GRAAFE: GRaph Anomaly Anticipation Framework for Exascale HPC systems

Martin Molan, Mohsen Seyedkazemi Ardebili, Junaid Ahmed Khan, Francesco Beneventi, Daniele Cesarini, Andrea Borghesi, Andrea Bartolini

Abstract: Highlights•Based on telemetry data, the GRAAFE framework predicts the compute node availability.•It is the first HPC anomaly prediction framework based on graph neural networks.•GRAAFE is a full-scale ML-ops framework for anomaly prediction in HPC.•It requires an additional 30% CPU and 5% more RAM compared to monitoring only.