Carbon-Aware Scheduling

• Carbon-aware scheduling is the practice of aligning workload timing and placement with real-time carbon signals to minimize greenhouse gas emissions.
• It integrates spatial and temporal carbon intensity data into resource allocation decisions, balancing energy usage, cost, and performance.
• Applications span cloud computing, manufacturing, and federated learning, achieving significant emissions reductions with controlled latency and cost trade-offs.

Carbon-aware scheduling is the algorithmic practice of aligning computational workload placement and timing with real-time or forecasted carbon intensity signals, with the goal of minimizing total operational greenhouse gas emissions across diverse domains such as cloud computing, data processing clusters, serverless platforms, web services, manufacturing, edge systems, and federated learning. Rather than viewing all electricity as equivalent, carbon-aware schedulers integrate carbon intensity measures—often at the granularity of specific times and geographies—into the optimization of resource allocation, job dispatch, and autoscaling. This approach leverages both spatial (geographic) and temporal (time-period) variability in carbon intensity, exploits the elasticity and flexibility of modern systems, and introduces new trade-off frontiers involving performance, energy, cost, and sustainability objectives.

1. Fundamental Principles and Formal Models

At the core, carbon-aware scheduling seeks to solve constrained optimization problems that minimize total carbon emissions while satisfying correctness and performance requirements—such as deadlines, Service Level Objectives (SLOs), latency, and throughput. The general model is:

$$\min_{S}\; \sum_{k\in J} \int_{s_k}^{c_k} E_k\,I_{r_k}(t)\,\mathrm{d}t$$

subject to job allocations, capacity, and SLO constraints, where $E_k$ is the energy demand of job $k$, $I_{r_k}(t)$ is the carbon intensity of region (or resource) $r_k$ at time $t$, and $(s_k, c_k)$ are the scheduled start and completion times.

Frameworks such as GreenCourier (Chadha et al., 2023) extend Kubernetes scheduling to include real-time marginal emission scores per region, integrating filtering and node-scoring via a custom plugin; CASPER (Souza et al., 21 Mar 2024) jointly optimizes distributed web-service provisioning and load-balancing subject to latency SLOs, using a mixed-integer program over routing and server provisioning variables. CarbonFlex (Hanafy et al., 23 May 2025) adopts a case-based learning approach to mimic near-optimal cluster-wide resource provisioning and elastic scheduling, using historical job and carbon traces to set runtime thresholds.

Dependency-rich workflows and data processing DAGs necessitate more elaborate models, with variables for task start times, resource assignments, precedence constraints, and machine capacities. Flexible job shop and permutation flow-shop formulations are found in both manufacturing (Yang et al., 6 Dec 2025, Mencaroni et al., 3 Mar 2025) and data center batch-scheduling (Bostandoost et al., 8 Dec 2025), allowing explicit trade-off controls between makespan and carbon as a function of allowed schedule slack.

2. Architectures and Carbon Signal Integration

Architectural patterns for carbon-aware schedulers involve multi-component systems that interface with workload orchestrators (Kubernetes, Spark, CI/CD runners), real-time or forecasted carbon APIs (such as WattTime or Electricity Maps), and telemetry sources for regional energy measurement. Metrics are periodically collected, normalized, and exposed via REST endpoints (e.g., GreenCourier's Metrics Server exposes region→score mappings every 5 minutes (Chadha et al., 2023)) or integrated via case-bases and KD-trees for fast matching (e.g., CarbonFlex (Hanafy et al., 23 May 2025)).

Schedulers use these carbon intensity signals to filter candidate placements, score nodes for function or container allocation, and make temporal start-time or deferral decisions. Practical implementations handle synchronization, peering, and offloading (e.g., Liqo/Virtual Kubelet for multi-cluster serverless (Chadha et al., 2023)), or probabilistic load routing and dynamic autoscaling for SLO/carbon trade-offs (CASA (Qi et al., 31 Aug 2024), CASPER (Souza et al., 21 Mar 2024)).

