Enhancing Graph Execution for Performance and Energy Efficiency on NUMA Machines

Graph analysis, crucial in fraud detection and social networks, often execute in a non-optimized fashion on Non-Uniform Memory Access (NUMA) machines. To enhance performance while providing energy-savings, tuning thread and data placement is crucial, which also may involve reducing active threads for scalability-limited applications. In this paper, we propose PtGraph, a graph processing framework that optimizes Thread and Page Mapping, while automatically configuring thread count. Leveraging unique graph features, PtGraph trains Artificial Neural Networks (ANNs) for predictions, combining the precision of offline methods with online adaptability. PtGraph demonstrates significant Energy-Delay Product (EDP) improvements, outperforming the Linux OS by 68 % and achieving up to 1.18 × improvement over eight strateaies.