As we assume the uniformly distributed network, this hop scale is almost proportional to the transmission range. Here, the number of nodes h-hops apart from the sink, N(h), is expressed by NðhÞ ¼qðh 2 ðh 1Þ 2 Þp ¼ qð2h 1Þp: Now, let us assume the occurrence rate of the sensing event to be reported is very low and equal for all nodes, and the event occurrence rate is denoted as k. In our scenario, every sensing event should be delivered to the sink; therefore, the total number of event messages that all h- hop nodes receive in a time unit, Rx(h), is the number of the messages generated at nodes whose hop count to the sink is higher than h. ð3þ ð4þ RxðhÞ ¼ XH x¼hþ1 qkpð2x 1Þ ¼qkpðH 2 h 2 Þ: In the same way, the total number of the event messages that all h-hop nodes transmit in a time unit, Tx(h), is given by TxðhÞ ¼ XH x¼h Fig. 8. Example topology. qkpð2x 1Þ ¼qkpðH 2 ðh 1Þ 2 Þ: Energy consumptions due to sensing and processing are relatively much smaller than those of transmitting and receiving packets. Thus, the main sources of energy consumption are transmission and reception of packets, as well as idle listening. If the duty cycle is ideally configured with the finest granularity, we can disregard energy consumption due to idle listening. That is, the wakeup period is spent only for transmissions and receptions. Then, the energy consumption rate for each node h hops from the sink, E(h), is calculated as EðhÞ ¼ E totðhþ NðhÞ ¼ qkpððh 2 h 2 ÞE rx þðh 2 ðh 1Þ 2 ÞE tx Þ ð7þ qð2h 1Þp ¼ kððh 2 h 2 ÞE rx þðh 2 ðh 1Þ 2 ÞE tx Þ ; 2h 1 where E tot, E tx,ande rx denote total energy consumption, energy consumption due to transmission and reception, respectively. In S-MAC, every node has the same duty cycle. Thus, the network lifetime of S-MAC is dependent on the energy consumption of the nodes 1-hop apart from the sink, because from Eq. (7), 1-hop nodes have the largest energy consumption rate. If both S-MAC and A-MAC have the same network lifetime, the duty cycle of each node in S- MAC should be configured to that of the node 1-hop apart from the sink. The rest of the nodes other than 1-hop nodes will have less load but have the same duty cycles, so that ð5þ ð6þ

9 254 Y. Nam et al. / Computer Communications 3 (27) the remaining energy will be wasted in idle listening. By defining as the ratio of E rx to E tx, we express the delay of S-MAC over H hops as E½D S-MAC ðhþš ¼ XH h¼1 kððh 2 1ÞE rx þ H 2 E tx Þ ¼ kððh 2 1ÞE tx þ H 2 E tx ÞH ð1þþke tx H 3 : In A-MAC, from Eq. (3), the delay is inversely proportional to the duty cycle and the duty cycle is also inversely proportional to the energy consumption rate. Then we calculate the delay of A-MAC over H hops as E½D A-MAC ðhþš ¼ XH h¼1 kððh 2 h 2 ÞE rx þðh 2 ðh 1Þ 2 ÞE tx Þ 2h ð1 þ ÞkE txh 2 logh: From Eqs. (8) and (9), the orders of growth of S- MAC s and A-MAC s delays are O(H 3 ) and O(H 2 logh), respectively. This shows the advantage of A-MAC over S-MAC in terms of scalability. We disregard the delay due to queueing and collision in our analysis because we assume the events happen in a low frequency. Therefore, when the event occurrence rate increases, the difference between the delays of the two MAC protocols may be reduced due to packet collision and the queueing delay. 6. Performance evaluation We carry out simulation experiments with comprehensive network conditions to compare A-MAC with S- MAC. We do not compare the results with recent MAC protocols including T-MAC, because a majority of them are tailored to make the node with more traffic wakeup more, which incurs even shorter network lifetime than S- MAC. Time to first node s death and end-to-end latency are chosen as key performance metrics. The simulation study is conducted using NS-2 [18]. By default, we use 25-node topology, forming a 5 5 grid. The distance between adjacent nodes is 2 m and we do not assume mobility of nodes. Other simulation parameters are similar to those in [2], as shown in Table 1. In terms of energy consumption, we adopt the energy model in [11]. We set the required network lifetime of the