P-MAC is proposed for WSNs deployed for rare events detection with prompt reporting. Applications including fire or other hazards detection fall into this category. Such a network consists of many sensor nodes randomly deployed in a sensing area with one (or a few) sink(s) collecting information for an outside system. All nodes are assumed to be homogeneous with the same transmission range.

Before starting data transmission, each node need to find a grade it belongs to, and choose a schedule for periodically listening and sleeping. This is implemented by operating the GDSA mechanism at the network layer of each node.

GDSA is initiated from the sink. In GDSA, each node maintains grade information (denoted by G_n) with an initial value of −1, except that the sink’s grade is zero (G_n = 0) at all times. The sink first chooses a schedule according to its grade (the schedule choosing rule will be introduced in detail in Section 3.5). Then it generates a GRADE message containing a field denoted by G_m. After setting G_m to one, the sink broadcasts this message. A node receiving a GRADE message with G_m = i sets its grade to i (G_n = i), chooses a schedule corresponding with its grade, and rebroadcasts the message after increasing G_m by one, unless it has already joined an equal or a lower grade. The pseudo-code of the algorithm used by a node for processing the received GRADE message is shown in Algorithm 1.

After the grade division using the above scheme, the whole network is divided into several annular grades similar to concentric circles with the center at the sink, as shown in Figure 1(a). But this is not absolute. For example, if node A and A′ do not exist, the network may be divided into the formation shown in Figure 1(b).

As stated above, each node simultaneously establishes a periodical sleep/wakeup schedule according to its grade information during the grade division process. Those nodes in the same grade keep the same schedule, but schedules are staggered between two adjacent grade nodes. Suppose a node is in grade i (i > 0), it repeatedly experiences three periods: receiving data from the (i+1)th grade node (the RECEIVE DATA period), sending data to the (i-1)th grade node (the SEND DATA period), and sleeping (the SLEEP period). If the upper grade nodes are in the SEND DATA period, their adjacent lower grade nodes must be in the RECEIVE DATA period (the duration of each period will be analyzed in Section 3.4). Thereby, data frames can be forwarded continuously in a pipeline way from the source node to the sink, thus making the delivery latency acceptable. Those adjacent nodes in the same grade contend with each other for the shared medium while those in different grades cooperate with each other for data transmission.

The pipelining concept cited in this paper is defined as follows:

Definition 1 (Pipelining): In duty-cycling MAC protocols, a node generally can complete receiving a data frame from one of its upstream nodes and then sending this data frame to one of its downstream nodes within one cycle. This is called pipeline data transmission.

After completing GDSA, a sensor node with pending data will not be aware of any concrete routing paths to the sink, but it can use a variation of the RTS/CTS handshake mechanism at the MAC layer to determine the next-hop node from its adjacent lower grade nodes. There are two differences between the variation and the original RTS/CTS handshake in IEEE 802.11. First, in P-MAC, the RTS sent by a source node contains the node’s grade information passed down from the network layer instead of a concrete next-hop address. All nodes in the adjacent lower grade can reply to the source node with CTS if they receive the RTS. Second, the Contention Window (CW) is also used when a node replies with CTS in P-MAC, because several lower grade nodes will simultaneously receive the RTS, and there will be contention for data relaying among these nodes.

Consider the network shown in Figure 1(a). When nodes in grade 3 are in the SEND DATA period, those nodes in grade 2 are in the RECEIVE DATA period. Suppose node S has data to send to the sink. S first broadcasts the RTS frame in its SEND DATA period after contending with its neighboring nodes in the same grade (e.g., S′) and winning the medium. Both node A and A′ can receive this RTS in their RECEIVE DATA periods. They contend with each other for replying with CTS. If node A’s CTS is first received by node S, S will send its data to A. After A receives the data, it sends an ACK frame to S. Then A waits to enter its SEND DATA period. After receiving the ACK, S enters sleep mode. Those nodes (e.g., S′ and A′) that failed in the channel contention will go to sleep and wait to enter their subsequent periods. In the same manner, node A finds its next relaying node B and sends data to B in its SEND DATA period. Figure 2 shows this data transmission process.

