Greentooth: Robust and Energy Efficient Wireless Networking for Batteryless Devices

Batteryless systems capitalize on their external environment for energy harvesting, storing this harvested energy in small capacitors, and then strategically executing tasks once an adequate energy reserve has been accumulated [35]. As depicted in Figure 1(a) and (b), intermittent operation, stemming from inconsistent energy harvesting and the presence of limited energy storage (orders of magnitude smaller than supercapacitors or batteries [33, 57]), leads to sporadic and inevitable power outages. This situation can give rise to unpredictable control-flow, uncertain timing, as well as the loss of all volatile memory (including stacks, program counters, and registers), which significantly complicates network communication [34, 35, 57] and introduces a variety of challenges discussed below. Communication is expensive. Transmitting data packets or listening for messages from other sensors are among the most energy-intensive tasks performed by low-power sensors [77]. For applications that require infrequent transmissions, transmission costs can be amortized over time, but listening for messages remains prohibitively expensive. For instance, intermittent battery-free devices [32], using traditional low-power radios like the TI CC1101 [41], need to charge their small capacitors (often tens or hundreds of \(\mathrm{\mu }\mathrm{F}\) to encourage quick charging) for a few seconds depending on the harvesting situation, to listen for 1–50 \({\rm ms}\) before depleting their stored energy. Therefore, effective listening among batteryless nodes—a crucial component in relaying messages and acknowledgments—is impractical without knowing in advance when the message will arrive, which is challenging under unpredictable intermittent operation.

Uncertain timekeeping makes communication more expensive. Intermittent power outages can last anywhere from milliseconds to minutes and are often both unavoidable and hard to predict. During a power outage, system clocks stop working, and techniques like TARDIS [75], CusTARD [36], and CHRT [17] can provide rough estimates of outage duration but with errors of up to 10%. With less-accurate timekeeping, batteryless nodes must listen for longer periods to ensure they hear transmissions, while consuming more energy that could have been used for gathering, processing, and transmitting new data, and often causing new power outages, which further increase timing uncertainty. When faced with these timing and energy constraints, most batteryless intermittent networks have shifted the energy burden to the receiver—leaving the receiver on and listening continuously, so that intermittent nodes can transmit whenever they have data to send and the energy to do so.

What about the receiver? Shifting the energy costs to the receiver may not be an option for some applications. In smart buildings, autonomous vehicles, and other applications, receivers may have convenient access to wired power and can afford to listen continuously, as long as the owner does not mind running extra wires and paying the additional electricity costs. In other scenarios, like the cattle farming application [56] mentioned earlier, agricultural sensing [67, 87] and ecological monitoring [80], receivers are often battery-powered, untethered, mobile, or deployed in remote locations and extreme environments [62, 91]. Wired power may be expensive, inconvenient, or even impractical, and whether these receivers run on harvested energy or need periodic battery replacements, the cost of constant listening is significant. An energy-efficient receiver can be smaller, cost less, collect and relay more data and operate longer without human intervention.

A WuR is an extremely low-power radio receiver (orders of magnitude lower consumption than existing main radio transceivers) that continuously monitors the wireless medium and generates an interrupt signal whenever it detects the presence of a WuPkt [23, 71]. A typical WuR supports two modes of operation: broadcast mode, in which a single WuPkt activates multiple WuRs, and addressable mode, in which each WuR is activated individually by a WuPkt containing its address pattern or ID [73]. Often, the WuR is used with a main radio (dual-radio architecture), because main radios are capable of generating WuPkts in addition to actual data transmission, and this obviates the need for a separate WuPkt transmitter. As shown in Figure 2, a WuR is configured to monitor network channel continuously, while the main radio is off or in a sleep mode. The main radio is then activated by an interrupt (for data communication) whenever the WuR receives an addressed (ID-based) or broadcast WuPkt, thereby eliminating idle listening overheads on the main radio.

