SUSTAIN

Summary

The SUSTAIN research group met at IETF 125 to discuss multidisciplinary research into the sustainability and resilience of the internet. The session covered five main technical areas: high-resolution sub-national carbon mapping, server lifecycle and traffic testbed measurements, the impact of generative AI on web protocols and storage, service-level sustainability accounting in multi-domain networks, and the development of a holistic API for evaluating path energy.

Key Discussion Points

Towards a Sub-National Carbon Map

Presenter: Wen Cai Slides

Methodology: The research proposes moving beyond national-level carbon intensity data (like Electricity Maps) to NUTS-3 and LAU (Local Administrative Units) levels. It uses OpenInfraMap, ENTSO-E, and TSO data to calculate local carbon intensity based on local production and imports from parent regions.
Findings: National averages often mask significant local variations. For example, in mainland Spain, solar production in specific municipalities creates hourly carbon intensity profiles that differ drastically from the national average.
Networking Implications: This data is critical for carbon-aware routing. A path to a "greener" country may not actually be lower-carbon if the specific path traverses carbon-intensive sub-regions.
Discussion: Diego Lopez inquired whether the maps focus on production or consumption; Wen Cai clarified it is a combination of both per hour. Sebastian asked about applications in other industries, which Wen Cai noted could include infrastructure site selection.

Traffic testbed and Carbon topography representation

Presenter: Sebastian Slides

Testbed (TRAFFIC): The study measured various hardware (Dell servers to Raspberry Pis) at the power plug under modular loads.
Key Finding: CPU usage is a poor proxy for power consumption, with deviations up to 25% depending on the service (e.g., AI inference vs. counting).
Carbon Topography: A visualization tool to determine whether fabrication (embodied carbon), static power, or dynamic power dominates a server's footprint based on regional carbon intensity and load.
Optimization Strategies:
1. Use low-carbon energy.
2. Maximize server load (efficiency increases with load).
3. Extend equipment lifespan (amortizing embodied carbon over 10 years instead of 4).
Discussion: Michael Welzl asked about scaling testbeds; Sebastian emphasized the need for internal hardware measurements (RAM, fans) rather than just software proxies.

Small World Web of AI

Presenter: Noa Zilberman Slides

Concept: Proposes a shift where servers send prompts/metadata instead of raw content (images/video), and the client generates the content locally via GenAI.
Protocol Changes: Suggests a "gen-ability" indicator in HTTP/2 SETTINGS frames (value 0x07) and a new HTML class for generated content.
Sustainability Trade-offs: While it significantly reduces network traffic and CDN storage (reducing embodied carbon of disks), current generation energy on end-user devices is much higher than transmission energy.
Discussion: Jen-Hsuan noted similarities to agent-to-agent scenarios. Dirk Trossen questioned the latency and energy efficiency of generating at multiple locations versus caching. Jari Arkko raised concerns about brand consistency and "hallucinations" in critical contexts like surveillance. Noa Zilberman emphasized that standards must evolve now to steer future AI integration sustainably.

EXIGENCE View on Sustainability of ICT

Presenter: Artur Hecker Slides

Problem: ICT services are increasingly disaggregated across domains, making end-to-end measurement difficult.
EXIGENCE Approach: A three-pillar methodology: Measurement, Optimization, and Incentivization.
Service-Level Accounting: Proposes an "energy overlay" where "ecodata" (Joules, CO2) is reported along the Service Function Chain (SFC).
Incentives: Suggests providers can incentivize users to accept lower SLAs (e.g., lower video resolution) to achieve energy savings beyond what is possible through hardware efficiency alone.
Discussion: Jen-Hsuan asked if this requires underlay protocol modifications; Artur Hecker replied that while rough models can work over existing protocols, precise attribution requires more complex mapping within domains.

Sustainability Holistic API for Path Energy Evaluation (SHAPE)

Presenter: Luis Contreras Slides

Overview: SHAPE is an evolution of the PETRA API (originally discussed in the GREEN WG as draft-petra-green-api).
Scope: Focuses on sustainability metrics for network paths, including energy mix, greenness degree, sleep mode availability, and anomaly factors.
Architecture: Positioned to interface between network controllers and service orchestrators to allow customers to request "Decarbonization Level Agreements."
Use Cases: SD-WAN reporting, multilayer energy management (IP+Optical), and 5G edge selection.
Discussion: Jen-Hsuan asked about the generic nature of the API; Luis Contreras confirmed the intent is to keep parameters generic while implementation remains technology-specific.

Decisions and Action Items

Data Sharing: Wen Cai to share GitHub repository links for the sub-national carbon map data for RG review.
Cross-WG Coordination: Presenters (Artur Hecker, Luis Contreras) to coordinate with the GREEN working group regarding overlapping API and architectural models.

Next Steps

One-Year Review: The SUSTAIN-RG is due for its one-year review. The chairs will solicit formal feedback on the mailing list.
IETF 126: The group plans to meet in person in Vienna. Participants are encouraged to review the SHAPE API draft and provide feedback on the proposed sustainability parameters.

Automatic IETF Minutes