WaterWise: Co-optimizing Carbon- and Water-Footprint Toward Environmentally Sustainable Cloud Computing CCS: Social and professional topics Sustainability CCS: Computer systems organization Cloud computing

Yankai Jiang, Rohan Basu Roy, Raghavendra Kanakagiri, Devesh Tiwari · 2025

Read WaterWise: Co-optimizing Carbon- and Water-Footprint Toward Environmentally Sustainable Cloud Computing CCS: Social and professional topics Sustainability CCS: Computer systems organization Cloud computing by Yankai Jiang, Rohan Basu Roy, Raghavendra Kanakagiri, Devesh Tiwari (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-17944.txt.

Introduction

Yankai Jiang and co-authors introduce WaterWise in a 2025 paper on greener cloud computing. Large data centers use much power and much water for cooling, and both leave a mark on the planet. The team found a hard trade-off: cutting the carbon mark alone tends to raise the water mark, and cutting water alone does the reverse.

WaterWise is a scheduler that assigns jobs across data centers in many places while weighing both marks at once. By choosing where and when work runs, it aims to lower the joint harm rather than shift it from one form to the other. The paper frames the problem, shows why single-goal tools fall short, and presents the scheduler as a path to cloud work that is kinder on both counts.

Catalog record

CREATORYankai Jiang, Rohan Basu Roy, Raghavendra Kanakagiri, Devesh Tiwari
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE65 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2501-17944

