Efficient Bitruss Decomposition on GPU

Shunyang Li

doi:10.1109/TKDE.2025.3569804

Return

Efficient Bitruss Decomposition on GPU

2025-08-01

PRE

Shunyang Li

王

王凯 (Kai Wang)

Wenjie Zhang (Wenjie Zhang)

林

林学民 (Xuemin Lin)

Yizhang He

DOI：10.1109/TKDE.2025.3569804

Original

Original request for help

Save

Abstract

En 中文

Cohesive subgraph computation on bipartite graphs has drawn significant research interest recently. As a popular cohesive subgraph model, $k$-bitruss is defined as the maximal subgraph where each edge is contained in at least $k$ butterflies (i.e., a (2, 2)-biclique). The bitruss decomposition problem is widely studied, which aims to compute all $k$-bitrusses for $k \geq 0$. The state-of-the-art CPU-based solutions require extensive costs to construct an index structure for grouping butterflies, leading to scalability challenges on large bipartite graphs. In this paper, we explore bitruss decomposition with GPU by leveraging the parallel computing capabilities of GPU architectures. As the index-based approach requires extensive space and the memory resources of GPUs are limited, we propose GBiD, which is a peeling-based algorithm on GPUs that utilizes a block-centric computation scheme to enable space-efficient bitruss decomposition without any indexing structure. In addition, cost-aware common neighbor exploration and neighbor list accessing optimizations are proposed to enhance GBiD by reducing the cost of enumerating butterflies and accessing the graph structure during the peeling process. Extensive experiments conducted on 10 real-world datasets demonstrate that our proposed techniques significantly surpass existing CPU-based solutions in terms of both space and time efficiency.

Keywords:

Bipartite graph

cohesive subgraph

GPU