Serverless computing has become a promising model for building scalable and flexible applications, particularly in the context of Multi-access Edge Computing (MEC). MEC involves bringing computation closer to the end users by leveraging edge nodes, such as 5G base stations, to meet the stringent latency requirements of modern applications like autonomous driving, mobile gaming, and augmented reality. By decentralizing computation, MEC significantly reduces latency and enhances performance for latency-sensitive applications. However, deploying serverless functions on MEC environments presents considerable challenges due to the inherent resource constraints of edge nodes. These nodes often have limited CPU, memory, and storage, which complicates efficient resource allocation and function placement. Moreover, the complexity of managing execution inter-dependencies among functions adds an additional layer of difficulty. Furthermore, the highly fluctuating and unpredictable nature of workloads, driven by users' geographic movements, exacerbates the issues of resource allocation and function placement. Existing solutions often fail to adequately address these challenges, leading to inefficient resource usage, increased latency, and suboptimal performance. This thesis proposes two dependency-aware approaches, Allocation and Placement, which aim to optimize resource allocation and function placement in MEC environments, respectively. Both approaches are built upon the NEPTUNE framework, which was originally designed to manage serverless functions in edge environments, but with limited consideration for function dependencies. Allocation introduces improvements to the resource allocation process by explicitly considering function dependencies, leading to up to a 67% reduction in CPU core usage when compared to the original NEPTUNE framework, as well as baseline approaches from the state-of-the-art while maintaining comparable response times and complying with service-level objectives (SLOs). In addition, Placement places functions with complex dependencies in MEC environments. In a comparison with five state-of-the-art placement strategies, Placement demonstrated clear outperformance, reducing latency by up to 90% and memory usage by up to 87% compared to the other approaches and showcasing its ability to optimize the deployment of serverless functions in resource-constrained edge environments. Through extensive simulations using both synthetic and real-world workloads, the thesis demonstrates that Allocation and Placement effectively optimize resource usage and improve the performance of serverless applications in MEC environments. By addressing the critical challenges associated with function dependencies, resource allocation, and placement, these approaches enable serverless applications to be more efficient, scalable, and performant in edge computing scenarios, especially in the face of fluctuating and unpredictable workloads.
Il calcolo serverless è diventato un modello promettente per la creazione di applicazioni scalabili e flessibili, in particolare nel contesto del Multi-access Edge Computing (MEC). Il MEC consiste nel portare il calcolo più vicino agli utenti finali sfruttando i nodi edge, come le stazioni base 5G, per soddisfare i rigorosi requisiti di latenza delle applicazioni moderne come la guida autonoma, il mobile gaming e la realtà aumentata. Decentralizzando il calcolo, il MEC riduce significativamente la latenza e migliora le prestazioni delle applicazioni sensibili alla latenza. Tuttavia, il deployment di funzioni serverless negli ambienti MEC presenta notevoli sfide a causa dei vincoli intrinseci delle risorse dei nodi edge. Questi nodi spesso dispongono di CPU, memoria e spazio di archiviazione limitati, il che complica l'allocazione efficiente delle risorse e il posizionamento delle funzioni. Inoltre, la complessità della gestione delle dipendenze di esecuzione tra le funzioni aggiunge un ulteriore livello di difficoltà. La natura altamente fluttuante e imprevedibile dei carichi di lavoro, guidata dai movimenti geografici degli utenti, aggrava ulteriormente i problemi di allocazione delle risorse e posizionamento delle funzioni. Le soluzioni esistenti spesso non riescono ad affrontare adeguatamente queste sfide, portando a un utilizzo inefficiente delle risorse, un aumento della latenza e prestazioni subottimali. Questa tesi propone due approcci consapevoli delle dipendenze, chiamati Allocation e Placement, che mirano rispettivamente a ottimizzare l'allocazione delle risorse e il posizionamento delle funzioni negli ambienti MEC. Entrambi gli approcci si basano sul framework NEPTUNE, originariamente progettato per gestire funzioni serverless negli ambienti edge, ma con una considerazione limitata per le dipendenze delle