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

WaterWise: Co-optimizing Carbon- and Water-Footprint Toward Environmentally Sustainable Cloud Computing CCS: Social and professional topics Sustainability CCS: Computer systems organization Cloud computing Yankai Jiang, Rohan Basu Roy, Raghavendra Kanakagiri, Devesh Tiwari Abstract . The carbon and water footprint of large-scale computing systems poses serious environmental sustainability risks. In this study, we discover that, unfortunately, carbon and water sustainability are at odds with each other – and, optimizing one alone hurts the other. Toward that goal, we introduce, WaterWise, a novel job scheduler for parallel workloads that intelligently co-optimizes carbon and water footprint to improve the sustainability of geographically distributed data centers. 1. Introduction Problem Motivation and Insights. The rapidly increasing carbon footprint of large-scale computing systems is posing serious environmental sustainability risks – and, has consequently received significant attention from the computer systems research community ( Gupta et al. 2021 ; Gupta et al. 2022 ; Eeckhout 2022 ; Li et al. 2023a ; Wu et al. 2022 ; Acun et al. 2023 ; Sukprasert et al. 2024 ; Wang et al. 2024 ; Hanafy et al. 2024 ; Anderson et al. 2023 ) . Unfortunately, the environmental impact of computing systems on another critical environmental sustainability factor — water footprint – has not been well-explored. A recent study has shown the dramatic side effects that parallel computing workloads, including ML training, place on our water resources ( Li et al. 2023b ) . Mitigating the water footprint challenge is critical to making large-scale cloud computing more sustainable. Without a focus on the water footprint of the computing infrastructure, we cannot achieve holistic sustainability of large-scale cloud computing systems. Motivated by these factors, in this study, we investigate the temporal and spatial interaction between the carbon and water footprints of the computing systems. Our experimental characterization reveals several insights (Sec. 2 and 3 ). First, we show that carbon-friendly energy sources (renewable energy sources) reduce the carbon footprint of computing systems – which is desirable, but the availability depends on the geographical location and energy sources may also require water to generate electricity used to power the data centers (offsite water usage). Unfortunately, some carbon-friendly energy sources (renewable sources) may need proportionally larger amounts of water to generate electricity – creating a conflict between carbon and water co-optimization (carbon-friendly energy sources may not be water-sustainable). Furthermore, water is also required to cool down the data centers (onsite water usage) which is significantly impacted by the weather conditions of the data center region – introducing geographical trade-offs for water sustainability . Motivated by these insights, we design a carbon- and water-aware job scheduler that aims to navigate these complex trade-offs in a geographically distributed data center serving parallel workloads. Contributions. The key contributions of WaterWise include the following highlights. Novel Carbon- and Water-Sustainability Job Scheduling Framework. WaterWise is a novel job scheduler for parallel workloads that identifies the need and opportunity for co-optimizing carbon and water footprint for large-scale systems, and designs a simple yet effective approach to improve the sustainability of geographically distributed data centers. WaterWise reveals that allowing a small amount of delay tolerance can bring additional significant carbon- and water-footprint savings – leveraging the characteristics that batch jobs can inherently tolerate delays ( Ambati et al. 2021 ; Zhang et al. 2022 ; Chen et al. 2019 ) . WaterWise Key Ideas and Approach. To navigate the competing carbon and water trade-offs, WaterWise designs and implements a Mixed Integer Linear Programming (MILP) based optimization – due to MILP’s inherent known advantage of configurability of objectives, low-overhead, and easier explainability ( Floudas and Lin 2005 ) . However, MILP methods as-is cannot operate most effectively in WaterWise’s context; to address this limitation, we augment the MILP design with the concept of delay tolerance, soft constraints, and slack management (Sec. 4 ). WaterWise dynamically and opportunistically exploits spatial and temporal variations in carbon and water intensity characteristics, while ensuring that jobs minimize the violations of their delay tolerance and consider inter-region movement transfer latency. Notably, WaterWise recognizes that one unit of water does not have the same value across all geographical regions (one liter of water in a water-stressed region may be more precious than in a water-abundant region ( Li et al. 2023b ) ). WaterWise systematically captures this effect of water scarcity factor ( Ritchie et al. 2024 ) in its scheduling decisions – filling a critical gap in computer system sustainability studies. Evaluation and Open-source Framework. WaterWise is evaluated using widely-used production traces such as Google Borg ( Wilkes 2020 ) and Alibaba ( Tian et al. 2019 ) , and parallel workloads from the PARSEC benchmark ( Zhan et al. 2017 ) . WaterWise provides approx. 21% carbon-footprint and approx. 14% water-footprint savings compared to the carbon- and water-unaware scheduler. WaterWise is available at https://zenodo.org/records/14219862 – the first open-source framework to enable exploration of carbon- and water-aware scheduling. 2. Background 2.1. Carbon Footprint in Data Center The carbon footprint in the data center consists of two parts: operational and embodied carbon footprint ( Gupta et al. 2022 ) . Operational Carbon Footprint. The operational carbon footprint refers to the carbon emissions incurred while executing the workloads in the data center. The electricity consumed by the data center may come from a mix of energy sources, including both fossil fuels and renewables. The operational carbon footprint can be estimated by multiplying the energy consumption (unit kWh) and the real-time carbon intensity (unit gCO 2 /kWh) of the energy mix used by the data center (lower is better). Carbon intensity ( \text{CO}_{2}^{\text{Intensity}} ) depends on the energy mixture of the electricity grid. Embodied Carbon Footprint. The embodied carbon footprint is a one-time carbon emission that occurs during server design and manufacturing. This footprint is amortized over the hardware’s lifespan. Essentially, the footprint attributed to a single job is the server’s embodied carbon footprint scaled by the job’s execution time over the server’s lifetime. Total Carbon Footprint Estimation. The carbon footprint for a job j is the sum of its operational and embodied carbon footprint and can be expressed as: (1) \displaystyle\begin{split}\text{CO}_{2,\textit{j}}&=\text{CO}_{2,\textit{j}}^{\text{operational}}+\text{CO}_{2,\textit{j}}^{\text{embodied}}\newline \\ \text{CO}_{2,\textit{j}}&=E_{\textit{j}}\cdot\text{CO}_{2}^{\text{Intensity}}+\frac{t_{j}}{T_{\text{lifetime}}}\cdot\text{CO}_{2,\text{server}}^{\text{embodied}}\end{split} Here, \text{CO}_{2}^{\text{Intensity}} represents the real-time carbon intensity of the regional power grid, and it’s multiplied by the energy ( E_{j} ) of the job to get the operational carbon. t_{j} denotes the execution time of the job, and T_{\text{lifetime}} is the overall lifetime of the server hardware. \text{CO}_{2,\text{server}}^{\text{embodied}} is the total embodied carbon footprint of the server. The embodied carbon footprint for job execution is calculated by multiplying the time proportion \frac{t_{j}}{T_{\text{lifetime}}} by the total embodied carbon footprint of the server. 2.2. Water Footprint in Data Center In 2023, Google published the data center’s water consumption usage of 5.6 billion gallons (equivalent to irrigating 37 golf courses annually) ( Google 2023 ) , and it observed a 24% increase compared to 2022. This sharp rise highlights the growing reliance on water-intensive cooling technologies to meet computing demands. If this trend continues without innovation in water footprint scheduling techniques, it could exacerbate water stress in regions where data centers are located, particularly in drought-prone areas. Moreover, reducing water usage in energy generation and cooling is crucial for preserving aquatic ecosystems and maintaining biodiversity. Water consumption during

Read the complete text