A new class of asynchronous protocols, called WuR MACs have used the dual-radio architecture for improved performance. Like traditional MAC protocols, WuR MACs can be classified into transmitter initiated, receiver initiated, and bi-directional, based on which end initiates the communication. Transmitter-initiated WuR MACs like WuR-TICER [47], PTW [1], WuR-MAC [61], and DCW-MAC [63], though implemented and evaluated in simulation tools, often operate in an on-demand fashion, where a node initiates communication by sending a WuPkt to activate the sink, followed by an actual data transfer. Receiver-initiated protocols [19, 29, 49, 50, 79] however, place the task of initiating communication on the receiver or sink node. Because of the sink’s capability to manage collisions, they are mostly suited for applications with high reliability requirements; coupled with scheduled and infrequent transmissions. Finally, in Bi-directional WuR MACs [5, 15, 44, 45, 64, 72], both transmitter and receiver can initiate communication as both have WuR capabilities, making them suitable for multihop communication.

Despite the low-power listening and low latency features of the dual-radio architecture, wake-up packet transmission can be energy consuming, particularly the ID-based WuPkts used for selective node awakening. Moreover, the bulk of existing receiver-initiated WuR MACs [15, 19] have only been tested in simulation, without any practical results based on real hardware or testbeds [73]. Evaluating the performance of these protocols using simulation-based testing is reasonable, but further evaluation on a real testbed (or in the wild) will provide more credibility to the performance claims. Furthermore, these protocols are designed for ENO—those that adapt to avoid power failures [8, 69, 70]. Unfortunately, when harvested energy becomes insufficient compared to the nodes’ energy demands during their active periods, power failures become an inevitable challenge for numerous devices. When conventional WuR MAC protocols and methods are deployed in intermittent batteryless networks experiencing frequent and unpredictable power disruptions along with timing issues, it adversely affects the networks’ performance, reliability, and overall energy efficiency.

The communication cycle begins with a WuPkt broadcast from the receiver that wakes and synchronizes all batteryless nodes in the network. The WuPkt—transmitted by the receiver’s main radio—serves both as an advertisement packet for new nodes looking to pair and a synchronization packet for already paired nodes. The WuPkt transmission period, \(T_w\) , can also vary depending on whether pattern correlation is enabled (requires a longer wake-up pattern) or additional data is included in the WuPkt. Enabling pattern correlation is beneficial for asynchronous WuR MAC protocols that prioritize selective ID-based wake ups, as it reduces false wake-up events. After broadcasting the WuPkt, the receiver proceeds to the Discovery phase.

Prior to receiving a WuPkt, all nodes in the network are expected to either be in a sleep state or harvesting energy to startup (for nodes that just joined the network or have depleted their energy). After waking up, connected nodes synchronize their times with the receiver and return to sleep until the start of their time slots, while unconnected nodes proceed to the Discovery phase.

The Discovery phase provides an opportunity, \(T_d\) , for unconnected nodes to connect. During this phase, the receiver actively listens (Rx) for new nodes. As described in Figure 5(a), a node that wants to connect and has enough energy will send a Connection Request packet to the receiver after a small random delay (configurable via \(dd_{t}\) ). The random delay along with clear channel assessment helps avoid collisions when two or more nodes try to pair with the receiver at the same time.

After receiving a Connection Request packet, the receiver adds the new node to the schedule by assigning it a transmission slot \(t_{i}\) and then sends back a Connection Response packet with time slot if the pairing process was successful. The node saves the received connection information, in non-volatile memory (FRAM) for use across power failures, and returns to sleep until the start of its time slot. If pairing fails either due to a collision or a lost packet, then the node will try to rediscover in the next communication cycle—provided it has harvested enough energy.

The discovery period ( \(T_d\) ) on the receiver presents a tradeoff between energy efficiency and responsiveness. Longer discovery periods allow more nodes to joint the network more quickly. Shorter discovery periods improve the energy efficiency of the receiver. To help developers tune their systems to specific application conditions and needs, we support two discovery modes: fixed and adaptive discovery.

