[Paper] Coordinating Power Grid Frequency Regulation Service with Data Center Load Flexibility

Source: arXiv - 2601.22487v1

Overview

The paper investigates how GPU‑heavy data centers can help stabilize power‑grid frequency—traditionally a job for fossil‑fuel generators—by flexibly adjusting their compute load. By quantifying the “exogenous carbon” saved on the grid side, the authors show that data centers can actually reduce overall carbon emissions, even after accounting for the extra energy they consume while providing regulation services.

Key Contributions

• Exogenous Carbon metric – a new way to measure carbon reductions that occur outside the data center, i.e., on the grid, when the center participates in frequency regulation.
• EcoCenter framework – an optimization platform that schedules GPU workloads and frequency‑regulation bids to maximize the amount of regulation a data center can safely provide.
• Empirical validation – real‑world traces from modern GPU clusters demonstrate that the carbon savings on the grid often exceed the additional operational emissions incurred by the data center.
• Economic insight – the study quantifies potential revenue streams for operators who sell regulation capacity, showing a win‑win for both the grid and data‑center owners.

Methodology

1. Modeling the Grid‑Data Center Interaction – The authors model the power grid’s frequency regulation market and the data center’s power‑draw flexibility (e.g., throttling GPU clocks, shifting batch jobs).
2. Defining Exogenous Carbon – They calculate the carbon intensity of the marginal generation that would otherwise be dispatched for regulation (often a fossil plant) and compare it to the carbon intensity of the data center’s marginal load.
3. EcoCenter Optimization – Using mixed‑integer linear programming, EcoCenter jointly decides:
   • Which GPU workloads can be delayed or accelerated,
   • How much regulation capacity to bid into the market, and
   • The resulting power‑profile that satisfies both compute SLAs and grid constraints.
4. Trace‑Driven Evaluation – Real workload traces from a production GPU cluster and publicly available grid frequency data are fed into the optimizer to simulate day‑long operation under various renewable penetration scenarios.

Results & Findings

Key takeaways

• The net carbon benefit is positive in > 80 % of simulated hours, even when the data center’s own power use rises.
• The framework can adapt to high renewable penetration periods, where the marginal grid carbon intensity is low, by scaling back regulation bids to avoid “negative” carbon savings.
• Workload latency constraints are respected; most batch jobs see ≤ 5 % increase in completion time.

Practical Implications

• Data‑center operators can monetize idle GPU capacity by participating in ancillary services markets, turning a traditionally “cost‑only” asset into a revenue generator.
• Grid operators gain a flexible, fast‑responding source of regulation that is environmentally cleaner than keeping fossil peaker plants on standby.
• Cloud service providers can embed EcoCenter‑style schedulers into their orchestration layers (e.g., Kubernetes, Slurm) to automatically expose regulation capacity as a service offering.
• Policy makers can consider incentives for “grid‑friendly” data centers, accelerating the transition to low‑carbon ancillary services.
• Developers can design GPU workloads with built‑in flexibility (e.g., checkpoint‑restart, dynamic batch sizing) to make them more amenable to regulation participation.

Limitations & Future Work

• The study assumes perfect knowledge of short‑term grid frequency regulation prices; real markets may have more volatility and latency.
• Only GPU‑centric workloads were examined; extending the approach to CPU‑heavy or storage‑intensive services may require different flexibility models.
• The optimization currently runs offline on daily traces; real‑time or near‑real‑time implementations would need faster solvers or heuristic approximations.
• The carbon intensity data is based on regional averages; finer‑grained, sub‑hourly emissions factors could improve accuracy.

Future research directions include integrating renewable forecasting, exploring distributed coordination among multiple data centers, and prototyping a production‑grade EcoCenter service in a public cloud environment.

Authors

• Ali Jahanshahi
• Sara Rashidi Golrouye
• Osten Anderson
• Nanpeng Yu
• Daniel Wong

Paper Information

• arXiv ID: 2601.22487v1
• Categories: cs.DC
• Published: January 30, 2026
• PDF: Download PDF