Source-linked AI summary
Let's Wait Awhile: How Temporal Workload Shifting Can Reduce Carbon Emissions in the Cloud
Philipp Wiesner, Ilja Behnke, Dominik Scheinert, Kordian Gontarska, Lauritz Thamsen
TL;DR
Data-center emissions depend on when electricity is consumed, but regional variation in grid carbon intensity and limited prior evidence complicate carbon-aware scheduling. This paper characterizes delay-tolerant workloads, analyzes temporal shifting across four regions in 2020, and simulates scheduling scenarios. The results show substantial savings from shifting toward cleaner periods, with interruptibility and forecast quality affecting scheduling outcomes.
Problem
Existing carbon-aware research emphasizes geographic workload migration or grid stability and costs, leaving temporal shifting toward cleaner grid energy insufficiently examined.
Method
The paper categorizes delay-tolerant workloads, estimates regional carbon intensity from generation and imports, analyzes four regions in 2020, and simulates two shifting scenarios.
Results
Interrupting scheduling outperforms Non-Interrupting scheduling even with 10% forecast errors, while weekend shifting can produce more than 20% savings in most regions.
Takeaways & Limitations
Flexible execution windows and workload snapshots can support carbon-aware temporal shifting without requiring geographic migration between data centers.
Takeaways & Limitations
Forecast-error analyses use uniform random noise rather than realistic correlated errors, limiting their validity and motivating evaluation with actual forecasts.
Abstract
from arXiv · showhide
Depending on energy sources and demand, the carbon intensity of the public power grid fluctuates over time. Exploiting this variability is an important factor in reducing the emissions caused by data centers. However, regional differences in the availability of low-carbon energy sources make it hard to provide general best practices for when to consume electricity. Moreover, existing research in this domain focuses mostly on carbon-aware workload migration across geo-distributed data centers, or addresses demand response purely from the perspective of power grid stability and costs. In this paper, we examine the potential impact of shifting computational workloads towards times where the energy supply is expected to be less carbon-intensive. To this end, we identify characteristics of delay-tolerant workloads and analyze the potential for temporal workload shifting in Germany, Great Britain, France, and California over the year 2020. Furthermore, we experimentally evaluate two workload shifting scenarios in a simulation to investigate the influence of time constraints, scheduling strategies, and the accuracy of carbon intensity forecasts. To accelerate research in the domain of carbon-aware computing and to support the evaluation of novel scheduling algorithms, our simulation framework and datasets are publicly available.
1 INTRODUCTION
Data-center electricity use drives both operating expenses and cloud carbon emissions, while grid carbon intensity varies by energy source, region, season, and hour. The paper investigates shifting delay-tolerant workloads toward periods when public-grid energy is expected to be cleaner.
- Motivation: Data-center energy consumption is a major concern because it drives operational expenses and largely determines cloud-computing carbon emissions.Worldwide data centers consumed an estimated 205 TWh in 2018, approximately 1% of global energy consumption.
- Motivation: Carbon-aware scheduling must consider energy sources, not only electricity quantity, because source availability varies over time and across regions.The paper notes variability across regions, seasons, and hours due to weather, installed generation capacity, and inter-regional imports.
- Research scope: The paper targets temporal shifting of delay-tolerant workloads toward periods when the public grid is expected to provide low-carbon energy.It explicitly distinguishes this goal from saving energy or integrating local generation.
- Terminology: The paper uses CO2eq as its common term when referring to carbon or CO2 emissions.CO2eq represents the amount of CO2 needed to produce an equivalent warming effect for a given greenhouse gas.
- Contributions: The study identifies workload characteristics, estimates regional grid carbon intensity, analyzes four regions, and evaluates shifting scenarios through simulation.The regions are Germany, Great Britain, France, and California; the simulation examines time constraints, scheduling strategies, and forecast accuracy.
2 SHIFTABLE WORKLOADS
Workload-shifting potential depends chiefly on timing constraints, duration, execution pattern, and interruptibility. The paper distinguishes short-, long-, and continuously running workloads, and separates ad hoc jobs from scheduled workloads.
