Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

The Effect of the Network in Cutting Carbon for Geo-shifted Workloads

Yibo Guo, Amanda Tomlinson, Runlong Su, George Porter

TL;DR

This work tackles the problem of reducing the carbon footprint of cloud workloads by intelligently migrating computation across geographies (space-shifting) while accounting for the often-significant carbon cost of WAN data transfers. It advances the state of the art by improving CIDT estimation through combining traceroute data with inferred physical routes, and by providing a public route dataset and a linear-time optimization algorithm that jointly considers compute and transfer emissions. Through evaluation on representative workloads, time- and region-varying carbon intensities, and load-balancing strategies, the study shows that space-shifting can deliver substantial carbon savings—often outperforming time-shifting alone—especially for low-data workloads, and that careful route- and region-selection is essential to realize these benefits. The findings highlight practical implications for carbon-aware cloud orchestration, suggesting that operators should account for WAN carbon costs, utilize distributed data-center inventories, and consider timezone and seasonal effects to maximize reductions in real-world deployments.

Abstract

Organizations are increasingly offloading their workloads to cloud platforms. For workloads with relaxed deadlines, this presents an opportunity to reduce the total carbon footprint of these computations by moving workloads to datacenters with access to low-carbon power. Recently published results have shown that the carbon footprint of the wide-area network (WAN) can be a significant share of the total carbon output of executing the workload itself, and so careful selection of the time and place where these computations are offloaded is critical. In this paper, we propose an approach to geographic workload migration that uses high-fidelity maps of physical Internet infrastructure to better estimate the carbon costs of WAN transfers. Our findings show that space-shifting workloads can achieve much higher carbon savings than time-shifting alone, if accurate estimates of WAN carbon costs are taken into account.

The Effect of the Network in Cutting Carbon for Geo-shifted Workloads

TL;DR

This work tackles the problem of reducing the carbon footprint of cloud workloads by intelligently migrating computation across geographies (space-shifting) while accounting for the often-significant carbon cost of WAN data transfers. It advances the state of the art by improving CIDT estimation through combining traceroute data with inferred physical routes, and by providing a public route dataset and a linear-time optimization algorithm that jointly considers compute and transfer emissions. Through evaluation on representative workloads, time- and region-varying carbon intensities, and load-balancing strategies, the study shows that space-shifting can deliver substantial carbon savings—often outperforming time-shifting alone—especially for low-data workloads, and that careful route- and region-selection is essential to realize these benefits. The findings highlight practical implications for carbon-aware cloud orchestration, suggesting that operators should account for WAN carbon costs, utilize distributed data-center inventories, and consider timezone and seasonal effects to maximize reductions in real-world deployments.

Abstract

Organizations are increasingly offloading their workloads to cloud platforms. For workloads with relaxed deadlines, this presents an opportunity to reduce the total carbon footprint of these computations by moving workloads to datacenters with access to low-carbon power. Recently published results have shown that the carbon footprint of the wide-area network (WAN) can be a significant share of the total carbon output of executing the workload itself, and so careful selection of the time and place where these computations are offloaded is critical. In this paper, we propose an approach to geographic workload migration that uses high-fidelity maps of physical Internet infrastructure to better estimate the carbon costs of WAN transfers. Our findings show that space-shifting workloads can achieve much higher carbon savings than time-shifting alone, if accurate estimates of WAN carbon costs are taken into account.

Paper Structure

This paper contains 49 sections, 10 equations, 9 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Motivations and Background
  3. Motivations
  4. Increasing renewable deployment
  5. Lower cost in WAN vs power grid
  6. Background and Carbon Modeling
  7. Carbon Intensity of Power Grids and ISOs
  8. CIDT: Carbon Intensity of Data Transfer
  9. Carbon Emissions of Workload Migration
  10. Network Transfer Costs
  11. Endpoint Transfer Costs
  12. Motivating example
  13. Compute Heavy Example:
  14. Related Work
  15. Network Carbon Costs and Routes
  16. ...and 34 more sections

Figures (9)

  • Figure 1: The carbon intensity of different regions in the northern hemisphere on June 1st, 2023. The routes between AWS datacenters us-west-2, us-east-1, and eu-north-1 are shown with their CIDTs.
  • Figure 2: Comparison of all route accounting methods
  • Figure 3: Comparison of time shifting vs. space shifting for two different datacenters. Time shifting is most effective for shorter jobs with long deadlines. Space and time shifting can match or beat time shifting in every scenario, with up to 90% reduction in some cases.
  • Figure 4: Migration overhead using different CIDT values
  • Figure 5: Analysis of AWS regions that can benefit from space shifting. For each origin region, the min, average, and max number of regions that lead to an emissions reduction are shown. Regions with high average carbon intensity benefit the most from migration.
  • ...and 4 more figures