Note that the Network Allocation Vector (NAV) in IEEE 802.11-style MAC protocols for virtual carrier sense is not used in P-MAC, because each node in P-MAC only receives data during its RECEIVE DATA period and only sends data during its SEND DATA period, respectively. If a node fails in the contention for receiving/sending data or it has waited a long enough time without receiving RTS/CTS, the node will go to sleep. Thus, the virtual carrier sense used for collision avoidance is not necessary in P-MAC.

An analysis on the duration of each period in the P-MAC cycle follows:

(1) SEND/RECEIVE DATA period: As shown in Figure 2, the duration of the SEND DATA period in node S is identical to the duration of the RECEIVE DATA period in node A, if node S and A are involved in the current data transmission. The maximum time of the SEND/RECEIVE DATA period, T_S/R, is calculated as follows: (1) T S / R = 2 CM + 2 DIFS + 2 SIFS + durRTS + durCTS + durDATA + durACK ,where durRTS, durCTS, durDATA and durACK are the transmission duration of RTS, CTS, DATA and ACK, respectively.

(2) SLEEP period: In P-MAC, the duration of the SLEEP period is determined by the following equation: (2) T SLEEP = sleep _ factor · T S / R ,where sleep_factor is a positive integer. The whole cycle duration is given by: (3) T cycle = ( sleep _ factor + 2 ) · T S / R = τ · T S / R ,where τ is called cycle coefficient. Next we discuss how to determine the value of sleep_factor.

Figure 3 shows an example used for analysis. For simplicity, the RTS/CTS handshake process is not shown. There are four nodes, where node 0’s grade is zero (it is the sink node), node 1’s grade is one, and so on. Node 3 has data to send to node 0. T_d is the actual duration time for a node to finish its one-hop data transmission process. Note that T_d ≤ T_S/R, since the node may not select the last slot in CW. Thus the node’s actual sleep time is ΔT + T_SLEEP, where ΔT = T_S/R – T_d. Anyway, each node can sleep for at least T_SLEEP time in each cycle.

If sleep_factor is a positive integer, for example, sleep_factor = 1, when the ith (i > 0) grade nodes wake up, the (i – 1)th grade nodes are still in the SLEEP period, and the (i + 2)th grade nodes will go to sleep for at least sleep_factor · T_S/R = T_S/R time. Therefore, the communications between the ith grade nodes and the (i + 1)th grade nodes will not be interfered with by the (i – 1)th and (i + 2)th grade nodes. Considering that the interference range is about two times the transmission range, the value of sleep_factor needs to be at least 2. Figure 3 shows the case of sleep_factor = 2.

Suppose the time needed to finish the GDSA process is within T_g time. The transmission duration of a GRADE message is denoted by durGRADE. Once a node receives a GRADE message and updates its grade to i, it uses the following rules to choose an initial schedule:

if i % τ= 0, the node will enter the RECEIVE DATA period after (T_g – i · durGRADE) time.

Similar to S-MAC [3] and RMAC [5], synchronization is also required among neighboring nodes in P-MAC to solve the problem of clock drift as time advances. However, the difference is that P-MAC doesn’t use an extra period (e.g., SYNC) to achieve this goal. In P-MAC, each frame contains the relative start time of the current period. When a sensor node receives a frame, it will adjust its schedule if the clock drift is too serious. If a node fails to receive correct data for a long enough time, it will stay active and listen to the channel for some time to revise its grade and/or schedule. This loose synchronization mechanism is used because in P-MAC, a node with pending data to send may have several adjacent lower grade nodes, any of which may become the relay node. We suppose this mechanism will guarantee reliability for data forwarding.

P-MAC is a cross-layer duty cycle MAC protocol seamlessly integrated with routing function. It not only reduces the protocol overhead, but also helps enhance scalability when there are changes in network topology:

If a new node is added to the network, it will identify which grade it should join and the corresponding schedule it should choose after listening to its neighbors for some time. Subsequently, it can join the network for data transmission.

No further elaboration on these issues is presented in this paper, but we believe these issues are reasonable and provide a guideline for future exploration.

