Reducing false wake-up in contention-based wake-up control of wireless LANs

In the multiple access uplink of a WLAN, nodes contend to transmit to the same AP (access point). The performance heavily depends on the collision probability, which is controlled by the CW. Following the binary exponential backoff policy [2], CW starts from \(CW_{min}\), doubles per transmission failure until it reaches \(CW_{max}\), and resets to \(CW_{min}\) per transmission success. Usually \(CW_{min}\) and \(CW_{max}\) are fixed. This, however, does not work well when the network is congested. Thus, dynamic adjustment of CW is necessary. System throughput depends on the setting of CW [18], and it is possible to adjust CW, based on the number of nodes, to maximize the throughput. System throughput also depends on collision probability, which is correlated with average contention interval [19] and inter-frame space (IFS). In order to reduce collisions, each WLAN module needs to sense the channel continuously, which wastes much power. In comparison, in this work, a WuR is used to measure in an energy efficient way the length of IFS, based on which CW is adjusted to reduce both collisions and false wake-up events.

Power waste due to idle listening is a big issue in sensor networks where there is little traffic and each node is almost always idle. Low power listening [20] is a conventional technique for reducing the power waste due to idle waiting, which shifts some overhead from the receiver to the transmitter. The transmitter transmits a long preamble (B-MAC [13]) or periodically repeats the transmission of a short preamble (X-MAC [21]) before each packet, so that a receiver can detect the arrival of this packet. Instead of always being active, a receiver periodically wakes up to sample the channel, and if the energy due to the preamble is detected, remains active to receive the following packet.

A similar power-save mode is adopted in WLANs for the downlink reception. An AP periodically broadcasts a beacon which carries a traffic notification map. Nodes periodically wake up to receive beacons, and if there are packets buffered at the AP, stay awake to receive and otherwise go to sleep immediately. However, duty-cycling [22], which realizes low power consumption by using a long period, usually cannot meet the real-time requirement.

A packet carries a MAC (media access control) header which includes an address specifying the receiver and a duration field indicating when the communication will end. Then, in a further improvement [23, 24], each node at first only decodes the MAC header of a packet, and if this packet is destined to itself, remains active to receive the payload, and otherwise sleeps for a period indicated by the duration field. This avoids the complete decoding of non-relevant packets sent to other nodes.

Event-triggered sensing requires a real-time response, where a receiver should respond immediately to the transmitter that has detected an event. To this end, very-low power WuR [5] has been extensively studied. Usually it is envelope (energy) detection instead of coherent detection that is used in the WuR to reduce power consumption.

Wake-up control is also applied to WLANs. A low-power WuR, sharing the same channel as its co-located WLAN module, is used to monitor the WLAN channel. In [25], a ZigBee device as a WuR monitors the channel, and activates its WLAN module if the energy is detected. Without knowing the specific receiver, this usually causes many false wake-ups. A ZigBee module is even used as a secondary radio to obtain the fingerprint of signal strength from nearby APs, and on this basis, provide an energy-efficient background scanning of APs [26].

An elegant method uses a wake-up ID to specify a specific receiver so as to avoid false wake-up. When using a WuR for energy detection of WLAN signals, this can be realized by changing the physical length of a packet [11, 14‐16] or applying the on-off-keying modulation on WLAN signals [7]. In the flow-level wake-up control, e.g., an AP is remotely activated by a node initiating a new flow [10, 15], the target (AP) will stay awake until the flow is finished, and then enter the sleep state again. In such cases, the wake-up latency, only at the initial stage, is not a big problem.

Some literature tried to integrate the wake-up control into the MAC protocol, especially for sensor networks [27]. Different from the flow-based wake-up control, here, it is assumed that a node is usually in the sleep state to avoid short-term idle listening, besides long-term idle listening. Then, a wake-up signal is transmitted before each data packet, activating the receiver on demand. The sender node performs carrier sense to avoid potential collision, and if the channel is idle, sends the wake-up signal, which carries the receiver’s wake-up ID [28]. The actual data communication is realized by CSMA/CA. The receiving node performs the indicated task according to the received packet and after that enters the sleep state again. A comparison is performed between duty-cycling MAC protocols and the WuR-based MAC protocol, using the parameter of a real wake-up radio [9]. The evaluations under different use cases confirm that the WuR-based MAC protocol reduces power consumption, improves packet delivery rate and reduces delay, compared with duty-cycling based methods.

