Machine Learning–Driven Optimization of Graph Processing Performance on HPC Systems Using the Graph500 Benchmark

As large-scale artificial intelligence (AI) models such as GPT-4 and Gemini demand increasingly complex computations, optimizing graph processing performance in high-performance computing (HPC) systems has become essential. Unlike traditional benchmarks focusing on numerical operations, graph-based workloads emphasize connectivity and traversal efficiency, which are critical for AI training, data analytics, and large-scale knowledge modeling. This study develops an XGBoost-based performance prediction model using the Graph500 benchmark dataset to identify and optimize the factors affecting GTEPS (giga-traversed edges per second). By analyzing key system variables—such as memory size, problem scale, and node-core allocation—the model predicts graph processing performance with high accuracy (R² = 0.96). The results demonstrate that memory capacity and problem scale have the most significant influence, suggesting that balanced resource allocation can yield substantial performance gains without hardware expansion. This research contributes to the field by introducing a machine learning-driven approach for HPC optimization, enhancing both performance prediction accuracy and operational efficiency. The findings provide a practical framework for data-driven HPCresource management in future AI and graph analytics environments.