The effect of the network in cutting carbon for geo-shifted workloads
Yibo Guo, Amanda Tomlinson, Runlong Su, George Porter · 2025
Read The effect of the network in cutting carbon for geo-shifted workloads by Yibo Guo, Amanda Tomlinson, Runlong Su, George Porter (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-2504-14022.txt.
Introduction
Yibo Guo and co-authors weigh in 2025 the network cost of moving computer work across the globe to cut carbon. Loose-deadline jobs can shift to data centers with cleaner power. Yet the network haul itself burns carbon, and past maps of that cost were crude.
Fine maps of real internet hardware give sharper haul cost grades. With true haul costs in hand, place shifting beats time shifting alone by a wide margin. Picking the wrong site or hour can erase the gains, the runs show. The work guides cloud users who chase real carbon cuts, not paper ones.
Catalog record
| CREATOR | Yibo Guo, Amanda Tomlinson, Runlong Su, George Porter |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 67 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2504-14022
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
The effect of the network in cutting carbon for geo-shifted workloads Yibo Guo, Amanda Tomlinson, Runlong Su, George Porter 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 ( 37 ) 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. 1. Introduction Organizations have embraced cloud computing to a significant degree. In 2020, spending on cloud platforms grew by 35% to reach almost $130 billion, whereas spending on on-prem datacenters dropped by 6% to under $90 billion ( 1 ) . A major reason that organizations migrate their compute and data to these platforms is to achieve features and properties that which would be difficult or cost-prohibitive for them to implement themselves ( 38 , 17 ) . Migrating work to the cloud permits scaling up (or down) resources in response to changes in user behavior or need. Migrating data geographically can enable replication for improved availability and fault tolerance. And more recently, cloud adoption can help organizations meet climate-related goals such as limiting the carbon footprint of their compute needs. Cloud providers are also using the flexibility of their platforms to reduce their own carbon footprints. Researchers have proposed ( 43 ) , and Google has publicly shown ( 32 ) , how to shift workloads with relaxed deadlines forwards or backwards in time to align with the availability of lower-carbon energy sources (such as wind and solar photovoltaic). The net result is delivering a lower carbon footprint for the same underlying computation. Given that low-carbon energy sources are heavily affected by environmental conditions (the sun in the case of solar, and wind speed in the case of wind power), there is significant potential in migrating data and workloads geographically to match compute demands with the status of energy grids and environmental conditions ( 22 , 31 , 27 , 47 ) . Already, geographic migration and replication is a key enabler of reducing wide-area latency ( 22 , 44 ) , better managing WAN bandwidth ( 14 ) , or lowering monetary cost ( 34 ) . Any optimization task requires accurate information on the system’s underlying resource utilization. Reducing the CPU requirements of a network-bound application is unlikely to deliver results in the same way that increasing the IO resources of a CPU-bound application will likely be unproductive. For geographic migration of cloud workloads, accurately estimating the carbon impact of the wide-area network (WAN) transfer is critical to reducing the overall carbon footprint of the computation. We observe that inaccuracies or omissions in estimating the carbon cost of the WAN lead to migration decisions that fail to achieve carbon optimization goals, and could even be entirely counterproductive. For example, as we’ll show in Section 2.3 , a network-heavy job can achieve 75\% carbon savings without considering the network carbon cost, but if we account for the network transfer cost, we may not even want to migrate at all. In this paper, we show that previous approaches to estimating the carbon impact of WAN transfers underestimate the real impact by up to 2-3x due to inaccurate understandings of the underlying routes, and we describe improvements to these approaches to better capture a higher fidelity picture of the carbon costs of WAN transfers. In particular, the contributions we claim in this paper are: • We improve upon state-of-the-art WAN carbon transfer estimation approaches based on traceroute data (Tabaeiaghdaei et al. ( 37 ) ), as shown in Table 1 and Figure 2 . • We have built a dataset that maps components of WAN paths between public cloud locations to their underlying energy grids (called ISOs), and will make this publicly available for other researchers to build upon and to use to verify and replicate our network transfer cost results (see Appendix B ). • We developed an highly optimized and efficient algorithms for carbon-aware workload scheduling between geographically distributed data centers, quantitatively analyzing the potential additional benefit of space-shifting solutions over previously published time-shifting solutions ( 33 , 43 ) . 2. Motivations and Background In this section, we first describe the high-level motivations of our work on carbon-aware workload migration, and then define the terms and definitions that we use throughout the paper; finally we show a concrete example that highlights some of the factors impacting a successful workload migration. 2.1. Motivations We first describe the motivations for general carbon-aware computing in Section 2.1.1 , and then the specific reason advocating for space-shifting in Section 2.1.2 . 2.1.1. Increasing renewable deployment A key trend in carbon-aware computing is to adapt to the availability of renewable energy, which varies in both time and space domain ( 33 , 43 , 47 , 32 ) . This means that some data centers may have access to a lot of low-carbon renewables (e.g. hydro- or nuclear-based locations) and others may have none (e.g. coal- or gas-based locations); or they may only have access to low-carbon during certain times of day or certain months of the year (e.g. solar- or wind-based locations) ( 32 ) . As the amount of renewable deployment continues to grow, existing grid infrastructure already lacks the capacity to store (in battery) or move (over long-distance grid lines) these excess renewable energy, due to various practical reasons like technological limitations, regulations and monetary costs. This leads to an increasing amount of curtailment, or wasted renewable energy, especially in solar powered region like California ( 16 ) . As data centers are now accounting for an increasing and non-negligible share of global electricity usage ( 3 , 2 ) , researchers and data center operators have looked into how data centers can help alleviate this problem by shifting their power demand in time ( 32 , 33 , 43 ) and space ( 31 , 47 ) , which is a key mechanism to achieve the 24/7 carbon-free climate goals of many major companies. Additionally, such trend has led to closer integration of renewable power and data centers, as shown in both academic ( 14 ) and industry ( 10 ) worlds. This further demands cost-efficient carbon-aware solutions in not only time-shifting, but more importantly space-shifting. 2.1.2. Lower cost in WAN vs power grid A key infrastructure that we leverage in carbon-aware space-shifting is the wide area network (WAN). Although expensive in absolute numbers and being a scare resource, WAN serves a similar role in connecting data centers as power grid serves in connecting electricity consumers, and is better in two major ways: 1) it’s easier to add