When Configuration Verification Meets Machine Learning: A DRL Approach for Finding Minimum k-Link Failures

Network configuration verifier is widely used for checking configuration updates before deployment. Finding the minimum number of failed links that breaks the reachability (i.e., minimum k-link failures) as fast as possible is an important feature of the configuration verifier for network maintenance. However, existing tools cannot check this property efficiently due to its NP-hardness. In this paper, we formally define the min-k-link-failure problem as finding min-cut edges in a route propagation graph, and are the first to prove its NP-hardness via rigorous analysis. We tackle this problem by emerging machine learning techniques for NP-hard problems. Specifically, we leverage the deep reinforcement learning approach to build a tool, KFinder, with the following steps: 1) using the message-passing neural network (MPNN) to embed the graph information; 2) incrementally constructing the failed links via MPNN and checking whether the current network is still reachable by a new algorithm for the special route propagation graph. We extensively evaluate the performance of KFinder using training datasets of random graphs and testing datasets of synthesized graphs derived from typical Internet topology. Simulation results show that our tool outperforms the state-of-the-art tool in terms of running time significantly despite of small sacrificed accuracy.