University of Thessaly (UTH)
Intelligent orchestration in heterogeneous networked systems
Abstract
dc:descriptionThe increasing complexity, heterogeneity, and scale of modern digital infrastructures have redefined the operational boundaries of networking and computing systems. Environments spanning dense urban wireless deployments, mobile ad hoc networks, disaggregated cloud datacenters, federated 5G testbeds, and the distributed cloud-to-things continuum now demand intelligent, efficient, and resilient orchestration mechanisms. These systems must adapt in real time, manage constrained and fluctuating resources, and remain energy-aware, while still delivering performance guarantees. The central hypothesis of this thesis is that intelligent orchestration in such environments is feasible through a combination of cross-layer observability, software-defined control, and learning-based automation. Rather than proposing a single monolithic solution, this thesis introduces several specialized orchestration frameworks that are tailored to the architectural and operational constraints and needs of their respective environments. These frameworks are based on the same foundational concept: orchestration must be context-aware, data-driven, and dynamically adaptive. We begin with wireless local area networks, where contention, interference, and hidden terminal effects lead to persistent performance degradation. We design and implement a MAC-layer-based diagnosis system that utilizes driver-exposed statistics from commercial access points to identify and classify WLAN pathologies. Without requiring any hardware modifications, this lightweight framework operates in both active and passive modes, achieving high accuracy and offering practical deployment potential. Next, we turn to mobile ad hoc networks (MANETs), where centralized orchestration appears incompatible with volatile topologies. To address this, we design a fault-tolerant Software Defined Networking (SDN) framework that supports distributed controller election, in-band signaling, and adaptive routing based on cross-layer link quality metrics. We further extend this system with reinforcement learning agents that dynamically tune SDN control-plane parameters to maintain network stability under mobility and channel fluctuations. Our experiments show significant improvements in throughput, latency, and resilience. In the datacenter domain, we investigate orchestration challenges in disaggregated systems, where compute and memory are decoupled and interconnected over programmable optical fabrics. We design a graph-based orchestration plane, the Software Defined Memory (SDM) controller, that enables efficient scheduling, dynamic attachment of memory resources, and fast topology reconfiguration. Our implementation, integrated with OpenStack, demonstrates substantial gains in scheduling latency and energy efficiency. Building upon the growing need for distributed experimentation across cloud-native, federated (5G) testbeds, we propose a multi-cluster orchestration framework that leverages state-of-the-art open-source solutions to securely expose services and establish inter-cluster connectivity across geographically dispersed administrative domains. Deployed over a high-speed fabric, the system enables cloud-native 5G workloads to be orchestrated transparently across distinct infrastructures, while preserving isolation and maintaining performance guarantees.Shifting to the computing continuum, we explore machine learning operations (MLOps) and the deployment of AI workflows across heterogeneous cloud, edge, and far-edge resources. First, we introduce an orchestration framework that decomposes ML pipelines into modular components and schedules them according to their resource and latency requirements. Then, we design EELAS, an energy-efficient and latency-aware scheduler for ML inference workloads that integrates into Kubernetes and is validated in real-world testbeds, achieving notable reductions in energy consumption while meeting application-level latency constraints. In overall, these contributions offer a comprehensive and realistic approach to orchestrating performance- and energy-critical systems in heterogeneous networking environments. They demonstrate that the principles of cross-layer design, programmability, and learning-based adaptation can be effectively applied to address the coordination challenges posed by next-generation infrastructures.
Degree
thesis:*- Grantor dc:publisher
- University of Thessaly (UTH)
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Syrigos, Ilias
- Συρίγος, Ηλίας
Subjects
dc:subject × 16- Ενορχήστρωση Δικτύων
- Ετερογενείς Υποδομές
- Δικτύωση Λογισμικού
- Ενισχυτική μάθηση
- Αποσυσχετισμένα Κέντρα Δεδομένων
- Network Orchestration
- Heterogeneous Infrastructures
- Software-defined networking (SDN)
- Reinforcement learning
- Disaggregated Datacenters
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Επιστήμες Μηχανικού και Τεχνολογία
- Μηχανική και συστήματα επικοινωνιών, Τηλεπικοινωνίες
- Electrical Engineering, Electronic Engineering, Information Engineering
- Engineering and Technology
- Communication engineering and systems, Telecommunications
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/60337