利用虛擬節點實現GPU上密集圖高效GNN聚合

Abstract

圖神經網路（GNN）已廣泛應用於圖結構資料的學習任務，但由於其鄰居聚合階段具備不規則記憶體存取與低運算密度的特性，使得在 GPU 上高效執行仍具挑戰性。雖然現有方法如 binpack（如 PCKGNN）與冗餘消除（如 HAG）在加速GNN 上已取得成果，但前者在 bin 內仍存在重複記憶體存取，後者則需高昂的預處理成本。本研究提出 BinHAG，一種在 bin 層級進行階層式聚合的策略，透過引入虛擬節點（vnode）來消除 bin 中的冗餘計算。實驗結果顯示，BinHAG 在密集圖上相較於先前方法最高可達 31% 的效能提升。

Graph Neural Networks (GNNs) have become essential for learning on graphstructured data, yet their efficiency remains a challenge due to irregular memory access and low arithmetic intensity. While techniques such as binpacking (e.g., PCKGNN) and redundancy elimination (e.g., HAG) improve GPU performance, each faces limitation: bin-based methods suffer from intra-bin redundancy, and global redundancy reduction incurs high preprocessing costs. In this work, we propose BinHAG, a bin-level hierarchical aggregation strategy that introduces virtual nodes (vnodes) to eliminate redundant memory accesses within dense bins. Experiments show that BinHAG achieves up to 31% speedup over prior approaches on dense graphs.