Carbon-Aware Computing

Carbon-aware computing is an approach to computing resource management that takes into account the carbon intensity of the electricity powering these resources, with the goal of reducing overall carbon emissions. This approach acknowledges that the same computation can have significantly different carbon impacts depending on when and where it is performed.

Core Concepts

Definition and Principles

Carbon-aware computing is based on several key principles:

1. Carbon Intensity Awareness: Recognizing that the carbon emissions per unit of electricity vary significantly based on:

   • Time (hour, day, season)
   • Location (region, country, grid)
   • Energy sources powering the grid

2. Temporal and Spatial Flexibility: Leveraging the flexibility in when and where computing is performed to minimize carbon emissions

3. Workload Classification: Identifying which workloads can be shifted in time or location without compromising functionality or performance

4. Prioritization: Making carbon impact a primary consideration alongside traditional factors like cost, performance, and reliability

Carbon Intensity of Electricity

Carbon intensity is the amount of carbon dioxide equivalent (CO₂e) emitted per unit of electricity:

• Measured in grams of CO₂e per kilowatt-hour (gCO₂e/kWh)
• Varies dramatically by location: from ~10 gCO₂e/kWh (hydro/nuclear) to >800 gCO₂e/kWh (coal)
• Changes throughout the day based on:
  • Renewable generation (e.g., solar during daytime)
  • Demand patterns
  • Grid management decisions

Types of Carbon Intensity Signals

Two main types of carbon intensity signals are used in carbon-aware computing:

Average Carbon Intensity

• Reflects the overall carbon emissions of the electricity mix
• Based on the weighted average of all generation sources
• Useful for reporting and long-term trend analysis
• Limitations: May not reflect marginal impact of additional consumption

Marginal Carbon Intensity

• Reflects the emissions from the next unit of electricity to be generated
• Indicates the actual impact of increasing or decreasing consumption
• More relevant for real-time decision making
• Challenges: More complex to calculate and forecast

Carbon-Aware Computing Strategies

Temporal Shifting (Time-Shifting)

Moving computing workloads to times when electricity has lower carbon intensity:

Workload Types Suitable for Time-Shifting:

• Batch Processing: ETL jobs, data analytics, scientific computing
• ML Training: Non-urgent machine learning model training
• Maintenance Operations: Backups, upgrades, indexing
• Content Delivery: Pre-generating and caching content

Implementation Approaches:

• Delay Scheduling: Holding jobs until carbon intensity drops below a threshold
• Carbon-Aware Windows: Defining preferred execution windows based on forecasted intensity
• Opportunistic Computing: Dynamically scaling up when renewable generation is high

Spatial Shifting (Location-Shifting)

Moving workloads to locations with lower-carbon electricity:

Workload Types Suitable for Location-Shifting:

• Distributed Processing: Map-reduce jobs, data processing
• Regional Services: Services with global redundancy
• Content Delivery: Content with multiple hosting locations
• Data Processing: Analysis that isn’t tied to data location

Implementation Approaches:

• Geographic Load Balancing: Directing traffic to regions with lower carbon intensity
• Follow-the-Sun (or Wind): Moving compute loads to follow renewable generation
• Carbon-Weighted Autoscaling: Preferentially scaling in regions with lower carbon intensity

Workload Efficiency Optimization

Adapting workload execution based on carbon intensity:

• Quality Adaptation: Adjusting quality/precision based on carbon intensity
• Resource Allocation: Allocating more resources when carbon intensity is low
• Execution Paths: Choosing different algorithms based on carbon availability
• Service Levels: Varying service levels based on carbon intensity

Implementation Mechanisms

Carbon Intensity Data Sources

Sources for carbon intensity information:

• Electricity Maps: Real-time and forecast data for various regions
• WattTime: Marginal carbon intensity data and forecasting
• Grid Operators: Direct data from electricity system operators
• Carbon Intensity API: UK’s National Grid ESO API
• Cloud Provider Tools: Google Cloud Carbon Footprint, Microsoft Sustainability Calculator

Technical Approaches

Methods for implementing carbon-aware computing:

1. Carbon-Aware Schedulers:

   • Enhanced job schedulers that consider carbon intensity
   • Examples: Google Carbon-Intelligent Computing, Microsoft GEAR

2. Carbon-Aware Middleware:

   • Software layers that make carbon-aware decisions transparent to applications
   • Examples: Carbon-Aware Kubernetes Scheduler, SLURM Sustainable Plugin

3. Carbon-Aware Applications:

   • Applications directly integrating carbon awareness
   • Examples: Carbon-aware video streaming, adaptive ML training frameworks

4. Carbon-Aware Infrastructure:

   • Infrastructure designed to operate preferentially on low-carbon electricity
   • Examples: Carbon-aware data center power management

Real-World Applications and Results

Case Studies

Google’s Carbon-Intelligent Computing Platform

Google implemented a carbon-aware computing system that:

• Shifts non-urgent compute tasks to times of day with lower-carbon electricity
• Achieved 50% increase in using lower-carbon energy for compute tasks
• Required no user intervention or application changes
• Prioritized existing tasks with deadline requirements

