Themis: A passive-active hybrid framework with in-network intelligence for lightweight failure localization

The fast and efficient failure detection and localization is essential for stable network transmission. Unfortunately, existing schemes suffer from a few drawbacks such as significant resource consumption, lack of support for fast online failure localization, and limited applicable topologies. In this paper, we design Themis, a lightweight learning-based failure localization scheme for general networks. In the data plane, Themis achieves line-speed high performance failure detection using in-network classifiers and fine-grained traffic features. To reduce communication overhead, only coarse-grained traffic features are reported to the control plane for localization when a failure occurs. In the control plane, we propose a two-stage passive-active hybrid failure localization approach to accurately locate the failure without incurring excessive probing traffic. First, passive detection is conducted through the lightweight model XGBoost to infer a Potential Failure Link Set (PFLS). Then, active detection is done by only sending out probing packets to locations in the PFLS for precise failure localization. Comprehensive experiments demonstrate that Themis achieves ms-level failure localization with at least 95.63% accuracy, while saving 87.41% of bandwidth and 41.88% of hardware resource overhead on average compared with the state-of-the-art schemes.