high performance computing on graphics processing units: hgpu.org

Ensemble problems uses multiple models generated from a data set to improve the correctness and ensure faster convergence. The use of multiple models makes ensemble problems computationally intensive. In this paper, we explore the parallelization of ensemble problems on modern multicore hardware like CPUs and GPUs. We use the K-means clustering algorithm as a case study to explain our parallelization methodologies. We introduce the novel concatenated parallelization methodology and detail the performance tweaks to be considered when developing a parallel algorithm for modern hardware. We benchmark our implementations on multi-core hardware from different vendors and our approach gives significant performance improvement over traditional parallelization methodologies.