SWIFT: Enabling Large-Scale Temporal Graph Learning on a Single Machine

Temporal graph neural networks (T-GNNs) are crucial for modeling dynamic graphs, capturing evolving structures and interactions to address complex temporal properties in applications like event prediction, dynamic social network analysis, and temporal knowledge graph reasoning. However, training T-GNNs is hampered by the massive scale of graphs and complex temporal dynamics, leading to significant runtime and memory efficiency challenges. To tackle these challenges, this paper proposes SWIFT, the first secondary memory-based T-GNN training system for large-scale temporal graph learning on a single machine. SWIFT employs a novel bucket-based pipeline parallelism strategy to efficiently manage data flows across GPU, main, and secondary memories, addressing the computation and memory bottlenecks that hinder scaling for large-scale temporal graphs. Remarkably, SWIFT surpasses its main memory-based counterparts in runtime efficiency while requiring significantly less main memory. Extensive experiments demonstrate that SWIFT achieves up to a 4.3× speedup and a 7.9X reduction in main memory usage compared to state-of-the-art baselines on large temporal graphs.