pith. sign in

arxiv: 2607.01246 · v1 · pith:3CNGXHLMnew · submitted 2026-06-01 · 💻 cs.CY

Measure Once, Model Everywhere: Model-Based Per-Request Resource Consumption for HTTP

Geerd-Dietger Hoffmann , Verena Majuntke This is my paper

Pith reviewed 2026-07-04 00:48 UTC · model grok-4.3

classification 💻 cs.CY
keywords energy modelingHTTP requestsCO2e estimationsustainability disclosureoffline benchmarkingnginx extensionper-request impactresource consumption
0
0 comments X

The pith

Models from offline benchmarks estimate per-request energy and CO2e for HTTP endpoints at runtime.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper addresses the gap in HTTP sustainability disclosures by showing how to generate per-request resource and emissions values without live power telemetry in production. Endpoints are benchmarked offline under controlled conditions to derive compact models from observable request features, which are then evaluated online when requests arrive. These models take constant, linear, or piecewise forms and can incorporate application-layer signals such as token counts. Implementation as an nginx extension demonstrates that the approach runs with low overhead and remains feasible for operational use.

Core claim

We present a model-based approach for estimating resource consumption and CO2e per HTTP request without requiring fine-grained production power telemetry. The approach benchmarks endpoints offline under controlled conditions, derives compact endpoint-specific energy models from observable request features, and evaluates these models online at the HTTP server boundary. We show that heterogeneous request classes can be represented with constant, linear, and piecewise models, and that the same approach extends to endpoints whose dominant cost driver is only visible at the application layer through inputs such as token counts. Our evaluation indicates that the approach is operationally feasible

What carries the argument

Endpoint-specific energy models (constant, linear, or piecewise) built from offline benchmarks of request features, stored in a JSON registry, and evaluated at the nginx server boundary to produce energy, grid intensity, embodied emissions, and total impact metadata.

If this is right

  • Heterogeneous request classes can be represented with constant, linear, and piecewise models derived from request features.
  • The modeling approach extends to endpoints where the dominant cost driver appears only at the application layer, such as through token counts.
  • An nginx extension can load the JSON model registry and emit per-request metadata for energy, grid intensity, embodied emissions, and total impact.
  • The method remains operationally feasible while introducing only low runtime overhead.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Services could attach these per-request estimates to HTTP response headers to meet emerging sustainability disclosure requirements without new hardware.
  • Models would likely require periodic re-benchmarking after hardware upgrades or major software changes to maintain accuracy.
  • The same offline-to-online pattern could apply to other measurable resources such as memory bandwidth or network bytes beyond energy.
  • Request-level granularity might allow finer carbon accounting for multi-tenant or serverless deployments.

Load-bearing premise

Resource consumption measured in controlled offline benchmarks will produce models that accurately predict consumption for the same request features when deployed in production environments with varying loads and hardware.

What would settle it

Compare model predictions against direct power measurements collected on a production server under varying load for the same request feature values; systematic large errors would show the models do not transfer.

Figures

Figures reproduced from arXiv: 2607.01246 by Geerd-Dietger Hoffmann, Verena Majuntke.

Figure 1
Figure 1. Figure 1: Fitted models for one endpoint from each class. [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
read the original abstract

Recent proposals for HTTP-based sustainability disclosure focus on \textbf{what} environmental information should be transmitted at the protocol boundary, for example through response headers, but leave open the practical question of \textbf{how} such per-request values can be generated in realistic deployments. This paper addresses that implementation gap. We present a model-based approach for estimating resource consumption and $CO_2e$ per HTTP request without requiring fine-grained production power telemetry. The approach benchmarks endpoints offline under controlled conditions, derives compact endpoint-specific energy models from observable request features, and evaluates these models online at the HTTP server boundary. We implement this mechanism as an nginx extension that loads a JSON model registry and emits per-request metadata for energy, grid intensity, embodied emissions, and total request-level impact. We show that heterogeneous request classes can be represented with constant, linear, and piecewise models, and that the same approach extends to endpoints whose dominant cost driver is only visible at the application layer through inputs such as token counts. Our evaluation indicates that the approach is operationally feasible and introduces only low runtime overhead.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper claims to address the implementation gap in HTTP-based sustainability disclosure by presenting a model-based approach for estimating per-request resource consumption and CO2e. Endpoints are benchmarked offline under controlled conditions to derive compact endpoint-specific energy models (constant, linear, and piecewise) from observable request features, including application-layer inputs such as token counts. These models are then evaluated online at the HTTP server boundary using an nginx extension that loads a JSON model registry and emits per-request metadata for energy, grid intensity, embodied emissions, and total impact. The authors state that heterogeneous request classes can be represented with these models and that the approach is operationally feasible with low runtime overhead.

