CUDA-DClust+: Fast DBSCAN on GPU
A massively parallel DBSCAN clustering algorithm implemented on CUDA, published at HiPC 2021. Achieves a 22-fold speedup over CPU implementations by parallelizing both index construction and cluster expansion entirely on the GPU.
Key Contributions
GPU-parallel index construction — CUDA-DClust built the index on CPU; DClust+ moves it entirely to GPU, eliminating a non-negligible serial bottleneck
Correction merge routine — replaces the fixed seed list with a merge pass that recovers discarded border points, improving both accuracy and performance
Minimized CPU/GPU synchronization — seed list managed entirely within the expansion kernel until lists are empty, reducing round-trip overhead
22× speedup over baseline CPU DBSCAN on benchmark datasets
Publication
Presented at the 28th IEEE International Conference on High Performance Computing, Data, and Analytics (HiPC 2021). Research paper included in the repository.