Microsoft’s Carbon-Aware Azure

Microsoft’s approach involves:

• Intelligent workload placement across regions
• Time-shifting workloads within and between data centers
• Matching renewable energy generation with cloud workloads
• Reported 100,000 metric tons of CO₂ reduction in initial implementation

Academic Research Projects

Several research initiatives have demonstrated:

• 10-30% carbon reductions through simple time-shifting strategies
• Up to 45% reduction through combined time and location shifting
• Minimal impact on performance for appropriate workloads

Simulation Results

Research simulations show significant potential carbon reductions:

1. Periodic Jobs Scenario:

   • Time-shifting nightly builds, integration tests, and recurring business reports
   • Allowing flexible scheduling windows of ±8 hours
   • Results: 30-45% carbon reduction with minimal operational impact

2. Ad Hoc Jobs Scenario:

   • Flexible scheduling of machine learning training jobs
   • Based on dataset from NVIDIA research project with 3387 training jobs
   • Results: 15-20% carbon reduction with delay tolerance of only 3 hours

Challenges and Considerations

Technical Challenges

1. Data Quality and Availability:

   • Carbon intensity data not available for all regions
   • Forecasting accuracy varies
   • Granularity issues with large geographic reporting areas

2. Integration Complexity:

   • Legacy systems not designed for carbon awareness
   • Compatibility with existing schedulers and orchestrators
   • Network and infrastructure limitations

3. Performance Trade-offs:

   • Balancing carbon reduction with performance requirements
   • Meeting service-level agreements while optimizing for carbon
   • User experience considerations

Grid-Level Considerations

1. Renewable Energy Curtailment:

   • Periods when renewable energy exceeds demand and must be curtailed
   • Carbon-aware computing could utilize this otherwise wasted energy
   • In 2022, California curtailed approximately 7% of its solar production

2. Grid Stability:

   • Large-scale workload shifting could impact grid stability
   • Potential for “herding” behaviors if many systems respond to the same signals
   • Need for coordination with grid operators

3. Grid-Aware Computing:

   • Evolution beyond carbon-aware to grid-aware computing
   • Understanding how computing decisions affect grid operations
   • Avoiding negative impacts of simultaneous load shifting

Policy and Organizational Challenges

1. Metrics and Reporting:

   • Standardizing how carbon savings are measured and reported
   • Integrating with existing sustainability reporting frameworks
   • Validating actual carbon impact

2. Incentives and Priorities:

   • Aligning carbon reduction with business objectives
   • Developing internal carbon pricing mechanisms
   • Communicating trade-offs to stakeholders

3. Organizational Boundaries:

   • Coordinating between IT, sustainability, and business units
   • Addressing data sovereignty and compliance requirements
   • Balancing carbon considerations with other organizational priorities

Renewable Energy Integration

Renewable Excess Energy Utilization

Using cloud resources to consume excess renewable energy:

• Curtailment Problem: When renewable generation exceeds demand, energy may be wasted
• Opportunity: Compute resources can consume this otherwise curtailed energy
• Datacenter Locations: Strategic placement near renewable generation sources
• Dynamic Resource Allocation: Scaling up compute during periods of excess renewables

Carbon-Aware vs. Energy-Efficient Computing

Important distinctions between approaches:

1. Energy Efficiency: Using less energy to perform the same computation

   • Focus: Reducing overall energy consumption
   • Metric: Performance per watt

2. Carbon Awareness: Timing or relocating computation for lower emissions

   • Focus: Reducing carbon emissions per computation
   • Metric: Carbon per computation

3. Complementary Approaches:

   • Energy efficiency reduces the baseline consumption
   • Carbon awareness optimizes the timing and location of that consumption
   • Both are necessary for comprehensive emissions reduction

Future Directions

Emerging Research Areas

1. Machine Learning for Carbon Prediction:

   • Improved forecasting of carbon intensity
   • ML-based workload characterization for shifting potential
   • Predictive scheduling algorithms

2. Carbon-Aware Edge Computing:

   • Distributing computation between cloud and edge based on carbon signals
   • Edge devices powered by local renewable generation
   • Location-specific carbon optimization

3. Carbon-Aware Hardware:

   • Dynamic power scaling based on carbon intensity
   • Hardware-level support for workload shifting
   • Power-proportional computing with carbon awareness

Integration with Broader Sustainability Initiatives

Carbon-aware computing as part of holistic approaches:

1. Circular Economy:

   • Integration with equipment lifecycle management
   • Carbon-aware decisions on hardware refresh cycles
   • Balancing embodied carbon with operational efficiency

2. Green Software Engineering:

   • Designing software with carbon awareness from the beginning
   • Carbon metrics as first-class software design considerations
   • Standardized tools and frameworks for carbon-aware development

3. Climate-Positive Computing:

   • Moving beyond carbon neutrality to climate positivity
   • Using computation to enable broader carbon reductions
   • Supporting climate science and mitigation technologies

Jevons’ Paradox:

As technology makes resource use more efficient, demand increases, so resource use overall often increases