Distributed systems diagnosis using belief propagation

In this paper, we focus on diagnosis in distributed computer systems using end-to-end transactions, or probes. Diagnostic problem is formulated as a probabilistic inference in a bipartite noisy-OR Bayesian network. Due to general intractability of exact inference in such networks, we apply belief propagation (BP), a popular approximation technique proven successful in various applications, from image analysis to probabilistic decoding. Another attractive property of BP for our application is it natural parallelism that allows a distributed implementation of diagnosis in a distributed system to improve diagnostic speed and robustness. We derive lower bounds for diagnostic error in bipartite Bayesian networks, and particularly in noisy-OR networks, and provide promising empirical results for belief propagation on both randomly generated and realistic noisy-OR problems.