This paper is focused on the analysis of the Kmeans clustering algorithm and on the study of its performance. In particular, in addition to the sequential version, we developed other implementations of the algorithm exploiting the parallelism of processors and GPUs through the programming languages Omp and CUDA.
The Kmeans clustering algorithm's performance has been improved by multi-processors parallelism obtained through the Omp framework, which allowed to obtain a speedup close to the number of processors avaible. Subsequently, through the parallelism of the GPUs and the parallel computing platform CUDA, we developed a parallel version of the algorithm achieving significantly higher speedups, in the order of hundreds times, compared to both the sequential and the Omp parallel version. Finally, to achive even higher speedup results, we developed a version of the algorithm making use of the shared memory in CUDA and compare it with the previous versions. Refer to the document Parallel_relazione_kmeans.pdf for more details.