Significance. If the offline-derived models prove accurate in production environments, this approach could enable practical, low-overhead per-request environmental impact disclosure at the protocol level without requiring continuous power telemetry in production. This directly tackles a key practical barrier in recent sustainability proposals for HTTP.

major comments (2)
  1. [Abstract] The statement that 'our evaluation indicates that the approach is operationally feasible and introduces only low runtime overhead' is not supported by any quantitative data, error metrics, model equations, or details on benchmark conditions in the provided abstract. This absence undermines assessment of the central feasibility claim.
  2. [Evaluation] There is no reported direct validation of the models' prediction accuracy against ground-truth measurements under production variability, such as concurrent load, hardware differences, thermal throttling, or background processes. This is critical because offline benchmark models may not transfer to production, which is load-bearing for the sustainability disclosure use case.
minor comments (1)
  1. The abstract mentions 'piecewise models' but does not specify the breakpoints or how they are determined from the data.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments on our manuscript. We address each of the major comments point by point below, indicating where revisions will be made.

read point-by-point responses

  1. Referee: [Abstract] The statement that 'our evaluation indicates that the approach is operationally feasible and introduces only low runtime overhead' is not supported by any quantitative data, error metrics, model equations, or details on benchmark conditions in the provided abstract. This absence undermines assessment of the central feasibility claim.

    Authors: We agree that the abstract lacks quantitative support for the feasibility claim. We will revise the abstract to include key quantitative findings from the evaluation, such as the reported runtime overhead and any error metrics or benchmark details, to substantiate the statement. revision: yes

  2. Referee: [Evaluation] There is no reported direct validation of the models' prediction accuracy against ground-truth measurements under production variability, such as concurrent load, hardware differences, thermal throttling, or background processes. This is critical because offline benchmark models may not transfer to production, which is load-bearing for the sustainability disclosure use case.

    Authors: We acknowledge the absence of production-environment validation in the current manuscript. Our evaluation focuses on controlled offline benchmarking to derive and test the models, demonstrating low overhead in the implementation. We will add a dedicated limitations paragraph in the revised manuscript to discuss the potential effects of production variability and the need for further validation studies. This addresses the concern without overclaiming the current results. revision: partial

Circularity Check

0 steps flagged

No circularity: explicit fitting to benchmarks is the stated method, not a hidden reduction

full rationale

The paper's core mechanism is to run controlled offline benchmarks on endpoints, fit compact constant/linear/piecewise models to observable request features (including application-layer inputs such as token counts), and then evaluate the resulting models at runtime inside an nginx extension. This fitting step is presented as the intended engineering process rather than a derivation that reduces to its own inputs by construction. No self-citation load-bearing steps, uniqueness theorems, or ansatzes imported from prior author work appear in the provided text; the evaluation of runtime overhead is reported separately from the model construction. The central claim therefore remains self-contained against external benchmarks and does not collapse into any of the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central claim rests on the domain assumption that controlled benchmarks yield generalizable models and that observable request features are sufficient predictors; model coefficients are free parameters fitted per endpoint. No new entities are postulated.

free parameters (1)
  • endpoint-specific model coefficients
    Parameters in constant, linear, or piecewise energy models fitted to benchmark measurements for each request class.
axioms (2)
  • domain assumption Resource consumption of an endpoint can be adequately captured by a function of observable request features
    Invoked when deriving compact models from benchmark data.
  • domain assumption Offline controlled benchmarks produce models representative of production behavior
    Required for online evaluation at the HTTP boundary to be valid.

