On the Integration of Machine Learning Models within the Linux Kernel through eBPF-based Development

So far, eBPF has being used in combination with ML models by employing eBPF programs as mere data collectors at kernel level and feeding a Machine Learning (ML) algorithm running in user space. In this talk, we present a solution that integrates advanced ML processing within the Linux kernel to provide intelligent fast packet inspection. Concretely, by exploiting the TinyML paradigm and the characteristics of eBPF, we have implemented a complex model such as a Multi Layer Perceptron (MLP) in kernel space. This proposal allows to significantly reduce processing latencies compared to similar solutions implemented outside the kernel, as well as enabling the deployment of advanced network functions in a flexible and dynamic way on devices not specialized in network management tasks.