The exponential growth of connected devices, alongside the widespread adoption of Machine Learning (ML) applications, has significantly increased computational and communication demands on modern telecommunication networks. Addressing these challenges requires integrating advanced ML techniques like Federated Learning (FL) with emerging networking paradigms such as Multi-access Edge Computing (MEC) and Software-Defined Wide Area Networks (SD-WANs). Therefore, an urgent, thorough evaluation of intelligent resource management and ML's impact in MEC-based SD-WAN networks is necessary. This work conducts a detailed analysis of enhancing FL through the integration of MEC and SD-WAN architectures. We introduce an innovative methodology incorporating dynamic global aggregator node selection and intermediate aggregation techniques to boost FL implementations' efficiency and scalability. By leveraging FL's distributed nature in conjunction with MEC and SD-WAN technologies, our approach significantly reduces latency, and optimizes network resource utilization. Additionally, integrating an inference phase within training rounds facilitates real-time model updating and evaluation, markedly improving system responsiveness and accuracy. Our results demonstrate that the proposed enhancements not only address the limitations of traditional centralized FL systems but also provide a robust framework for deploying machine learning models in edge computing environments. Specifically, the Decentralized Federated Learning (DFL) approach shows a significant improvement in average resource utilization, with an increase of up to 25% compared to other solutions. This improvement highlights DFL's efficiency and scalability in handling distributed data and computational resources. This research lays a solid foundation for future studies on decentralized machine learning architectures and their applications across various real-world scenarios.
La crescita esponenziale dei dispositivi connessi, insieme alla diffusione delle applicazioni di Machine Learning (ML), ha aumentato significativamente le esigenze computazionali e di comunicazione delle moderne reti di telecomunicazioni. Affrontare queste sfide richiede l'integrazione di tecniche ML avanzate come il Federated Learning (FL) con paradigmi di rete emergenti come il Multi-access Edge Computing (MEC) e le Software-Defined Wide Area Networks (SD-WANs). Pertanto, è necessaria una valutazione urgente e approfondita della gestione intelligente delle risorse e dell'impatto del ML nelle reti SD-WAN basate su MEC. Questo lavoro conduce un'analisi dettagliata dell'integrazione del FL con le architetture MEC e SD-WAN. Introduciamo una metodologia innovativa che incorpora la selezione dinamica del nodo aggregatore globale e tecniche di aggregazione intermedia per migliorare l'efficienza e la scalabilità delle implementazioni FL. Sfruttando la natura distribuita del FL in combinazione con le tecnologie MEC e SD-WAN, il nostro approccio riduce la latenza e ottimizza l'utilizzo delle risorse di rete. Inoltre, l'integrazione di una fase di inferenza nei cicli di addestramento facilita l'aggiornamento e la valutazione in tempo reale del modello, migliorando notevolmente la reattività e l'accuratezza del sistema. I nostri risultati dimostrano che i miglioramenti proposti non solo affrontano le limitazioni dei sistemi FL centralizzati tradizionali, ma forniscono anche un quadro robusto per il dispiegamento di modelli di machine learning negli ambienti di edge computing. In particolare, l'approccio Decentralized Federated Learning (DFL) mostra un miglioramento significativo nell'utilizzo medio delle risorse, con un incremento fino al 25% rispetto ad altre soluzioni. Questo miglioramento evidenzia l'efficienza e la scalabilità del DFL nella gestione di dati distribuiti e risorse computazionali. Questa ricerca pone una solida base per futuri studi sulle architetture di machine learning decentralizzate e le loro applicazioni in vari scenari reali.
Adaptive resource allocation strategies for distributed machine learning in MEC driven SD WANs
Spatocco, Carlo
2023/2024
Abstract
The exponential growth of connected devices, alongside the widespread adoption of Machine Learning (ML) applications, has significantly increased computational and communication demands on modern telecommunication networks. Addressing these challenges requires integrating advanced ML techniques like Federated Learning (FL) with emerging networking paradigms such as Multi-access Edge Computing (MEC) and Software-Defined Wide Area Networks (SD-WANs). Therefore, an urgent, thorough evaluation of intelligent resource management and ML's impact in MEC-based SD-WAN networks is necessary. This work conducts a detailed analysis of enhancing FL through the integration of MEC and SD-WAN architectures. We introduce an innovative methodology incorporating dynamic global aggregator node selection and intermediate aggregation techniques to boost FL implementations' efficiency and scalability. By leveraging FL's distributed nature in conjunction with MEC and SD-WAN technologies, our approach significantly reduces latency, and optimizes network resource utilization. Additionally, integrating an inference phase within training rounds facilitates real-time model updating and evaluation, markedly improving system responsiveness and accuracy. Our results demonstrate that the proposed enhancements not only address the limitations of traditional centralized FL systems but also provide a robust framework for deploying machine learning models in edge computing environments. Specifically, the Decentralized Federated Learning (DFL) approach shows a significant improvement in average resource utilization, with an increase of up to 25% compared to other solutions. This improvement highlights DFL's efficiency and scalability in handling distributed data and computational resources. This research lays a solid foundation for future studies on decentralized machine learning architectures and their applications across various real-world scenarios.File | Dimensione | Formato | |
---|---|---|---|
Adaptive_Resource_Allocation_Strategies_for_Distributed_Machine_Learning_in_MEC_driven_SD_WANs.pdf
non accessibile
Descrizione: Thesis
Dimensione
4.86 MB
Formato
Adobe PDF
|
4.86 MB | Adobe PDF | Visualizza/Apri |
Executive_Summary.pdf
non accessibile
Descrizione: Executive Summary
Dimensione
812.76 kB
Formato
Adobe PDF
|
812.76 kB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/223580