MAC-level wake-up control can also be included in the WLAN module, if part of its circuit can be used as a low power WuR. An example is illustrated in [8]. A WLAN module in its idle state reduces its clock rate to reduce power consumption. The preamble of each packet is extended so that a node at a low clockrate can also detect the preamble and the wake-up ID inside it, which is similar to B-MAC [13]. If the wake-up ID in the preamble matches its own, a WLAN module recovers its clock-rate to receive the following data packet, using the normal CSMA/CA sequence. This, however, still consumes more power than an optimized WuR, and only applies to remote wake-up control.

Currently, researchers mainly focus on the power waste due to idle listening at a receiver. In contrast, we investigate the case where multiple nodes contend for the channel and study how to reduce the power consumption due to the carrier sensing (idle waiting) before the actual transmission. In this problem, first, low power listening [20] via duty-cycling is not feasible because the slot for carrier sense is very short. Although overhearing the packet header and going to sleep when a packet is non-relevant help to reduce the power consumption [23, 24], decoding the header of all packets is still expensive. In addition, wake-up latency is not considered in these works.

When activating a WLAN module on-demand, wake-up latency is a big problem. In the remote wake-up control [5, 8‐12], the wake-up latency of WLAN modules, although not clearly addressed, can be solved by letting the transmitter hold the channel via a long preamble [13]. But when using a WuR for carrier sensing at the transmitter side, the wake-up is contention-based and the wake-up latency leads to false wake-ups, which has not been studied before. To this end, in this work, we analyzes the impact of wake-up latency, considering both the power consumption of WLAN modules in the wake-up period and the power waste due to false wake-ups of non-relevant WLAN modules, and propose to reduce false wake-ups from different aspects. In this way, it is expected that only one node is activated, only for its actual transmission.

Figure 1 shows the system model. Conventionally (Fig. 1a), a WLAN module consists of a power-hungry RF transceiver and a carrier sense function based on the RF receiving. A WLAN module is idle during most of the time, but its carrier sense function (and the RF receiving) runs all the time and consumes much power. In the proposed framework (Fig. 1b), the carrier sense function of a WLAN module is separated as an independent part, and is implemented by a low power WuR. A WuR shares the same channel as a WLAN module and continuously detects the energy of a WLAN signal, based on which carrier sense is conducted. Compared with a WLAN module, a WuR consumes much less power because it performs energy detection without requiring a stable carrier. Although in concept a WuR and a WLAN module are independent, in the actual implementation, a WuR can be realized as a part of a WLAN module for a better integration [8].

A WuR and a WLAN module interact as shown in Fig. 1(c). When a WLAN module has a packet to send but the channel is not ready yet, it sets the backoff counter (using current CW) and starts its WuR. Carrier sense by a WuR is performed by energy detection. A WuR detects the envelope of a received signal, and compares it with the carrier sense threshold (\(CS_{th}\)) per slot (with a length \(T_{S}\)). A WuR decreases its backoff counter by 1 if the channel is idle in a slot, and leaves it unchanged otherwise. When its backoff counter is decreased to 0, a WuR activates its co-located WLAN module. A WLAN module manages the CW based on the number of slots between two successive transmissions estimated by the WuR, besides initializing the backoff counter.