In this section, we evaluate the P-MAC design in comparison with S-MAC and RMAC using the OPNET modeler. For fairness, we give the simulation result of both the basic P-MAC without the routing function and the full P-MAC that is seamlessly integrated with routing. In comparison with the full P-MAC, the difference is that in the basic P-MAC, the RTS is sent to a concrete next node, which has its address passed down by the networking layer. So only the node for which the RTS is destined will reply with CTS without waiting for CW time. Figure 5 illustrates the basic P-MAC, and the maximum time of the SEND/RECEIVE DATA period, T_S/R, is shown below. This result differs from that calculated by Equation (1): (5) T S / R = CW + DIFS + 3 SIFS + durRTS + durCTS + durDATA + durACK .

Each node uses the two-ray ground radio propagation model and has a single omni-directional antenna. Table 1 shows the key networking parameters used in our simulation. These parameters are the default settings in the standard S-MAC simulation module distributed with the ns-2.29 package. The sizes and transmission latencies of different types of packets are shown in Table 2, and the settings are the same as in [5].

PION relaying number N in RMAC defines the distance (hops) a PION frame can be forwarded in the DATA period. In our simulation, we set N = 4, as in [5]. The cycle related parameters are shown in Table 3. Both S-MAC and RMAC keep the same duty cycle (about 6%). Because the cycle division in P-MAC is different from that of both S-MAC and RMAC, the full P-MAC is set with sleep_factor = 14 to have the same cycle duration as RMAC and the basic P-MAC is set with sleep_factor = 21 to have a similar cycle duration to the full P-MAC.

We use two types of topologies in our simulations: chain topology and random topology. Figure 6 shows the chain topology. All nodes are equally spaced in a straight line with a 200 meter interval between neighboring nodes. Node 0 sends packets to node n through a single CBR (constant bit rate) flow. The hop length of the chains varies from 1–24 hops. For S-MAC, RMAC and the basic P-MAC, the routes for data transmission are assigned manually.

Figure 7 shows an example of the random topology, which consists of 200 sensor nodes and a sink node (not shown in the figure). The 200 sensor nodes are randomly distributed in a 2,000 × 2,000 m² square area, and the sink node is located at the top-left corner of the square.

The full P-MAC integrated with routing needn’t consider the routing issue. For fairness, we propose a routing mechanism for S-MAC, RMAC and the basic P-MAC by modifying the GDSA process in P-MAC, since it is not convenient to assign routes manually in the random topology. In the modified GDSA process, each node maintains a routing table containing only one field to record its next-hop nodes’ IDs. The process has the following steps:

The sink node with its grade set to zero initiates this process by generating a GRADE message packet. The packet contains two fields denoted by G_m and ID_ph. After setting ID_ph to its own ID and G_m to one, the sink node broadcasts this message.

The completion of this process will result in one or more entries in the routing table of each node. During data transmission, the current node randomly chooses one entry and extracts the ID from the entry as the next-hop ID. Also, each node maintains a grade representing the path length from it to the sink. Figure 8 shows the number of nodes in each grade.

In addition, we make the following assumptions. For S-MAC and RMAC, all nodes have already been synchronized to use a single schedule. There is no synchronization traffic during the simulation, but the SYNC period is still contained in the cycle. The original or modified GDSA process is executed only once during the initial phase and the corresponding overhead is not considered during the comparison among S-MAC, RMAC and P-MAC.

We first evaluate the performance of the end-to-end delivery latency. For the chain topology, node 0 generates a CBR flow at the rate of 1 packet every 10 seconds. The simulation time is set as 1,200 seconds. For the random topology, a random sensor node is selected every 10 seconds to send one packet to the sink. The simulation time is set as 19,200 seconds.

Figure 9 shows the growth trend of the average packet delivery latency with respect to the increase in path length in the chain topology and random topology, respectively. Both figures show that the delivery latency in P-MAC increases at a much lower rate than S-MAC and RMAC. This is because S-MAC only forwards a packet to a 1-hop distance in each operational cycle and RAMC has to wait for the start of the next DATA period to forward data again if the path length exceeds the PION relaying number, while P-MAC can forward data in a pipeline fashion.

Provided the path length does not exceed the PION relaying number, data transmission can be completed within a single cycle in RMAC, spending about T_DATA time. For P-MAC, the data frame moves to a 1-hop distance per SEND DATA period, spending about T_S/R time. In our simulation, T_DATA (168.0 ms) is 66.0 ms less than T_S/R (234.0 ms), so Figure 9 shows that RMAC has slightly less delivery latency when the path length is within 4 hops, as N = 4 in our simulation.