funzioni. Allocation introduce miglioramenti nel processo di allocazione delle risorse considerando esplicitamente le dipendenze delle funzioni, portando a una riduzione dell'uso dei core della CPU fino al 67% rispetto al framework NEPTUNE originale e agli approcci di riferimento dello stato dell'arte, mantenendo tempi di risposta comparabili e rispettando gli obiettivi di livello di servizio (SLO). Inoltre, Placement posiziona funzioni con dipendenze complesse negli ambienti MEC. In un confronto con cinque strategie di posizionamento all'avanguardia, Placement ha dimostrato un chiaro vantaggio, riducendo la latenza fino al 90% e l'uso della memoria fino all'87% rispetto agli altri approcci, evidenziando la sua capacità di ottimizzare il deployment delle funzioni serverless negli ambienti edge con risorse limitate. Attraverso simulazioni estese utilizzando carichi di lavoro sintetici e reali, la tesi dimostra che Allocation e Placement ottimizzano efficacemente l'uso delle risorse e migliorano le prestazioni delle applicazioni serverless negli ambienti MEC. Affrontando le sfide critiche associate alle dipendenze delle funzioni, all'allocazione delle risorse e al posizionamento, questi approcci consentono alle applicazioni serverless di essere più efficienti, scalabili e performanti negli scenari di edge computing, specialmente di fronte a carichi di lavoro fluttuanti e imprevedibili.
Dependency-aware placement and resource allocation in serverless edge computing
Ticongolo, Inacio Gaspar
2024/2025
Abstract
Serverless computing has become a promising model for building scalable and flexible applications, particularly in the context of Multi-access Edge Computing (MEC). MEC involves bringing computation closer to the end users by leveraging edge nodes, such as 5G base stations, to meet the stringent latency requirements of modern applications like autonomous driving, mobile gaming, and augmented reality. By decentralizing computation, MEC significantly reduces latency and enhances performance for latency-sensitive applications. However, deploying serverless functions on MEC environments presents considerable challenges due to the inherent resource constraints of edge nodes. These nodes often have limited CPU, memory, and storage, which complicates efficient resource allocation and function placement. Moreover, the complexity of managing execution inter-dependencies among functions adds an additional layer of difficulty. Furthermore, the highly fluctuating and unpredictable nature of workloads, driven by users' geographic movements, exacerbates the issues of resource allocation and function placement. Existing solutions often fail to adequately address these challenges, leading to inefficient resource usage, increased latency, and suboptimal performance. This thesis proposes two dependency-aware approaches, Allocation and Placement, which aim to optimize resource allocation and function placement in MEC environments, respectively. Both approaches are built upon the NEPTUNE framework, which was originally designed to manage serverless functions in edge environments, but with limited consideration for function dependencies. Allocation introduces improvements to the resource allocation process by explicitly considering function dependencies, leading to up to a 67% reduction in CPU core usage when compared to the original NEPTUNE framework, as well as baseline approaches from the state-of-the-art while maintaining comparable response times and complying with service-level objectives (SLOs). In addition, Placement places functions with complex dependencies in MEC environments. In a comparison with five state-of-the-art placement strategies, Placement demonstrated clear outperformance, reducing latency by up to 90% and memory usage by up to 87% compared to the other approaches and showcasing its ability to optimize the deployment of serverless functions in resource-constrained edge environments. Through extensive simulations using both synthetic and real-world workloads, the thesis demonstrates that Allocation and Placement effectively optimize resource usage and improve the performance of serverless applications in MEC environments. By addressing the critical challenges associated with function dependencies, resource allocation, and placement, these approaches enable serverless applications to be more efficient, scalable, and performant in edge computing scenarios, especially in the face of fluctuating and unpredictable workloads.| File | Dimensione | Formato | |
|---|---|---|---|
|
PhD_Thesis_FinalVersion(InacioTicongolo).pdf
accessibile in internet per tutti
Descrizione: DEPENDENCY-AWARE PLACEMENT AND RESOURCE ALLOCATION IN SERVERLESS EDGE COMPUTING
Dimensione
44.98 MB
Formato
Adobe PDF
|
44.98 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/10589/238537