Layer-Aware MAC Energy Model

Updated 28 November 2025

Layer-Aware MAC Energy Model is a rigorous cross-layer framework that defines and integrates protocol states, PHY parameters, and hardware characteristics to quantify energy usage.
It employs analytical, Markovian, and queueing models to capture transitions like sleep, listen, transmit, and processing, ensuring detailed energy accounting.
The framework supports comparative metrics and cross-layer optimization strategies, enhancing energy efficiency in wireless sensor networks, IoT devices, and deep learning accelerators.

A Layer-Aware MAC Energy Model provides a rigorous, cross-layer analytical framework for quantifying and optimizing the energy expenditure associated with medium access control (MAC) protocols in networked systems. Such models make explicit the per-layer, protocol-dependent, and hardware-specific contributors to total energy consumption, often linking MAC behavior to the physical layer (PHY), queueing/buffer constraints, and system-level dynamics. This approach enables fine-grained evaluation and optimization of wireless sensor networks (WSNs), IoT nodes, multiple access channels (MAC) in communication theory, and even data-parallel hardware like systolic accelerators in deep neural networks. The following sections synthesize the state-of-the-art in layer-aware MAC energy modeling, drawing on analytical, Markovian, and cross-layer optimization frameworks used in modern literature.

1. Fundamental Principles of Layer-Aware MAC Energy Models

Layer-aware MAC energy models are characterized by explicit modeling of protocol-level states and events—such as transmission, reception, idle listening, polling, contention, backoff—and directly incorporate PHY-layer parameters like transmit/receive power, packet durations, preamble formats, modulation/coding parameters, and duty/resource cycles. In addition, these models systematically account for protocol logic (e.g., carrier sensing, backoff, duty cycling, sleep scheduling, and stochastic traffic arrival) and buffer constraints (e.g., finite queues, retransmissions, overflow/loss rates).

At their core, these models partition the node’s operation into a set of states $i$ (e.g., sleep, listen, transmit, receive, processing), assign to each a well-defined power consumption $P_i$ , and accumulate the expected duration $T_i$ spent in each state per protocol cycle. This yields the basic MAC energy per cycle:

$E_{\mathrm{cycle}} = \sum_{i} P_i T_i$

Extensions include stochastic duty cycles, variable per-packet processing, retransmission policy, contention resolution, and the explicit modeling of multi-user interactions (e.g., via Markov chains or game-theoretic analysis) (Corbellini et al., 2011, Callebaut et al., 2018, Miao et al., 2016, Siddik et al., 15 Jan 2024).

2. Closed-Form Analytical Models for Preamble Sampling and CSMA/CA Protocols

The probabilistic framework for preamble-sampling MAC protocols (e.g., LA-MAC, B-MAC, X-MAC) conditions total cycle energy on the precise interplay of sender/receiver wake-ups, preamble transmission, early-ACK presence, polling duration, and multi-node overhearing. Assuming a star topology with Poisson packet arrivals, the expected one-cycle energy in LA-MAC is:

$E_{LA-MAC} = E_{listen} + E_{preamble} + E_{data,tx} + E_{data,rx} + E_{ACK} + E_{SCHED}$

where each term accounts for the expected energy spent by sender, receiver, and overhearers in polling; number of preambles transmitted (accounting for overlap probability and expected strobe count); data transmission and reception; ack overheads; and schedule message (Corbellini et al., 2011).

For CSMA/CA (as in IEEE 802.15.6 or IEEE 802.15.4), a two-dimensional Markov chain is employed to model the MAC contention process across all retransmission and backoff stages. Layer-aware energy per packet or per time-slot for node $i$ (priority class UP- $i$ ) is:

$\begin{align*} E_i\ = &\ T_{\rm slot}P_{IDLE}P_{i,idle} + (T_{DATA}P_{TX} + T_{ACK}P_{RX} + T_{SIFS}P_{IDLE})P_{tran}P_{i,succ} \ & + T_{succ}P_{IDLE}P_{tran}(2P_{i,succ}-P_{i,succ}) \ & + (T_{DATA}P_{TX} + T_{ACK}P_{RX} + T_{SIFS}P_{IDLE})P_{tran}P_{i,coll} \ & + T_{coll}P_{IDLE}P_{tran}(2P_{i,coll}-P_{i,coll}) \ & + (T_{DATA}P_{TX} + T_{ACK}P_{RX} + T_{SIFS}P_{IDLE})P_{tran}P_{i,error} \ & + T_{succ}P_{IDLE}P_{tran}(P_{i,error}-P_{i,error}) \end{align*}$

where all protocol, queueing, and PHY-layer parameters (contention windows, error rates, payload size, channel access phase lengths) are explicit (Siddik et al., 15 Jan 2024).

3. Cross-Layer MAC Energy Efficiency: Multiple Access Channels

In multi-user MAC channels, layer-aware modeling extends beyond MAC protocol to encompass queueing behavior, buffer size, random arrivals, and hardware energy overheads. The cross-layer model by Mhiri et al. introduces the generalized per-user energy-efficiency metric:

$\chi_i(\mathbf{p}) = \dfrac{R q [1-\Phi(\gamma_i(\mathbf{p}))]}{b + q p_i [1-\Phi(\gamma_i(\mathbf{p}))]/f(\gamma_i(\mathbf{p}))}$

where $R$ is link rate, $q$ packet arrival probability, $b$ is nonzero fixed circuitry power cost, $p_i$ transmit power, $f(\gamma)$ is physical-layer success probability, and $\Phi(\gamma)$ models total (buffer overflow + decoding) packet loss (Mhiri et al., 2016). This model links the MAC energy efficiency to both storage-induced losses and hardware inefficiency, permitting evaluation of protocol improvements both at the MAC (scheduling, power control) and across layers (buffer sizing, retransmission policy).

