Carbon-Aware Computing for Datacenters

Overview

The CO2 emitted per kilowatt-hour on an electricity grid varies substantially by time of day and by location due to the types of generation, so hyperscale computing emits more carbon than needed if operated without regard to these variations. The paper introduces Google's system, which minimises electricity-based carbon footprint and power infrastructure costs by delaying temporally flexible workloads. Since datacentres are planned based on peak power and resource usage, smaller peaks reduce the need for more capacity. The core of the system is a suite of analytical pipelines that gather the next day's carbon intensity forecasts and train day-ahead demand prediction models.

Flexible workloads (for example data compaction, ML, simulation, video processing) tolerate delays as long as work completes within 24 hours; user-facing services and cloud VMs are not shifted. Day-ahead forecasts of carbon intensity and of flexible and inflexible compute demand feed a risk-aware optimisation that produces hourly Virtual Capacity Curves (VCCs) limiting resources for flexible jobs in each cluster. Cluster power is estimated from CPU usage with piecewise-linear power models, which achieve daily MAPE below 5% for more than 95% of power domains. Day-ahead carbon-intensity forecast MAPE across Google grid locations ranges between 0.4% and 26% for horizons of 8-32 hours.

Operational data show the VCCs cut flexible load at high-carbon hours and reduce peaks: in one cluster the VCC dropped flexible load by roughly 50% during peak carbon-intensity hours, giving an 8% drop in power during those hours. In another cluster forecast uncertainty pushed the average VCC about 33% above average load demand, so the roughly 8% power reduction lasted only 3 hours.

Limits: benefits depend on forecast uncertainty, the share of flexible load and the carbon-intensity profile, and measurement choices such as average versus marginal carbon affect results. The measurements are from one company's own fleet and the flexibility is limited to a 24-hour window.

Key points

1. Flexible workloads (data compaction, ML, simulation, video processing) tolerate delays as long as work completes within 24 hours; user-facing services and cloud VMs are not shifted.
2. A piecewise-linear model of power versus CPU usage achieves daily MAPE below 5% for more than 95% of power domains.
3. Day-ahead carbon-intensity forecast MAPE across Google grid locations ranges between 0.4% and 26% for horizons of 8-32 hours.
4. In one cluster the VCC dropped flexible load by roughly 50% during peak carbon-intensity hours, giving an 8% drop in power during those hours.
5. In another cluster forecast uncertainty pushed the average VCC about 33% above average load demand, so the roughly 8% power reduction lasted only 3 hours.

Using it in design

• Consider shifting temporally flexible workloads (completable within 24 hours) to hours of lower carbon intensity and lower power peak as an option in data-centre grid-connection negotiation and power capacity planning.
• Design carbon-aware scheduling risk-aware with day-ahead carbon-intensity (MAPE 0.4-26%) and demand forecast uncertainty; high uncertainty can shorten the benefit of flexibility.
• Define the flexible-load share and that user-facing services are not shifted as boundary conditions.

Who should read it

Operators and researchers developing carbon-aware workload scheduling and data-centre power peak reduction strategies.

Document contents

1. I. Introduction
2. II. Power architecture
3. III. Mechanism for carbon- and cost-aware load shaping
4. IV. Power modeling
5. V. Day-ahead forecasting
6. VI. Optimization framework
7. VII. Discussion

Read the document

Publication details

Authors

Ana Radovanović; Ross Koningstein; Ian Schneider; Bokan Chen; Alexandre Duarte; Binz Roy; Diyue Xiao; Maya Haridasan; Patrick Hung; Nick Care; Saurav Talukdar; Eric Mullen; Kendal Smith; MariEllen Cottman; Walfredo Cirne

Publisher

IEEE Transactions on Power Systems (Institute of Electrical and Electronics Engineers (IEEE))

Year

2021

Document number

arXiv:2106.11750

Volume

38

Issue

2

Pages

1270-1280

Document language

English

Source type

Scientific

Access

Open licence — downloadable

DOI

10.1109/TPWRS.2022.3173250

Licence

CC BY 4.0 — share with attribution

Licence evidence

arxiv.org

Citation