Work

RTDBStream: Efficient DBSCAN in Data Streams with Ray Tracing Cores

Year: 2026

Type: conference-paper

Abstract: Density-based clustering has been widely used in various applications such as hot spot detection, image segmentation, and geo-social network analysis. DBSCAN, as a representative density-based clustering approach, is known for expensive computation overhead, especially when the data volume is large. Specifically, performing DBSCAN on high-speed continuous incoming data with stringent realtime requirement poses challenges for modern applications, such as traffic monitoring and industry sensor monitoring. This paper proposes RTDBStream, a high-performance DBSCAN approach that utilizes ray tracing cores (RT cores) in modern GPUs for acceleration. RTDBStream transforms DBSCAN in data streams into an efficient ray tracing job in a three-dimensional space. To efficiently utilize the RT cores, two optimization techniques are designed in RTDBStream, namely Identifying Cores by Gridding and Clustering by Creating Mixed-radius Spheres, to significantly accelerate the clustering process. Experimental results show that RTDBStream outperforms existing implementations on both CPUs and GPU CUDA cores, and it achieves up to 16.4× over existing RT-based implementations.

Source: ACM International Conference on Information and Knowledge Management (CIKM)

DOI: 10.1145/3799682.3841089