Toward sustainable diffusion-based AIGC: Design and online orchestration in distributed edge networks☆

Artificial intelligence-generated content (AIGC) is an automated method that generates the content according to its knowledge and the intent information. Deploying the AIGC model in edge networks unlocks new possibilities. Unfortunately, realizing AIGC in distributed edge networks faces critical challenges, including the interdependent control decisions, the complex trade-offs between energy consumption and total variation (TV) distance, the extension of TV distance due to insufficient denoising steps, the restriction of AIGC inferencing deadlines, and the uncertainty of AIGC task arrivals. In this paper, targeting AIGC tasks, we design polynomial- time online algorithms to overcome all these challenges. Firstly, we formulate distributed AIGC as a non-linear mixed-integer program for long-term total cost optimization. Subsequently, we propose a novel algorithmic approach that generates candidate inferencing schedules, reformulates the original problem into anew schedule selection problem, and solves this new problem using an online primal-dual-based algorithm. Moreover, we rigorously prove that our approach leads to a constant competitive ratio for the long-term total cost. Through extensive evaluations using real-world data, the superior practical performance of our approach is demonstrated, reducing the total cost by more than 50% compared to various alternative methods.