Cloud applications and processes have become a part of our daily lives: think of web apps, working on documents in real time seamlessly, or backing up your data via a cloud storage. But cloud computing also impacts the climate as data centers consume around 1% of global electricity, tendency rising (Freitag et al., 2021). Methods reducing the environmental impact of data centers will be essential for achieving climate neutrality in the EU by 2050 (European Climate Law, 2021). On the software side, the idea of load shifting can help reduce greenhouse gas emissions. Briefly speaking, this concept means shifting (moving) a load (task) either from one data center to another or in time. The emission reduction potential of load shifting lies in the differences in carbon intensity of the power grid supplying energy: Performing a task when a lot of renewable energy is generated results in lower emissions.
Understanding the Prerequisites
When researching the carbon reduction potential, it is important to focus on the actual carbon emissions of the energy used. A task emits the amount of carbon relating to the current carbon intensity of the energy used, even if the cloud provider promises net-zero carbon emissions. The term ”net-zero carbon” only means that the cloud provider offset any emission, for example by investing in reforestation projects, not necessarily reducing emissions. However, the potential of load shifting lies in the actual reduction of emissions caused by a workload.
While it might sound easy to just use load shifting, there are some prerequisites and technical limitations. Temporal load shifting inherently requires flexibility of when the results are expected, as part of the process is to defer jobs to a less carbon intensive time. For example, a backup that is run every night can be shifted in a specific time frame, as the job only needs completion the next morning. In contrast, verifications of a payment cannot be shifted in time, as the vendors expect results just in time.
Geographical or spatial load shifting instead requires flexibility across locations and jurisdictions. Here, the expected latency is crucial, as latency increases when the data center is further away from the user which might be important for interactive real-time workloads. When shifting across countries, data privacy laws are particularly important but other aspects such as the country of jurisdiction or political stability of the involved countries must not be overlooked.
Carbon Savings and their Responsibility
This flexibility in execution is the base for carbon emission reduction. For temporal load shifting, depending on the time of the day, the carbon emissions produced by electrical energy production varies drastically [1]. When a job offers enough flexibility to wait until the carbon intensity of the power grid is lower, emissions can be reduced. Geographical load shifting is also based on the carbon intensity of the grid, but utilizes the differences between countries.
Let us shortly dive into findings regarding the carbon impact of both types of load shifting. While temporal load shifting has received limited research, several studies indicate its potential. For example, Wiesner et al. (2021) demonstrated that shifting workloads can achieve emissions reductions of up to 33.7% in California and around 3-11% in European countries, depending on flexibility. Similarly, simulated workloads like machine learning tasks revealed further potential for savings, particularly when users shift interruptible jobs. Radovanovi et al. (2023) highlighted how automated systems can reduce emissions by 1-2%, bypassing the need for user involvement. However, Sukprasert et al. (2024) pointed out that long-running jobs, which dominate data center resources, offer limited savings, reducing the overall impact of temporal load shifting. Hanafy et al. (2024) also noted the trade-offs involved, as reducing emissions often comes with increased costs or reduced performance. All in all, while temporal load shifting has potential, its real-world benefits are constrained and often modest.
On spatial load shifting, Sukprasert et al. (2024) explored the theoretical and realistic potential of spatial load shifting, revealing that while a 96% reduction could be achieved in theory, real-world factors like underutilization of origin data centers, operational costs, and emissions reduce its effectiveness. Furthermore, migrating to the greenest data center provides the greatest benefit, as frequent hopping between centers yields marginal gains. Hence, the analysis of use cases can help saving emissions. Souza et al. (2023) examined latency-sensitive applications, showing that allowing for increased latency can cut emissions by up to 70% without significantly compromising performance. While spatial load shifting offers significant potential, particularly when moving workloads from carbon-intensive regions to greener ones, it faces technical and non-technical limitations. Factors like latency, data privacy laws, and regional restrictions complicate its application, and as more countries adopt renewable energy, the carbon reduction potential will diminish. However, spatial load shifting remains a key strategy for minimizing emissions in a more volatile energy landscape.
These findings yield an interesting take on responsibility. While the potential for temporal load shifting is low when continuing as known, there is high potential when rethinking how computing jobs are designed. Short duration jobs and interruptible jobs have the highest savings, and long duration interruptible jobs that only need to be finished within a week offer great savings as well. If the product, job or software allows for such parameters to be tweaked, software developers can significantly reduce carbon emissions of their product but may need to accept higher costs and reduced performance depending on the expectations. To make this clear: Cloud providers who just get the temporal circumstances of a job as requirements will not reduce carbon emissions drastically when applying temporal load shifting. The costumer companies buying those jobs can make a difference by relaxing the requirements and adjusting temporal needs to actual business behavior instead of requesting completion as quick as possible. In contrast, in the case of spatial load shifting the responsibility lays mostly with the cloud service providers as they know best about technical and legal boundaries as well as energy consumption and marginal changes in carbon intensity. The data center providers can maximize their carbon reductions by intelligently mixing those parameters. In situations where latency is crucial, the costumer must decide how much relaxation in requirements they can accept, hence, depending on the use case some responsibility must be undertaken by the costumer themselves.