Both a WuR and a WLAN module have a non-zero wake-up latency due to the hardware constraint. A WLAN module, transiting from the sleep state to the fully functional active state, has to wait until its clock and carrier frequency get stable. Its wake-up period has a relatively long time \(T_{WU}=N_{WU}\cdot T_{S}\) (equal to \(N_{WU}\) slots), which may be greater than 100 \(\mu s\), but its transition from the active state to the sleep state can be finished within a short time \(T_{SL}\). A WuR has a small latency,² which usually only occurs at the beginning of a flow. Although the interaction between a WuR and a WLAN module also causes a delay, this delay can be made small by circuit design. Therefore, for the simplicity, in this work we focus on the wake-up latency of a WLAN module and neglect that of a WuR. In the analysis we take two assumptions. (i) Each WLAN module has the same wake-up latency. (ii) A WuR has the same sensitivity as \(CS_{th}\) of a WLAN module.

Without a wake-up latency, a WLAN module is expected to transmit immediately when receiving a wake-up instruction from its WuR. But in the presence of the wake-up latency (\(T_{WU}=N_{WU}\cdot T_{S}\)), the first WLAN module that is activated cannot transmit immediately. It is because this WLAN module has to first transit from sleep to the active state. In this wake-up period, the channel, although expected to be busy by the carrier sense mechanism, remains idle. Other WuRs detect the channel idleness and continue the decrease of their backoff counters. Then, multiple backoff counters may reach 0, and their WuRs will activate the co-located WLAN modules. All but the first node are falsely activated.

Figure 2(a) illustrates an example of the multiple access uplink, using a simple WLAN consisting of an AP and three nodes. At first, A and B set their backoff counters with random initial values 8 and 4, respectively. After the channel is idle for a period of DIFS (Distributed coordination function IFS), WuRs of A and B decrease their backoff counters per idle slot. WuR B first decreases its backoff counter to 0, and activates its WLAN module to transmit a data frame \(P_{BU}\) to the AP. When WLAN module of node B transits from sleep to the awake state, the channel remains idle. As a result, WuR A decreases its backoff counter to 0 per idle slot before node B starts transmitting, and activates its WLAN module. Node A faces a false wake-up because the channel is already busy due to the transmission from node B and stays awake. Next, node C contends with node A to transmit a frame \(P_{CU}\) to the AP, and randomly selects a small value 3 for its backoff counter. Though WLAN module of node A selects a relatively large value 6 for its backoff counter, it starts transmission earlier because it is already active and the wake-up latency does not happen. In this round, node C faces a false wake-up. Here, the hybrid use of WuRs and WLAN modules for the backoff control worsens the false wake-up problem.

The impact of wake-up latency is reduced from three aspects, as follows.

By using the backoff freezing method in the wake-up period, the classical Bianchi’s Markov model [29] can be applied here. In the saturation scenario, the probability \(\tau\) that a node transmits in a randomly chosen slot and the collision probability p under the condition of a transmission can be solved by the following equations.Then, \(b_{i,k}\), the probability that a node is at the \(i\mathrm{th}\) backoff stage with a backoff counter value k, can be computed accordingly from \(\tau\) and p [29].

Considering all possible backoff stages, the probability that a backoff counter has a value k is equal to\(B_{k}\) determines the false wake-up probability, and \(B_{0}\) is equal to the transmission probability \(\tau\).

In the slot-based contention, there are three cases at each slot.

The transmission probability per slot is \(P_{tr}=1-P_{I}\). Nodes that transmit per contention round in the conventional CSMA are regarded as normal wake-up, and their numbers as successful transmissions (\(N_{S}\)) and collisions (\(N_{C}\)) are computed asOther nodes that get active due to the wake-up latency are regarded as false wake-up and their number is equal toOn average, total energy consumption of WLAN modules per contention round is composed of three parts: successful transmission (\(E_{S}(W)\)), transmission collisions (\(E_{C}(W)\)), and false wake-up (\(E_{F}(W)\), without early sleep).Numbers of successful transmissions (\(N_{S}\)), collisions (\(N_{C}\)) and false wake-ups (\(N_{F}\)) are shown in Fig. 4(a), under different values of \(CW_{min}\). With a fixed number of nodes and a fixed wake-up latency, increasing \(CW_{min}\) effectively reduces the number of collisions and the number of false wake-ups. An interesting fact here is that \(N_{F}\) is usually much greater than \(N_{C}\). Actually, the ratio of \(N_{F}\) to \(N_{C}\) increases with \(CW_{min}\) and approaches a steady value, which is nothing but \(N_{WU}\) (refer to the appendix for more details). This is because at large CW, \(B_{k}\) is almost the same when \(k<N_{WU}\) and the wake-up period is \(N_{WU}\) times of a collision slot. \(N_{WU}\) is typically much greater than 1, which infers that there are many more false wake-ups than collisions.