Fixed discovery mode, shown in Figure 6, uses a constant \(T_d\) value (which could be High or Low) every communication cycle throughout the entire deployment lifetime of the network. A Low \(T_d\) value shortens the active discovery listening time, thereby conserving energy on the receiver. But this allows only a small number of new nodes to connect each cycle, and may extend the time it takes to completely discover all participating nodes (discovery duration). A Low discovery duration is best when sensed data changes slowly, new nodes join the network less frequently, and the energy efficiency of the sink node is critical (i.e., a smart city air-quality sensing deployment with a mobile sink node or cellphone as the receiver). A High \(T_d\) value allows more nodes to connect more quickly due to the extended discovery listening time, but this leads to an increase in the receiver’s power draw and reduces its battery life. This discovery mode is more suited for deployments where rapidly changing data must be collected as soon as new nodes join the network. Overall, the right settings for \(T_d\) will depend on deployment requirements and application needs, which may change over time and may be difficult to predict before the network is deployed. Adaptive discovery mode automatically sets ( \(T_d\) ) to match current network conditions at runtime. In this mode, Greentooth dynamically scales \(T_d\) up or down depending on whether there are new nodes waiting to join the network or not respectively. This is aimed at increasing the likelihood of a node getting discovered in the first few discovery cycles. For instance, the value of \(T_d\) increases incrementally (in step size defined by \(d_{ss}\) ) up to the max discovery period as long as new discoveries were made or when collisions happen during discovery. Conversely, \(T_d\) decreases gradually to the min discovery period (taken as \(d_{ss}\) ) as soon as the nodes are completely discovered (no new discoveries). Increasing the \(T_d\) value helps in achieving a more responsive network through faster discovery, while decreasing it helps optimize for overall power savings. So adaptive discovery benefits from the strengths of both the High and Low discovery modes without significant effort on the part of the developer. To ensure that batteryless nodes adapts their random delay threshold \(dd_{t}\) in line with receiver’s discovery period, the nodes first set their \(dd_{t}\) to \(d_{ss}\) and then increase it incrementally (by \(d_{ss}\) ) whenever a node experiences packet collision while trying to pair with the receiver.

During the Transmission phase, connected nodes have the opportunity to exchange messages with the receiver. This phase consists of \(n_{max}\) transmission slots, of time \(t_s\) , where \(n_{max} = T_t/t_s\) . Each connected node is assigned to a single transmission slot. The slot duration \(t_s\) and other parameters in Table 1 are customizable based on payload size and application needs.

As shown in Figure 5(b), if a connected node, assigned slot i, has data to send and enough energy to send it, then that node wakes up from sleep after receiving the broadcasted WuPkt, activates its main radio, and transmits its data at time, \(t_i = t_0 + (i-1)t_s\) , where \(t_0\) is the start of the transmission period and \(t_i\) is the time delay before the start of each node’s time slot. After receiving a data packet, the receiver responds with an ACK that contains the receiver’s estimate of the node’s current time drift (how far beyond \(t_i\) the packet was received) to allow the node to adjust its timing.

After communicating with all connected nodes, the receiver goes to sleep to conserve energy until the next communication cycle. Each sensor node also returns to sleep at the end of its time slot, turns off its main radio, and waits for the next communication cycle while listening for a WuPkt. If a WuPkt fails to arrive, then a backup timer wakes up the node at its assigned time slot and the communication cycle continues. If either the sensor node or receiver fails to communicate for a user-configurable amount of time ( \(CONNECTION\_TIMEOUT\) ), then the connection is considered lost. Nodes can also specifically end a connection by sending a connection termination message during their assigned time slots.

While the scalability of Greentooth depends on the maximum communication cycle period achievable and the duration of each allocated time slot, Greentooth is capable of handling up to 10 times more concurrent connections than existing protocols.

To enable robust, reliable, and energy efficient connection-oriented networking for intermittent batteryless devices, Greentooth implements the following set of features along with the adaptive discovery mode mechanism presented in Section 3.2. These key features are tailored distinctively to address the fundamental challenges of batteryless networking described in Section 2.1.

Lightweight synchronization and connections - Sol1: While existing synchronous protocols incur considerable energy overhead and complexity for neighbor discovery and synchronization [21, 25, 48], Greentooth uses a simple broadcast WuPkt (about 32 times less payload compared to ID-based WuPkts) in conjunction with the dual-radio architecture to achieve efficient synchronization and significant power savings on battery-free sensor nodes. The nodes in the network usually remain in sleep mode (like LPM3 on an MSP430 MCU [85]), with their main radios powered off, while using their WuR to listen for a WuPkt from the coordinating receiver—WuR listening costs 3–5 orders of magnitude less power than the main radio. When a WuPkt is received, each node wakes from sleep and aligns itself with the receiver’s global time. This efficient synchronization mechanism helps Greentooth maintain inexpensive but reliable connections between the sink and nodes that are plagued with unpredictable failures and timing inaccuracies. The resulting connection-oriented communication method helps mitigate congestion and overhearing problems and is significantly more energy efficient and achieves better throughput compared to asynchronous WuR MAC protocols that poll individual nodes using ID-based WuPkt.