application at 4 time units by default. We simulate both the single-flow and multiple-flow cases, varying the packet inter-arrival time, and compare A-MAC with S-MAC with diverse duty cycles. During the simulations, the Max-Avg routing metric is used by default Single-flow network In the single-flow network, the source node is located in the top-left corner and the sink in the bottom-right, which ð8þ ð9þ Table 1 Simulation parameters Radio transmission range 3 m Radio interference range 6 m Radio bandwidth 2 kbps Data packet size 5 bytes Duration of listen interval 115 ms Contention window for SYNC 15 slots Contention window for data 31 slots Minimum duty cycle 1% Maximum duty cycle % Difference upperbound.1 Difference lowerbound Transmit power.66 W Receive power.395 W Idle power.35 W Sleep power.1 W Initial energy 3 J implies at least four hops from source to sink. Fig. 9a exhibits time to first node s death versus the packet interarrival time. A-MAC prevents network partitioning or sensing hole s occurrence until the pre-configured lifetime, 4, regardless of the traffic load. Furthermore, the first node dies shortly after the configured lifetime. On the other hand, in S-MAC, time to first node s death relies on both the traffic load and the duty cycle. Only S-MAC with 2% duty cycle maintains time to first node s death above 4. End-to-end latency under the same condition is shown in Fig. 9b. S-MAC with the high duty cycle (6%) shows lower delay than A-MAC overall; however, it does not satisfy the pre-configured network lifetime by a large margin. Likewise, S-MAC with the low duty cycle (2%) satisfies the pre-configured lifetime, but it shows much higher latency than A-MAC. In medium and low traffic loads, end-to-end latency of S-MAC is mostly dependent on the duty cycle, not on the traffic load, whereas in A-MAC, latency diminishes as the traffic load decreases, because the residual energy is utilized to reduce the delay. Fig. b plots the number of the remaining nodes as time goes on, where packets are transmitted every five seconds. A-MAC shows the steepest curve of decreasing remaining nodes, which implies almost every node dies around the end of the pre-configured network lifetime. It also signifies A-MAC has the fairest distribution of energy consumption as shown in Fig. a Multiple-flow network In the multiple-flow case, every node except the sink periodically transmits packets to the sink, which reflects the monitoring application scenario. Every node in the network transmits sensory data to the sink. Fig. 11a plots time to first node s death versus the packet inter-arrival time. A- MAC shows stable results of time to first node s death slightly above the pre-configured lifetime, while the performance of S-MAC depends on the traffic load, and in most

10 Y. Nam et al. / Computer Communications 3 (27) A-MAC 7 6 A-MAC Time to first node s dying (timeunit) End-to-end Latency (sec) Packet inter-arrival time (sec) (a)time to first node s death Packet inter-arrival time (sec) (b) End-to-end latency Fig. 9. Performance comparison w.r.t. packet inter-arrival time (single-flow) A-MAC 35 3 A-MAC Number of remaining nodes Standard deviation Time (time unit) (a) Number of remaining nodes Time (time unit) (b) Standard deviation Fig.. Performance comparison w.r.t. time (A-MAC/S-MAC). (a) Number of remaining nodes. (b) Standard deviation. cases, nodes start dying much earlier than the pre-configured lifetime. Each curve corresponds to the one in Fig. 11b, which shows the end-to-end delay of A-MAC and S-MAC. S-MAC with the 4% duty cycle satisfies the required network lifetime only with the light traffic load. In those cases, however, A-MAC shows lower delay than S-MAC. Also, A-MAC outperforms S-MAC with 2% duty cycle in terms of the end-to-end delay. Under the same condition, we also evaluate the performance of A-MAC, varying the pre-configured network lifetime from 3 time units to 5 time units. The results are shown in Figs. 12a and b. In both graphs, the number next to A-MAC in the legend stands for the configured network