The Sunk Carbon Fallacy: Rethinking Carbon Footprint Metrics for Effective Carbon-Aware Scheduling

Overview

The paper argues that job-level metrics that add amortised embodied carbon to operational carbon, such as the Green Software Foundation's SCI, lead to worse scheduling decisions. Embodied emissions are already sunk once hardware is bought. In heterogeneous fleets at intermediate utilisation, SCI favours old, inefficient servers with low embodied carbon and raises total cluster emissions. The paper is from 2024, appeared on pages 542-551 of the Proceedings of the ACM Symposium on Cloud Computing, and is open access as arXiv:2410.15087.

SCI = ((E x I) + M) per R, where the embodied share is M = TE x (T/EL) x (RR/TR): total embodied emissions TE amortised linearly over the expected lifespan EL and apportioned by the resources reserved (RR) out of the server's total resources (TR). In the example, a newer Xeon Gold 6538N server has 4.09 times the processor embodied carbon of an older Xeon E-2286G (101.89 versus 8.04 kgCO2) but uses 32.5% less energy; SCI picks the old server and raises cluster carbon by 24.10%.

With processor-only estimates, embodied carbon is 11.7% and 44.5% of life-cycle emissions for the two example servers (average 28.1%). When operational efficiency is all that matters, the wrong choice adds up to 48% to system carbon. Servers are not power proportional: they often draw more than 30% of peak power at 0% utilisation, and datacenter-level utilisation is 30% to 60% even in state-of-the-art facilities.

The case study covers 740 servers of 15 Intel processor types, with an average age of 9.5 years and only 236 (31.9%) less than five years old. Processor embodied carbon ranges from 9.44 to 101.0 kgCO2 per server, 17,633 kgCO2 in total. Scheduling by SCI raised datacenter carbon by nearly 30%, with a penalty over 15% at 40-80% utilisation. Amortising over years beyond a 5-year expected life double-counts embodied carbon. The authors recommend using embodied carbon only for procurement and replacement decisions. Only the record's summary and key facts were used here.

Key points

1. SCI = ((E x I) + M) per R, with embodied share M = TE x (T/EL) x (RR/TR).
2. A newer Xeon Gold 6538N has 4.09x the processor embodied carbon of an older Xeon E-2286G (101.89 vs 8.04 kgCO2) but uses 32.5% less energy; SCI picks the old server and raises cluster carbon by 24.10%.
3. With processor-only estimates, embodied carbon is 11.7% and 44.5% of life-cycle emissions for the two example servers (average 28.1%); the wrong choice adds up to 48% to system carbon.
4. Servers are not power proportional: they often draw more than 30% of peak power at 0% utilisation; datacenter utilisation is 30-60%.
5. Case: 740 servers of 15 Intel types, average age 9.5 years, only 236 under five years old; processor embodied carbon 9.44-101.0 kgCO2 per server, 17,633 kgCO2 in total.
6. Scheduling by SCI raised datacenter carbon by nearly 30%; the penalty was over 15% at 40-80% utilisation.
7. Amortising beyond a 5-year expected life double-counts embodied carbon; the longer the assumed lifespan, the larger the SCI scheduling error.

Using it in design

• Keep embodied-carbon amortisation (TE x T/EL) in procurement, refresh and design-option comparisons, and use operational carbon only for runtime placement and scheduling recommendations.
• Make the amortisation lifetime an explicit, user-editable field (default 5 years) and stop amortising once that lifetime has passed.
• Warn that per-job allocation of embodied carbon ignores idle servers and depends on the chosen lifetime.
• When modelling fleets or capacity, include idle power (often over 30% of peak) and typical 30-60% utilisation.

Who should read it

Sustainability and operations engineers who calculate carbon footprints and set server refresh and workload scheduling strategy.

Related standards

• GHG Protocol Scope 2 Guidance
• GHG Protocol Scope 3 Standard
• Green Software Foundation SCI specification

Publication details

Authors

Noman Bashir; Varun Gohil; Anagha Belavadi Subramanya; Mohammad Shahrad; David Irwin; Elsa Olivetti; Christina Delimitrou

Publisher

Proceedings of the ACM Symposium on Cloud Computing (ACM)

Year

2024

Document number

arXiv:2410.15087

Pages

542-551

Document language

English

Source type

Scientific

Access

Free to read

DOI

10.1145/3698038.3698542

Licence

CC BY-NC-SA 4.0 — share with attribution for noncommercial purposes under the same licence

Citation