CUDA-DClust+: revisiting early GPU-accelerated DBSCAN clustering designs

Density-based clustering algorithms are widely used unsupervised data mining techniques to find the clusters of points in dense regions that are separated by low-density regions. This algorithm is inherently sequential and has limitations in its parallel implementation. There have been several parallel algorithms presented in the literature for multi-core CPUs and many-core GPUs. One such algorithm for the GPU is CUDA-DClust. In this paper, we propose a new GPU-accelerated DBSCAN algorithm with several optimizations. In comparison to prior work, our algorithm, CUDA-DClust+: (i) computes the indexing structure on the GPU, (ii) uses kernel fusion to combine the index search and cluster expansion kernels, which reduces communication and synchronization overhead with the host, and (iii) seed list management control is primarily given to the GPU rather than the CPU, which further decreases CPU-GPU communication overhead. We compare our algorithm to three state-of-the-art parallel algorithms in the literature on six real-world datasets. We find that our algorithm achieves a speedup of up to ~23x over the fastest GPU algorithm