ETD

• cloud-edge continuum
• energy efficiency
• FLOPs estimation
• Kubernetes
• MLOps
• multi-criteria ranking
• ONNX
• orchestration
• resource allocation
• sustainable AI
• task placement

This thesis addresses the efficient management of Machine Learning workloads across the Cloud–Edge continuum, where infrastructure heterogeneity and the growing complexity of models create an information gap that traditional orchestrators struggle to bridge. The goal is to propose a resource-aware approach that combines model awareness and context awareness. The solution consists of (i) a static FLOPs counter based on ONNX, capable of estimating a model’s computational cost a priori, and (ii) a hierarchical, multi-criteria orchestrator that, by combining node telemetry with computational estimates, optimizes placement by balancing end-to-end latency, energy consumption, and economic cost, while enforcing localization and feasibility constraints. Validation was performed through simulations on large-scale topologies and through experimentation on a real Kubernetes-based testbed. The results show high estimation accuracy for standard architectures and confirm that there is no one-size-fits-all strategy: heavyweight tasks (e.g., training) and micro-tasks (e.g., deployment) require different placement logics, influenced by queueing effects and network latency.