Graph Neural Networks for Anomaly Anticipation in HPC Systems

This thesis investigates the potential of Graph Neural Networks (GNNs) for anticipating anomalies in high performance computing (HPC) systems. The proposed approach is based on the structure of the HPC system, which involves compute nodes in close proximity to each other. This is possible due to GNNs. GNNs learn to represent each node in the graph by aggregating information from its neighbors, allowing them to capture the graph’s topology and structural information. The task is framed as a supervised prediction problem, where the GNN predicts the probability of a compute node failing within a fixed-length future window. Experiments were carried out on a production Tier-0 supercomputer hosted at CINECA datacenter facilities, the largest Italian provider of HPC, and they demonstrated promising results in anomaly prediction, especially when tested on real, production data.