Network Energy Saving (NES) Strategies

• NES is a set of strategies combining algorithmic, architectural, and protocol-driven methods to reduce network energy consumption without compromising service quality.
• Techniques include traffic scheduling, cell sleep control, and SDN architectures that dynamically adapt to changing network loads using deep learning and optimization models.
• Evaluations show energy savings ranging from 9% to 88%, balancing reductions in power use with minimal impacts on throughput, delay, and overall network performance.

Network Energy Saving (NES) scenarios encompass a diverse set of algorithmic, architectural, and protocol-driven strategies for minimizing energy consumption in networked systems, including telecommunication backbones, cellular radio access networks (RANs), cloud/data centers, and enterprise LANs. NES techniques leverage both hardware features (e.g., Energy Efficient Ethernet, cell sleep modes) and software control (e.g., SDN, deep reinforcement learning, multi-agent coordination) to optimize resource allocation with explicit constraints on performance and quality of service. The following sections systematically review the representative methodologies, mathematical models, and empirical trade-offs in NES, as documented across multiple domains.

1. Algorithmic Foundations: Traffic Scheduling, Resource Packing, and Cell Sleep Control

Early NES methods focus on concentrating network traffic on a minimal subset of resources—whether switch ports, servers, or radio cells—thus allowing unused resources to enter low-power or idle states:

• Ethernet Aggregates (SDN/ONOS, IEEE 802.3az): Traffic is assigned to as few EEE ports as necessary, determined at each interval by per-flow byte counters and a port-packing ("water-filling") algorithm. Flows are reordered and allocated to ports (up to capacity $C$), with variants like Greedy (fullest port first), Bounded-Greedy (with loss-reducing slack), or Conservative packing, which includes a safety margin $\alpha$ to limit packet drops and delay increases. Only used ports are kept active; others enter "Low Power Idle" (LPI) (Ferreiro et al., 2018).
• Multi-resource Data Center Routing: Here, the NES objective is extended to multi-resource settings (CPU, RAM, IO) via NP-hard MIP minimization of the number of active nodes (servers), subject to vector bin packing constraints. Flows are greedily routed to exhaust residual resources on already active nodes, using inversion-counting heuristics or fast topology-aware packing for structured Fat-tree networks (Wang et al., 2015).
• Cell Sleep Scheduling (RAN): In cellular RANs, NES is commonly implemented by turning off higher-frequency "capacity" layers during low-traffic periods, maintaining only the coverage layer. Advanced scenarios include ON/OFF decisions for small cells or gNB sectors, coordinated to avoid coverage holes and guarantee paging reliability, sometimes leveraging digital twin models or hybrid game-theoretic controllers for robust operation under traffic uncertainty (Mariegaard et al., 2023, Chakraborty et al., 1 Feb 2025).

2. Mathematical Models for Energy Consumption

Accurate models of the energy state are crucial for both theoretical bounds and practical controller tuning. Two common types:

• Ethernet/IEEE 802.3az Model: Per-port normalized power $\sigma(\rho)$, where $\rho$ is port utilization:

$$\sigma(\rho) = 1 - (1-\sigma_{\text{off}})(1-\rho)\cdot\frac{E_{\text{off}}(\rho)}{E_{\text{off}}(\rho)+S+W}$$

with $S$ the sleep entry time, $W$ the wake-up, and $E_{\text{off}}(\rho)$ a function of arrival rate and mean packet size.

• RAN/BS Model: Power consumption is decomposed into static and dynamic terms per BS or cell:

$$P = P_{\text{static}} + P_{\text{dynamic}}(\text{load}, N_{\text{act}})$$

For IEEE 802.11/3GPP/6G, models incorporate transmit power, PA efficiency, and the number of active antennas. Repeaters (NCRs) add receive-side and transmit-side terms, weighted by usage, with energy efficiency (EE) described as the throughput-to-power ratio (Azzino et al., 27 Sep 2024).

3. SDN Architectures and Real-Time Dynamic Adaptation

Network energy optimization in programmable environments leverages SDN controllers to implement traffic steering and sleep control logic:

• ONOS SDN Platform: Energy-aware apps register as packet-in listeners, dynamically sample flow tables and reprogram flow rules based on real-time traffic estimates, ensuring flows are concentrated and ports/cells can sleep accordingly. Controller-initiated rule updates and sampling intervals (e.g., $T \simeq 0.5$–$1$ s) achieve rapid convergence during traffic bursts or drops (Ferreiro et al., 2018).
• Metro Ethernet Networks (SPB): Forwarding topologies are pruned by electing "exporter" bridges (low-impact nodes), importing shortest-path trees at importer bridges, and turning off unused links unless their load exceeds a safety threshold $Th$. A hysteresis mechanism avoids rapid topology oscillations; activation/deactivation is orchestrated by the management plane (Maaloul et al., 2015).

