Abstract: An online cluster anomaly job prediction method (OCAJP) is proposed. Firstly, a calculation of dynamic features of sub-tasks in the job was designed. Secondly, an improved gated recurrent unit (IGRU) neural network was designed according to the dynamic features. Then, the IGRU was used to predict whether the sub-task's final status was abnormal according to its dynamic features. Finally, the anomaly job was obtained based on the status relevance between the job and its sub-tasks, so as to complete prediction of abnormal jobs. The experimental results showed that OCAJP had a significant improvement in prediction sensitivity, error rate, accuracy, and prediction time compared with other prediction methods; this method had applicability in protecting the security of the cluster platform.

Key words: cluster anomaly job, dynamic features, real-time prediction, gated recurrent unit, status relevance