lifetime. In Fig. 12a, A-MAC always satisfies each configured network lifetime. In Fig. 12b, A-MAC shows varying delay performance as the pre-configured network lifetime changes. When the configured network lifetime is short, i.e., 3 time units, A-MAC almost always shows similar end-to-end delay values to S-MAC with 6% duty cycle, whereas with the long lifetime, i.e., 5 time units, A-MAC shows slightly higher delay values than S-MAC with 2% duty cycle when the traffic load is high. This is because the main goal of A-MAC is to guarantee the network lifetime and hence the delay can only be reduced with the surplus energy Performance versus node density We evaluate the performance of A-MAC and S-MAC with regard to the node density. Simulation parameters are the same as those of the multiple-flow case and we fix the packet inter-arrival time at 2 s. We vary the distance between adjacent nodes from to 3 m, each of which represents the dense and the sparse network. S-MAC

11 2542 Y. Nam et al. / Computer Communications 3 (27) Time to first node s dying (timeunit) A-MAC End-to-end Latency (sec) A-MAC Packet inter-arrival time (sec) (a)timet of first node s death Packet inter-arrival time (sec) (b) End-to-end latency Fig. 11. Performance comparison w.r.t. packet inter-arrival time (multiple-flows). Time to first node s dying (timeunit) A-MAC (3) A-MAC (4) A-MAC (5) End-to-end Latency (sec) A-MAC (3) A-MAC (4) A-MAC (5) Packet inter-arrival time (sec) (a) Time to first node s death Packet inter-arrival time (sec) (b) End-to-end latency Fig. 12. Performance comparison w.r.t. packet inter-arrival time and pre-configured network lifetime (multiple-flows). shows the increasing network lifetime and decreasing endto-end delay as the node density decreases. This is because the probability of the packet collision diminishes as the node density decreases. A-MAC shows the similar endto-end latency values but still maintains the pre-configured network lifetime Fig Performance versus routing metrics In order to evaluate the performance of A-MAC with regard to routing metric, we use AODV with three routing metrics: minimum hop count (Min-Hop), Max-Min metric, and Max-Avg metric. The simulation is conducted with the single-flow network with five seconds of packet inter-arrival time. Figs. 14a,b shows time to first node s death and end-to-end latency with respect to the routing metrics. All the metrics achieve network partition time over the pre-configured network lifetime, 4. However, the metric, Max-Avg, showed lowest latency than other two routing metrics as expected Latency versus hop counts To evaluate the delay performance of A-MAC and S- MAC, we perform simulations under the linear topology as in Fig. 15. We vary the number of sensor nodes, which is equivalent to the number of hops, from one to ten; we do not count the sink as the sensor node. The sink is located at the edge, and each node periodically, i.e., every seconds, sends messages to the sink. To find out the optimal duty cycle of S-MAC, we vary the duty cycle from % to 5% depending on the number of nodes; we chose

12 Y. Nam et al. / Computer Communications 3 (27) Time to first node s dying (timeunit) A-MAC (3) A-MAC (4) A-MAC (5) End-to-end Latency (sec) A-MAC (3) A-MAC (4) A-MAC (5) Distance between adjacent nodes (m) (a) Time to first node s death 2 3 Distance between adjacent nodes (m) (b) End-to-end latency Fig. 13. Performance comparison w.r.t. node density. Time to first node s dying (timeunit) End-to-end Latency (sec) MIN-HOP MAX-MIN MAX-AVG Routing metrics (a)time to first node s death MIN-HOP MAX-MIN MAX-AVG Routing metrics (b) End-to-end latency Fig. 14. Performance comparison w.r.t. routing metrics. the maximum duty cycle value that satisfies the network lifetime for each plotting. Fig. 16 plots the average delay of A-MAC and S-MAC with regard to the number of hops. The curve of S-MAC shows a different slope from that of [2] because every node transmits packets to the sink in our scenario; whereas there is only one source node in [2]. The delay increases rapidly in