An artificial neural network based model for online prediction of potential deadlock in multithread programs

In this paper we introduce a novel approach for online potential deadlock detection in multithread programs. Our approach is based on reasoning about deadlock possibility using the prediction of future behavior of threads. Predicting the future behavior of threads is not a trivial task. Due to the nondeterministic nature of multithread programs the future behavior of these programs, in most of cases, cannot be easily specified. In this work we extracted some specific behaviors of threads at runtime and then we converted extracted behaviors into a predictable format. Time series is a proper choice to this conversion. Many Statistical and also Artificial Intelligence techniques have been developed to predict the future members of time series. Among all the prediction techniques, artificial neural networks showed applicable performance and flexibility in predicting complex behavioral patterns which are the most usual cases in real world applications. We experimented our approach on some Java multithread programs which was deadlock prone. Applying our approach to this test suit, about 74% of deadlocks were predicted.