Heterogeneous Graph Neural Networks for I/O Performance Bottleneck Diagnosis

This proposal aims to address the challenge of I/O bottlenecks in High-Performance Computing (HPC) systems by developing a Graph Neural Network (GNN) model for accurate and interpretable I/O performance predictions. The approach includes creating a data pre-processing pipeline for I/O logs, constructing a GNN model tailored to the intricacies of HPC I/O, and implementing SHAP value analysis for insightful feature contribution understanding. Deliverables encompass the GNN model, integration with the AIIO framework, and a SHAP-based evaluation methodology to enhance diagnostic precision and accountability in I/O performance assessments.