EGGPU: Enabling Efficient Large-Scale Network Analysis with Consumer-Grade GPUs
Year: 2024
Type: conference-paper
Abstract: With the increasing volume of data in the information age, existing network analysis tools are increasingly struggling to handle large-scale networks, for example, social networks. Considering the high parallel performance and wide spread adoption of consumer-grade graphics processing units (GPUs), we aim to better leverage the power of GPUs to accelerate network analysis. We build a new GPU-based network analysis library, called EGGPU, on top of the representative EasyGraph library. We conduct comparative experiments with nx-cugraph, Gunrock, and igraph, covering three key functions. Benefiting from the well-designed system architecture, parallel execution flow, and native CUDA C/C++ implementation, the results demonstrate that EGGPU consistently outperforms these counterparts across all tested network analysis functions, achieving speedups of up to 94.10x, 126.42x, and 28.49x in calculating between-ness centrality, single-source shortest path, and k-core centrality, respectively.
Institutions: Fudan University, Aalto University
Language: English
Topic: Graph Theory and Algorithms
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
Funder: Universitas Brawijaya
Award: G8841469610
OpenAlex ID: W4403728351