Green Orchestration: Energy-Aware Scheduling

• Green orchestration is a framework that applies policy-driven, energy-aware control to optimize workload placement, migration, and power-state management.
• It employs constrained optimization techniques, feasibility filtering, and real-time telemetry to balance renewable availability with stringent service-level agreements.
• Empirical studies demonstrate significant non-renewable energy reduction alongside measured trade-offs in job completion times and quality-of-service metrics.

Searching arXiv for papers on green orchestration and closely related orchestration frameworks. Searching arXiv for "green orchestration" and related workload/network orchestration papers. Green orchestration denotes a class of policy-driven, energy-aware control and scheduling methods that dynamically route, place, migrate, activate, deactivate, or reshape computation and network functions so as to minimize total energy consumption, greenhouse-gas emissions, non-renewable energy use, or grid power reliance while respecting service-level agreements, quality-of-service constraints, capacity limits, fairness criteria, or real-time requirements. In enterprise cloud settings, it is the process of dynamically routing and executing workloads across local machines, private clouds and public clouds to minimize total energy consumption while respecting SLAs, QoS requirements and budgetary constraints (Hulkury et al., 2012). In telecom systems, it is the systematic, real-time control of heterogeneous ICT network elements—spanning access, transport, core and data centers—via a policy-driven management layer that continuously adjusts operating modes to minimize overall energy consumption while preserving required service quality metrics (Gati et al., 2019). Recent work extends the same idea to geo-distributed data centers with multi-factor environmental profiles (Attenni et al., 14 Jul 2025), renewable-powered micro-datacenters with live workload migration (Tomei et al., 20 Nov 2025), cloud-edge deployment plans guided by green constraints (D'Iapico et al., 20 Feb 2026), O-RAN cell on/off control (Catalan-Cid et al., 3 Jun 2026), federated learning resource allocation under renewable availability (Panagea et al., 31 Mar 2026), and energy-harvesting IoT application embedding (Zhan et al., 22 May 2026).

1. Definitions, lineage, and problem scope

Two canonical formulations anchor the term. The Integrated Green Cloud Architecture defines green orchestration at the client edge as dynamic workload routing among local, private, and public execution venues through a middleware that computes estimated energy consumption for each venue, filters infeasible options by security, SLA, QoS, or budget, and selects the feasible minimum-energy option (Hulkury et al., 2012). The operator-centric telecom view defines green orchestration as a closed “observe–decide–act” loop in which Governance stores operator policies, Coordination/Orchestration schedules SON functions, and Knowledge aggregates monitored data across multiple timescales (Gati et al., 2019).

Later work broadens both the optimization target and the system boundary. Environmentally-conscious cloud orchestration formalizes per-data-center environmental-impact profiles over carbon, water, land, and e-waste, then couples these profiles to user-specific preference vectors (Attenni et al., 14 Jul 2025). Renewable-aware distributed AI training replaces best-effort migration heuristics with a feasibility-domain model linking checkpoint size, wide-area bandwidth, and renewable-window duration (Tomei et al., 20 Nov 2025). Constraint-based deployment in the cloud continuum introduces automatically learned green-aware constraints such as AvoidNode and Affinity that prune the deployment search space using monitoring-derived energy and carbon information (D'Iapico et al., 20 Feb 2026). In radio access networks, green orchestration spans cell sleep, transmitter switching, spectrum aggregation, and AI-driven O-RAN control loops (Han et al., 2023, Catalan-Cid et al., 4 Jun 2026). In mission-critical UAV networks, it also incorporates service equity through Max-Min Fairness and Jain’s Fairness Index (Lai et al., 10 Feb 2026).

This suggests that “green orchestration” is not a single algorithmic pattern. It is a systems concept spanning workload placement, migration control, power-state management, environmental profiling, and intent-to-action translation.

2. Core mathematical formulations and control abstractions

Across domains, green orchestration is typically posed as constrained optimization with binary placement or activation variables, continuous resource-allocation variables, or mixed integer formulations. A recurring structure is an environmental objective combined with hard feasibility constraints.

In geo-distributed cloud orchestration, each data center $d$ is assigned an environmental profile $p(d)\in\mathbb{R}^4$ over carbon, water, land, and e-waste, and the weighted cost of running job $j$ at $d$ is

$C(j,d)=p(d)^{T}\theta\bigl(u(j)\bigr),$

where $\theta(u)$ is the user’s weight vector. The optimization minimizes

$\min_x \sum_{j\in J}\sum_{d\in D} C(j,d)x_{j,d}$

subject to exact placement, capacity, migration-improvement, and binary constraints (Attenni et al., 14 Jul 2025).

