Πανεπιστήμιο Μακεδονίας
AI-assisted resource orchestration for network services and distributed cloud-native applications
Abstract
dc:descriptionNetwork Function Virtualization (NFV) has paved the way for the migration of Virtual Network Functions (VNFs) into multi-tenant datacenters, lowering the barrier for the introduction of new processing functionality into the network. In this context, we investigate the challenging problem of VNF placement in large-scale networks, particularly for Service Function Chains (SFCs), which represent sequences of VNFs. To this end, we initially present a state space complexity analysis of the orchestration problems associated with the placement of SFCs onto substrate networks and reason for the use of approximate methods for addressing the computational hardness of the problem. Genetic algorithms (GA) have emerged as a promising alternative to the proliferation of heuristic and exact methods for the SFC embedding problem. Motivated by this observation, we design a GA that computes efficient embeddings with runtimes on par with other approximate and state-of-the-art methods. Recent trends for resource orchestration across the entire compute continuum raise the need for decision making at low timescales, a requirement which can be hardly met by centralized resource optimizers that rely either on exact methods or Machine Learning (ML). In this respect, we present an approach based on Distributed Deep Learning tailored to the embedding of SFCs onto large-scale network infrastructures. In order to confront the computational hardness of the SFC embedding problem, we utilize a clustering method for the partitioning of the solution space, empowering the search for efficient solutions in parallel across all clusters. The convergence of Internet of Things and edge/cloud computing technologies empowers the provisioning of Hyper-Distributed Applications (HDA), encompassing components that can be deployed at both the edge and core part of the cloud infrastructure. The placement of HDAs across the so-called compute continuum poses significant challenges stemming from the vast search space, the coexistence of resource and spatial demands, as well as the potential heterogeneity among the various cloud Points-of-Presence (PoPs). To address these challenges, we propose a novel hierarchical resource optimization framework for HDA placement across the compute continuum. Our approach decomposes the complex placement problem into two core steps. At first the HDA-graph is partitioned into segments and using the aforementioned GA based method the segments are assigned to cloud PoPs. Subsequently, a multi-agent distributed deep learning method computes the segment mapping within the PoPs. Our evaluation demonstrates that our placement strategy significantly improves request acceptance rates and resource utilization, while effectively minimizes path stretch across cloud PoPs.
Degree
thesis:*- Grantor dc:publisher
- Πανεπιστήμιο Μακεδονίας
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Rodis, Panteleimon
- Ρόδης, Παντελεήμων
Subjects
dc:subject × 18- Εικονικοποιημένες Υπηρεσίες Δικτύου
- Υπερ-κατανεμημένες Εφαρμογές
- Κατανεμημένη Βαθιά Μάθηση
- Γενετικοί αλγόριθμοι
- Ενορχήστρωση πόρων
- Virtualized Network Services
- Hyper-distributed applications
- Distributed deep learning
- Genetic algorithms
- Resource orchestration
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Φυσικές Επιστήμες
- Δίκτυα ηλεκτρονικών υπολογιστών και Επικοινωνίες
- Τεχνητή νοημοσύνη
- Computer and Information Sciences
- Natural Sciences
- Computer Networks and Communications
- Artificial Intelligence
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/60518