3. Algorithmic Techniques and Trade-Offs

Carbon-aware scheduling algorithms span a rich methodological spectrum:

• Greedy and Score-based Plugins: Real-time filtering and scoring of nodes by normalized carbon metrics (GreenCourier, CASA, EcoLife (Jiang et al., 3 Sep 2024)).
• Mathematical Programming: ILP/MILP formulations for fixed mapping, scheduling, and provisioning tasks in cloud, manufacturing, and workflow domains (Mencaroni et al., 3 Mar 2025, Bostandoost et al., 8 Dec 2025, Schweisgut et al., 11 Jul 2025).
• Case-based and Historical Learning: CarbonFlex records prior optimal decisions and retrains parameters daily to mimic oracle behavior (Hanafy et al., 23 May 2025).
• Metaheuristics: Memetic algorithms (random-key + local search), PSO variants (with dynamic inertia and perception-response), and greedy list scheduling, augmented by local search (Jiang et al., 3 Sep 2024, Mencaroni et al., 3 Mar 2025, Schweisgut et al., 11 Jul 2025).
• Reinforcement Learning and GNN/LLM Integration: RL-based and graph-augmented policies optimize multi-objective makespan/carbon trade-offs in manufacturing and DAG scheduling (Yang et al., 6 Dec 2025), and DRL agents govern adaptive microgrid or CPN load in coordinated power-compute frameworks (Luo et al., 6 Aug 2025, Zhao et al., 22 Jul 2025).
• Wrapper Approaches: CAP throttles resource quotas dynamically in response to carbon thresholds (Lechowicz et al., 13 Feb 2025).
• LP-based Temporal Carbon-Optimal Transfer Scheduling: LinTS provides competitive allocations for inter-datacenter transfers (Rodrigues et al., 4 Jun 2025).
• Multi-agent and Multi-objective Extensions: Weighted sum and Pareto-front approaches balance carbon with performance metrics (e.g., latency, energy, cost), with explicit theoretical stretch factors and trade-off guarantees (Lechowicz et al., 13 Feb 2025, Chadha et al., 2023, Bostandoost et al., 8 Dec 2025).

Trade-off curves are empirically characterized, e.g., a 13.25% CO₂ reduction vs 10.26% function latency penalty for serverless (GreenCourier), up to 70% carbon savings at no mean-latency cost at large SLOs (CASPER), and up to ~47% reduction in flow-shop emissions at <2% makespan penalty (manufacturing (Mencaroni et al., 3 Mar 2025)).

4. Empirical Results and Quantitative Impact

Experimental evaluations across infrastructure types and workloads demonstrate:

System/Domain	CO₂ Reduction	Latency/Makespan Impact	Notable Approaches
Serverless FaaS	13.25% (GreenCourier); up to 2.6× (CASA)	Latency ↑10% (GreenCourier); 1.4× lower SLO violations (CASA)	(Chadha et al., 2023, Qi et al., 31 Aug 2024)
Web Services	Up to 70% (CASPER)	No SLO degradation	(Souza et al., 21 Mar 2024)
Batch/Elastic Jobs	57% (CarbonFlex)	Within 2.1% of clairvoyant oracle	(Hanafy et al., 23 May 2025)
Manufacturing	47.6% (flow-shop)	<2% makespan penalty	(Mencaroni et al., 3 Mar 2025)
Workflow DAGs	25% (DAG/jobshop)	No makespan increase	(Bostandoost et al., 8 Dec 2025)
Data Transfer	13–66% (LinTS vs. baselines)	Full deadline adherence	(Rodrigues et al., 4 Jun 2025)
Federated Learning	20–60% (via slack)	Improved accuracy at budget limits	(Arputharaj et al., 10 Sep 2025)
Microgrids	29% (DiffCarl)	2–30% lower cost, CVaR control	(Zhao et al., 22 Jul 2025)

Empirical data highlight that simple spatial shifting or region-aware heuristics can capture a majority of the attainable savings (Sukprasert et al., 2023, Claßen et al., 2023), while additional complexity yields diminished returns except under high flexibility or low utilization.