Backup timing and drift resolution - Sol2: The wake-up radio may occasionally miss WuPkts, perhaps due to physical barriers, RF-interference, or node mobility. In the absence of a WuPkt, a connected Greentooth node uses a local timer to continue using its time slot during subsequent communication cycles as long as it can. This local timer is configured to wake the node in a recurring fashion (every \(T_c\) seconds) as soon as the node receives a Connection Response packet at the end of the initial pairing process. So, WuPkts are always helpful, but only necessary during initial pairing or when a node completely drifts out of sync. The local timer on the batteryless nodes is prone to clock drift that can disrupt subsequent slot alignment with the receiver. Consequently, the receiver accounts for node clock drift by monitoring node timing errors at every connection event and provides corrective feedback containing the node ID and drift ( \(packet\ arrival\ time - slot\ start\ time\) ) in its ACKs. This allows sensor nodes to automatically compensate for time drift even when WuPkts are missed or transmitted less frequently. Power outages also introduce timing errors, as batteryless nodes lose track of global time right after reboot. Therefore, Greentooth nodes use a timekeeper to estimate outage duration, which is then used to stay on schedule if WuPkts are missed after reboots. The nodes can keep time with any remanence-based timekeeping method, but we use CuSTARD [36] in our implementation because of its simplicity and low-power benefits. In general, remanence-based timekeeping techniques, like CusTARD, tend to be less precise compared to their counterparts that require active power.

Adaptive neighbor discovery and dynamic slot recycling - Sol3: In addition to the detailed description of the novel adaptive neighbor discovery mechanism provided in Section 3.2, Greentooth’s TDMA-style protocol also allows time slots to be managed dynamically, allowing sensor nodes to be added and removed from the network without disrupting existing connections or setup. For example, a node will be kicked off the network (disconnected) if it has been inactive for a period longer than CONNECTION_TIMEOUT. This helps maintain efficient use of allocated time slots and energy conservation at the receiver, as the receiver goes to sleep during unused time slots, including those belonging to recently evicted nodes. These unused slots are assigned to newly connected nodes, or reconnected ones that were previously disconnected.

Preserving connection state across outages - Sol4: Frequent and erratic power failures can significantly impact the operation and quality-of-service of intermittent networks. Greentooth addresses this fundamental issue by preserving connection state information across power failures. Each node saves its connection state in persistent memory after successfully pairing with the receiver during the discovery phase. This checkpointing technique is lightweight as writing and reading operations in the MSP430 MCUs (with built-in persistent memory) [86] are very energy efficient. By preserving connection state in non-volatile memory (FRAM), a depleted node can resume communication with the receiver after each reboot without spending additional energy to establish a new connection.

Variable WuPkt transmission - Sol5: Sending a WuPkt every single communication cycle is often an overkill as the batteryless nodes can stay synchronized using their backup timers. Thus, Greentooth further reduces the energy consumption on both the batteryless nodes and the receiver by supporting two different WuPkt transmission modes: fixed mode and dynamic mode.

In fixed mode, WKUP_TX_REPS (a user defined variable) defines how often a WuPkt is broadcasted by the receiver. When WKUP_TX_REPS is set to 1, a WuPkt is transmitted at the start of every communication cycle, when set to 2, it is transmitted every other communication cycle, and so on. This flexibility allows programmers to define WuPkt transmission repetition that suits their application requirements. For example, a higher WKUP_TX_REPS value leads to exceptional power savings but may limit how quickly new nodes are discovered and added to the network. However, a small WKUP_TX_REPS value is crucial for poor harvesting conditions, where nodes are kicked off the network due to inactivity (exceeds CONNECTION_TIMEOUT period) or in conditions where nodes enter and leave the network frequently.

In dynamic mode, rather than defining the value of WKUP_TX_REPS statically, the receiver dynamically adjusts the value at runtime based on node_coverage—which tracks the percentage of connected nodes that actively interact with the receiver. This is critical as network conditions are subject to change over time because of the arbitrary nature of harvested energy. So, adapting WKUP_TX_REPS at runtime in response to changes in network environment provides the benefits of faster discovery and power savings on the receiver. We compute node_coverage every communication cycle using

The connected_active_nodes is computed by recording the number of connected Greentooth nodes that delivered data to the receiver during a given communication cycle period. We utilize a circular buffer to keep track of the last few values of node_coverage. So, before inserting the current node_coverage in to the buffer, the value is compared with the minimum value in the buffer. This is to confirm whether batteryless nodes are maintaining their interaction with the receiver or dropping out of the network at a significant rate due to poor harvesting conditions. If node_coverage is greater than or equal to buffer min, then the WKUP_TX_REPS is doubled to save power, else it is set to 1—which signifies that some of the nodes are either unavailable or disconnected, thus requiring continuous transmission of WuPkts.

