MATADOR: ML-based Cloud Gaming Traffic Detection entirely in Programmable Hardware

Cloud gaming (CG) is gaining increasing interest as a new approach to delivering high-quality gaming anywhere, anytime, and on diverse devices. To improve the gaming experience, CG traffic needs to be monitored and prioritized to avoid high latency and jitter. Recent attempts to classify CG traffic on cloud nodes are time- and resource-consuming. Improving CG detection time and accuracy is critical to ensuring a responsive and seamless experience. This paper proposes MATADOR, a machine learning system that classifies CG traffic entirely in P4 programmable hardware. Our design and implementation using a Tofino switch overcome P4 challenges for hardware targets. The performance is analyzed for Tofino and a Python-based simulator. The obtained results demonstrate that, with an accuracy of around 97%, our approach significantly reduces the CG detection time when compared against alternative approaches.