Designing fault tolerant networks to prevent poison message failure
Xiaojiang Du, Mark A. Shayman, Ronald A. Skoog · Security and Communication Networks · 2008
Abstract Poison message failure is a mechanism that has been responsible for large‐scale failures in both telecommunications and IP networks. We design a fault management framework that integrates passive diagnosis and active diagnosis to identify the poison message and prevent network instability. Passive diagnosis uses real‐time inference and reasoning techniques to analyze network information and generates a probability distribution of the poison message, and the probability distribution is used in active diagnosis for further failure identification. In active diagnosis, message filtering is used to block suspect message types. Blocking messages affects network performance and service. The tradeoff of message filtering is formulated as a Markov Decision Process (MDP). The large size of the state space makes it impractical to use traditional techniques to solve the MDP. Con sequently, we use a combination of reinforcement learning and feature‐based function approximation to obtain a suboptimal policy. Extensive simulations demonstrate the effectiveness of passive diagnosis, and show that the suboptimal policy performs significantly better than a well‐known heuristic policy. Copyright © 2008 John Wiley & Sons, Ltd.