Research on Anomaly Detection Methods for HPC Systems

Abstract

Detecting anomalies in high-performance computing (HPC) systems is essential to keeping them reliable. However, this task is difficult because modern systems are very complex and there is little labeled failure data. Because of this, we need unsupervised methods that can learn only from normal operating data. Many existing methods mainly look at static system states. They often ignore the dynamic behaviors that show how faults start to develop. As a result, they may miss the first signs of problems and only detect failures after they have grown. This paper introduces a Difference Analysis (DA) method to solve this problem. This unsupervised method measures how reconstructed data from an LSTM-based autoencoder (RUAD) changes over time. Unlike methods that use only reconstruction error, DA directly models the instability of system behavior. We evaluated our method on monitoring data from the Marconi100 supercomputer. DA achieved an ROC AUC of 0.881, while the conventional reconstruction error method scored 0.770. These results show that studying dynamic state changes is a more effective way than relying only on static reconstruction error to improve the reliability of large-scale HPC systems. DA may also provide useful signals for anomaly anticipation, which we leave for future study.