Harnessing federated learning for anomaly detection in supercomputer nodes

High-performance computing (HPC) systems are a crucial component of modern society, with a significant impact in areas ranging from economics to scientific research, thanks to their unrivaled computational capabilities. For this reason, the worldwide HPC installation is steeply trending upwards, with no sign of slowing down. However, these machines are both complex, comprising millions of heterogeneous components, hard to effectively manage, and very costly (both in terms of economic investment and of energy consumption). Therefore, maximizing their productivity is of paramount importance. For instance, anomalies and faults can generate significant downtime due to the difficulty of promptly detecting them, as there are potentially many sources of issues preventing the correct functioning of computing nodes.

In recent years, several data-driven methods have been proposed to automatically detect anomalies in HPC systems, exploiting the fact that modern supercomputers are typically endowed with fine-grained monitoring infrastructures, collecting data that can be used to characterize the system behavior. Thus, it is possible to teach Machine Learning (ML) models to distinguish normal and anomalous states automatically. In this paper, we contribute to this line of research with a novel intuition, namely exploiting Federated Learning (FL) to improve the accuracy of anomaly detection models for HPC nodes. Although FL is not typically exploited in the HPC context, we show that FL can boost several types of underlying ML models, from supervised to unsupervised ones. We demonstrate our approach on a production Tier-0 supercomputer hosted in Italy. Applying FL to anomaly detection improves the average f-score from 0.46 to 0.87. Our research also shows FL can reduce the data collection time required to develop a representation data set, facilitating faster deployment of anomaly detection models. ML models need 5 months of training data for efficient anomaly detection performance while using FL reduces the training set by 15 times to 1.25 weeks.