FSGraph: fast and scalable implementation of graph traversal on GPUs

Graph is one of the best ways to express and process association relationship. It is widely used in various applications, including social networks, fraud detection, Internet of things, etc. As a typical graph traversal algorithm, the Breadth-First Search (BFS) performance on GPU is not desirable, due to strong data dependency, intensive irregular memory access and low computation intensity. On GPUs, the situation is even worse with unbalanced data partitioning and high communicationto-computation ratios. In this paper, we implement FSGraph that is a fast and scalable BFS implementation on GPUs. In FSGraph, we propose three optimizing techniques: GPU-friendly Compressed Sparse Row (CSR) structure, bidirectional one-dimensional (1d) partition and UM-aware communication. We have evaluated our work with extensive experiments on four T4 and four V100 GPUs. The average performance of BFS on four T4 GPUs is 132.67 Giga-Traversed Edges per Second (GTEPS), which delivers up to 1.44x improvement than that on single T4. In terms of four V100 GPUs, the BFS performance achieves 392.35 GTEPS and outperforms existing CPU-based cluster with 1024 nodes on November 2022 Graph500 list.