Publication
PaGraph significantly reduces the data loading time in sampling-based GNN training by exploiting available GPU resources to keep frequently accessed graph data in cache. It also employs a lightweight yet effective computation-aware graph partitioning strategy to reduce data movement during training.
Zhiqi Lin, Cheng Li, Youshan Miao, Yunxin Liu, Yinlong Xu
PaGraph reduces data loading time in GNN training by caching frequently accessed graph data in GPU memory and using computation-aware graph partitioning, achieving significant speedups over existing approaches.