Identifying and Analyzing Pitfalls in GNN Systems

Explore critical system design flaws in graph neural network (GNN) implementations through this 15-minute conference presentation from USENIX ATC '25. Examine how the widespread omission of training accuracy results in GNN system papers creates cascading problems across system design, implementation, framework integration, and evaluation processes. Discover how researchers from William & Mary conducted an in-depth analysis revealing fundamental pitfalls that question the practicality of many proposed system optimizations and affect the validity of conclusions in GNN research. Learn about the development of hypotheses, recommendations, and evaluation methodologies for addressing these issues, along with future research directions. Understand the quantitative impact of these pitfalls through GRAPHPY, a new prototype that establishes baseline memory consumption and runtime metrics for GNN training while introducing optimization strategies that can be integrated into existing frameworks to solve system-design problems efficiently and practically.