- Workload characteristics: Time constraints are the most important determinant of whether and how a workload can be shifted.Duration, execution time, and interruptibility also substantially affect shifting potential.
- Short-running workloads: Short-running workloads are common, and their shifting potential depends strongly on how strictly completion timing is enforced.More than 90% of Alibaba batch jobs run less than 15 minutes; examples include FaaS executions and CI/CD runs.
- Long-running workloads: Long-running workloads can have substantial absolute shifting potential because they are often energy-intensive and may tolerate completion overnight or the following morning.Examples include machine-learning training, scientific simulations, and big-data analysis; most production-priority jobs in Google traces were long-running.
- Continuously running workloads: Continuous services and workloads lasting weeks or months are outside the paper’s shiftable-workload scope.The study covers workloads up to several days because available carbon-intensity forecasts extend only a few days.
- Execution timing: Ad hoc workloads can be deferred only after issuance, whereas scheduled workloads planned for future execution can potentially shift in either temporal direction.Scheduled workloads include recurring batch jobs such as nightly builds, backups, database-index updates, and reports.
2.3 Interruptibility
Interruptibility determines whether a scheduler can divide a workload and place its pieces in lower-carbon periods. Checkpointing enables pausing and resuming, whereas setup costs or consistency requirements can make interruption impractical.
- Interruptible workloads: Interruptible workloads can be divided into chunks and scheduled separately to better align execution with low-carbon energy.This scheduling opportunity follows from exploiting workload interruptibility.
- Interruptible workloads: Checkpoint mechanisms allow long-running jobs such as iterative machine-learning training and discrete-event simulations to pause and resume later.Checkpoints support later analysis, resumption from earlier states, and error handling.
- Non-interrupting workloads: Some workloads should execute without interruption because startup and shutdown costs outweigh expected benefits or interruption risks inconsistency.Examples include certain CI/CD and compile jobs, database migrations, backups, and many test suites.
3 REGIONAL CARBON INTENSITY
The study estimates regional carbon intensity from electricity generation and imports, using openly available 2020 data for Germany, Great Britain, France, and California. It compares regional and temporal variation while noting limits in data coverage and marginal-intensity estimation.
- 3.1 Region Selection: The analysis covers Germany, Great Britain, France, and California, selected for cloud-provider availability, data access, and regional energy diversity.Open production data with at least hourly reporting for 2020 was required, and the selected regions represent Europe and the United States.
- 3.1 Region Selection: Open-data constraints limited geographic coverage: the study could not include southern-hemisphere or several emerging-market regions, so regional diversity is only partially achieved.All selected regions are located in Europe or the United States.
- 3.3 Carbon Intensity of Regions: Regional carbon intensity is calculated by weighting each energy source’s generation by its life-cycle carbon intensity and incorporating imported electricity.The weighted total is divided by all generated and imported electricity, with imports assigned neighboring regions’ average carbon intensity.
- 3.3 Carbon Intensity of Regions: The dataset covers all of 2020, combines ENTSO-E data for the European regions with CAISO data for California, and standardizes observations to 30-minute resolution.Energy sources were mapped to common categories, while cross-border flows used neighboring regions’ yearly average carbon intensity.
- 3.4 Average vs. Marginal Carbon Intensity: Average carbon intensity is used instead of marginal carbon intensity because the marginal energy source is difficult to identify and available estimates have high uncertainty.Supplier dispatch decisions are decentralized and depend on prices, weather, demand, and neighboring-region conditions.
4 ANALYSIS OF THEORETICAL POTENTIAL
The analysis examines temporal carbon-intensity patterns in four regions and identifies when shifting workloads could reduce their carbon intensity. The best opportunities vary by region and generally expand with longer flexibility windows.
- Regional carbon-intensity patterns: Germany has the highest mean carbon intensity at 311.4 gCO2/kWh, with high variation linked to its mixed renewable and fossil generation.Wind and solar comprise 24.7% and 8.3% of production, while lignite, coal, and gas account for substantial remaining generation.
- Regional carbon-intensity patterns: Great Britain averages 211.9 gCO2/kWh, while France averages 56.3 gCO2/kWh with comparatively low variability.Great Britain is cleanest at night; France remains low and steady throughout the year.