When early sleep is used, the wake-up time of a falsely activated node is reduced from \(T_{WU}\) to \(T_{WU}-C_{BO}\cdot T_{S}\), where \(C_{BO}\) is the backoff counter value just before the node enters the wake-up period, as shown in Fig. 3(b). \(C_{BO}\) can be between 1 and \(N_{WU}\). Let the \(C_{BO}\) be k (with a probability \(B_{k}\)). Then, the number of false wake-up slots is equal to \(N_{WU}-k\). On average, the number of false wake-up slots is equal toThen, a factor, \(\alpha _{ES}=(N_{ES}\cdot T_{S}+T_{SL})/(T_{WU}+T_{SL})\), should be multiplied to the right side of Eq. (13) when early sleep is used.

It should be noted that increasing CW affects the throughput and channel efficiency because it may lead to more idle slots at a larger \(CW_{min}\). Therefore, there is a tradeoff between throughput and energy overhead (due to collision and false wake-up).

Average throughput is computed via Eq. (15) as a ratio of the number of transmitted bits to the average slot length (\(T_{AVE}\)),in a similar way as [29], but involves the effect of wake-up latency, where L is packet length in bits.

Spectral energy efficiency \(\xi (W)\), defined as the ratio of the number of transmitted bits to the time and the energy consumed in the transmission, is equal toChannel efficiency \(\eta (W)\) is the percentage of time that a channel is actually used.\(CW_{min}\) should be adjusted in such a way that throughput and channel efficiency are improved (reducing collision and idle slots), but energy consumption is reduced (reducing collision and false wake-up). But according to Fig. 4(b), the peak of throughput (\(\varGamma (W)\)) appears in the range where energy consumption (\(E_{S}(W)+E_{C}(W)+E_{F}(W)\)) is relatively high. Then, throughput decreases, but at small \(CW_{min}\), energy consumption decreases faster than throughput. As a result, \(\xi (W)\) is maximized at a relatively large \(CW_{min}\). To better stress spectral efficiency, \(\eta (W)\) is used together with \(\xi (W)\) to find the optimal \(CW_{min}\) by the following equation\(\xi (W)\cdot \eta (W)\) and \(\xi (W)\) are shown as Metric and Metric2, respectively, in Fig. 4(c) under different \(CW_{min}\). Here, by considering channel efficiency, Metric reaches its peak at smaller \(CW_{min}\), where higher throughput can be achieved.

By changing the wake-up latency \(T_{WU}\) and finding corresponding optimal \(CW_{min}\), the impact of wake-up latency and the effect of CW adjustment in different methods are investigated. Here, the backoff freezing method (BOF) is regarded as a base. On this basis, functions such as ES (early sleep), OC (optimal CW by applying Eq. (18)) or their combination (ESOC) are studied. MaxTh which finds CW by maximizing the throughput (\(\varGamma (W)\)) in Eq. (15) and MaxEF which finds CW by maximizing Eq. (18) but without considering the false wake-up energy (i.e. \(E_{F}=0\)) are used for comparison.

