Authors: Zihao Fan and Yunzhuo Liu (Shanghai Jiao Tong University), Tian Guo (Worcester Polytechnic Institute (WPI)), and Bo Jiang (Shanghai Jiao Tong University)
Abstract: Dynamic graph neural networks (DGNNs) enable learning from evolving graphs by capturing both structural and temporal dependencies. However, scaling DGNN training remains challenging due to three conflicting objectives: maximizing resource utilization, minimizing data movement, and maximizing data reuse. Existing systems typically optimize one objective at the expense of the others, leading to communication bottlenecks, load imbalance, or redundant computation.
We present REAL, an efficient distributed system for DGNN training that jointly addresses these challenges. REAL minimizes data transfers through snapshot-group scheduling and introduces an incremental aggregation mechanism that decouples computation from structural updates, enabling snapshot- and group-level reuse to eliminate redundant I/O and FLOPs. We further design two complementary schedulers: Real-L, an ILP-based optimizer for globally optimal planning, and Real-G, a scalable greedy heuristic. Experiments on six dynamic graph datasets and four DGNN models show that Real-L and Real-G achieve up to 3.26× and 3.16× speedup over state-of-the-art baselines.
Presentation: file