The Failure Prediction of Cluster Systems Based on System Logs

摘要

The failure prediction of cluster systems is an effective approach to improve the reliability of the cluster systems, which is becoming a new research hotspot of high performance computing, especially with the growth of cluster systems and applications both in scale and complexity. A classification sequential rule model is proposed to predict cluster system failures. The system logs of BlueGene/L, Red Storm, and Spirit are used as experimental datasets to predict cluster system failures. The results show that sequential rule approach outperforms SVM and HSMM in terms of precision and F-measure in 5hr prediction window, and in 1hr or 12hr prediction window, sequential rules, SVM and HSMM have their own strengths and weaknesses respectively.