Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Towards Environmentally Equitable AI via Geographical Load Balancing

Pengfei Li, Jianyi Yang, Adam Wierman, Shaolei Ren

TL;DR

This paper tackles AI’s regional environmental inequity by introducing equity-aware geographical load balancing (eGLB) for AI model inference across 10 data centers. It formulates a minimax objective that minimizes total energy cost while constraining the worst-case regional carbon and water footprints using convex equity functions, and it develops an online dual mirror descent algorithm (eGLB) with auxiliary variables to enable online optimization. Empirical results on trace-based BLOOM workloads show that eGLB dramatically reduces the maximum regional footprints and achieves performance close to an offline optimum (eGLB-Off), whereas equity-oblivious GLB baselines can worsen inequity. The work demonstrates the potential of fairness-aware scheduling in data-center operations and outlines future paths for integrating non-IT resources and environmental science tools to further advance environmental justice in AI deployment.

Abstract

Fueled by the soaring popularity of large language and foundation models, the accelerated growth of artificial intelligence (AI) models' enormous environmental footprint has come under increased scrutiny. While many approaches have been proposed to make AI more energy-efficient and environmentally friendly, environmental inequity -- the fact that AI's environmental footprint can be disproportionately higher in certain regions than in others -- has emerged, raising social-ecological justice concerns. This paper takes a first step toward addressing AI's environmental inequity by balancing its regional negative environmental impact. Concretely, we focus on the carbon and water footprints of AI model inference and propose equity-aware geographical load balancing (GLB) to explicitly address AI's environmental impacts on the most disadvantaged regions. We run trace-based simulations by considering a set of 10 geographically-distributed data centers that serve inference requests for a large language AI model. The results demonstrate that existing GLB approaches may amplify environmental inequity while our proposed equity-aware GLB can significantly reduce the regional disparity in terms of carbon and water footprints.

Towards Environmentally Equitable AI via Geographical Load Balancing

TL;DR

This paper tackles AI’s regional environmental inequity by introducing equity-aware geographical load balancing (eGLB) for AI model inference across 10 data centers. It formulates a minimax objective that minimizes total energy cost while constraining the worst-case regional carbon and water footprints using convex equity functions, and it develops an online dual mirror descent algorithm (eGLB) with auxiliary variables to enable online optimization. Empirical results on trace-based BLOOM workloads show that eGLB dramatically reduces the maximum regional footprints and achieves performance close to an offline optimum (eGLB-Off), whereas equity-oblivious GLB baselines can worsen inequity. The work demonstrates the potential of fairness-aware scheduling in data-center operations and outlines future paths for integrating non-IT resources and environmental science tools to further advance environmental justice in AI deployment.

Abstract

Fueled by the soaring popularity of large language and foundation models, the accelerated growth of artificial intelligence (AI) models' enormous environmental footprint has come under increased scrutiny. While many approaches have been proposed to make AI more energy-efficient and environmentally friendly, environmental inequity -- the fact that AI's environmental footprint can be disproportionately higher in certain regions than in others -- has emerged, raising social-ecological justice concerns. This paper takes a first step toward addressing AI's environmental inequity by balancing its regional negative environmental impact. Concretely, we focus on the carbon and water footprints of AI model inference and propose equity-aware geographical load balancing (GLB) to explicitly address AI's environmental impacts on the most disadvantaged regions. We run trace-based simulations by considering a set of 10 geographically-distributed data centers that serve inference requests for a large language AI model. The results demonstrate that existing GLB approaches may amplify environmental inequity while our proposed equity-aware GLB can significantly reduce the regional disparity in terms of carbon and water footprints.
Paper Structure (24 sections, 4 theorems, 22 equations, 3 figures, 5 tables, 1 algorithm)

This paper contains 24 sections, 4 theorems, 22 equations, 3 figures, 5 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Problem Formulation
  3. Environmentally Equitable GLB
  4. Objective
  5. An Online Algorithm
  6. Experiments
  7. Methodology
  8. Workload Trace
  9. Data Centers
  10. Energy Price, Carbon Intensity, and WUE
  11. Offline and Online Optimization
  12. Metrics
  13. Baseline Algorithms
  14. Results
  15. GLB with Full Flexibility
  16. ...and 9 more sections

Key Result

Lemma 1

The transformed problem eqn:objective_dual--eqn:constraint_gateway_dual_2 and the original problem eqn:objective--eqn:constraint_gateway have the same optimal GLB decisions.

Figures (3)

  • Figure 1: The energy cost, carbon and water footprint of eGLB with different learning rates $\eta$ under full GLB flexibility. The results for eGLB-Off, GLB-Energy, GLB-Carbon, and GLB-Water are shown for comparison.
  • Figure 2: The energy cost, carbon footprint, and water footprint of eGLB with different $(\mu_w,\mu_c)$ shown in the legend under full GLB flexibility. The results for eGLB-Off, GLB-Energy, GLB-Carbon, and GLB-Water are also shown for comparison.
  • Figure 3: The energy cost, carbon and water footprint of eGLB with different learning rates $\eta$ under partial GLB flexibility. The results for eGLB-Off, GLB-Energy, GLB-Carbon, and GLB-Water are shown for comparison.

Theorems & Definitions (4)

  • Lemma 1
  • Theorem 1
  • Lemma 2
  • Lemma 3