Flexible software-defined networks

• Autor(a) principal: Santos, Alexandre Veiga de Almeida
• Data de Publicação: 2021
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10773/33696
• topic: Next-generation software defined networks; P4; Programmable data plane; Offloading; Zenoh

The fifth generation of mobile networks (5G) are able to offer better services than its predecessors mainly through the usage of software defined networks (SDN) and network functions virtualization (NFV) However, after multiple solutions developed using OpenFlow, the conclusion was that the even after several years of the first version released, OpenFlow fails to offer full flexibility and cannot handle unknown protocols. With that in mind, the community got together and created what is known today as P4. P4 is a language designed to program the data plane behavior, that, with the help of P4Runtime, the alternative of OpenFlow to P4 enabled devices, it allows the management of the data plane behavior regarding the target or the protocol. All of that because, unlike OpenFlow, P4Runtime does not assume that network devices have a fixed and well defined behavior, usually described by the ASIC chip. In this work, P4 ecosystem is used to implement offloading of functions to the network devices and evaluate whether that is impactful for the network performance. Given the low amount of work developed with P4 regarding publish-subscribe systems, that traditionally rely on brokers, it was decided to offload several functions of such systems to the dataplane with P4, leading that the overall solution can be comparable to distributed broker ones. However, P4 is limited regarding the management of state related data, just like of TCP sessions, which many publish-subscribe system rely on. Zenoh, a new publish-subscribe protocol that is still in early phases and directed to IoT, is also able to run over UDP and therefore is a great candidate to be implemented in P4 to overcome such issues. It is then used to show the advantages of doing offloading of processing to the dataplane. The conceptualized system was then compared to two more traditional ones, that do not make use of offloading. The overall results achieved are promising. Results show that there are benefits in the offloading of certain tasks to the dataplane and therefore be closer to the end user and with that improve latency. However, regarding the pure Zenoh, the results achieved are poorer. That can be explained by the usage of software switches that are not production grade ready and whose performance is highly impacted by several data plane factors. That makes it necessary to do more tests on expensive hardware equipment for a more concrete conclusion.