Our paper titled "GNNDrive: Reducing Memory Contention and I/O Congestion for Disk-based GNN Training" is accepted by the 53rd International Conference on Parallel Processing (ICPP 2024). In the paper, we studied the severe memory contention and I/O congestion issues that affect the training efficiency for large-scale GNN models. We accordingly proposed GNNDrive, a disk-based GNN training algorithm that minimizes memory footprint and avoids I/O congestion. More details can be found in the paper.

This work was solely completed by Toast Labmates. Two postgraduate students, Mr. Qisheng Jiang and Mr. Lei Jia, conducted it in their Master's programs. We are grateful to Dr. Cheng Chen for an early discussion. We also express sincere gratitude to Mr. Tianming Wen and Mr. Qun Xu for their support of a machine with multi-GPUs for us to do comprehensive experiments.