Mobile operators face the challenge of deploying a flexible network to handle the emerging use cases. These applications, typically divided into enhanced Mobile BroadBand (eMBB), ultra Reliable Low Latency Communication (uRLLC), and massive Machine Type Communication (mMTC), present very specific requirements in terms of bandwidth, device density and latency. A new Radio Access Network (RAN) for 5G was proposed to provide more dynamic, low-cost and energy-efficient network management, resource allocation, and service provisioning compared to 4G. One of its main features is the deployment of functional splits. They define the separation of RAN baseband functions that were previously performed next to the antennas and the further distribution into different radio-network units: Radio Unit (RU), Distributed Unit (DU), Centralized Unit (CU). Therefore, operators can take advantage of this functionality together with virtualization technologies to improve the network management and their overall costs. Nevertheless, this architecture brings new challenges to network operators. The optimal placement of these units is non-trivial: it depends on the application requirements, on the expected traffic volume, and on the costs and power derived by the placement. In addition, application requirements and expected traffic may vary in time due to users instantaneous needs. Therefore, performing hourly reconfigurations to comply with the traffic fluctuation may lead to service disruptions. Consequently, it is important to plan the baseband Virtual Network Functions (VNF) placement and service provisioning in advance to ensure efficient network reconfiguration. This PhD dissertation optimizes the resource allocation in mobile networks to minimize cost and guarantee compliance with the quality of service required by users. We propose several optimization techniques to enable static and dynamic resource allocation, including mathematical programming, heuristic, black-box, machine learning, and auction. We developed mixed integer linear programming models and heuristic algorithms to solve the baseband virtual network function problem in metro-area networks considering different scenarios. There is a significant effort in the literature to solve it. Nevertheless, aspects related to the costs and to the functional split constraints were not fully considered. The goal of the models proposed in this work is to minimize the overall system power consumption, including both network and node components. This optimization is subject to all split-related constraints and to the service being carried. In order to plan the baseband VNF placement in advance, the works available in the literature typically deploy machine learning algorithms, because they can provide accurate traffic forecast and improves the applicability of optimization frameworks in real-time scenarios. However, unpredictable events may create perturbations in these patterns. Therefore, the traditional techniques used to predict the traffic and then optimize the placement fail to provide solutions that ensures feasibility in real-time. Indeed, they do not guarantee that the real traffic could be carried by the anticipated placement. Instead, we proposed two different techniques to perform early and efficient planning of the baseband VNF placement. We propose machine-learning technique that exploits a fine-grained two-step multitask algorithm. By forecasting the mean and quantile expected traffic, the deployment of the artificial capacity is no longer necessary. Furthermore, we propose a novel framework that uses black-box optimization to train a traffic prediction algorithm based on the optimization outcomes. The goal is to minimize a loss function related to power consumption and constraint violation to ensure that the predicted placement is feasible and that its consumption is close to optimal. The results of these works confirm that machine learning techniques cannot be blindly used into optimization models. However, by applying intelligent mechanisms that overestimate the demands, we can ensure the feasibility of the placement in real time.
Gli operatori mobili devono affrontare la sfida di implementare una rete flessibile per gestire le nuove applicazioni, tipicamente suddivise in enhanced Mobile BroadBand (eMBB), Ultra Reliable Low Latency Communication (uRLLC) e Massive Machine Type Communication (mMTC). Queste applicazioni presentano requisiti molto specifici in termini di larghezza di banda, densità del dispositivo e latenza. È stata proposta una nuova rete di accesso radio (Radio Access Network - RAN) per il 5G per migliorare la gestione della rete, l’allocazione delle risorse e la fornitura di servizi, in modo ad avere più dinamicità, costi più bassi ed efficienza dal punto di vista energetico rispetto al 4G. Una delle sue caratteristiche principali è l'implementazione di functional splits, che definiscono la separazione delle funzioni di banda base RAN che venivano precedentemente eseguite nelle antenne e l'ulteriore distribuzione in diverse unità di rete radio: Radio Unit (RU), Distributed Unit (DU), Centralized Unit (CU). Pertanto, gli operatori possono sfruttare questa funzionalità insieme alle tecnologie di virtualizzazione per migliorare la gestione della rete ei costi complessivi. Tuttavia, questa architettura pone nuove sfide agli operatori. Il posizionamento ottimale di queste unità dipende dai requisiti dell'applicazione, dal volume di traffico previsto e dai costi e dalla potenza derivanti dal posizionamento. Inoltre, i requisiti dell'applicazione e il traffico previsto possono variare nel tempo a causa delle esigenze istantanee degli utenti. Pertanto, l'esecuzione di riconfigurazioni periodiche per conformarsi alla fluttuazione del traffico può causare interruzioni del servizio. Di conseguenza, è importante pianificare in anticipo il posizionamento delle funzioni di banda base virtualizzate e la fornitura dei servizi per garantire una riconfigurazione efficiente della rete. Questa tesi di dottorato ottimizza l'allocazione delle risorse nelle reti mobili per ridurre i costi e garantire il rispetto della qualità del servizio richiesta dagli utenti. Proponiamo diverse tecniche di ottimizzazione per abilitare l'allocazione statica e dinamica delle risorse, tra cui programmazione matematica, euristica, black-box, machine learning e aste. Abbiamo sviluppato modelli di programmazione matematica e algoritmi euristici per risolvere il problema dell’allocazione delle funzioni di banda base virtuali nelle reti metropolitane considerando diversi scenari. C'è uno sforzo significativo in letteratura