Serverless Computing for NFV: Is it Worth it? A Performance Comparison Analysis

Network Function Virtualization has established itself as one of the most important paradigms towards software-based networking. While today Virtual Network Functions (VNFs) are typically deployed in the form of serverful virtual-machine or container-based applications, the emergence of serverless computing opens the door to the possibility of implementing them as serverless functions, with benefits in terms of scalability and resource efficiency. This paper aims to assess whether this really makes sense or not, given the system-level overheads that a serverless computing platform naturally brings. We propose an open source platform designed to optimize the execution of network-intensive VNFs and we implement a data-plane and a control-plane function (i.e., NAT and DHCP responder, respectively) as serverless functions. We carry out extensive benchmarking of performance with their serverful counterparts, implemented as stand-alone containerized applications. Our experience makes it possible to conclude that serverless computing is beneficial for the execution of short-lived and request-based control-plane VNFs, while it should be avoided for the execution of data-plane traffic-intensive VNFs.