Figure 9 also shows that the basic P-MAC has lower delivery latency than the full P-MAC with routing, since the RTS is sent to a certain node during data transmission in the basic P-MAC, which reduces network congestion and improves the data delivery ratio.

Next, we evaluate the network throughput and energy efficiency for P-MAC. The network throughput is recorded in terms of the average number of packets successfully received by the sink per second. For the chain topology, we keep n = 24, and the data input interval for node 0 varies from 10–1 seconds. For the random topology, the interval for randomly selecting a sensor node to send data varies from 10-1 seconds. In both topologies, the simulation time is set as 1,200 seconds.

In the chain topology, Figure 10(a) shows that when the input interval is less than 8 seconds, the output rate in S-MAC decreases rapidly, since S-MAC has poor traffic contention handling due to its weakness of forwarding a packet to a 1-hop distance in each operational cycle. Because the network throughput of S-MAC decreases when the input interval is less than 8 seconds, the corresponding energy consumption also decreases, as shown in Figure 10(b).

For P-MAC and RMAC, Figure 10(a) shows that the output rate follows the input rate until the input interval is less than 5 seconds, and finally reaches the steady state. When the network throughput reaches its peak point, the incoming injected packets cannot continue to be sent, and the network energy consumption will not be increased, as shown in Figure 10(b). When the input interval is smaller than 4 seconds, P-MAC with routing will have higher energy consumption than RMAC because of its higher throughput. The basic P-MAC has almost the same throughput as the full P-AMC but has lower energy consumption, since it doesn’t consider the routing issue.

In the random topology, Figure 11(a) shows that RMAC has better performance than the P-MAC integrated with routing in terms of throughput. But the basic P-MAC has better throughput than RMAC when the input interval is less than 4 seconds. Figure 11(b) shows that P-MAC is more energy efficient than RMAC and S-MAC in the random topology. For all protocols, the range of improvement of the average power consumption is not obvious as the input interval decreases. This is because each point on the curve represents the average of 200 nodes, many of which did not participate in packet relaying as much as the nodes in the chain topology.

Finally, we evaluate how the variation of sleep_factor in P-MAC integrated with routing affects the network performance. The value of sleep_factor determines the length of the SLEEP period (see Equation 2). Increasing sleep_factor increases the length of the SLEEP period time and the whole cycle time. Increasing the sleeping time of sensor nodes can save more energy. However, when a packet is not generated at the start of the SEND DATA period in the current cycle, a longer cycle time makes the packet wait longer for the next cycle.

For the chain topology, it is still the case that n = 24. The input interval of a CBR flow in node 0 varies from 10–1 seconds. Each simulation runs for 1,200 seconds. Table 4 shows the different values of sleep_factor we used in our simulation, as well as their corresponding SLEEP period time and cycle time. The average packet delivery latency, the average power consumption per sensor node and the data throughput are observed, and the simulation results are shown in Figure 12.

Note that Figure 12(a) shows the delivery latency that only varies with respect to the input interval from 10–5 seconds. This is because when the input interval is less than 5 seconds, the latency in the network with sleep_factor = 17 will be too large to be displayed. Figure 12(a) implies that the packet delivery latency increases as the sleep_factor increases for the reason stated above. But a greater sleep_factor ensures that nodes save more energy, as shown in Figure 12(b). When the network throughput is considered, a small sleep_factor is expected if the network traffic load is high, as shown in Figure 12(c). Figure 12(c) also shows that the cycle time is almost equal to the lowest input interval, which can still ensure that the output rate follows the input rate. For example, when sleep_factor = 17, the cycle duration is about 5 seconds, which is the lowest input interval ensuring that the network has 100% throughput, as shown in Figure 12(c).

Figure 13 shows the impact evaluation of sleep_factor in the random topology. The result is in accord with that in the chain topology. Note the evaluation doesn’t include a delivery latency evaluation, which is not convenient to be given for the random topology. Designers can select the value of sleep_factor according to the network traffic load. For example, if the network input rate doesn’t exceed 0.2 packets/second (the input interval is not less than 5 seconds), sleep_factor = 17 should be chosen, since the throughput can reach 100% with the lowest energy consumption. But if the input rate reaches 1 packet/second or higher, the minimum value of sleep_factor = 2 should be chosen.