- Weekly patterns: Weekend carbon intensity decreases by 25.9% in Germany, 20.7% in Great Britain, 22.2% in France, and 6.2% in California.The paper attributes the drop to lower weekend demand and reduced fossil-fuel generation.
- Best times of day for shifting: 44% of days in 2020 offered California workloads scheduled at 6 am a reduction exceeding 80 gCO2/kWh by shifting them to 6–8 am.California also shows substantial potential for shifting from after sunset to before sunset, while other regions differ by time of day.
- Best times of day for shifting: Larger forecast windows substantially increase shifting potential, but optimal times differ: early morning is favored in solar-rich regions, whereas evening is favored where fossil generation is reduced at night.Shifting into the past can complement future shifting, although it requires workloads scheduled for future execution.
5 EXPERIMENTAL EVALUATION
The experimental evaluation moves from theoretical potential to two simulated workload-shifting scenarios. It examines how time constraints, scheduling strategies, and forecast errors affect carbon savings.
- Experimental evaluation: Two realistic load-shifting scenarios are evaluated experimentally because openly available cloud datasets describing workload delay tolerance are unavailable.The scenarios feature short-running periodically scheduled jobs and long-running workloads.
- Experimental evaluation: The simulations examine the effects of workload time constraints, scheduling strategies, and carbon-intensity forecast errors.
- Experimental evaluation: The evaluation uses simulation to assess temporal workload shifting under realistic workload scenarios.
5.1 Scenario I: Nightly Jobs
The nightly-jobs scenario models delay-tolerant periodic work whose scheduling window can expand around a fixed baseline. Carbon savings increase with flexibility, but the magnitude depends strongly on region, timing, and forecast accuracy.
- Experimental setup: The jobs represent delay-tolerant nightly builds, integration tests, or database migrations that can run outside working and high-traffic hours.
- Experimental setup: The scenario simulates 366 non-interruptible, 30-minute jobs in 2020, with baseline execution fixed at 1 am.Experiments expand the scheduling window by 30 minutes in both directions at each step.
- Experimental setup: A 5% forecast error is simulated by adding normally distributed noise, based on a 2020 48-hour forecast mean absolute error of 10.
- Results: 3.0% savings in France and 4.3% in Great Britain are achieved with only ±2-hour flexibility, while ±8 hours yields 4.1% and 7.4%, respectively, at 5% forecast error.
- Results: 11.2% lower emissions are achieved in Germany with ±8 hours under forecast error, while California reaches 13.1% savings at ±6 hours and 33.7% at ±8 hours.Germany’s savings are more than 2 percentage points higher with optimal forecasts; California’s improve by only 1–1.5 percentage points.
- Discussion: Carbon-aware scheduling can reduce emissions by more than 10% in the other regions without negative impact on data-center operations.The discussion connects these results to time-window-based service agreements and middleware behavior.
5.2 Scenario II: Machine Learning Project
The machine-learning scenario evaluates temporal workload shifting under two deadline constraints and two scheduling strategies. Longer flexibility windows and interruptibility produce greater carbon savings, while forecast errors affect interrupting scheduling more strongly.
- Scenario setup: The scenario models 3387 machine-learning jobs, randomly distributed across 262 workdays, with jobs averaging almost two days and typically using eight GPUs.The workload is based on the StyleGAN2-ADA project’s reported energy statistics.
- Time constraints: Next Workday allows jobs ending outside working hours to shift until 9 am the next working day, while Semi-Weekly allows shifting until the next Monday or Thursday at 9 am.Under Next Workday, 20.4% of jobs are not shiftable, 51.2% are shiftable until the next morning, and 28.4% are shiftable further.
- Scheduling strategies: Interrupting scheduling selects the lowest-carbon 30-minute intervals and splits jobs, whereas Non-Interrupting scheduling selects one coherent window with the lowest average carbon intensity.The baseline starts jobs immediately when issued.
