A Parallel EM Algorithm for Gaussian Mixture Models Implemented on a NUMA System Using OpenMP

In the paper the problem of estimation of Gaussian mixture model parameters is considered. A shared memory parallelization of the standard EM algorithm, based on data decomposition, is proposed. Our approach uses a rowwise block striped decomposition of large arrays storing feature vectors and posterior probabilities. Additionally, some NUMA optimizations, which allow threads to use as much local memory as possible, exploiting the first-touch memory allocation policy of the Linux operating system, are described. The proposed method was implemented in OpenMP and tested on a 64-core system based on four AMD Opteron 6272 (codenamed "Interlagos\´\´) processors. The experimental results indicate, that on large datasets, the algorithm scales very well with respect to the number of cores, and NUMA optimizations significantly improve its performance.