Joint Energy Management and Coordinated AIGC Workload Scheduling for Distributed Data Centers: A Diffusion-Aided Reward Shaping Approach

2026-07-08 · Source: stat.ML updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cloud Computing & IT Infrastructure, Emerging Technologies & Innovation · Depth: Expert, extended

Summary

A novel framework addresses joint energy management and coordinated AIGC workload scheduling across distributed data centers. This framework tackles challenges like model heterogeneity and implicit service quality by mathematically characterizing service quality and integrating diverse energy resources, including GPU DVFS, Battery Energy Storage Systems (BESS), and cooling control. It formulates a system utility maximization problem, balancing AIGC service revenue with operational penalties and costs. To overcome severe reward sparsity in deep reinforcement learning (DRL) for scheduling, a diffusion model-aided reward shaping approach synthesizes complementary reward signals. Experiments using real-world AIGC models like Stable Diffusion v1-5, SDXL, and SD3.5, along with real electricity price data, demonstrate the scheme's effectiveness. It accommodates electricity price fluctuations and model heterogeneity, achieving superior learning convergence and increasing system utility by over 30% compared to benchmark methods. The reward shaping improves cumulative reward up to 1.5x.

Key takeaway

For MLOps Engineers or Cloud Architects managing distributed AIGC services, this research demonstrates a viable path to optimize operational costs and service quality. You should consider implementing DRL-based scheduling frameworks that incorporate diffusion model-aided reward shaping to handle complex, sparse-reward environments. This approach enables adaptive job transfers, fine-grained inference control via denoising steps, and dynamic energy management, significantly improving system utility and learning convergence, even with heterogeneous AIGC models and fluctuating electricity prices.

Key insights

Diffusion model-aided reward shaping enhances DRL for joint AIGC workload scheduling and energy management in distributed data centers.

Principles

• AIGC service quality can be explicitly quantified using metrics like BRISQUE and CLIP.
• Coordinated job transfer and adaptive denoising steps are crucial for AIGC workload efficiency.
• Generative diffusion models can synthesize effective complementary rewards for sparse DRL environments.

Method

The approach decomposes the problem into DRL-driven job scheduling (ASP selection, denoising steps) and a heuristic-based energy management subproblem (GPU DVFS, cooling, BESS). Diffusion models condition denoising on state-action pairs to generate complementary rewards for SAC.

In practice

• Dynamically adjust AIGC denoising steps based on model and electricity prices.
• Offload delay-sensitive AIGC jobs to data centers with faster models.
• Integrate BESS and renewable energy to buffer against electricity price spikes.

Topics

• AIGC
• Diffusion Models
• Distributed Data Centers
• Workload Scheduling
• Energy Management
• Deep Reinforcement Learning
• Reward Shaping

Code references

• openai/CLIP
• alibaba/clusterdata
• google-research/parti

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, MLOps Engineer

Related on AIssential

• Deep Reinforcement Learning for Flexible Job Shop Scheduling with Random Job Arrivals — DRL with PPO and MLPs effectively solves FJSP with random arrivals by learning dispatching rule selection.
• DataCenterGym: A Physics-Grounded Simulator for Multi-Objective Data Center Scheduling — DataCenterGym simulates geo-distributed data center scheduling, integrating thermal, power, and workload dynamics for comprehensive policy evaluation.
• Real-Time AI Service Economy: A Framework for Agentic Computing Across the Continuum — Dependency graph topology dictates real-time AI service market stability and efficiency in agentic computing environments.
• Reinforcement Learning from Denoising Feedback — RLDF enhances dLLM reasoning by efficiently estimating policy loss via weighted sampling of denoising feedback and clean state prediction.
• GOAL: Graph-based Objective-Aligned Diffusion Solvers for Dynamic Multi-Objective Optimization — GOAL is a graph-based diffusion solver for dynamic multi-objective optimization, achieving high feasibility and speed.
• Libra: Efficient Resource Management for Agentic RL Post-Training — Libra efficiently manages resources for agentic RL post-training by dynamically optimizing GPU allocation and scheduling complex workloads.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by stat.ML updates on arXiv.org.