In renewable-aware AI migration, feasibility is explicitly separated from utility maximization. Transfer time is

$T_{\text{transfer}}=\frac{8\cdot C}{B},$

total time cost is

$$T_{\text{cost\_time}}=T_{\text{transfer}}+T_{\text{load}}+T_{\text{down}},$$

and the renewable-energy breakeven time is

$T_{BE}=\frac{E_{cost}}{P_{node}}.$

Migration is “practically achievable” only if both

$p(d)\in\mathbb{R}^4$0

with $p(d)\in\mathbb{R}^4$1, and

$p(d)\in\mathbb{R}^4$2

The orchestrator then evaluates feasible destinations using

$p(d)\in\mathbb{R}^4$3

where $p(d)\in\mathbb{R}^4$4 is normalized renewable surplus and $p(d)\in\mathbb{R}^4$5 is congestion/queueing (Tomei et al., 20 Nov 2025).

In telecom orchestration, the control objective is often expressed directly over power models. The operator survey minimizes

$p(d)\in\mathbb{R}^4$6

subject to throughput, latency, and availability constraints, and also gives a reinforcement-learning reward

$p(d)\in\mathbb{R}^4$7

for sleep-mode selection (Gati et al., 2019). O-RAN cell control uses binary activity variables $p(d)\in\mathbb{R}^4$8 and minimizes total energy over time slots subject to coverage, QoS, transition limits, and timing budgets (Catalan-Cid et al., 3 Jun 2026). Federated learning orchestration replaces placement variables with per-worker CPU frequency, transmit power, and bandwidth allocations $p(d)\in\mathbb{R}^4$9, minimizing total discounted grid energy until the global model reaches the target performance $j$0 (Panagea et al., 31 Mar 2026).

Domain	Decision structure	Objective/feasibility basis
Geo-distributed cloud (Attenni et al., 14 Jul 2025)	$j$1	Minimize weighted environmental impact with capacity and migration-improvement constraints
Renewable AI migration (Tomei et al., 20 Nov 2025)	Destination choice after feasibility filter	Temporal and energetic feasibility, then maximize $j$2
O-RAN cell control (Catalan-Cid et al., 3 Jun 2026)	$j$3	Minimize time-slotted energy subject to coverage and QoS
Federated learning (Panagea et al., 31 Mar 2026)	$j$4	Minimize discounted grid energy under synchronization, device, and bandwidth constraints

A plausible implication is that green orchestration has evolved from venue selection to multi-timescale, multi-objective control in which feasibility tests, policy thresholds, and learned constraints are as important as the objective function itself.

3. Renewable-aware compute placement and migration

A prominent current direction is to align AI execution with renewable availability rather than with static data-center boundaries. Distributed AI training across renewable-powered micro-datacenters uses live workload migration and a formal feasibility domain. In the trace-driven evaluation, 5 micro-datacenters connected by 10 Gbps links were simulated using CAISO-based 7-day renewable traces with average 2.5 h windows and a mix of Class A/B/C checkpoints. Compared with Static, the Energy-only policy reduced non-renewable energy by 38% but increased JCT by 35% and attempted migrations for 18% of jobs. The Feasibility-aware policy reduced non-renewable energy by 52%, reduced JCT by 18%, and kept migration overhead below 2% (Tomei et al., 20 Nov 2025). The paper’s central claim is that energy constraint is almost always satisfied and the temporal constraint dominates.

Cloud systems for real-time workloads pursue a different renewable-aware mechanism. Instead of migrating jobs across sites, openstack-gc switches CPU cores between real-time and low-power profiles according to renewable availability $j$5, maintains an inventory of “GreenActive,” “GreenUsed,” “RegularActive,” and “RegularUsed” cores, and performs criticality-aware VM evictions when the real-time core budget $j$6 falls below pinned demand (Hewage et al., 2024). On an experimental server, after an energy-loss signal, the system evicted one VM and deep-slept 6 cores, causing peak power to drop by 22%. In the 14-day data center-scale simulation, the proposed scheme achieved harvest +34.8% over Crit-Aware while cutting evictions by 79.6% vs. Best-Fit (Hewage et al., 2024).

