GREEN: Carbon-efficient Resource Scheduling for Machine Learning Clusters
summary
GREEN monitors per-job power with NVML readouts (plus CPU estimates) and integrates carbon intensity every 30s, uses an energy efficiency optimizer to scale GPUs for progress-per-joule and a carbon efficiency optimizer to shift high-power jobs to cleaner hours under a multi-level feedback queue. On an 800-job 24‑hour university cluster trace it cuts cluster carbon by up to 41.2% and peak power by 12% while incurring a 3.6%–5.9% JCT tradeoff; shifting aggressiveness uses mu=2.
AI-written summary · about
chapters (9)
- 0:00Carbon-efficient ML scheduling
- 1:40Energy-saving limits
- 3:45GREEN design goals
- 6:57Real-time carbon tracking
- 10:04Energy optimizer explained
- 11:29Carbon optimizer & job shifting
- 14:20Multi-level feedback queue and fairness
- 15:24Evaluation & trade-offs
- 17:13Conclusion & future
related talks
25mGreen AI with JobRunr: Sustainable Training and Inference on the JVM
07mLightning Talk: Environmentally Sustainable AI via Power-Aware Batch Scheduling
29mHow AI and Deep Learning could effort the Climate Change
17mGreen AI with JobRunr: Sustainable training and inference on the JVM
32mOptimizing AI/ML workloads for sustainability
21mExperience Productionizing and Operating GB200 Cluster
27mGreening the AI Cloud: Validating Power Models for Kubernetes Containers
30mGreen AI in Cloud Native Ecosystems: Strategies for Sustainability and Efficiency