An Auto-tuning Solution to Data Streams Clustering in OpenCL

Due to its applicability to numerous types of data, including telephone records, web documents, and click streams, the data stream model has recently attracted attention. For analysis of such data, it is crucial to process the data in a single pass, or a small number of passes, using little memory. This paper provides an OpenCL implementation for data streams clustering, and then presents several optimizations for it, which make it more efficient in terms of memory usage. In order to maximize performance for different problem sizes and architectures, we also propose an auto-tuning solution. Experimental results show that our fully optimized implementation can perform 2.1x and 1.4x faster than the native OpenCL implementation on NVIDIA GTX480 and AMD HD5870, respectively, it can also achieve 1.4x to 3.3x speedup relative to the original CUDA implementation solution on GTX480.