To mathematically characterize and analyze the behavior of Greentooth, we adopt a formal approach by treating the operation of the batteryless nodes as a stochastic process in discrete time and state. This enables us to model the high-level behavior using a Markov process, with state transition probability from state i to state j defined as \(p_{ij} = P[X_{n+1} = j \mid X_n = i]\) , where \(\lbrace X_0, X_1, X_2, \ldots , X_n\rbrace\) denotes a vector of discrete random variable at various timesteps. Figure 7 illustrates the representative Markov chain showing different states of the node and transition probabilities among the states. The following are the key states that each batteryless sensor node can be in the following:

Given the finite and countable nature of the Markov chain (Figure 7) with these five states, we consolidate the transition probabilities into a probability matrix denoted as P. The matrix P conveys the probabilities associated with transitioning from one state to another within a single timestep, and a transition probability of 0 denotes that a state is not accessible from another state.

Since the transition probabilities leaving a state are mutually exclusive and forms a universal set, each row of P represents the transition probabilities departing from a state and add up to 1. For instance, all transition probabilities departing from the \(Power\ failure\) state \(S_1\) equals \(p_{11} + p_{12} = 1\) . Conversely, each column represents the transition probabilities terminating at a state. Transitions terminating at the \(WuRx\ sleep\) state \(S_2\) is given as \(p_{12} + p_{22} + p_{32} + p_{42} + p_{52}\) . In general, the Chapman–Kolmogorov equation given as \(P_{ij}(n+m) = \sum _{k \in K} p_{ik}(n) p_{kj}(m)\) , allows us to recursively compute the transition probability of moving from state i to state j after n steps, where k is an intermediate state between i and j. Furthermore, the probability that the node is in state i at timestep n is given as \(\pi _i(n) = P(X_n = i)\) , which can be grouped as a state probability vector \(\boldsymbol {\pi }(n) = [ \pi _0(n), \pi _1(n), \pi _2(n), \ldots ]\) . So, we can represent the state probability vector of the node at timestep n as \(\boldsymbol {\pi }(n) = \boldsymbol {\pi }(n-1) \mathbf {P} = \boldsymbol {\pi }(0) \mathbf {P}^n\) , where \(\mathbf {P}^n\) is the n-step transition probability matrix for moving between different states and \(\boldsymbol {\pi }(0)\) is the state probability vector at the initial state (timestep 0, signifying when the batteryless node is initially introduced into the network). As n continues to increase, the states of the nodes reaches a steady state i.e., \(P_{\text{ss}} = \lim _{{n \rightarrow \infty }} P^n\) , and we can compute the steady-state probability vector \(\boldsymbol {\pi }\) using: \(\boldsymbol {\pi } = \boldsymbol {\pi }(0) P_{\text{ss}}\) or \(\boldsymbol {\pi } = \boldsymbol {\pi }P\) , through a system of linear equations derived fromWe can solve the following equations to obtain the steady-state probabilities \(\pi _1\) , \(\pi _2\) , \(\pi _3\) , \(\pi _4\) , \(\pi _5\) :The computed steady-state probability vector \(\boldsymbol {\pi } = [\pi _1\ \pi _2\ \pi _3\ \pi _4\ \pi _5]\) can be used to evaluate key network metrics like throughput, energy consumption, reliability, and so on.