pith-pipeline@v0.9.1-grok · 5721 in / 1394 out tokens · 39083 ms · 2026-07-04T00:48:35.462404+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

34 extracted references · 34 canonical work pages

  1. [1]

    2023.Insolvent: How to Reorient Computing for Just Sustain- ability

    Christoph Becker. 2023.Insolvent: How to Reorient Computing for Just Sustain- ability. MIT Press

    work page 2023

  2. [2]

    2026.The Sustainability HTTP Response Header Field

    Andrei Nicolae Besleaga. 2026.The Sustainability HTTP Response Header Field. Internet-Draft draft-besleaga-green-sustainability-header-00. Internet Engineer- ing Task Force. https://datatracker.ietf .org/doc/draft-besleaga- green- sustainability-header/00/ Work in Progress

    work page 2026

  3. [3]

    Francisco Caravaca, Ángel Cuevas, and Rubén Cuevas. 2025. From Prompts to Power: Measuring the Energy Footprint of LLM Inference.arXiv preprint arXiv:2511.05597(2025). doi:10.48550/arXiv.2511.05597

    work page doi:10.48550/arxiv.2511.05597 2025

  4. [4]

    Kris De Decker. 2018. How to build a low-tech website? Low-tech Magazine. https://solar.lowtechmagazine.com/2018/09/how-to-build-a-low-tech-website/ Accessed 2026-05-25

    work page 2018

  5. [5]

    Radosvet Desislavov, Fernando Martínez-Plumed, and José Hernández-Orallo

    work page

  6. [6]

    doi:10.1016/j.suscom.2023.100857

    Trends in AI inference energy consumption: Beyond the performance-vs- parameter laws of deep learning.Sustainable Computing: Informatics and Systems 38 (2023), 100857. doi:10.1016/j.suscom.2023.100857

    work page doi:10.1016/j.suscom.2023.100857 2023

  7. [7]

    Electricity Maps. [n. d.]. Electricity Maps: Live and historical electricity carbon intensity by region. Online service. https://www.electricitymaps.com/ Accessed 2026-05-25

    work page 2026

  8. [8]

    Green Coding Solutions. 2026. Best Practices – Green Metrics Tool Documenta- tion. Online documentation. https://docs.green-coding.io/docs/measuring/best- practices/ Accessed 2026-04-01

    work page 2026

  9. [9]

    Green Software Foundation Standards Working Group. 2024. Software Carbon Intensity (SCI) Specification. Specification. https://sci.greensoftware.foundation/ Version 1.1.0; accessed 2026-04-01

    work page 2024

  10. [10]

    Fernando, Markus Funke, Jens Gröger, Lorenz M

    Achim Guldner, Rabea Bender, Coral Calero, Giovanni S. Fernando, Markus Funke, Jens Gröger, Lorenz M. Hilty, Julian Hörnschemeyer, Geerd-Dietger Hoffmann, Dennis Junger, Tom Kennes, Sandro Kreten, Patricia Lago, Franziska Mai, Ivano Malavolta, Julien Murach, Kira Obergöker, Benno Schmidt, Arne Tarara, Joseph P. De Veaugh-Geiss, Max Westing, Volker Wohlgem...

    work page doi:10.1016/j.future.2024.01.033 2024

  11. [11]

    Lee, and Udit Gupta

    Leo Han, Jash Kakadia, Benjamin C. Lee, and Udit Gupta. 2025. Fair-CO2: Fair Attribution for Cloud Carbon Emissions. InProceedings of the 52nd Annual Interna- tional Symposium on Computer Architecture (ISCA ’25). Association for Computing Machinery, New York, NY, USA, 646–663. doi:10.1145/3695053.3731023

    work page doi:10.1145/3695053.3731023 2025

  12. [12]