Figure 5(a) shows the number of wake-up per transmission under different methods, where S, C and F correspond to \(N_{S}\), \(N_{C}\) and \(N_{F}\), respectively. \(N_{C}\) is relatively large in BOF and ES without adjusting CW, but becomes almost negligible when CW is adjusted in the other 4 methods. \(N_{F}\) quickly increases with \(N_{WU}\). Adjusting CW helps to reduce \(N_{F}\) to different degrees. Without considering the energy due to false wake-up in MaxTh and MaxEF, \(N_{F}\) is still relatively large. In comparison, \(N_{F}\) is greatly reduced in OC and ESOC. False wake-up probability, conditioned on a transmission, is computed as \(N_{F}/N\). This value is 0.540 in BOF, 0.238 in MaxEF, 0.123 in OC, and 0.149 in ESOC, when \(N_{WU}\) is equal to 20. ESOC reduces false wake-up probability by 72.4% and 37.2% compared with BOF and MaxEF, respectively. By adopting the early sleep policy, ESOC takes a more aggressive policy than OC and allows more false wake-ups.

Figure 5(b) compares energy consumption per transmission under different methods, where S, C and F correspond to \(E_{S}\), \(E_{C}\) and \(E_{F}\), respectively. Here, energy consumption of WuR, being the same in all methods, is not included. Unsurprisingly, overall energy increases with \(N_{WU}\) and false wake-up becomes the dominant energy overhead when CW is adjusted in OC, ESOC, MaxTh and MaxEF. MaxEF is a little better than MaxTh by considering the energy due to collisions. But the impact of collision is almost negligible. Therefore, their difference is not large. In contrast, energy overhead due to false wake-up is greatly reduced in OC and ESOC, and ES helps to further reduce energy consumption by letting nodes sleep early. When \(N_{WU}\) is equal to 20, ESOC reduces energy overhead (due to collision and false wake-up) by 85.2%, 62.8%, 28.0% compared with BOF, MaxEF, OC, respectively.

Spectral energy efficiency \(\xi\) is what we care most. Figure 6 shows spectral energy efficiency under different numbers of nodes (without considering energy consumption of WuR). As the number of nodes increases, spectral energy efficiency decreases in BOF and ES because of more collisions. Adjusting CW to reduce both collisions and false wake-ups, spectral energy efficiency in the other four methods approaches steady values. OC and ESOC outperform MaxTh and MaxEF, with ESOC being the best.

Table 2 shows the values of optimal \(CW_{min}\)(W) and energy overhead per transmission (\(E_{C}+E_{F}\)) of BOF, ESOC and MaxEF. Compared with MaxEF, ESOC uses larger CW. Although this degrades the throughput of ESOC a little (the difference is approximately 1Mbps), it helps to reduce the energy overhead greatly by over 60%. This explains why ESOC outperforms MaxEF in spectral energy efficiency in Fig. 6.

It should be noted that a large W may also cause large transmission delay. Average transmission delay (\(T_{D}\)) is equal to the product of the expected number of transmission slots (\(N_{AVE}\)) and the average slot length (\(T_{AVE}\)).Using \(1/\tau\) to reflect the number of slots required for a transmission without collision, and using \(1/(1-p)\) to account for the number of transmissions under collisions, \(N_{AVE}\) can be computed as [30]The delay results are also shown in Table 2. Although ESOC has a larger average slot length (\(T_{AVE}\)) than BOF due to a larger contention window, it reduces the collision probability and \(N_{AVE}\) effectively. As a result, its delay is even less than that of BOF in the saturation state.

The analysis in Sect. 4.3 gives the optimal CW under the saturation scenario. Let \(f_{X}(x)\) represent the distribution \(B_{x}\) of backoff counters (\(X=x\)) under the saturation scenario, and \(F_{X}(x)\) be its cumulative distribution. Then, the minimal value \(X_{min}=min(X_{i})\) of backoff counters of all nodes, follows the distributionaccording to the order statistics [31]. \(X_{min}\) decides the number of idle slots, and its expectation \(E(X_{min})\) represents the average number of idle slots per channel contention. These values are very small compared with \(CW_{min}\).

