Carbon Aware Workload Management for Sustainable Computing

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2188 papers408 conceptsfrozen 2026-10-04

Traditional green computing focuses on energy efficiency—doing more work with fewer Joules. However, as Dustin Wright argues in Efficiency is Not Enough: A Critical Perspective of Environmentally Sustainable AI, efficiency alone is insufficient. Because the electrical grid’s carbon intensity fluctuates based on the time of day and geographic location, a highly efficient model run during a "brown" energy peak can emit more CO2 than a less efficient model run during a "green" peak.

The emerging research direction of carbon-aware computing shifts the paradigm from demand-driven execution to supply-driven scheduling. Instead of treating the grid as a static resource, researchers are developing systems that treat carbon intensity as a dynamic, fluctuating constraint.

The Dynamic Nature of Carbon Intensity

The core intuition behind this field is that the "cleanliness" of a kilowatt-hour is not constant. Lasse F. Wolff Anthony and colleagues introduced Carbontracker to quantify this, showing that carbon intensity varies significantly based on the local energy mix (wind, solar, coal, gas) and the time of day.

To exploit these fluctuations, systems must move beyond annualized carbon offsets—which Noman Bashir and Prashant Shenoy critique in Sustainable Computing - Without the Hot Air—and move toward real-time matching of compute demand with renewable supply.

Temporal Shifting: Waiting for the Sun

The most straightforward lever for carbon reduction is temporal flexibility. Many high-intensity tasks, such as training machine learning models or running batch scientific workflows, are delay-tolerant.

In Let's wait awhile: how temporal workload shifting can reduce carbon emissions in the cloud, Ilja Behnke and Lauritz Thamsen demonstrate that pausing workloads during high-intensity periods and resuming them when renewable energy is abundant can reduce emissions by up to 33%. This requires "interruptible" scheduling, where the system monitors real-time grid signals to fit execution into optimal "green windows." Similarly, Bingsheng He and colleagues proposed ReinDB, a database system that uses a multi-queue structure to buffer queries until green energy is available, exploiting the fact that users are often willing to tolerate latency for sustainability.

Spatial Shifting: Following the Sun

When a task cannot wait, it can often move. Spatial flexibility involves migrating workloads across geographically distributed data centers to regions where the grid is currently cleanest.

Zhanwei Yu and Lei You explore this in Less Carbon Footprint in Edge Computing by Joint Task Offloading and Energy Sharing, showing that combining spatial migration with local battery storage can slash emissions by over 80%. In the context of decentralized AI, Jieming Bian and Shaolei Ren developed CAFE, which dynamically selects federated learning participants based on their local carbon intensity, ensuring the global model converges within a strict carbon budget.

Dynamic Computational Intensity

Beyond shifting when and where a task runs, researchers are exploring how to adjust the intensity of the task itself based on environmental signals.

  • Neural Architecture Search (NAS): Yiyang Zhao and Tian Guo introduced CE-NAS, which switches between "cheap" low-fidelity evaluations and "expensive" full-model training depending on grid cleanliness.
  • Generative AI: In Sprout: Green Generative AI with Carbon-Efficient LLM Inference, Baolin Li and Yankai Jiang propose "generation directives" that nudge LLMs toward conciseness during high-carbon periods, reducing the tokens generated (and thus energy consumed) without losing semantic quality.

The Embodied Carbon Trade-off

A critical nuance in this field is the tension between operational savings and embodied carbon—the emissions generated during the manufacturing of hardware.

Bilge Acun and Benjamin C. Lee point out in A Holistic Approach for Designing Carbon Aware Datacenters that temporal shifting often requires over-provisioning server capacity to handle backlogs when the sun comes out. If the emissions from manufacturing those extra servers exceed the operational carbon saved by waiting, the net environmental impact is negative. Young Geun Kim and Udit Gupta address this in GreenScale by incorporating hardware lifecycle data into scheduling decisions, ensuring that the system optimizes for the total carbon footprint rather than just immediate energy use.

Papers in this version

Showing 60 of 188 papers.

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 1, frozen text):

@misc{lacuna_carbon_aware_workload_management_for_sus_44da6fc5_v1,
  title = {Carbon Aware Workload Management for Sustainable Computing},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/carbon-aware-workload-management-for-sustainable-computing/rdr_44da6fc5c5454446b5c6ba6e15a42f23/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}