Autor: Lucas de Angelo Martins Ribeiro (Currículo Lattes)
Resumo
Graph processing can be applied in various areas of society, technology, industry, and science to extract knowledge from real-world data. With the growth in the understanding and application of Artificial Intelligence, derived fields have emerged to explore the application of graphs using intelligent mechanisms, such as recommendation systems, social networks, among others. In this context, numerous models and frameworks for Graph Neural Networks have emerged, expanding the capabilities of these mechanisms. Despite significant advances in the study of Graph Neural Networks aimed at achieving better accuracy in models, the analysis of resources for processing these models is still an area that can be further explored to gain a deeper understanding of these systems from an architectural and execution environment perspective. In this work, a quantitative analysis of hardware resource consumption during the training phase of graph neural networks was conducted, taking into account potential benefits from using GPU accelerators when evaluating datasets of different compositions. The GCN and GraphSAGE models were used as the subject of study in this work, with implementations in the PyTorch Geometric framework. As a result, a significant improvement in the processing time of these models was observed when GPU accelerators were used, along with significant variations in the points of high resource utilization. It was also noted that datasets with different structural compositions (e.g., the number of edges and average node degree) can also exhibit significant variations in results.