Scalable GNN Training for Track Finding

Abstract

Graph Neural Networks (GNNs) are widely used for particle track finding in High-Energy Physics but are computationally expensive to train on large graph datasets. We study Distributed Data Parallelism (DDP) for accelerating GNN training across multiple GPUs and analyze its impact on runtime and convergence. We evaluate both strong and weak scaling behavior and show that while DDP substantially reduces training time, speedup saturates at larger GPU counts due to communication overhead. In addition, increasing the number of GPUs degrades validation efficiency due to growth in effective batch size. We demonstrate that learning-rate scaling partially mitigates this degradation. Results on the TrackML dataset highlight a trade-off between throughput and model quality that must be addressed for scalable GNN training.