    Geerd-Dietger Hoffmann. 2026. Green Metrics Tool run comparison: per-endpoint benchmarking measurements for per-request energy labels. Green Metrics Tool measurement run. https://metrics.green-coding.io/compare.html?ids=4d0b0afa- a994-446e-b10c-a9afdcd0cfb4,22947267-6714-4fd2-9261-e86f2e17dea6,6 9192a65-366f-4298-ae44-2a3dbd1913d2,b73289cb-c090-4190-ab2e-...

    work page 2026

  13. [13]

    Geerd-Dietger Hoffmann. 2026. Green Metrics Tool run comparison: runtime overhead measurements with nginx module disabled. Green Metrics Tool mea- surement run. https://metrics.green-coding.io/compare.html?ids=87d1724b- 81ec- 44d7- a400- 444f 663b76e3, 4d0b0af a- a994- 446e- b10c- a9af dcd0cf b4, 22947267- 6714- 4f d2- 9261- e86f 2e17dea6, 69192a65- 366f ...

    work page 2026

  14. [14]

    Geerd-Dietger Hoffmann. 2026. Green Metrics Tool run comparison: runtime overhead measurements with nginx module enabled. Green Metrics Tool mea- surement run. https://metrics.green-coding.io/compare.html?ids=37223dd2- 9481- 4009- a9cc- 25944c40f677, 328f8dc3- b7ed- 44be- 91d6- 6db2a8962d 88,c243bddb-3224-400e-b77f -5099c6bc2b1b,0c64123d-3649-47b7-aeb3- 4...

    work page 2026

  15. [15]

    Geerd-Dietger Hoffmann. 2026. Green Metrics Tool run comparison: ten-run reference energy for click-through validation. Green Metrics Tool measurement run. https://metrics.green-coding.io/compare.html?ids=30ed795e-04ff-4555-82 a3-863cbbdac0c3,072645a6-0dac-4952-a7f1-0f191cfbfa6e,17fee139-3d17-4610- 8d46-449e7c500914,64c35e72-d728-42cb-a342-3e7807a6c006,97...

    work page 2026

  16. [16]

    Geerd-Dietger Hoffmann. 2026. Green Metrics Tool run: no-header user click- through validation walkthrough. Green Metrics Tool measurement run. https:// metrics.green-coding.io/stats.html?id=fca8c396-7591-4f16-9fd5-c3f87c2a424f Accessed 2026-05-25

    work page 2026

  17. [17]

    Geerd-Dietger Hoffmann. 2026. Green Metrics Tool run:/login workload-scaling benchmark (1–1000 calls). Green Metrics Tool measurement run. https://me trics.green-coding.io/stats.html?id=6c40a5b8-79b7-4372-a1c4-3ae7185451d2 Accessed 2026-05-25

    work page 2026

  18. [18]

    International Organization for Standardization. 2006. ISO 14040:2006 Environ- mental management — Life cycle assessment — Principles and framework. Inter- national Standard. https://www.iso.org/standard/37456.html Edition 2; official standard page accessed 2026-04-01

    work page 2006

  19. [19]

    International Organization for Standardization. 2006. ISO 14044:2006 Environ- mental management — Life cycle assessment — Requirements and guidelines. International Standard. https://www.iso.org/standard/38498.html Edition 1; official standard page accessed 2026-04-01

    work page 2006

  20. [20]

    2023.HTTP Response Header Field: Carbon-Emissions-Scope-2

    Bertrand Martin. 2023.HTTP Response Header Field: Carbon-Emissions-Scope-2. Internet-Draft draft-martin-http-carbon-emissions-scope-2-00. Internet Engi- neering Task Force. https://datatracker.ietf.org/doc/draft-martin-http-carbon- emissions-scope-2/00/ Work in Progress

    work page 2023

  21. [21]

    David Mytton. 2020. Hiding greenhouse gas emissions in the cloud.Nature Climate Change10, 8 (2020), 701. doi:10.1038/s41558-020-0837-6

    work page doi:10.1038/s41558-020-0837-6 2020

  22. [22]