S-MAC as the number of hops increases, while A-MAC yields a much slower curve. This gap approximately corresponds to our analysis result in Section 5. A-MAC and S-MAC show almost similar delay performance when the number of hops is below four. Moreover, A-MAC shows slightly higher latency than S-MAC in one Fig. 15. Example linear topology. and two hop cases. This is because we manually configured the optimum duty cycle for S-MAC, while the duty cycle adaptation of A-MAC could not find the optimum duty cycle at once; it adaptively makes the duty cycle approach to the optimum value by doubling or halving it long term. Consequently, in a very small network, the latency of A- MAC may not be as small as that of S-MAC. However, in most cases, the latency of A-MAC is not only less than S-MAC, but also grows more slowly than S-MAC as the network size grows. 7. Conclusions Energy-efficient MAC protocols reduce unnecessary energy consumption, thus prolong the network lifetime of the WSNs. However, switching the nodes into periodic sleep state does not guarantee that a network will survive for a pre-configured network lifetime. Moreover, energy-

13 2544 Y. Nam et al. / Computer Communications 3 (27) End-to-end Latency (sec) efficiency is achieved at the cost of increased end-to-end latency. We discussed the existing CSMA-based MAC protocols in this perspective. Making a sensor network survive for a pre-configured network lifetime is a viable issue in terms of network connectivity and the quality of surveillance. With the objective of satisfying a given network lifetime, we propose an adaptive MAC (A-MAC) protocol that guarantees the network lifetime and reduces end-toend latency. Analysis and simulation results reveal that A-MAC meets the network lifetime requirements while showing substantially less delay than S-MAC. Acknowledgements This research is supported by the Ubiquitous Autonomic Computing and Network Project, the Ministry of Information and Communication (MIC) 21st Century Frontier R&D Program in Korea. References A-MAC S-MAC number of hops Fig. 16. Latency comparison w.r.t. hop counts. [1] IEEE Computer Society : Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications, June [2] W. Ye, J. Heidemann, D. Estrin, Medium access control with coordinated adaptive sleeping for wireless sensor networks, in: IEEE/ ACM Transactions on Networking, vol. 12, No. 3, June 24. [3] J. Pollastre, J. Hill, D. Culler, Versatile low power media access for wireless sensor networks, in: Proc. ACM SenSys, 24. [4] T. Zheng, S. Radhakrishnan, V. Sarangan, PMAC: an adaptive energy-efficient mac protocol for wireless sensor networks, in: Proc. IEEE IPDPS, 25. [5] T. van Dam, K. Lamgemdoen, An adaptive energy-efficient mac protocol for wireless sensor networks, in: Proc. ACM SenSys, 23. [6] G. Lu, B. Krishnamachari, C.S. Raghavendra, An adaptive energyefficient and low-latency mac for data gathering in wireless sensor networks, in: Proc. IEEE IPDPS, 24. [7] K. Balacandran, J.H. Kang, W.C. Lau, Adaptive sleeping and awakening protocol (ASAP) for energy efficient adhoc sensor networks, in: Proc. IEEE ICC, 25. [8] S. Singh, M. Woo, C.S. Raghavendra, Power aware routing in mobile ad hoc networks, in: Proc. ACM MobiCom, [9] C. Gui, P. Mohapatra, Power conservation and quality of surveillance in target tracking sensor networks, in: Proc. ACM MobiCom, 24. [] C.E. Perkins, E.M. Royer, Ad hoc on-demand distance vector routing, in: Proc. IEEE Workshop on Mobile Computing Systems and Applications, [11] M. Stemm, R.H. Katz, Measuring and reducing energy consumption of network interfaces in hand-held devices, in: IEICE Transactions on Communications, vol. E8-B, No. 8, Aug [12] W.R. Heinzelman, A. Chandrakasan, H. Balakrishnan, Energy-efficient communication protocol for wireless microsensor networks, in: Proc. Hawaii International Conference on System Sciences, Jan. 2. [13] C. Schurgers, M.B. Srivastava, Energy efficient routing in wireless sensor networks, in: Proc. IEEE MILCOM 21. [14] C. Intanagonwiwat, R. Govindan, D. Estrin, J. Heidemann, F. Silva, Directed