Going forward, these aspects on responsibility of emissions can and should be used to define a framework on carbon reductions regarding cloud applications. With a broad framework the savings can be maximized as every job shifted reduced emissions just by a small percentage. The use of such a framework could start by adding recommendations to the Code of Conduct2 (CoC) or requiring it within the Green Public Procurement (GPP) criterias [3]. Ultimately, such a framework can lead to reducing the idle workload of servers for connected server locations. In this scenario the overall available workload of connected server locations could be compared, and then decided which locations to be shut down, and the workload shifted to one location, for example with the lowest carbon emissions.
Looking forward
Both temporal and spatial load shifting have the potential to reduce carbon emissions of cloud computing jobs, particularly when combining both concepts. Of course, it will not be suitable for each and every application to utilize the full potential of carbon reductions. Let’s think about some examples: When a new AI model is trained or researched, this task can easily be shifted to a data center with less carbon emissions. A user sending a text message does not recognize if the latency is some milliseconds longer, but resolving this in a less carbon intensive region can reduce emissions drastically.
Though, one must be aware that the whole concept of load shifting is based on differences in the carbon intensity of the power consumed. Once the whole world or even just a region like Europe has shifted to 100% renewable energies, there will be no more savings. Until this transition is finally done, the concept of load shifting can help reduce carbon emissions. Utilizing even small scale effects to reduce carbon emissions from data centers will be essential in tackling climate change, especially considering findings by Freitag et al. (2021) that gains in energy efficiency alone cannot counter the growth in demand in capacity.
Footnotes
[1] See: Electricity Maps, https://www.electricitymaps.com/data-portal
[2] See: Data Centres Code of Conduct, https://e3p.jrc.ec.europa.eu/communities/data-centres-code-conduct
[3] See: Green Public Procurement, Criteria for data centres, server rooms and cloud services, https://circabc.europa.eu/ui/group/44278090-3fae-4515-bcc2-44fd57c1d0d1/library/24bf5149-d99b-4bc9-a7fc-132b711c46ce/details
References
European Climate Law (2021). Regulation (EU) 2021/1119 of the European Parliament and of the Council of 30 June 2021 establishing the framework for achieving climate neutrality and amending Regulations (EC) No 401/2009 and (EU) 2018/1999 (European Climate Law).
Freitag, C., Berners-Lee, M., Widdicks, K., Knowles, B., Blair, G. S., & Friday, A. (2021). The real climate and transformative impact of ICT: A critique of estimates, trends, and regulations. Patterns, 2(9), 100340.
Hanafy, W. A., Liang, Q., Bashir, N., Souza, A., Irwin, D., & Shenoy, P. (2024). Going Green for Less Green: Optimizing the Cost of Reducing Cloud Carbon Emissions. In Proceedings of the 29th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 3, ASPLOS ’24 (pp. 479–496). New York, NY, USA: Association for Computing Machinery.
Radovanovi, A., Koningstein, R., Schneider, I., Chen, B., Duarte, A., Roy, B., Xiao, D., Haridasan, M., Hung, P., Care, N., Talukdar, S., Mullen, E., Smith, K., Cottman, M., & Cirne, W. (2023). Carbon-Aware Computing for Datacenters. IEEE Transactions on Power Systems, 38(2), 1270–1280.
Souza, A., Shruti, J., Chakrabarty, B., Bridgwater, A., Lundberg, A., Skogh, F., Ali-Eldin, A., Irwin, D., & Shenoy, P. (2023). CASPER: Carbon-Aware Scheduling and Provisioning for Distributed Web Services. In Proceedings of the 14th International Green and Sustainable Computing Conference (IGSC), Toronto, ON, Canada.
Sukprasert, T., Souza, A., Bashir, N., Irwin, D., & Shenoy, P. (2024). On the Limitations of Carbon-Aware Temporal and Spatial Workload Shifting in the Cloud. In Proceedings of the Nineteenth European Conference on Computer Systems, EuroSys ’24 (pp. 924–941). New York, NY, USA: Association for Computing Machinery.
Wiesner, P., Behnke, I., Scheinert, D., Gontarska, K., & Thamsen, L. (2021). Let’s wait awhile: how temporal workload shifting can reduce carbon emissions in the cloud. In Proceedings of the 22nd International Middleware Conference, Middleware ’21 (pp. 260–272). New York, NY, USA: Association for Computing Machinery.