    Patterson, Jay Chen, Daniel Pargman, Barath Raghavan, and Birgit Penzenstadler

    Bonnie Nardi, Bill Tomlinson, Donald J. Patterson, Jay Chen, Daniel Pargman, Barath Raghavan, and Birgit Penzenstadler. 2018. Computing Within Limits. Commun. ACM(Oct. 2018). https://cacm.acm.org/research/computing-within- limits/ Accessed 2026-04-01

    work page 2018

  23. [23]

    Mark Nottingham. 2023. Communicating carbon emissions. GitHub issue #52 in httpwg/admin. https://github.com/httpwg/admin/issues/52 Opened 2023-04-12; accessed 2026-04-01

    work page 2023

  24. [24]

    Lucas Pardue. 2023. Re: Introducing a new HTTP response header for Carbon Emissions calculation. IETF HTTP Working Group mailing list archive. https: //lists.w3.org/Archives/Public/ietf-http-wg/2023AprJun/0020.html Message dated 2023-04-11; part of the thread initiated by Bertrand Martin; accessed 2026- 04-01

    work page 2023

  25. [25]

    Daniel Pargman and Barath Raghavan. 2014. Rethinking sustainability in com- puting: From buzzword to non-negotiable limit. InProceedings of the 8th Nordic Conference on Human-Computer Interaction (NordiCHI ’14). Association for Com- puting Machinery, 638–647. doi:10.1145/2639189.2639228

    work page doi:10.1145/2639189.2639228 2014

  26. [26]

    Matti Pärssinen, Marko Kotila, Rubén Cuevas, Akshay Phansalkar, and Jukka Man- ner. 2018. Environmental impact assessment of online advertising.Environmental Impact Assessment Review73 (2018), 177–200. doi:10.1016/j.eiar.2018.08.004

    work page doi:10.1016/j.eiar.2018.08.004 2018

  27. [27]

    Soham Poddar, Paramita Koley, Janardan Misra, Niloy Ganguly, and Saptarshi Ghosh. 2025. Towards Sustainable NLP: Insights from Benchmarking Infer- ence Energy in Large Language Models. InProceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Pa...

    work page doi:10.18653/v1/2025.naacl-long.632 2025

  28. [28]

    Barath Raghavan and Justin Ma. 2011. The energy and emergy of the Internet. InProceedings of the 10th ACM Workshop on Hot Topics in Networks (HotNets-X). Association for Computing Machinery. doi:10.1145/2070562.2070571

    work page doi:10.1145/2070562.2070571 2011

  29. [29]

    ribalba. 2026. easytodo: A simple ToDo REST API benchmark application. GitHub repository. https://github.com/ribalba/Limits-2026/tree/main/code/todoapp

    work page 2026

  30. [30]

    Antonello Sala, Lorenzo Barbetti, and Andrea Rosini. 2024. Green Web Meter: Structuring and Implementing a Real-Time Digital Sustainability Monitoring System.Sustainability16, 17 (2024), 7627. doi:10.3390/su16177627

    work page doi:10.3390/su16177627 2024

  31. [31]

    Arne Tarara, Dan Mateas, and Geerd-Dietger Hoffmann. 2026. Green Metrics Tool. GitHub repository. https://github.com/green-coding-solutions/green- metrics-tool

    work page 2026

  32. [32]

    WattTime. [n. d.]. WattTime: Marginal emissions and grid carbon intensity data. Online service. https://www.watttime.org/ Accessed 2026-05-25

    work page 2026

  33. [33]

    World Resources Institute (WRI) and World Business Council for Sustainable Development (WBCSD). 2011. Product Life Cycle Accounting and Reporting Standard. Standard. https://ghgprotocol.org/product-standard

    work page 2011

  34. [34]

    Yuqing Yang, Yuedong Xu, and Lei Jiao. 2024. A Queueing Theoretic Perspective on Low-Latency LLM Inference with Variable Token Length.arXiv preprint arXiv:2407.05347(2024). doi:10.48550/arXiv.2407.05347

    work page doi:10.48550/arxiv.2407.05347 2024