Usually the network runs in a non-saturation case. Here we consider a heuristic algorithm for adjusting CW under the non-saturation case, where less traffic is generated. When the same CW is used as in the saturation scenario, more idle slots will be detected, and CW needs to be decreased in order to improve channel efficiency. The number of consecutive idle slots per contention round can be measured by using a WuR and is denoted as \(I_{avg}\) after removing the impact of wake-up latency (\(N_{WU}\)) and taking moving average. Then, we adjust \(W=CW_{min}\) as follows: increasing W by \(W_{delta}\) if \(I_{avg}\) is less than \(E(X_{min})-\delta\) and decreasing W by \(W_{delta}\) if \(I_{avg}\) is greater than \(E(X_{min})+\delta\).In this way, the number of idle slots per contention, on average, is maintained around \(E(X_{min})\).

First, we investigate the impact of wake-up latency, using a saturation scenario where the number of nodes is fixed to 10. Figure 7(a) shows different components of energy per packet when a WuR is used, where S, C, and F correspond to success, collision, false wake-up, respectively. Energy per packet in WuR-CS increases quickly when the wake-up latency is small, which confirms the necessity of reducing false wake-ups. But it then approaches a constant value, which is a little counter-intuitive. This can be explained as follows: Falsely activated nodes stay awake in WuR-CS, waiting to transmit; Wake-up latency increasing within a certain range decreases the waiting time, and the overall false active time does not change greatly. But a further increase in wake-up latency will lead to the increase of energy per packet. In WuR-BOF, a falsely activated WLAN module is put into sleep again. When \(N_{WU}\) is small, this works relatively well, although the wake-up overhead increases with \(N_{WU}\). At a large \(N_{WU}\) (e.g., \(N_{WU}=50\)) which is much greater than default \(CW_{min}\), in each wake-up period all nodes are activated. As a result, energy per packet in WuR-BOF might even get greater than that of WuR-CS. WuR-OC effectively reduces false wake-up events by adjusting CW, and WuR-ESOC further reduces false wale-up time, and their energy overhead due to false wake-up can be kept very low when \(N_{WU}\) is no more than 20. WuR-MaxEF reduces collisions very well by adjusting CW, but the energy overhead due to false wake-up is much larger than that in WuR-ESOC because energy overhead of false wake-up is not considered in the optimization. When \(N_{WU}=20\), WuR-ESOC reduces energy overhead (collision and false wake-up) by 85.0%, 62.4%, 26.8% compared with WuR-BOF, WuR-MaxEF, and WuR-OC, respectively. This is almost consistent with the analysis result in Sect. 4.4.

When the wake-up latency is relatively large, per-packet sleep/wake-up causes much overhead and degrades the performance of wake-up control. It is possible to further reduce per-packet overhead, e.g., by introducing frame aggregation to enable longer idle time for sleep [32].

Figure 7(b) shows throughput of different methods. Throughput of CSMA is irrelevant of \(N_{WU}\) because each node is almost always active, and throughput of other methods except WuR-CS decreases as \(N_{WU}\) increases. Throughput of WuR-CS even increases a little with \(N_{WU}\) at a small \(N_{WU}\). This is because the wake-up latency leads to false wake-up in WuR-CS. Contention mainly occurs among the falsely activated nodes, which both reduces the impact of wake-up latency and controls the collision degree, although at the cost of more active time. When \(N_{WU}\) further increases, severe collision among almost all nodes also degrades the throughput of WuR-CS. Throughput of WuR-MaxEF is a little greater than that of WuR-ESOC.

Combining the results in Fig. 7(a) and (b) leads to spectral energy efficiency in Fig. 7(c). Because throughput decreases while energy per packet increases as \(N_{WU}\) increases, spectral energy efficiency decreases very fast. At small wake-up latencies, energy overhead is very small and the superiority of WuR-ESOC over WuR-MaxEF in spectral energy efficiency is not very obvious. The gap between WuR-ESOC and other methods increases with the wake-up latency.

Next, we investigate how system performance varies with the number of nodes. Here, the wake-up latency is fixed to \(N_{WU}=22\) slots, and a saturation scenario is used.

