Stats 101 in P4: Towards In-Switch Anomaly Detection (Presentation Video)

Data plane programmability is greatly improving network monitoring. Most new proposals rely on controllers pulling information (e.g., sketches or packets) from the data plane. This architecture is not a good fit for tasks requiring high reactivity, such as failure recovery, attack mitigation, and so on. Focusing on these tasks, we argue for a different architecture, where the data plane autonomously detects anomalies and pushes alerts to the controller. As a first step, we demonstrate that statistical checks can be implemented in P4 by revisiting definition and online computation of statistical measures. We collect our techniques in a P4 library, and showcase how they enable in-switch anomaly detection.
Sam Gao (University College London), Mark Handley (University College London), Stefano Vissicchio (University College London),
Created with Midspace: midspace.app/