Emerging trends such as cloud computing, the internet of things, and augmented and virtual reality demand highly responsive, available, secure, ubiquitous, and scalable networks to meet users' quality of experience expectations. Operators currently manage these networks and protocols using a variety of ad‐hoc tools and scripts; however, the unpredictable and complex interactions between network conditions and workloads make such manual tuning difficult. What is needed is an autonomous approach to network management, i.e., networks should self‐drive: sensing, inferring, and deciding without human intervention.
In this project, we present the design and implementation of Taurus, an intelligent data plane capable of running machine learning inference at line rate. Taurus modifies existing programmable network devices, such as switches and NICs, by adding custom hardware based on a map‐reduce abstraction; the new block uses SIMD and pipeline parallelism for fast inference. Our evaluation shows that Taurus achieves an average latency of 270 ns‐‐‐three orders of magnitude better than traditional control‐plane, server‐based approaches‐‐‐while adding at most 24% more chip area using anomaly detection and traffic scheduling models. We believe Taurus is the first step toward realizing a self‐driving network, where data‐plane devices make intelligent decisions and operate autonomously using machine learning.