4. Markovian and Queueing Approaches in LPWAN and IoT Networks

Layer-aware MAC models for low power wide area networks (LPWAN) or IoT nodes systematically build energy expressions by aggregating per-state and per-operation costs:

$E_{\mathrm{MAC}} = \mathbb{E}[N_{\rm tx}] \cdot (E_{\rm UL} + E_{\rm DL}) + P_{\rm sl} T_{\rm off} (\mathbb{E}[N_{\rm tx}]-1)$

Here, $E_{\rm UL}$ , $E_{\rm DL}$ are the total uplink/downlink energies per attempt, and $\mathbb{E}[N_{\rm tx}]$ is the expected number of (re)transmissions per successfully delivered packet, obtained from the packet error process. The model explicitly incorporates wake/sleep durations, post-processing, bandwidth, coding rate, payload size, channel error dynamics, and MAC-level policies (retransmissions, duty cycle enforcement) (Callebaut et al., 2018).

MAC/PHY cross-layer frameworks further capture the impact of dynamic adaptation (e.g., ADR, coding rate scaling), and the statistical consequences of channel variation and buffer-induced overflow under realistic traffic arrival and system constraints (Callebaut et al., 2018, Miao et al., 2016).

5. Layer-Aware MAC Energy Models in Hardware Accelerators

For CNN/accelerator hardware, a “layer-aware” MAC energy model quantifies switching energy for multiply-accumulate (MAC) units at granularity of convolution layer, weight, and partial sum state transitions. The model explicitly partitions the energy cost by layer statistics:

Partition the 22-bit partial sum space into $G$ groups using MSB and Hamming weight binning.
Construct per-layer activation statistics $P_{\ell}^{act}$ and partial sum transition matrix $P_{\ell}^{PS}$ .
The MAC-level (per-operation) energy for weight $w$ in layer $\ell$ :

$E_{\mathrm{MAC}}^{(\ell)}(w) = \sum_{i=0}^{G-1} \sum_{j=0}^{G-1} P_{\ell}^{PS}(i \rightarrow j) E_{\text{switch}}(w, i, j)$

Total energy for layer $\ell$ :

$E_{\ell} = \sum_{w \in \mathcal{W}} p_{\ell}(w) E_{\mathrm{MAC}}^{(\ell)}(w) \cdot N_{\mathrm{MAC}}^{(\ell)}$

where $N_{\mathrm{MAC}}^{(\ell)}$ is total MAC count in the layer; $p_{\ell}(w)$ is the empirical frequency of weight $w$ (Fang et al., 21 Nov 2025). Tiled systolic array mapping extends to hardware-cycle granularity, based on actual switching activity as gathered via gate-level simulation for each tile and layer.

This per-layer statistical modeling corrects for the significant variance in dynamic range and sparsity across layers, enabling energy-optimal pruning/quantization protocols not attainable with network-agnostic or global average-based models.

6. Comparative Metrics and Cross-Layer Optimization Strategies

Layer-aware MAC energy models directly support the derivation and comparison of key system performance metrics, including:

Energy-per-packet/bit: Derived by normalizing total expected energy by the number of successfully delivered payload bits.
Energy-efficiency ( $\chi_i$ or $U_E(g)$ ): Bits-per-Joule, including both protocol-level inefficiency (e.g., collisions, backoff, retries) and hardware-level cost (idle power, circuit overhead) (Mhiri et al., 2016, Miao et al., 2016).
Node/network lifetime: Computed via the "first-energy-drain" metric, tying residual battery, cycle cost, and duty cycle to expected operational lifetime (Miao et al., 2016).
Tradeoffs: Explicit quantification of the relationships between energy, delay, collision probability, spectral efficiency, and reliability (e.g., via cluster size or number of CSMA phases $n$ ).

Orchestrated cross-layer optimizations—such as repeated games for distributed power control (Mhiri et al., 2016), dynamic cluster-head selection and reformation (Miao et al., 2016), and energy-aware weight restriction in neural nets (Fang et al., 21 Nov 2025)—can be carried out under accuracy, throughput, or fairness constraints, with enforceability/boundaries dictated by closed-form model outputs.

7. Extensions, Limitations, and Applications

Layer-aware MAC energy models have found broad application in:

Wireless sensor/IoT/LPWAN protocol benchmarking, cross-layer design, and optimization (Corbellini et al., 2011, Callebaut et al., 2018, Miao et al., 2016, Siddik et al., 15 Jan 2024).
Queue-aware, fairness-optimized multi-user MAC power control (Mhiri et al., 2016).
Accurate energy accounting and hardware-aware compression in deep neural network accelerators (Fang et al., 21 Nov 2025).

A plausible implication is that such models, by making explicit all dependencies, delimit the maximal benefit achievable through protocol or hardware optimization under real-world, resource-constrained settings. However, accuracy depends critically on the underlying statistical and behavioral assumptions (e.g., independence of transitions, memoryless retransmissions, uniform node deployment), and on correct calibration of all power/timing/traffic parameters to real devices and workloads.

The model frameworks can be augmented to include additional layer interactions (e.g., transport/network layer retransmissions, hop-by-hop forwarding, inter-node synchronization cost), and adapted to new domains such as hardware-level weight and switching selection for energy-efficient machine learning inference. Data-driven calibration (e.g., using gate-level simulation or hardware-in-the-loop profiling) increases the predictive accuracy and practical applicability to deployment-scale systems.

References: