high performance computing on graphics processing units: hgpu.org

A recent development in High Performance Computing is the availability of NVIDIA’s Fermi or the 20-series GPUs. These offer features such as inbuilt atomic double precision support and increased shared memory. This thesis focuses on optimizing and evaluating the new features offered by the Fermi series GPUs for data mining algorithms involving reductions. Using three data mining applications namely K-Means clustering, Principal Component Analysis(PCA) and k-nearest neighbor search(kNN), three approaches for parallelization were used. These were the full replication, the locking scheme and the hybrid scheme-a trade o between replication and locking. Experiments were conducted to evaluate the performance of these algorithms with the new inbuilt atomic floating point support. In addition, the effect of increased shared memory was tested on the full replication approach for sufficiently small reduction objects. Finally, several hybrid versions of the application were created to determine the optimal configuration for the new features. We show how the hybrid approach outperforms the other two but for smaller object sizes, the full replication in shared memory has the best performance.