Figure 8(a) shows different components of energy per packet when a WuR is used. Without adjusting CW (WuR-CS, WuR-BOF), energy per packet increases almost linearly with the number of nodes. By adjusting CW, both collisions and false wake-ups can be greatly reduced.

Figure 8(b) shows throughput of all methods. Throughput of CSMA decreases as the number of nodes increases, due to more collisions. Throughput of WuR-CS can even be greater than that of WuR-CF, because the wake-up latency leads to false wake-up in WuR-CS and the contention is mainly among the falsely activated nodes without wake-up latency, as explained before. As the number of nodes further increases, the contention among falsely activated nodes also degrades the throughput of WuR-CS. In WuR-BOF, it is always all nodes that contend for the channel, and the curve has a similar trend as in CSMA, but the throughput is lower due to the wake-up latency. WuR-OC, WuR-ESOC and WuR-MaxEF achieve stable and higher throughput than WuR-BOF by adjusting CW. The gap between WuR-ESOC and WuR-CF is mainly due to the idle contention slots, and is partially due to unresolved collisions in WuR-ESOC.

Figure 8(c) shows spectral energy efficiency of different methods. Except WuR-CF, WuR-ESOC achieves the highest performance, and has a clear superiority over other methods. Its superiority over WuR-OC confirms the effectiveness of the early sleep method. WuR-ESOC far outperforms WuR-MaxEF, which confirms the necessity of considering energy overhead of false wake-up in optimizing CW. As the number of nodes increases, the transmission interval of each node increases, so does the energy consumption of WuR per packet. As a result, this leads to a little decrease in spectral energy efficiency of WuR-ESOC.

Figure 8(d) shows duty ratio per node. Duty ratio per node in CSMA is nearly 100% because each WLAN module always performs carrier sense for the next transmission in the saturation scenario. As the number of nodes increases, this value deceases to nearly 0.5 in WuR-CS and decreases greatly to about 0.2 in WuR-BOF. It decreases to almost 0.025 (the inverse of the number of nodes) in WuR-ESOC, because each WLAN module is only used for actual transmissions. There is an obvious difference between WuR-BOF and WuR-OC/WuR-ESOC, which confirms the effect of dynamic adjustment of CW on reducing false wake-up events.

We also investigated the performance of these methods under different traffic rates. Here, the number of nodes is fixed to 10. The results are shown in Fig. 9, where the horizontal axis is the overall rate of traffic from all nodes. WuR is used in all methods except in CSMA. In times of light traffic (no more than 500 packet/s), spectral energy efficiency increases with the traffic rate, when a WuR is used. This is because the number of bits divided by the transmission time is nearly constant, but an increase in traffic rate reduces the transmission interval, which further reduces the energy consumption of WuR for each transmission. Spectral energy efficiency is higher in CSMA than in other methods, because a WLAN module is put into sleep in CSMA and it has no extra power consumption for WuR. The superiority of WuR-ESOC over CSMA will get more obvious if nodes in CSMA always listen to the channel in the idle state.

As traffic rate further increases, the network starts to saturate, spectral energy efficiency slightly decreases in WuR-ESOC because a large contention window is used, which increases the idle waiting. Spectral energy efficiency greatly degrades in CSMA due to severe collisions and degrades in WuR-BOF, WuR-CS due to both collisions and false wake-up.

According to the results in Figs. 7– 9, we give the following remarks: Wake-up control for reducing the power consumption of WLAN modules in the short-term idleness before actual transmission generally works well at a small wake-up latency. At a relatively large wake-up latency, however, false wake-up is a dominant factor of energy overhead compared with transmission collisions in the multiple access uplink of WLANs. In the proposed method, the adjustment of CW helps to reduce false wake-up events and early sleep helps to reduce false wake-up time. When the wake-up latency is very large compared with packet transmission time, the per-packet sleep/wake-up overhead will be too large and degrade the performance of wake-up control. The per-packet overhead can be reduced, e.g., by introducing frame aggregation to enable longer idle time for sleep [32].