GraphSnapShot: Graph Machine Learning Acceleration through Fast Arch, Storage, Caching and Retrieval

TMLR Paper4827 Authors

Abstract: Large-scale graph machine learning suffers from prohibitive I/O latency, memory bottlenecks, and redundant computation due to the complexity of multi-hop neighbor retrieval and dynamic topology updates. We present \textbf{GraphSnapShot: Graph Machine Learning Acceleration through Fast Arch, Storage, Caching and Retrieval}, a system that decouples graph storage layout from runtime cache management to maximize data reuse and access efficiency. GraphSnapShot introduces two key components: (1) \textbf{SEMHS}, a hop-aware storage layout that co-locates neighbors in contiguous disk slabs for efficient one-burst DMA access; and (2) \textbf{GraphSDSampler}, a multi-level variance-adaptive caching module that optimizes refresh policies based on gradient statistics. Together, they form a hybrid disk–cache–memory architecture that supports high-throughput training over billion-scale graphs. Experiments on ogbn-arxiv, ogbn-products, and ogbn-mag demonstrate that GraphSnapShot achieves up to \textbf{4.9×} loader throughput, \textbf{83.5\%} GPU memory savings, and \textbf{29.6\%} end-to-end training time reduction compared to baselines like DGL’s NeighborSampler and uniform samplers. These results establish GraphSnapShot as a scalable and efficient solution for dynamic graph learning at industrial scale.

Submission Length: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Giannis_Nikolentzos1

Submission Number: 4827