Jul 2026· Fall Joint Computer Conference· pp. 81-88· 0 citations· 37 references
Abstract
The increasing use of renewable energy in data centers creates an opportunity to reduce the carbon footprint of energy-intensive LLM inference workloads. Unlike traditional stable power supply, renewable generation fluctuates over time, making it difficult to match computation demand with available energy. However, existing LLM serving systems primarily optimize latency and throughput without considering energy supply dynamics, leading to underutilization of renewable energy and unnecessary reliance on thermal power, and consequently, higher carbon emissions. We present GreenAlign, a renewable-aware scheduling framework that addresses this mismatch by treating best-effort (BE) requests as temporally shiftable load. GreenAlign enforces a power-constrained policy that executes BE requests using only residual renewable energy under normal conditions, and introduces a backlog risk metric to selectively relax this constraint when deadline violations are imminent. To ensure responsiveness, it maintains standby capacity to absorb unpredictable latencycritical (LC) bursts and uses lightweight length estimation to handle request uncertainty. Simulation results show that GreenAlign significantly reduces thermal energy usage while preserving LC latency and BE deadline satisfaction.
This work introduces a novel LLM-based predictive scheduling system designed to enhance operational efficiency while reducing the environmental impact of data centers, using an LLM to predict key metrics such as execution time and energy consumption from source code.
Hanzhao Wang, Jingxuan Wu, Yumeng Li et al.· 0 citations
The rapid growth in artificial intelligence (AI) demand has significantly increased the electricity consumption and carbon emissions of computing centers. How to schedule AI requests across computing centers to reduce carbon emissions and electricity costs while maintaining low latency is an essential research problem. Existing schedulers reduce emissions by shifting workloads or balancing resources but usually simplify power system modeling, ignore transmission-side costs and carbon emissions, or make local decisions without batch-level coordination. To better address these problems, we first develop an ILP-based scheduler to get optimized results, but it faces scalability limitations. Then, we propose RAPID, a region-aware and power-informed scheduling framework that integrates static and online heuristic schedulers for large-scale AI request scheduling. Experiments based on real-world GenAI traces and Chinese regional power profiles show that RAPID significantly reduces carbon emissions, electricity costs, and total energy consumption compared to methods from previous works while maintaining zero Service Level Agreement (SLA) violations.
B. Ding, Cai-Ning Wang, Ka-Fei Tang et al.· Sustainability· 0 citations
Large Language Models (LLMs) have enabled diverse AI applications. However, LLM inference imposes unprecedented computational and memory overhead, creating an inherent tradeoff between latency Service Level Objectives (SLOs) and resource constraints. Serverless computing, with on-demand provisioning and pay-as-you-go billing, is becoming a promising paradigm for LLM serving. But existing solutions fail to integrate state-of-the-art inference optimizations, resulting in suboptimal GPU utilization and prolonged latency. While Prefill-Decode (PD) disaggregation combined with continuous batching has resolved such inefficiencies in traditional cloud deployments, migrating these techniques to serverless makes two challenges particularly pronounced: (1) SLO-constrained resource provisioning for independently scaling prefill and decode phase functions, and (2) function lifespan management to mitigate resource waste from continuous batching-induced prolonged instance lifespans. To tackle these issues, we propose CELLServe, an SLO-aware and cost-efficient serverless LLM serving system that pioneers integrating PD disaggregation and continuous batching into serverless platforms. CELLServe formalizes SLO-constrained joint resource provisioning as an optimization problem with a dedicated algorithm, and introduces an opportunistic instance merging strategy for decode phase functions to reclaim fragmented resources. Comprehensive evaluations on five mainstream LLMs and real-world traces show that CELLServe achieves 1.71-1.85×higher request throughput than baselines under identical SLOs and GPU budgets, while sustaining high resource efficiency under dynamic workloads.
Ze-Jian Wang, Nan Lin, Zinuo Cai et al.· ACM Transactions on Architec...· 0 citations
A new paradigm for satisfying the ever-growing demands of real-time Sixth Generation (6G) applications is Mobile Edge Computing (MEC). Additionally, base stations and Internet of Things devices that incorporate renewable energy harvesting capabilities have the potential to lower grid energy use. To maximize system potential and lower carbon emissions, it is crucial to make effective decisions about job offloading and resource allocation. A carbon-aware MEC architecture that uses both grid and renewable energy sources is proposed in this paper. Our goal is to jointly manage resource allocation and task offloading while monitoring carbon emissions and task queue delays to optimize system behavior under uncertainty, specifically for stochastic workloads and variable renewable generation. To balance these two cost components (emissions and queue length), we create a combined optimization problem. We develop a deep deterministic policy gradient (DDPG)-based joint optimization technique to address this issue in a constantly changing environment. In the optimization, we consider greedy policy (GP) and full offloading (FO), as well as time-average carbon emission (TACE) and time-average queue length (TAQL) as performance metrics, and time-average queue length (TAQL) and full execution (FE) as baseline strategies; we also evaluate normalized time-average cumulative reward (NTACR). This method uses continuous-action reinforcement learning to generate efficient, real-time control policies. For the proposed MEC network, numerical statistics show that our approach can lead to effective offloading and lower carbon emissions.
Mamoon M. Saeed, Rashid A. Saeed, M. A. Ahmed et al.· 2026 6th International Confe...· 0 citations
PRISM, a prediction-guided runtime framework that jointly selects model variants and CPU allocations for containerized edge microservices, and adapts each pipeline stage in place and minimizes predicted CPU-package energy under deadline, resource, and offline model-level Quality of Result constraints is presented.
Uwe Gropengießer, Thomas Reuter, Dominik Schön et al.· 0 citations
This work proposes an adaptive resource allocation framework that leverages Digital Twins for real-time system monitoring and integrates Large Language Models to support context-aware decision-making under multi-objective constraints, enabling intelligent workload orchestration across heterogeneous data center environments.
Pedro Henrique Sachete Garcia, A. F. Lorenzon, M. Luizelli et al.· SN Computer Science· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
