Spatio-Temporal Shifting to Reduce Carbon, Water, and Land-Use Footprints of Cloud Workloads

TL;DR

The paper tackles the environmental footprint of cloud workloads by introducing spatio-temporal shifting to reduce carbon, water, and land-use impacts. It presents a holistic, multi-dimensional optimization framework that accounts for grid mix, data-center efficiency, data migration costs, and a new land-use metric (LUE). Through realistic simulations using Azure and AWS traces for FaaS and Big Data workloads, the study shows substantial footprint reductions, particularly from spatial shifting (up to 85% in some metrics) and improved gains when combined with temporal adjustments. The results are robust to prediction errors and seasonal changes, offering practical guidance for sustainability-aware cloud orchestration and highlighting trade-offs with data sovereignty and migration overheads.

Abstract

In this paper, we investigate the potential of spatial and temporal cloud workload shifting to reduce carbon, water, and land-use footprints. Specifically, we perform a simulation study using real-world data from multiple cloud providers (AWS and Azure) and workload traces for different applications (big data analytics and FaaS). Our simulation results indicate that spatial shifting can substantially lower carbon, water, and land use footprints, with observed reductions ranging from 20% to 85%, depending on the scenario and optimization criteria. Temporal shifting also decreases the footprint, though to a lesser extent. When applied together, the two strategies yield the greatest overall reduction, driven mainly by spatial shifting with temporal adjustments providing an additional, incremental benefit. Sensitivity analysis demonstrates that such shifting is robust to prediction errors in grid mix data and to variations across different seasons.

Figures (7)

• Figure 1: Spatial shifting for Scenario 1; showing a $45-85\%$ improvement over local baseline during a winter week, with an MAE of $10\%$.
• Figure 2: Request distribution across different regions for different criteria for Scenario 1 during a Winter week, with an MAE of $10\%$.
• Figure 3: Grid intensities for each factor, and each region for Scenario 1 over a winter week, with an MAE of $10\%$.
• Figure 4: Temporal, spatial, and spatio-temporal shifting for Scenario 2 improvement over local baseline results, with an MAE of $10\%$. Temporal shifting shows $6-12\%$ improvement; spatial shifting comes in at $20-45\%$; while combining both gives the best result of $40-55\%$ improvement over the baseline.
• Figure 5: Request distribution across different regions for different criteria for Scenario 2 during a Winter week, with an MAE of $10\%$.