Boosting the performance of NFV services with SmartNIC

The world of mobile networking is dramatically changing due to the new 5G technologies, which will contribute to the development of a smarter and interconnected planet by means of a ubiquitous and faster network. These ambitious plans will establish new challenges in the networking field. In particular, one of the objectives to pursue is to find a way for increasing the performance of network middleboxes, by means of the virtualization of network functions and the smart orchestration of traffic flows. The introduction of NFV technologies may bring many benefits in terms of scalability and flexibility, but there are some drawbacks too, in terms of performance, compared to the solutions based on purpose-built hardware tightly coupled with network function software. In this context, the problem previously introduced can be mitigated by performing an offload of software-defined network functions, with the aim of reaching the true speed of the underlying hardware. In this work, we will investigate the new technology of a Smart-NIC, that promises to solve this problem by implementing offloading functionalities. In particular, we will focus on the feasibility of offloading eBPF code into the hardware of the Smart-NIC that has been considered in our study. At first, we will start to explore the performance of the Smart-NIC and its functionalities; we will also conduct some network performance tests with the aim of understanding the potentialities and the limitations of the hardware. Then, we will examine the DDoS mitigator use case and we will analyze the advantages introduced by using the Smart-NIC, in terms of dropped traffic, in order to exploit the possibility of reaching a line rate performance without affecting the host system. Finally, we will try to extend our approach in order to individuate a flexible solution that allows offloading a generic eBPF program, which could be a generic VNF. Moreover, the code will be analyzed with the goal of extracting known patterns and stateless decision paths that can be offloaded. The feasibility of performing the offload of stateful paths will be investigated too, focusing on the recognition of the paths determined by userspace inputs.