high performance computing on graphics processing units: hgpu.org

A great challenge for scientists is to execute their computational applications efficiently. Nowadays, parallel programming has become a fundamental key to achieve this goal. High-performance computing provides a solution to exploit parallel architectures in order to get optimal performance. Both parallel programming model and the system architecture will maximize the benefits if both together are suitable to the inherent parallelism of the problem. We compared three parallelized versions of our algorithm when applied to the study of the heat transport phenomenon in a low dimensional system. We qualitatively analyze the obtained performance data based on the own characteristics of multicore architecture, shared memory and NVIDIA graphical multiprocesors related to the traditional programing models provided by MPI and OpenMP, and Cuda programming environment. We conclude that GPUs parallel computing architecture is the most suitable programing model to achieve a better performance of our algorithm. We obtained an improvement of 15X, quite good for a program whose efficiency is strongly degraded by an integration process that essentially must be carried out in a serial way due to the dependence of the data.