Towards Hybrid Network-Oriented Workloads on Edge Data Centers

With the advent of Cloud Computing, general-purpose workloads have been transformed into a set of cooperating micro-services, meeting the rapid, frequent and reliable delivery of very large and complex applications. Telcos have been trying to apply the same approach for network-oriented workloads, keeping into account the complexity and the high requirements needed. Among the leading technologies in this direction, eBPF represents an interesting option. eBPF is an in-kernel framework that also allows to implement data-plane services but it poses some limitations that prevents its applicability in some cases, which can be found for example in the Rate Limiter network function, which should be always present if policies are to be applied on the traffic. This thesis studies the possibilities of designing and implementing such a network function by using two technologies which could run simultaneously on a Linux machine: eBPF and AF\_XDP. A hybrid prototype is proposed, highlighting the advantages and challenges coming from this integration. Finally, performance are compared with traditional Traffic Control in Linux at first and then by changing the amount of traffic handled by the two technologies. Results show that AF\_XDP, besides overcoming limits coming from eBPF, could even achieve higher performance thanks to its capability of bypassing the Linux kernel. This might be an interesting solution to be considered when running virtual network functions at edge data centers, where both the flexibility and scalability requirements should be fulfilled and performance should be kept reasonably high.