Hardware-Accelerated Edge AI Orchestration on the Multi-Tier Edge-to-Cloud Continuum

Hardware-Accelerated Edge AI Orchestration on the Multi-Tier Edge-to-Cloud Continuum

연구 분야: Software Development

학회: Journal of Network and Systems Management

Future 6 G networks demand low-latency and high-bandwidth capabilities supported by geographically distributed edge and cloud resources. In this context, effective orchestration of edge Artificial Intelligence (AI) is essential to address challenges such as distributed data sources and dependencies between inference and training processes. However, traditional orchestration solutions do not integrate AI-driven applications effectively. This paper introduces a multi-tier orchestration platform aligned with ETSI specifications, designed to seamlessly integrate AI Functions (AIFs) into the Multi-Access Edge Computing (MEC) ecosystem while supporting hardware acceleration for AIFs. The platform consists of a multi-tier orchestrator and multiple MEC orchestrators interconnected via a federation interface, enabling efficient lifecycle management and deployment. Key innovations include a novel AIF descriptor and a hardware-acceleration-aware placement algorithm to enhance AIF placement. Evaluation results show that the proposed platform outperforms state-of-the-art solutions such as MEC Host Selection Mechanism (MEC-HSM) and Kubernetes. It achieves a deployment success rate of 95.6% under high AIF placement loads, representing a 10.5% improvement over MEC-HSM and a 22.3% improvement over Kubernetes. Additionally, the platform reduces AI inference execution times by 50–220% and increases throughput by up to 21-fold.