Integrating Machine Learning Models into the Linux Kernel: Opportunities and Challenges

The advent of the next generation of communication networks demands deep changes within current infrastructures. In this regard, the softwarization of the network following the Software Defined Networking (SDN) and Network Function Virtualization (NFV) paradigms will be crucial to permit flexibility and programmability levels never seen before. This paper studies the opportunities that both the extended Berkeley Packet Filter (eBPF) and Machine Learning (ML) technologies bring to this field. Their convergence enables the integration of intelligent ML-powered programs into the Linux kernel hence permitting almost every device in the network to perform a plethora of tasks, e.g., traffic processing, security analysis, etc. However, this integration is not straight-forward and poses a series of challenges that are discussed and solved in this work. To this end, a methodology for integrating complex ML models within the Linux kernel is proposed introducing specific tools in its different phases. This integration paves the way for the development of diverse virtual network functions in commodity hardware in an efficient and secure way.