Performance analysis using steady-state probability vector \(\boldsymbol {\pi }\) : We provide an analytical evaluation of the following network metrics using the derived steady-state probabilities \(\pi _1\) to \(\pi _5\) :

Throughput. To evaluate network throughput, we first identify the relevant state ( \(\mathbf {S_5}\) ) and transitions ( \(p_{12}, p_{35}, p_{45}, p_{52}\) ) in the Markov chain (Figure 7) that contribute to successful data transmission from each batteryless node to the receiver. For instance, a node will need to harvest enough energy to start up ( \(p_{12}\) ), be paired already ( \(p_{35}\) ) or just finish pairing ( \(p_{45}\) ) with the receiver, and be able to deliver its data to the receiver successfully ( \(p_{52}\) and \(\pi _5\) ). We compute the average successful transmissions per communication cycle, \(T_c\) , as the product of the steady-state probability of being in the relevant state and the associated state transitions, i.e., \((p_{12} + p_{35} + p_{45}) \times \pi _5 \times p_{52}\) . The network throughput \(T_p\) is further computed as follows:

The overall network throughput is then evaluated by averaging \(T_p\) with respect to the total number of batteryless nodes deployed in the network.

Energy consumption. For a batteryless node, the average energy consumption per successful transmission can be evaluated by considering the amount of energy consumed for data transmission given as \(E_{\text{transmit}}\) or \(E_{52}\) and the aggregate probability of successful transmission from all relevant state transitions \(P_{\text{tx}} = (p_{12} + p_{35} + p_{45} + p_{52})\) . Hence,

This can also be extended to analyze the overall energy efficiency and lifetime of the receiver by considering the different states and transitions of the receiver Markov chain.

Reliability. The reliability (success probability) represents the likelihood of a successful data transmission from a batteryless node to the receiver, indicating network robustness. From the Markov chain, we consider the following states ( \(\mathbf {S_1, S_3, S_4, S_5}\) ) and transitions ( \(p_{12},\ p_{35},\ p_{45},\ p_{52}\) ) that contribute to the likelihood of a successful data transmission. Using the steady-state probabilities ( \(\pi _1,\ \pi _3,\ \pi _4,\ \pi _5\) ) corresponding to the identified states, we define the network reliability asThe overall network reliability can be generalized over the total number of sensor nodes deployed in the network.

To evaluate Greentooth’s efficacy, we upgrade the Flicker hardware platform [32] with a MSP430FR5994 microcontroller, and build custom dual-radios boards (WuRx and main radio) as Flicker peripherals. Flicker’s modular design allows several peripherals to be easily interfaced with the compute core—which consists of the TI MSP430FR5994 MCU, energy management unit, crystal oscillator, and CusTARD timekeeper, debugging, and programming interface.

Greentooth node prototype consists of the following subsystems as shown in Figure 8: energy harvesting and management, MCU, sensors, timekeepers, dual-radios, and programming interface.

Energy harvesting and management. Our prototype uses federated energy storage [31] where MCU and each peripheral (radio or sensor) on the board has a small dedicated capacitor for storing harvested energy from a small solar cell [20]. This energy storage approach helps harvest energy more efficiently, relaxes subsystem coupling, while simplifying task scheduling and peripheral prioritization. We use a 100- \(\mathrm{\mu }\mathrm{F}\) main storage capacitor to power the MCU via a 2V regulator, and two 220- \(\mathrm{\mu }\mathrm{F}\) ceramic capacitors connected in parallel for the main radio. The WuRx peripheral is powered directly from the main storage capacitor, since it requires only a few microamps to operate. We select these capacitor values, using empirical measurements, to balance two competing goals: having enough energy for communication opportunities and ensuring quick recovery (availability). For example, the 440- \(\mathrm{\mu }\mathrm{F}\) total capacitance selected for the radio peripheral is sufficient to execute the most expensive atomic operation performed by the main radio during the initial pairing process (TX + RX).

When the main capacitor stores enough energy, the MIC841 comparator chip [40]—which gates power to different parts of the system, turns on when the main capacitor voltage reaches 2.8 \({\rm V}\) , and off when it drops below 2.2 \({\rm V}\) to prevents the system from entering a brownout state. The MCU enters a sleep state (LPM3) after startup where it only draws 1 \(\mathrm{\mu }\mathrm{A}\) [85]. While sleeping, the WuRx listens for WuPkts, while other peripherals stay off but keep harvesting energy.

Sensors and Timekeepers. Our prototype includes two sensors—an internal temperature sensor and a soil moisture sensor [81], which we use in indoor temperature sensing and soil moisture monitoring applications. The prototype also features CusTARD [36], a remanence-based timekeeper that estimates power outage duration using capacitor voltage decay. The timer uses a 1- \(\mathrm{\mu }\mathrm{F}\) capacitor and a low leakage diode [14] that significantly extends timing duration by minimizing the reverse leakage current. This timer can estimate outage duration for up to 1 \(\mathrm{min}\) .