Constraint-based deployment in the cloud continuum addresses the same problem at the deployment-plan level. It learns green-aware constraints from monitoring data, real-time node carbon intensities, and service/network energy estimates. In the Online Boutique case study, the framework generated 6 candidate constraints, ranked down to 3 avoidNode constraints, and reported end-to-end reductions of approximately 25–30% in application carbon footprint with negligible performance loss when these constraints were fed into a green-aware scheduler (D'Iapico et al., 20 Feb 2026).

Older enterprise middleware already contained the essential placement logic: compute $j$7 for Local, Private, and Public, filter by security, SLA, QoS, and budget, and choose the feasible minimum-energy option. In the AXY case study, the storage scenario recommended PRIVATE with $j$8, compared with $j$9 and $d$0; the processing scenario recommended LOCALHOST because transport overhead dominated (Hulkury et al., 2012). This suggests that contemporary renewable-aware orchestration extends, rather than replaces, earlier energy-aware placement logic.

4. Network, radio, and edge orchestration

In communications systems, green orchestration operates through coordinated activation, sleep, steering, and scheduling. The operator survey frames this as a unified management platform across wireless, optical access, core, and data center domains, with a library of self-organizing functions such as sleep-mode activation, traffic steering, dynamic routing, and VM consolidation (Gati et al., 2019). Illustrative results include an overall 60% nightly energy reduction with end-user latency increase below 2 ms in an urban macro/pico policy-switching scenario, 57% energy saving with zero net delay increase for Q-learning with $d$1, and approximately 90% energy saving at very low load with $d$2, tolerating up to 5 ms extra delay (Gati et al., 2019).

Release-18 flexible spectrum orchestration for carrier aggregation expresses the problem through cell activation indicators $d$3, scheduling decisions $d$4, and UL transmitter-switching variables $d$5. The energy-saving mechanism includes SCell deactivation when DL load falls below threshold $d$6, SSB suppression, and uplink Wake-Up Signal. Quantitatively, energy saving under varying resource utilization reaches 22.6% at $d$7, 17.3% at $d$8, 10.2% at $d$9, and 4.3% at $C(j,d)=p(d)^{T}\theta\bigl(u(j)\bigr),$0, with concurrent UE throughput gains (Han et al., 2023).

O-RAN implementations introduce a layered intelligence split between Non-RT RIC, Near-RT RIC, and an AI Engine. In the BeGREEN demo, the Control rApp consumes KPM and AI outputs, generates A1 energy-saving policies, and the Energy-Savings xApp translates them into E2 “percent-power” instructions for progressive cell-sleep. The reported results include up to 40% reduction in total energy consumption during off-peak periods, maintenance of at least 99% of baseline throughput, sub-10 ms additional latency, A1 decision latency of approximately 500 ms, and E2 enforcement within 50 ms (Catalan-Cid et al., 3 Jun 2026). The companion Intelligent Plane paper generalizes this architecture to other beyond-5G energy-efficiency use cases through a unified control-loop fabric over R1, A1, E2, O1, and O2 (Catalan-Cid et al., 4 Jun 2026).

Intent-driven orchestration pushes the abstraction further. AGORA embeds a local tool-augmented LLM in the 5G data-plane control loop, using telemetry tools such as energy_mean_last_time(mec, Δ) and upf_set_target(mec) to translate natural-language sustainability goals into UPF steering decisions. The evaluation found a strong latency–energy coupling in tool-driven control loops, and among the tested models only Qwen2.5 reliably triggered non-zero migrations under high MEC2 power, with 28.6% policy migrations, 80.0% positive predictive value, and 15.4% false positive rate (Moreira et al., 8 Feb 2026).

In mission-critical air-ground integrated networks, ORCHID shows that the conventional efficiency–fairness trade-off need not hold. Using Max-Min Fairness within a two-stage MARL framework, ORCHID-MMF achieved normalized energy efficiency $C(j,d)=p(d)^{T}\theta\bigl(u(j)\bigr),$1, compared with $C(j,d)=p(d)^{T}\theta\bigl(u(j)\bigr),$2 for ORCHID-PF, while also improving rate fairness and load fairness (Lai et al., 10 Feb 2026). The paper explicitly characterizes this as a counter-intuitive efficiency-fairness synergy.