5. Limitations, Practical Constraints, and Best Practices

Key limitations include:

• Data Granularity and Forecast Error: Coarse or delayed carbon signal updates (e.g., 5-min intervals) limit responsiveness for short-lived jobs (Chadha et al., 2023).
• Hardware Heterogeneity and Embodied Carbon: Real-time scheduling should ignore embodied carbon for operational decisions, as including it induces suboptimal assignments (the "sunk carbon fallacy" (Bashir et al., 19 Oct 2024)). Operational carbon is the correct metric to minimize in fixed-fleet settings.
• Slack and Flexibility Requirements: High carbon savings rely on having temporal or spatial slack, sufficient idle capacity, or the ability to shift queues/deads; highly utilized or latency-critical services see sharply reduced benefit (Claßen et al., 2023, Bostandoost et al., 8 Dec 2025).
• Capacity and SLA Constraints: All evaluated frameworks enforce strict capacity, latency, or SLO constraints, with extensions to multi-objective and constraint-aware forms (e.g., Pareto scheduling, strict SLO enforcement).
• Forecast/Signal Uncertainty: Robust schedulers must tolerate carbon forecast errors (typically 3% loss per 14% forecast error) and dynamically retrain as usage patterns or grid decarbonization evolve (Sukprasert et al., 2023, Hanafy et al., 23 May 2025).
• Complexity vs. Savings: Analytical and empirical results consistently show that simple heuristics (one-migration, greedy deferral) attain ≥90% of possible savings (spatial shifting dominates temporal), and increased sophistication offers little incremental carbon benefit in most settings.

Best practices distilled from case studies and guidelines include decoupling procurement from scheduling decisions, directly integrating real-time carbon metrics into orchestrators (K8s, Slurm), and balancing performance overheads against expected carbon gain (Bashir et al., 19 Oct 2024, Yang et al., 8 Aug 2025, Yang et al., 6 Dec 2025).

6. Extensions, Research Directions, and Open Challenges

Current and emerging research directions encompass:

• Multi-objective, SLA- and Cost-Aware Scheduling: Weighted or Pareto-optimal selection balancing carbon with latency, price, and energy (Chadha et al., 2023, Yang et al., 6 Dec 2025).
• Learning-Augmented Schedulers: RL-based or GNN/LLM-augmented agents, adaptive to varying system topology, features, and real-world constraints (Yang et al., 6 Dec 2025, Zhao et al., 22 Jul 2025).
• Integration with Power Grids: Joint optimization of task scheduling and grid dispatch via hierarchical and closed-loop frameworks (Luo et al., 6 Aug 2025).
• Device- and Resource-Heterogeneous Systems: GPU-aware, multi-generation hardware pools, dynamically-allocated orchestration for federated and edge-distributed workloads (Jiang et al., 3 Sep 2024, Yang et al., 8 Aug 2025, Arputharaj et al., 10 Sep 2025).
• Robustness, Uncertainty, and Risk: Explicit handling of signal fidelity, distributional shift, and risk (e.g., CVaR-regularized objectives) in adaptive scheduling (Zhao et al., 22 Jul 2025).
• Industry-Standard Carbon Telemetry: Standardization and real-time reporting pipelines for container- and job-level operational carbon, supporting fine-grained scheduling (Yang et al., 8 Aug 2025).

7. Impact, Taxonomies, and Cross-Domain Synthesis

Recent surveys and meta-analyses categorize carbon-aware scheduling algorithms by optimization method (hardware-centric, software-centric) and sustainability objective (energy efficiency, carbon minimization), with approaches mapped to heuristic/metaheuristic, mathematical programming, RL, drift-plus-penalty control, and federated learning-based clusters (Yang et al., 8 Aug 2025).

Carbon-aware scheduling has shifted the paradigm of distributed resource allocation toward a sustainability-centric model, achieving emissions reductions from 10–70% across diverse system designs at relatively modest cost in throughput and latency, provided sufficient flexibility and digital infrastructure. As the energy grid continues to decarbonize, the marginal benefit of scheduling will diminish, suggesting a need for ongoing adaptation and integration with emerging grid and market signals (Sukprasert et al., 2023, Hanafy et al., 23 May 2025). Robustness, explainability, and alignment of operational with procurement-level metrics remain enduring challenges.