- Results: Under Next Workday, Non-Interrupting scheduling saved 2.5% to 6.3% and Interrupting scheduling saved 5.7% to 8.5% relative to regional baselines.The experiments used a 5% forecast error.
5.3 Limitations
The experiments omit explicit resource-capacity constraints and use simplified forecast-error models. These assumptions limit how directly the simulated savings and forecast-error findings transfer to production settings.
- Resource constraints: The experiments omit resource constraints such as available computational capacity, although scheduled activity exceeded baseline activity by no more than 42% (64 compared to 45).The authors note this assumption is more consequential for the machine-learning scenario than for Scenario I.
- Forecast errors: Forecast-error validity is limited because the simulations apply uniform random noise, whereas real errors can be correlated, forecast-horizon dependent, and larger during variable daylight periods.The authors call for analyses using actual forecasts across different regions.
5.4 Implications
The paper recommends exposing workload flexibility and carbon-intensity information to carbon-aware schedulers and service interfaces. It also emphasizes that regional conditions constrain the usefulness of these practices.
- Cloud and service providers: Providers should encourage users to declare workloads as temporally flexible or interruptible so schedulers can exploit carbon-intensity fluctuations.The paper points to preemptive or Spot VMs as an existing mechanism that can shape cloud load.
- Cloud and service providers: Carbon-aware load shaping may become financially attractive, but its usefulness is limited to some locations and requires regular reevaluation because regional carbon intensity and pricing vary.The implication is explicitly tied to regional conditions.
- Cloud and service providers: Execution windows such as nightly intervals can increase workload flexibility and carbon-saving potential compared with exact execution times.The potential depends strongly on the data-center region.
- Schedulers and middleware: Carbon-intensity and workload forecasts are core scheduler inputs, and scheduler performance depends on workload properties such as temporal constraints.The paper notes that short-term carbon-intensity forecasts can often be highly accurate.
- Schedulers and middleware: Middleware should provide interfaces for applications to declare temporal constraints and workload properties, while automatically detecting characteristics such as interruptibility where possible.Profiling stop-and-resume time is given as an example of automatic detection.
6 RELATED WORK
Prior work largely emphasizes renewable-energy integration, grid-stability demand response, or geographic workload migration rather than temporal shifting based on public-grid carbon intensity. This paper positions its contribution within that narrower gap.
- Renewable-aware scheduling: Renewable-aware scheduling literature mainly integrates on-site or off-site renewable installations, with relatively few studies considering public-grid carbon intensity.Single-data-center work similarly focuses on predicting and using local solar or wind generation.
- Temporal workload shifting: Temporal shifting was identified earlier as an emissions-reduction strategy, but Google’s CICS is described as the only prior work using it, without reporting impacts across regions.CICS shapes compute load using current and predicted power-grid conditions.
- Demand response: Demand response generally adjusts consumption to relieve grid capacity stress and is primarily evaluated for cost reduction rather than emissions reduction.Data centers are considered promising because their automated workloads are flexible and energy-intensive.
- Research gap: This paper evaluates temporal workload shifting specifically for carbon savings, contrasting with demand-response work focused on grid stability and cost effectiveness.The stated objective is to assess carbon savings rather than power-demand adaptation alone.
- Carbon-intensity forecasting: Short-term grid carbon-intensity prediction has received comparatively little research compared with long-term emissions forecasting for countries or industrial sectors.The related-work discussion identifies this as a separate forecasting gap.
7 CONCLUSION
Temporal workload shifting can reduce data-center emissions by directing delay-tolerant computation toward cleaner grid electricity. The paper’s findings motivate scheduler development and further study using the released simulator and datasets.
- Over 20% savings are possible when delay-tolerant workloads are shifted toward weekends in most regions.The analysis covers Germany, California, Great Britain, and France over 2020.
- Short-term shifting potential is often high before sunrise in solar-rich countries and during evening hours when fossil-fuel generation is reduced.
- More than 5% emissions reductions are achievable across all regions when workload results are not needed by the next working day.
- Future work will develop and evaluate schedulers that exploit these temporal-shifting findings, supported by the paper’s simulator and published datasets.The authors also identify combining temporal and geo-distributed scheduling as a future research direction.