Carbon- and Water-Aware Scheduling of AI Workloads in Emerging-Market Data Centres
Carbon- and Water-Aware Scheduling of AI Workloads in Emerging-Market Data Centres
Description
Details
Context and Problem Statement
AI workloads are increasing data-center electricity demand. Carbon-aware scheduling often assumes granular carbon signals and limited water constraints, while emerging markets may face coarse signals and severe water scarcity.
Research Question
What joint reduction in carbon emissions and water use can be achieved by shifting AI workloads across time and location under realistic deadlines and imperfect carbon-intensity signals?
Proposed Approach
Formulate multi-objective scheduling under latency constraints, forecast carbon intensity and wet-bulb temperature, and compare constrained optimization with reinforcement-learning policies using public workload traces and national electricity-mix profiles.
Expected Contribution
A quantitative characterization of the carbon-water-latency trade-off and its sensitivity to signal quality.
Expected Prototype
An experimental scheduler integrated with a container orchestrator and reporting avoided emissions, water savings, and induced delay.
Datasets
Public data-center workload traces, electricity-mix series, carbon-intensity signals, and weather data.
Challenges
Signal granularity, lack of direct water measurements, limited real-world validation, and risk of impact displacement.
Research Question
Innovation
Expected Deliverable
Technologies
Required Skills
- Distributed Systems and Container Orchestration
- Multi-Objective Optimization
- Energy and Environmental Analysis
- Python Programming and Instrumentation
Datasets
- Public data-center workload traces
- National carbon-intensity and electricity-mix series
- Wet-bulb temperature and meteorological data