4. Data-Driven and Learning-Based NES Controllers

Recent NES approaches utilize deep learning, reinforcement learning, and multi-agent methods to handle high-dimensional parameter spaces and stochastic traffic environments:

• Deep Q-Networks (DQN, DRL): The NES problem is cast as a constrained MDP, with state comprising on/off status, antenna parameters, per-user requirements. DQN agents jointly optimize power, tilt, and cell-activation variables, trading off total energy against user throughput and delay. Architectures typically include FC layers (128–64 ReLU) and output per feasible action. DRL controllers outperform greedy and fixed baselines, with up to 25–45% energy savings at near-constant QoS (Tran et al., 20 Aug 2024, Mao et al., 28 Jul 2025, Zhao et al., 29 Apr 2024).
• Federated Learning & Actor-Critic Frameworks: For ultra-dense networks, per-cell traffic prediction is performed locally via FL (bi-LSTM), then aggregated at a macro-BS acting as the central actor. Actor-critic networks (global+local parameters, per-cell sigmoidal outputs) select cell ON/OFF patterns while respecting resource constraints. Joint optimization achieves up to 77% improvement over legacy machine-learning heuristics, with negligible coverage loss (Abubakar et al., 2023).
• Graph Attention Networks (GATs): GAT-based user association in 5G wireless networks learns soft assignment matrices $\Theta$ that balance load and minimize cluster energy. The model infers per-UE/BS associations, reassigning users and switching off underutilized cells. Unsupervised objectives incorporate total EC and load-smoothing regularization. Reported savings of 60–90% over legacy RSRP-based policies and competitive with optimal benchmarks (Mirzaei et al., 22 May 2025).

5. Performance Evaluation and Trade-Offs

NES schemes must balance energy savings with quantified impacts on loss, delay, and fairness:

• Benchmarking: Most studies use real or trace-driven traffic (e.g., CAIDA, Milan CDR) and measure energy, throughput, packet-loss, delay, and fairness indices (e.g., Jain's). For Ethernet aggregates, conservative packing yields $\sim9\%$ savings over round-robin at negligible loss. For deep learning RAN controllers, DQN maintains $<1\%$ QoS degradation at up to $45\%$ energy saving (cell DTX/DRX) (Mao et al., 28 Jul 2025).
• Control Sensitivity: Safety margins ($\alpha$), link-load thresholds ($Th$), clustering parameters ($k$), and regularization weights in GATs ($\lambda_2/\lambda_1$) must be carefully tuned per-load regime. Aggressive policies can induce burst loss or PRB imbalances; conservative policies reduce energy savings (Kuang et al., 2015, Maaloul et al., 2015).
• Limits and Scalability: Some techniques (e.g., MILP, brute-force combinatorial search) scale poorly with network size, necessitating greedy heuristics, metaheuristics, or decentralized learning for real-time feasibility. In the largest systems (cellular, data center, supercomputers), NES may be guided by digital twin models (SRCON) that blend expert rules with machine-learning, achieving $>50\%$ accuracy gain over operator tools (López-Pérez et al., 2023).

6. NES in Specialized Domains: Enterprise LANs, Data Centers, Core Networks

• Spectral Clustering for LANs: Device-to-switch assignments are dynamically recomputed using spectral bisection of the affinity matrix derived from traffic patterns, recursively partitioning to minimize cross-cluster communication. Switches serving idle clusters are placed in hibernate or powered-down modes when predicted idle time exceeds wake-up threshold. Achieved savings range from 10–34% annually, dependent on topology size and traffic locality (Khan et al., 2016).
• Core Network Routing and Coding: NES is extended by reducing node and port activations through network coding (XOR) in bypass/non-bypass configurations. MILP models and fast min-hop heuristics deliver up to 33% savings, analytically scaling with average hop count and equipment ratios. Heuristic approaches attain 70–80% of the optimal model’s savings at orders-of-magnitude less computation (Musa et al., 2019).
• High-Speed Ethernet in HPC/Data Centers: EEE mechanisms such as DeepSleep and FastWake allow dynamic port power-down during detected idle periods. Techniques like PerfBoundCorrect adapt the power-down timer based on inactivity histograms and application miss rates, achieving energy reductions of up to 88% in idle-dominated traces (Monte Carlo transport), 8–12% in active training jobs (DNNs), with performance increases (latency, exec-time) bounded to user pre-specified thresholds (Rosa et al., 22 Oct 2025, Pérez et al., 2020).

7. Implementation Guidelines and Future Directions

Practical deployment of NES requires orchestrating hardware/software integration, management plane and control logic, as well as continuous monitoring and adaptation:

• SDN/NMS Integration: Collect flow counters, link energy impact profiles, and utilization at fixed intervals; implement controlled sleep/wake transitions with hysteresis to avoid oscillation; automate via policy modules (ONOS/SPB/Industry NMS) (Maaloul et al., 2015, Ferreiro et al., 2018).
• Hybrid Controllers: Leverage top-down decomposition of energy-saving intent (softgoal graphs), prune conflictive operations via statistical analysis, and present candidate operations to RL agents for rapid convergence (Zhao et al., 29 Apr 2024).
• ML/Digital Twin Modeling: Build local, data-driven models using live KPIs and UE reports, integrating agent-based and ML submodels for scalable carrier shutdown optimization, with interface to greedy heuristics or metaheuristics (López-Pérez et al., 2023, Chakraborty et al., 1 Feb 2025).
• Special Considerations: Account for hardware hysteresis (NICs), large-scale traffic surges (cellular), and application-specific latency sensitivity (HPC, campus LAN). Enable fine-grained sleep modes with optimized timers, and maintain coverage/SLA-protected fallback policies.

Continued advances in NES will depend on developing scalable, adaptable multi-agent and data-driven controllers that can handle ultra-dense, stochastic network environments in conjunction with programmable architectures and real-time telemetry. Robust performance under dynamic traffic and hardware constraints remains an open research challenge.