Accelerating cosmological data analysis with graphics processors

In this paper we describe a successful effort to accelerate the two-point angular correlation function-a basic statistics tool used in the field of cosmology to characterize the distribution of the matter and energy in the Universe-by using an NVIDIA GPU-based system. We demonstrate the use of GPUs to accelerate the calculation of histograms of angular separations for large datasets as we achieve over two orders of magnitude performance improvement over conventional microprocessors. We discuss the specific implementation details of GPU kernels for computing bin assignments and updating histogram bins. We also describe an MPI-based GPU-accelerated two-point correlation application that runs on a compute cluster with multiple GPUs. Finally, we discuss specific lessons we learned in using GPUs to implement this class of algorithms.