diffusion for wireless sensor networking, in: IEEE/ACM Transactions on Networking, vol. 11, No. 1, 23. [15] F. Ye, G. Zhong, J. Cheng, S. Lu, L. Zhang, PEAS: a robust energy conserving protocol for long-lived sensor networks, in: Proc. ICDCS, 23. [16] A. Cerpa, D. Estrin, ASCENT: adaptive self-configuring sensor networks topologies, in: IEEE Transactions on Mobile Computing, vol. 3, No. 3, July 24. [17] E.J. Duarte-Melo, M. Liu, Analysis of energy consumption and lifetime of heterogeneous wireless sensor networks, in: Proc. IEEE GLOBECOM, 22. [18] Network Simulator, [Online] Yongsub Nam. He received B.S. in computer science and engineering from Seoul National University in 26, and is currently a master student in Carnegie Mellon University. Since 26, he has been a recipient of CyLab Korea Fellowship and Information Networking Institute Scholarship of Carnegie Mellon University. Prior to joining Carnegie Mellon University, he had also worked at the network industry for five years. His research interests include wireless sensor networks, mobile ad hoc networks, and vehicular ad hoc networks. Taekyoung Kwon. He is an assistant professor in Multimedia and Mobile Communications Laboratory, School of Computer Science and Engineering, Seoul National University. He received his Ph.D., M.S., and B.S. degrees in computer engineering from Seoul National University in 2, 1995, and 1993, respectively. He was a visiting student at IBM T. J. Watson Research Center in 1998 and a visiting scholar at the University of North Texas in His recent research areas include radio resource management, wireless technology convergence, mobility management, and wireless sensor networks. Hojin Lee. He received a B.S. degree in computer science and engineering from Seoul National University in 23. He is currently working toward a Ph.D. degree at Multimedia and Mobile Communication Lab., School of Computer Science and Engineering, Seoul National University. He was a visiting researcher at Network Modeling Lab., Simon Fraser University, Canada. His research interests include routing protocols in VANETs, wireless broadband access networks, and energy-efficient MAC protocols in WSNs.

14 Y. Nam et al. / Computer Communications 3 (27) Hakyung Jung. He received a B.S. degree in the School of Computer Science and Engineering at the Seoul National University, Korea, in 25. He is now a M.S.-Ph.D. student in the Multimedia and Communications Lab., School of Computer Science and Engineering, Seoul National University. He was a visiting researcher at Advanced Wireless Networking Lab., City University of New York, USA. His research interest lies in the field of various wireless access networks. Yanghee Choi. He received B.S. in electronics engineering from Seoul National University, M.S. in electrical engineering from Korea Advanced Institute of Science, and Doctor of Engineering in Computer Science from Ecole Nationale Superieure des Telecommunications (ENST) in Paris, in 1975, 1977 and 1984, respectively. Before joining the School of Computer Engineering, Seoul National University in 1991, he has been with Electronics and Telecommunications Research Institute (ETRI) during , where he served as director of Data Communication Section, and Protocol Engineering Center. He was research student at Centre National d Etude des Telecommunications (CNET), Issy-les- Moulineaux, during He was also Visiting Scientist to IBM T.J. Watson Research Center for the year He is now leading the Multimedia and Mobile Communications Laboratory in Seoul National University. He is also director of Computer Network Research Center in Institute of Computer Technology (ICT). He is vice-president of Korea Information Science Society. He was editor-in-chief of KISS journals and also chairman of the Special Interest Group on Information Networking. He has been associate dean of research affairs at Seoul National University. He was president of Open Systems and Internet Association of Korea. His research interest lies in the field of multimedia systems and high-speed networking.