地理分散型AI推論のサービス維持型カーボン・水配慮資源配分
Service-preserving carbon- and water-aware resource allocation for geo-distributed AI inference (原題)
Zhang Q, Sheng S, Liu T
🤖 gxceed AI 要約
日本語
地理分散型AI推論の資源配分を、サービス維持を優先しつつ炭素と水のトレードオフを可視化する二段階辞書式フレームワークを提案。実トレースで26,392リクエストを解析し、水制約が炭素排出に与える影響を定量化。再現可能なベンチマークと資源管理インターフェースを提供。
English
A two-stage lexicographic framework for geo-distributed AI inference that preserves service while optimizing carbon and water. Trace-driven analysis of 26,392 requests reveals a 21-point carbon-water frontier, quantifying trade-offs. Provides a reproducible benchmark for service-preserving environmental constraints.
Unofficial AI-generated summary based on the public title and abstract. Not an official translation.
📝 gxceed 編集解説 — Why this matters
日本のGX文脈において
日本のデータセンターやAIインフラ事業者にとって、Scope 2排出削減と水資源管理の両立は重要。SSBJ開示やTCFD対応で環境KPIの可視化が求められる中、本手法はサービス品質を保ちながら環境負荷を最適化する実践的枠組みを提供する。
In the global GX context
As global AI infrastructure expands, data centers face pressure to reduce carbon and water footprints under frameworks like CSRD and SEC climate rules. This paper offers a reproducible method to balance service quality with environmental constraints, relevant for cloud providers and regulators.
👥 読者別の含意
🔬研究者:Provides a novel lexicographic optimization framework for carbon-water trade-offs in AI inference, with reproducible benchmarks.
🏢実務担当者:Offers a resource allocation interface to evaluate environmental constraints before orchestration, useful for data center operators.
🏛政策担当者:Demonstrates how service-preserving environmental constraints can be enforced, informing data center sustainability regulations.
📄 Abstract（原文）
<title>Abstract</title> <p>Geo-distributed AI inference requires resource allocation that preserves feasible service while coordinating heterogeneous capacity, latency, migration, carbon, and water conditions. We introduce a trace-driven, two-stage lexicographic framework that converts this systems problem into an inspectable carbon-water frontier. Stage 1 maximizes accelerator-hour-equivalent service; Stage 2 minimizes model-accounted carbon over the service-equivalent feasible set while enforcing a physical direct-water cap for regions screened High or Extremely high by Aqueduct. This structure keeps service priority explicit and retains carbon and water in auditable physical units. For 26,392 successful diffusion-serving requests over 554 h, the reference analysis produces a 21-point frontier. A zero screened-water cap avoids 2.4250 L relative to unconstrained carbon-first allocation at an additional 0.0582 kgCO2e, while remaining 13.083% below matched-service no migration. An exhaustive 243-cell federation stress test yields 1,215 optimal cap-constrained solves; zero-cap carbon changes span − 48.151% to numerical zero. Four thousand paired circular 24-h block resamples and 81 spatial WUE profiles further isolate temporal-composition and water-efficiency sensitivity. The framework provides a reproducible planning benchmark and a resource-management interface for evaluating service-preserving environmental constraints before integration into higher-level orchestration.</p>
🔗 Provenance — このレコードを発見したソース
- Research Square https://doi.org/10.21203/rs.3.rs-10621258/v1first seen 2026-08-22 04:22:38 · last seen 2026-09-03 04:44:31
🔔 こうした論文の新着を逃したくない方は キーワードアラート に登録（無料・3キーワードまで）。
gxceed は公開メタデータに基づく研究支援データセットです。要約・翻訳・解説は AI 支援で生成されています。 最終的な解釈・検証は利用者が原典資料に基づいて行うことを前提とします。