A Place-as-you-go In-network Framework by Modular Decomposition for Flexible Embedding to Software/Hardware Co-design

• Ano de defesa: 2022
• Autor(a) principal: Mafioletti, Diego Rossi
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Espírito Santo; BR; Doutorado em Ciência da Computação; Centro Tecnológico; UFES; Programa de Pós-Graduação em Informática
• Palavras-chave em Português: Computação em Rede; Co-design de Hardware/Software; Co-design de Hardware e Software; Virtualização de Funções de Rede; Computação de Borda; Paradigma 5G; subject.br-rjbn; Ciência da Computação
• Link de acesso: http://repositorio.ufes.br/handle/10/16570

Cloud computing has become very popular as a computation platform, being part of people’s daily life. In order to deliver the cloud-required super-computing power, the network must play a crucial role by connecting hundreds of thousands of machines within data centres. However, with the rise of a new massive set of cloud-native and intensive applications (e.g. 5G, cloud robotics, deep learning, etc) combined with network function virtualisation (NFV), a significant strain has been placed on the processing capacity of server CPUs, demanding even higher performance and turning the network into a bottleneck. Typically, the network interface card (NIC) has been used to connect servers to the network. Although, smart network interface cards (SmartNICs) are becoming an increasingly popular method of offloading intensive packet processing tasks from servers, thus freeing CPU cycles to drive application performance. The critical challenge towards using a SmartNIC is how to make efficient use of these in-network heterogeneous computing resources, as there is a significant gap between application software and the computing capabilities of programmable devices. First, this device lacks generic programming models or abstractions, being usually programmed using low-level primitives or proprietary APIs. Second, the network developer needs to deal with the internal complexity of hardware resources, as well as manage the balance on offloading workloads, trying to find out the tradeoff between additional overheads and offloading benefits. It is needed to figure out how to co-design application logic between programmable network hardware and end-host servers within the edge computing paradigm. This thesis presents a novel framework for prototyping and deploying in-network applications. The framework is structured into a set of components that rely on i) application functional decomposition; ii) identification of logic blocks; iii) aggregation of overlapped functions merged and mapped on network functionalities in P4 language; iv) intercepting, interacting and forwarding with data structures for balancing the offloading of network flows. In order to demonstrate the principle of co-designing on diverse applications, Virtual Network Functions (VNFs) are created, and some of their functionally decomposed elements are deployed as small embedded Network Functions (eNFs) on in-network processors in sorted use cases, reviewing the framework components raised previously: decomposing network functions (i, ii, iii) and unifying components to fit into a set of VNFs/eNFs, examining latency, throughput and vCPU usage against their software implementation counterpart; intercepting and interacting with cloud robotics (iv), raising security concerns when using programmable data planes; evolving to security network functions running at in-network computing (i, ii, iii, iv), checking the overhead added by them; and modifying a PON upstream scheduler mechanism (iv), providing low latency requirements for applications.