5. Multi-dimensional environmental metrics and domain-specific extensions

Green orchestration increasingly optimizes more than energy alone. In geo-distributed cloud placement, carbon-only optimization yields the lowest total CO$C(j,d)=p(d)^{T}\theta\bigl(u(j)\bigr),$3 but a 60% higher water footprint; water-only minimizes water but incurs 45% more carbon; land-only minimizes land but worsens both carbon and water by 20–30%. The Preference-Based model keeps carbon within 5% of the carbon-only optimum, water within 8% of the water-only optimum, and land within 10% of the land-only optimum (Attenni et al., 14 Jul 2025). This directly challenges the common reduction of “green” to carbon minimization.

Federated learning introduces another resource vector: renewable energy, battery state, compute energy, communication energy, and bandwidth contention. GreenFLag uses Soft-Actor Critic to jointly optimize computational and communication resources while accounting for dynamic renewable availability. Across three renewable scenarios, it reduces grid-energy by 94.8% on average compared to the best baseline, cuts total system energy by a factor of 1.5–2.4, and maintains convergence speed at approximately 11–12 global rounds (Panagea et al., 31 Mar 2026).

In green IoT networks, the objective is not throughput or JCT but the min-max Age of Service. The first MILP for scheduling and embedding applications on energy-harvesting nodes jointly optimizes sampling time, whether to run an application, and the energy usage of devices, gateways and servers. Relative to the MILP optimum, the receding horizon control approach yields 1.07x higher min-max AoS and the greedy approach 1.13x higher (Zhan et al., 22 May 2026). Here, green orchestration is coupled directly to freshness rather than only to energy or emissions.

Grid-Orch extends orchestration into distribution-grid simulation and analytics. Through a four-layer architecture and 36 domain-specific tools exposed via MCP, it supports multi-step renewable-integration workflows such as DER interconnection screening, voltage violation analysis, capacitor placement, and overvoltage mitigation. Workflow demonstrations show DER interconnection screening completing in under two minutes and a 24-hour QSTS simulation in approximately three minutes, with final voltages and losses matching direct OpenDSS scripting to within $C(j,d)=p(d)^{T}\theta\bigl(u(j)\bigr),$4 per unit (Liu et al., 12 May 2026). A plausible implication is that green orchestration also encompasses tool-mediated analytic workflows that accelerate planning for renewable integration.

6. Trade-offs, misconceptions, and emerging directions

A recurring misconception is that “green” policies can be reduced to always chasing the greenest resource. The renewable AI migration study explicitly shows the opposite: the Energy-only policy lowers non-renewable energy but incurs a 35% JCT increase because many migrations are failed or late, whereas feasibility filtering eliminates failed migrations and delivers both greener and faster training (Tomei et al., 20 Nov 2025). Likewise, the geo-distributed cloud study shows that optimizing a single sustainability factor can severely degrade other factors (Attenni et al., 14 Jul 2025).

Another common assumption is that energy efficiency and service equity necessarily conflict. ORCHID argues the opposite for its UAV setting, demonstrating that MMF not only guarantees service for cell-edge users but also achieves superior energy efficiency compared to PF (Lai et al., 10 Feb 2026). By contrast, the IoT scheduling study confirms a more conventional quality–complexity trade-off: MILP gives the best AoS but becomes infeasible for large nets, RHCOP is near-optimal but still high run-time, and GreedyOL is extremely fast with a 10–15% optimality loss (Zhan et al., 22 May 2026).

Several papers identify missing constraints that remain operationally important. Environmentally-conscious cloud orchestration ignores network latency, data-residency/legal constraints, and SLA deadlines; real-time deployment would require streaming electricity-mix updates, dynamic WUE models, and online reoptimization heuristics to bound migration overhead (Attenni et al., 14 Jul 2025). Renewable-aware distributed AI training points to grid-level control integration through utility curtailment and demand-response APIs, real-time AMI and SCADA signals, and support for partially migratable or distributed workloads such as ZeRO optimizer shards, checkpoint compression, delta-encoding, and hierarchical storage (Tomei et al., 20 Nov 2025). BeGREEN emphasizes continuous online learning, standardized DME schemas on R1, and end-to-end orchestration frameworks for multi-vendor deployments (Catalan-Cid et al., 3 Jun 2026). AGORA identifies standardized tool-call protocols, federated model swarms, carbon-intensity signals, and multi-agent coordination as necessary for scale (Moreira et al., 8 Feb 2026).

Taken together, these developments suggest that green orchestration is moving toward feasibility-aware, telemetry-grounded, multi-objective control in which renewable availability, environmental externalities, service guarantees, and human sustainability intents are treated as first-class orchestration inputs rather than after-the-fact reporting outputs.