Dual-radio communication. Greentooth uses a two-radio (main and wake-up radio) communication strategy. We use a common sub-1- \(\mathrm{G}\mathrm{Hz}\) CC1101 radio [41] as our main radio and design a custom wake-up receiver board that uses a low-frequency wake-up radio (AS3932 [6]) for low-power listening while the main radio is turned off to conserve energy. The CC1101 radio operates at 433 \({\rm MHz}\) and provides up to 500- \(\mathrm{m}\) transmission range with configurable MAC and PHY layer properties, while the AS3932 operates at 125 \(\mathrm{k}\mathrm{Hz}\) with a programmable 16-bit wake-up pattern. Multiple batteryless nodes running Greentooth communicate with a battery-powered receiver that is responsible for access control, coordination, and network synchronization. We implement our receiver using an MSP430FR5994 launchpad [42] connected to a CC1101 radio. We did not equip the receiver with a WuRx, because WuPkt transmission is expensive for battery-free nodes, consequently, we move key energy intensive operations to the receiver to enable the batteryless nodes to only perform useful sensing operations due to the scarce nature of harvested energy.

Although we use the CC1101 radio and MSP430 MCU, the Greentooth protocol, including WuPkt generation and detection is both radio and platform agnostic. The receiver generates a WuPkt using a two-stage modulation technique [24] that transmits a low-frequency WuPkt signal as a series of normal data packets at higher frequency using the CC1101 radio. First, the radio’s 433-MHz carrier is turned on and off (OOK) creating an envelope signal with a period of \(125 \,\mathrm{k}\mathrm{Hz}\) . The 125-kHz signal is further OOK-modulated by a 40-bit wake-up signal containing 15 carrier bits, 1 separation bit, 8 preamble bits and 16 wake-up pattern bits when pattern correlation is enabled. This produces a 320-byte payload, which when transmitted by the receiver’s main radio, will be detected by the wake-up receiver on the nodes.

Each node’s wake-up receiver has a demodulation circuit, shown in Figure 9. Whenever a WuPkt arrives at the sensor node, that circuit filters out the high-frequency carrier and forwards the remaining 125-Hz signal to a low-power wake-up radio chip (AS3932). The received signal passes through an envelope detector and data correlator circuit and wakes the main MCU from sleep if the detected wake-up pattern is correct. When data correlation is disabled, the wake-up receiver wakes the MCU anytime it detects a broadcast WuPkt with a carrier frequency that lasts for more than 550 \(\mathrm{\mu }\mathrm{s}\) . Greentooth’s broadcast WuPkt is made up of a 10-byte payload with each byte containing 0xAA—whose binary representation (0b10101010) reflects the OOK carrier preamble expected by the receiver). A minimum 10-byte payload is needed to generate sufficient carrier burst and preamble that can wake a sleeping node. Our WuRx design is based on the dual-radio design presented in Reference [24], but we optimize the wake-up process for energy and latency by using a broadcast WuPkt with simplified payload (10 bytes) instead of the 320-byte payload needed for ID-based wake-up.

We develop Greentooth firmware in C, for both the receiver and sensor node. The firmware allows developers to configure MAC and PHY layer properties like data rate, modulation technique, transmit power, communication cycle duration, discovery duration, and other parameters listed in Table 1. The Greentooth protocol is built as an abstraction layer over low-level CC1101 and AS3932 libraries. All hardware designs, source code, and energy traces will be released open source at the time of publication.

The Greentooth protocol makes use of the following key packets structure.

Wake-up Packet. The WuPkt, broadcast by the receiver at the beginning of a communication cycle is used for waking and synchronizing sleeping batteryless sensor nodes. While any amplitude modulation technique can be employed, we use OOK. The WuPkt simply carries the minimum 10 byte wake-up payload needed to activate the WuR on the wake-up receiver.

Connection Request and Response Packet. The connection request packet is sent to the receiver by a node that wants to join the network. It contains only two bytes, Node Id and Pkt Type (Discovery). The receiver replies with a connection response packet shown in Figure 10. This provides the node with specific time slot information and other connection state information used by the receiver in managing the network.

