A real-time deadlock detection system using a hybrid graph neural network and LSTM framework

Abstract

The present invention is related to a real-time deadlock detection system using a hybrid graph neural network and LSTM framework. It is an effective cost-effective system design to help in increasing the reliability and the system performance to the unknown computing condition. The current invention demonstrates one example of a mechanism and a way to predict and identify deadlocks of concurrent and distributed computing systems by using an artificial intelligence approach. It models process and resource states in the form of directed graph. The invention applies machine learning algorithms to labeled sets of combinations of process-resource allocation patterns to provide an approximation of the likelihood that a given system state will be safe or deadlocked. The invention can execute in real-time relative to the previous cycle-detection algorithms or trace-based monitoring. It achieves this by deriving structural graph information and performing light-weight AI inference with less computational and false alarm. The technology can be installed on databases or a cloud infrastructure or operating systems. It continues to keep track of allocation requests, anticipates potential deadlocks before the execution shutdown. The solution is a helpful cost-effective system design to assist in raising the reliability and the system performance to the unknown computing condition capable of actively intervening.