[Paper Review] Measuring IT Carbon Footprint: What is the Current Status Actually?

Tom Kennes|arXiv (Cornell University)|Jun 12, 2023

Green IT and SustainabilityEngineering3 citations

TL;DR

This paper identifies and categorizes the key challenges in measuring IT's carbon footprint into four domains: software energy consumption, server overhead, energy mix, and embodied carbon. It proposes a practical, CPU-based mathematical framework for quick carbon assessments and calls for standardized, bottom-up measurement tools to empower developers and improve transparency in sustainable software engineering.

ABSTRACT

Despite the new Corporate Sustainability Reporting Directive from the European Union, which presses large enterprises to be more transparent about their GHG emissions, and though large technology- or advisory firms might peddle otherwise, there are plenty of challenges ahead when it comes to measuring GHG emissions from IT activities in the first place. This paper categories those challenges into 4 categories, and explains the current status, shortcomings and potential future research directions. These categories are: measuring software energy consumption, server overhead energy consumption, Energy Mix and emissions from embodied carbon. Next to that, various non-profit and open-source initiatives are introduced as well as a mathematical framework, based on CPU consumption, that can act as a rule-of-thumb for quick and effortless assessments.

Motivation & Objective

• To address the critical gap in accurate, bottom-up measurement of IT carbon emissions despite growing regulatory and industry pressure.
• To identify and categorize the primary challenges in measuring software and infrastructure carbon footprints into four distinct domains.
• To propose a practical, rule-of-thumb mathematical framework based on CPU consumption for rapid, accessible carbon assessments.
• To stimulate alignment between open-source initiatives, practitioners, and academic research in sustainable software engineering.
• To advocate for actionable, granular measurement tools that empower developers and IT teams to reduce emissions effectively.

Proposed method

• Categorizes IT carbon footprint measurement into four core components: software energy consumption (Es), server overhead (Eo), momentary energy mix (I), and embodied carbon (M).
• Proposes a simplified mathematical model: Total Carbon = ((Es + Eo) × I) + M, aligning with the Green Software Foundation’s SCI specification but emphasizing separation of software and overhead.
• Introduces a rule-of-thumb framework based on CPU utilization to estimate energy consumption and emissions without complex instrumentation.
• Analyzes resource sharing dynamics using time-dependent consumption factors (f_c(τ, o_i)) to model multi-tenant environments.
• Applies life cycle assessment (LCA) principles to estimate embodied carbon emissions (M) from hardware, including factors like M_o_i, R_o_i, EoL_o_i, and T_o_i.
• Highlights the importance of scope definitions and idle time attribution in embodied carbon accounting, especially in underutilized infrastructure.

Experimental results

Research questions

• RQ1 How can software energy consumption be measured accurately and practically at scale, especially without relying on physical power meters?
• RQ2 What are the implications of dynamic resource sharing on energy and carbon accounting in multi-tenant IT environments?
• RQ3 How should embodied carbon emissions be allocated over time, particularly during periods of idle usage?
• RQ4 What criteria determine the appropriate scope for including hardware and infrastructure in carbon accounting?
• RQ5 Who should be held responsible for embodied carbon emissions during idle or underutilized periods of IT assets?

Key findings

• The current state of IT carbon footprint measurement remains highly inconsistent, with top-down estimates lacking the granularity needed for actionable decision-making at the developer or system level.
• Physical power monitoring and simulation-based models are the most accurate methods for measuring software energy consumption, but both face scalability and accessibility challenges.
• A CPU-based rule-of-thumb framework can serve as a practical, low-overhead method for estimating carbon emissions, enabling quick assessments without deep instrumentation.
• Dynamic resource sharing significantly complicates carbon accounting, requiring detailed time- and consumer-specific tracking, though this may be feasible in reserved or predictable environments.
• There is no consensus on how to allocate embodied carbon emissions over time, especially during idle periods, leading to significant controversy and inconsistency in reporting.
• Despite growing interest and initiatives like the Green Software Foundation’s SCI, widespread adoption is hindered by lack of standardized, detailed methodologies for the underlying components of carbon measurement.