Machine learning-powered traffic processing in commodity hardware with eBPF

Network softwarization is paving the way for the design and development of Next-Generation Networks (NGNs), which are demanding profound improvements to existing communication infrastructures. Two of the fundamental pillars of NGNs are flexibility and intelligence to create elastic network functions capable of managing complex communication systems in an efficient and cost-effective way. In this sense, the extended Berkeley Packet Filter (eBPF) is a state-of-the-art solution that enables low-latency traffic processing within the Linux kernel in commodity hardware. When combined with Machine Learning (ML) algorithms, it becomes a promising enabler to perform smart monitoring and networking tasks at any required place of the fog-edgecloud continuum. In this work, we present a solution that leverages eBPF to integrate ML-based intelligence with fast packet processing within the Linux kernel, enabling the execution of complex computational tasks in a flexible way, saving resources and reducing processing latencies. A real implementation and a series of experiments have been carried out in an Internet of Things (IoT) scenario to evaluate the performance of the solution to detect attacks in a 6LowPAN system. The performance of the in-kernel implementation shows a considerable reduction in the execution time (-97%) and CPU usage (-6%) of a Multi-Layer Perceptron (MLP) model in comparison with a user space development approach; thus positioning our proposal as a promising solution to embed ML-powered fast packet processing within the Linux kernel.