Xinbiao Gan
The scalability of graph neural networks (GNNs) is critically dependent on the efficiency of their sampling and feature aggregation steps, which are often bottlenecked by memory access patterns in large-scale graphs. While sampling algorithms like GraphSAGE reduce computational costs, they rely on underlying sparse storage formats such as the Compressed Sparse Row (CSR) format, treat all non-zero vertices equally, and fail to exploit the skewed degree distribution inherent to real-world graphs. To address these challenges, we introduce TianheWare, a degree-aware sparse storage format specifically engineered to optimize sampling during large-scale graph learning. TianheWare groups the numerous low-degree vertices found in real-world graphs, storing only a starting index for each group to minimize the memory footprint. This design not only reduces memory consumption but also, crucially, enables highly efficient batched memory access during the neighbor sampling phase in frameworks like GraphSAGE. A data-driven threshold automatically computed from the graph's degree distribution, adapts this compression to any graph structure. Integrated as a plugin into GraphSAGE, TianheWare demonstrates its impact on the end-to-end learning pipeline by accelerating the underlying sampling operations, achieving up to 3.37× speedup in sampling throughput and over 85% memory reduction compared to stateof- the-art methods, while maintaining full sampling fidelity and model accuracy. Our extensive evaluation, including deployment on a production-scale supercomputer, where it exceeded the topranked Graph500 benchmark performance, confirms that TianheWare serves as a foundational optimization, enabling faster, more scalable graph learning without compromising results.