per risolverlo. Tuttavia, gli aspetti relativi ai costi e ai vincoli del functional split non sono stati pienamente considerati. L'obiettivo dei modelli proposti in questo lavoro è minimizzare il consumo energetico complessivo del sistema, inclusi i componenti di rete e di elaborazione del segnale. Questa ottimizzazione è soggetta a tutti i vincoli relativi al functional split e al servizio svolto. Al fine di pianificare in anticipo il posizionamento delle funzioni di banda base virtuali, i lavori disponibili in letteratura utilizzano tipicamente algoritmi di machine learning, perché possono fornire previsioni di traffico accurate e migliorano l'applicabilità dei framework di ottimizzazione in scenari in tempo reale. Tuttavia, eventi imprevedibili possono creare perturbazioni in questi modelli. Pertanto, le tecniche tradizionali utilizzate per prevedere il traffico e quindi ottimizzare il posizionamento non riescono a fornire soluzioni che garantiscano la fattibilità in tempo reale. Invece, abbiamo proposto due diverse tecniche per eseguire una pianificazione anticipata ed efficiente del posizionamento delle funzioni di banda base virtuali. Proponiamo una tecnica di machine learning che sfrutta un algoritmo multitasking in due fasi. Prevedendo il traffico medio e quantile atteso, non è più necessario il dispiegamento della capacità artificiale. Inoltre, proponiamo un nuovo framework che utilizza l'ottimizzazione del tipo black-box per addestrare un algoritmo di previsione del traffico basato sui risultati dell'ottimizzazione. L'obiettivo è ridurre una funzione di costo correlata al consumo di energia e alla violazione dei vincoli per garantire che il posizionamento previsto sia fattibile e che il suo consumo sia vicino all'ottimale. I risultati di questi lavori confermano che le tecniche di machine learning non possono essere utilizzate alla cieca nei modelli di ottimizzazione. Tuttavia, applicando meccanismi intelligenti che sovrastimano le richieste, possiamo garantire la fattibilità del posizionamento in tempo reale.
Optimization techniques for virtual baseband function placement of 5G radio access in metro-area networks
Moreira Zorello, Ligia Maria
2021/2022
Abstract
Mobile operators face the challenge of deploying a flexible network to handle the emerging use cases. These applications, typically divided into enhanced Mobile BroadBand (eMBB), ultra Reliable Low Latency Communication (uRLLC), and massive Machine Type Communication (mMTC), present very specific requirements in terms of bandwidth, device density and latency. A new Radio Access Network (RAN) for 5G was proposed to provide more dynamic, low-cost and energy-efficient network management, resource allocation, and service provisioning compared to 4G. One of its main features is the deployment of functional splits. They define the separation of RAN baseband functions that were previously performed next to the antennas and the further distribution into different radio-network units: Radio Unit (RU), Distributed Unit (DU), Centralized Unit (CU). Therefore, operators can take advantage of this functionality together with virtualization technologies to improve the network management and their overall costs. Nevertheless, this architecture brings new challenges to network operators. The optimal placement of these units is non-trivial: it depends on the application requirements, on the expected traffic volume, and on the costs and power derived by the placement. In addition, application requirements and expected traffic may vary in time due to users instantaneous needs. Therefore, performing hourly reconfigurations to comply with the traffic fluctuation may lead to service disruptions. Consequently, it is important to plan the baseband Virtual Network Functions (VNF) placement and service provisioning in advance to ensure efficient network reconfiguration. This PhD dissertation optimizes the resource allocation in mobile networks to minimize cost and guarantee compliance with the quality of service required by users. We propose several optimization techniques to enable static and dynamic resource allocation, including mathematical programming, heuristic, black-box, machine learning, and auction. We developed mixed integer linear programming models and heuristic algorithms to solve the baseband virtual network function problem in metro-area networks considering different scenarios. There is a significant effort in the literature to solve it. Nevertheless, aspects related to the costs and to the functional split constraints were not fully considered. The goal of the models proposed in this work is to minimize the overall system power consumption, including both network and node components. This optimization is subject to all split-related constraints and to the service being carried. In order to plan the baseband VNF placement in advance, the works available in the literature typically deploy machine learning algorithms, because they can provide accurate traffic forecast and improves the applicability of optimization frameworks in real-time scenarios. However, unpredictable events may create perturbations in these patterns. Therefore, the traditional techniques used to predict the traffic and then optimize the placement fail to provide solutions that ensures feasibility in real-time. Indeed, they do not guarantee that the real traffic could be carried by the anticipated placement. Instead, we proposed two different techniques to perform early and efficient planning of the baseband VNF placement. We propose machine-learning technique that exploits a fine-grained two-step multitask algorithm. By forecasting the mean and quantile expected traffic, the deployment of the artificial capacity is no longer necessary. Furthermore, we propose a novel framework that uses black-box optimization to train a traffic prediction algorithm based on the optimization outcomes. The goal is to minimize a loss function related to power consumption and constraint violation to ensure that the predicted placement is feasible and that its consumption is close to optimal. The results of these works confirm that machine learning techniques cannot be blindly used into optimization models. However, by applying intelligent mechanisms that overestimate the demands, we can ensure the feasibility of the placement in real time.| File | Dimensione | Formato | |
|---|---|---|---|
|
Optimization_techniques_for_virtual_baseband_function_placement_for_5G_radio_access_in_metro_area_networks.pdf
non accessibile
Dimensione
46.5 MB
Formato
Adobe PDF
|
46.5 MB | 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/10589/188459