Carbon Aware API

^{Draft, Project In Progress}

Sections

1. Overview
2. Architecture
3. Tool Methodology
4. Tool Validation
5. Case Studies
6. Recommendations

Terminology

Term	Definition
Carbon Intensity	Carbon emitted per energy unit.
Grid-based carbon intensity	All entities who operate on a shared electrical grid, share a common emission rate.
Carbon Aware	Adjusted behavior in response to the carbon intensity of consumption.
Carbon delta	The difference in emissions between carbon aware and unaware actions.
Carbon counterfactual	The carbon delta had the carbon aware action been different.
Demand Shifting	Selectively changing the time/location of a compute's execution, to a time/location where the energy demands are met by cleaner energy production, resulting in a lower grid-based carbon intensity.

Other definitions for the Carbon Awareness found at:
Green Software Foundation Dictionary

Carbon Aware API

To enable organizations to make smart decisions about their environmental impact and carbon footprint, we have created the Carbon Aware API to minimize the carbon emissions of computational workflows. A few key features of the API are:

• Utilizes marginal carbon emission from WattTime to identify carbon intensity per region

• Generates a retrospective carbon emission analysis

• Provides forecasted demand shifting suggestions

• Supplies regional carbon intensity information over different scopes

Marginal Carbon Emissions: This grid-responsive metric with finer granularity than average emissions, allows for seasonality/diurnal trends captured in demand shifting (source).

Retrospective Analysis: Time series evaluation to assess the carbon emissions for a given energy profile. Also provides counterfactual analysis to expose the potential emissions of if the run had been shifted.

Demand Shifting Scheduler: Recommends region and/or time which would yield a less carbon intensive run.

• Temporal: Identifies the window of minimum carbon intensity for a specified run duration within a chosen region from a 24-hour forecast.

• Geographic: Finds the region with the current lowest average carbon intensity for an immediate run of a specified duration. Can filters available regions by available SKU and migration laws for workspaces with protected data.

Regional Carbon Intensity: Provides the carbon intensity for each data center supported by a WattTime- tracked balancing authority. The possible scopes are historic intensities (time series for prior 24-hours, week, and month), real-time marginal intensity, and forecast (mean intensity for upcoming user- defined window).

Architecture

Carbon Aware API

wsgi.py          app         

utils.py          data.py          services          routes

ci_data.py          shift.py

Tool Methodology

Retrospective Analysis: By mapping grid-based marginal carbon intensity measurements to energy consumption profiles, the carbon footprint of cloud workloads is identified and evaluated. Part of this evaluation is to find carbon counterfactual emissions of if carbon aware scheduling was used, reporting possible outcomes of temporal and geographic shifts.

Given the following:

Where:

• t = discrete time vector
  
• M(t) = marginal operating emission rate (MOER) for a given region (i) at a certain point in time (j) for J number of timestamps.
  
• E(t) = energy consumed per time interval
• C_i = carbon emitted for a given region and time window (C₀ is the result from the M(t) for when and where the run was executed)
• C = vector of carbon emissions resultant of different initial conditions for M(t)
  

Using C the carbon couterfactuals can be identified because each value is the outcome of a different potential workload shift.

Demand Shifting: Combining carbon intensity forecasts from a certified provider, with anticipated workload constraints such as runtime, governance, and needed hardware, a green window can be found in the forecasted time span.

Given the following:

Where:

• t = discrete time vector
  
• F(t) = forecasted marginal operating emission rate (MOER) for a given region (i) at a certain point in time (j) for J number of timestampes.
  
• CI_i = mean forecasted carbon intensity for a given region and time window
• CI = vector of mean forecasted carbon intensities
  

Followed by:

The green suggestion (G) is the time period in the forcast span at a specific data center that yields the minimum mean forecasted carbon intensity. The goal of the Carbon Aware API is to find which start time (t₀) and/or location that would result in the least carbon to be emitted.

For each permitted data center all possible windows in the forecast span are searched to find the lowest average emitter.

• Time-shifting: no data center variation permitted. Sliding search to find best start time for the green window.
• Geographic-shifting: no start time variation. Evaluates permitted data centers to find the greenest location for an immediate start.
• Complete Shifting: data center and start time variation. Sliding search to find best start time for the green window across all permitted data centers.

For complete and geographic shifting, permitted data centers can be filtered based on regional data governance laws or which GPU SKU's are available.

Tool Validation

Carbon intensity data for validation provided by: WattTime

Data Granularity

Details: For the purpose of demand shifting, the effectiveness of using a carbon intensity baseline is compared against using a granular search. The baselines are the annual and monthly marginal emissions rate averages for respective regions, and the data granularities used are 1-hour and 5-minutes.

Years Evalulated	Specific Months Evaluated for Each Year
2020	January
2021	April
	July
	October

Working Data

• Using the Historical Emissions endpoint, CI data is reported at 5-minute intervals in month batches by default for each of these regions.
• Monthly Average Baseline: Mean MOER over the calander month
• Annual Average Baseline: Mean of Monthly Average Baselines
• 1-Hour CI Reporting Granularity: Mean of aggregated MOER values from 5-minute interval .

Results Validation in-progress

Search Sensitivity

Details:

Results: Validation in-progress

For more information, please see: Tool Validation

Case Studies

Carbon intensity data for case studies provided by: WattTime

Demand Shifting at Scale of Organizational Operations

Details: Geographic shifting compute instances at scale would reduce carbon by X% over the month

Results: Analysis in-progress

For more information, please see the full case study: Demand Shifting at Scale of Organizational Operations

Carbon Segmented Cloud Computes

Details: Execute and progress cloud computes during low emission periods to reduce the carbon footprint by Y kg

Results: Analysis in-progress

For more information, please see the full case study: Carbon Segmented Cloud Computes

Bulk Workload Shifting for Data Centers

Details: Shifting 1% of a data centers workload over a day would reduce overall emissions by Z.

Results: Analysis in-progress

For more information, please see the full case study: Bulk Workload Shifting for Data Centers

Recommendations

GSF SCI Specification