Carbon Aware Workload Scheduling

252 concepts110 papers

Traditional green computing focuses primarily on energy efficiency—minimizing the total number of joules required to complete a computational task. However, as the electrical grid increasingly integrates intermittent renewable sources like wind and solar, a critical realization has emerged: not all joules are created equal. A megawatt-hour consumed when the sun is shining and wind is blowing has a near-zero carbon footprint, whereas the same megawatt-hour consumed during a peak demand period may rely heavily on fossil-fuel peaker plants.

This insight shifts the paradigm from energy efficiency to carbon awareness. Instead of merely reducing power consumption, carbon-aware systems dynamically adjust when and where computation occurs to align electricity demand with the real-time availability of clean energy.


Temporal Shifting: Time as a Virtual Battery

The most accessible lever in carbon-aware scheduling is temporal shifting: delaying or advancing execution to match valleys in grid carbon intensity. This mechanism relies on the inherent flexibility of delay-tolerant workloads, such as machine learning training, continuous integration (CI/CD) pipelines, and database maintenance.

Grid Carbon Intensity Over Time
High  |      _---_
      |     /     \         <- Pause or Defer Workloads
Low   |____/       \_______ <- Execute Workloads (Renewable Surplus)
      +---------------------
             Time

To exploit this, schedulers introduce a "deadline buffer" or flexibility window. Rather than executing a task immediately, the scheduler evaluates forecasts of the local grid's carbon intensity to identify the cleanest continuous window for execution. As demonstrated by Ilja Behnke et al. in Let's wait awhile: how temporal workload shifting can reduce carbon emissions in the cloud, expanding this scheduling window allows flexible batch applications to act as a computational buffer, effectively functioning as a form of virtual energy storage.

For longer-running tasks that exceed the duration of a single clean energy window, schedulers employ interruptible task scheduling. Instead of running continuously, workloads are dynamically paused when carbon intensity spikes and resumed when green energy returns. Jesse Dodge et al. show in Measuring the Carbon Intensity of AI in Cloud Instances that this pause-and-resume approach can significantly reduce the operational footprint of multi-day AI training runs, even when it extends the total wall-clock duration of the project. This strategy is highly effective across diverse automated workflows; for instance, H. Ward Classen et al. apply similar temporal shifting principles to non-urgent software engineering tasks like nightly builds within Carbon-Awareness in CI/CD, demonstrating substantial emission cuts without harming developer productivity.


Spatial Shifting: Following the Clean Energy Grid

When a workload cannot afford to wait for local grid conditions to improve, it can often be shifted across space rather than time. Geo-distributed data centers allow schedulers to exploit geographic variations in weather, climate, and regional energy markets.

Spatial shifting routes computational demand to data centers currently experiencing a surplus of renewable generation or lower ambient temperatures (which minimizes the energy required for cooling infrastructure). In Sustainable AIGC Workload Scheduling of Geo-Distributed Data Centers: A Multi-Agent Reinforcement Learning Approach, Siyue Zhang et al. utilize multi-agent reinforcement learning to navigate these spatial and temporal variations, migrating heavy generative AI fine-tuning jobs to regions where energy is both cheaper and cleaner.

However, moving workloads is not free; it introduces data transfer overhead and network latency. To balance these trade-offs, Shuang Wang demonstrates in Electricity Cost Minimization for Multi-Workflow Allocation in Geo-Distributed Data Centers that combining graph deep learning with reinforcement learning allows systems to allocate complex, multi-stage workflows across global locations while strictly adhering to user-defined performance deadlines.


Optimizing the Signal: Marginal Emissions and Sunk Carbon

The effectiveness of any carbon-aware scheduler depends entirely on the signal it optimizes. Early frameworks often relied on average carbon intensity metrics, but recent research highlights two critical nuances required for accurate optimization:

  1. Marginal vs. Average Emissions: Schedulers must optimize for marginal emission rates—the carbon intensity of the specific, incremental power plant that will be turned on (or turned off) in response to a change in demand. As explained by Carole-Jean Wu et al. in Beyond Efficiency: Scaling AI Sustainably, optimizing based on marginal rates ensures that workloads are shifted to periods when renewable energy is actively being curtailed (wasted), preventing the accidental activation of fossil-fuel backups.
  2. The Sunk Carbon Fallacy: Schedulers must isolate operational emissions from manufacturing (embodied) emissions during real-time decision-making. In The Sunk Carbon Fallacy: Rethinking Carbon Footprint Metrics for Effective Carbon-Aware Scheduling, Noman Bashir et al. point out that once hardware is purchased, its embodied carbon is a sunk cost. If a scheduler factors embodied carbon into real-time routing, it may paradoxically favor older, underutilized hardware with lower amortized manufacturing costs, even if that older hardware is highly energy-inefficient and draws more operational carbon from the grid than a newer alternative.

Domain-Specific Adaptations

Carbon-aware scheduling is moving beyond generic infrastructure management and becoming deeply integrated into specific software runtimes:

  • Machine Learning & Architecture Search: Automated machine learning pipelines are notoriously resource-intensive. To mitigate this, Yiyang Zhao et al. introduce CE-NAS: An End-to-End Carbon-Efficient Neural Architecture Search Framework, which dynamically adjusts its search strategy based on grid cleanliness—prioritizing cheap, proxy evaluations when the grid is dirty, and deferring expensive, full-model training to periods of high renewable availability.
  • Serverless Computing: In ephemeral, event-driven environments, systems must balance performance against idle energy consumption. In Green or Fast? Learning to Balance Cold Starts and Idle Carbon in Serverless Computing, Bowen Sun et al. use reinforcement learning to dynamically adjust the "keep-alive" duration of serverless functions, becoming more permissive with resource retention when the grid is powered by green energy and aggressively reclaiming resources when the grid turns carbon-heavy.
  • Space Edge Computing: The concept even extends beyond terrestrial grids. In low-Earth orbit satellite networks, where power is strictly bounded by solar arrays and battery degradation, Weisen Liu et al. propose sunlight-aware scheduling in In-Orbit Processing or Not? Sunlight-Aware Task Scheduling for Energy-Efficient Space Edge Computing Networks. By synchronizing heavy computational tasks with orbital light cycles, satellites process data when direct solar energy is abundant, minimizing battery stress and tripling the operational lifespan of the hardware.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 113 papers
  2. Version 1 · 2026-10-04 · 188 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_carbon_aware_scheduling_44da6fc5,
  title = {carbon-aware scheduling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/carbon-aware-scheduling/rdr_44da6fc5c5454446b5c6ba6e15a42f23},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_carbon_aware_scheduling_44da6fc5_v2,
  title = {carbon-aware scheduling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/carbon-aware-scheduling/rdr_44da6fc5c5454446b5c6ba6e15a42f23/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}

See something wrong? Report a mistake