Data and ACK Packet. To prevent accidental wake ups of WuRs during the data transmission phase, we use Binary Frequency Shift Keying (2-FSK) modulation for data exchange between sensor nodes and the receiver. As shown in Figure 11, a total of 58 bytes of payload can be sent to the receiver in addition to Node Id and Pkt Type (Data). The lightweight ACK packet notifies the node of a successful transmission while delivering a time drift report. Finally, a node can also disconnect from the receiver by sending a special packet containing Node Id and Pkt Type (Disconnect).

In addition to the system specifications of the AS3932 wake-up radio reported in the Datasheet [6], we perform additional experiments to evaluate the power consumption and wake-up range of our custom WuRx board and present the results summary in Table 2.

To enable reliable and energy efficient networking of intermittent battery-free devices, Greentooth utilizes the key features (Sol1 to Sol5) presented in Section 1 and described in Sections 3.4 and 3.2 to address critical challenges of intermittent networks. In this section, we evaluate these features and how they contribute to the overall performance of Greentooth.

To assess the robustness of Greentooth in the wild, we select soil moisture monitoring application (smart agriculture) and evaluate the protocols under varying and intermittent real-world harvesting conditions. To ensure fair evaluation, batteryless nodes running different protocols must be deployed in close proximity—to minimize variations in harvested energy. However, deploying multiple nodes running different protocols in close proximity in a garden results in WuPkt interference. To prevent WuPkts from one protocol from interfering with nodes running another protocol, we record real-world solar energy traces using an Ekho recorder (Figure 19(f)) from different locations in the garden where the sensor nodes were meant to be deployed. This enables us to replay the recorded I–V surfaces in a repeatable manner across the competing protocols (Greentooth, AWD-MAC, and RI-CPT-WuR MAC). We capture different parts of the day by recording a solar trace at sunrise, sunset, and midday. To ensure our evaluation is carried out under real intermittent harvesting conditions, we also record three traces at different locations on a windy day, where nearby leaves intermittently shade the solar harvester.

Using a 33 \({\rm min}\) recorded trace, we replay the I–V surface of the traces on three separate Ehko emulators to capture the most critical section of the recorded energy traces and also to ensure the spatial differences in deployment locations are maintained. For the recorded intermittent solar traces, we explore the number of power failures present in the first 5 minutes (Figure 18) to confirm the presence and the arbitrary nature of power failures. We utilize the same testbed setup described earlier, but each sensor node is now equipped with a SEN-13322 soil moisture sensor [81] for soil wetness level measurement. Since soil moisture level changes less rapidly, we set up the protocols to use the following configurations: 10 \(\mathrm{s}\) communication cycle period, 50 \(\mathrm{m}\mathrm{s}\) slot duration and 500 \(\mathrm{m}\mathrm{s}\) max discovery period. We collect data for three experimental runs for the I–V surfaces replayed for each protocol. This is to ensure our experimental results are reliable.

For the results presented in Figure 19(a) to (d), we group the number of packets delivered into 3 \({\rm min}\) buckets and compute the mean and standard deviation of each bucket to provide a clearer representation of the distribution. At sunrise (Figure 19(a)), Greentooth nodes begin transmission before AWD-MAC and RI-CPT-WuR. This is due to their capability to save connection state across power failures, which keeps communication going during periods of scarcity without reconnection overhead after reboots. Consequently, Greentooth achieves 32% and 230% better throughput compared to AWD-MAC and RI-CPT-WuR, respectively. Also at Sunset (Figure 19(b)), Greentooth persists a bit longer, thereby achieving up to 15% and 134% better throughput compared to AWD-MAC and RI-CPT-WuR, respectively, as harvested energy becomes increasingly scarce. Greentooth and AWD-MAC shows similar performance at midday (Figure 19(c)) when harvested solar energy is abundant. Despite the availability of surplus energy, RI-CPT-WuR still underperforms due to its channel access competition requirement, which degrades performance as collision becomes unavoidable. When harvested energy becomes intermittent (Figure 19(d)), Greentooth shows significantly improved performance as it achieves about 73% and 283% better throughput compared to AWD-MAC and RI-CPT-WuR, respectively. This shows that Greentooth is robust enough to handle the complexities of intermittent harvesting conditions while enabling reliable communication between an energy constrained receiver and many battery-free sensors. In general, Greentooth is able to achieve better throughput (Figure 19(e)) under different real-world energy harvesting conditions.