Встановлено проблемні аспекти виконання паралельних алгоритмів на графічному прискорювачі. На основі встановлених особливостей запропоновано алгоритм генерації програм для графічного процесора.
This paper describes the main problems in implementing parallel algorithms on graphics accelerators. Based on the established features proposed algorithm generating programs for the GPU.
Domain decomposition algorithm for solving problems of elasticity based on parallel computing is considered. The global system of equations for the entire domain is not formed and is represented by local matrices and vectors for subdomains using Boolean matrices of connectivity. The system of linear equations is solved by modified conjugate gradient method. The algorithm is implemented with C ++ using parallel MPI library. The results of testing proposed approach for modeling example are included.
This paper proposes a method for measuring the energy efficiency of computing on CPU and GPU, which does not require specialized instrumentation. The results of the experiments carried out to the comparison of the calculation effectiveness on the CPU and GPU in terms of energy consumption.
Energy consumption of calculations in CPU and GPU is considered in the article. The mathematical apparatus for calculating the share of energy efficiency and made a series of experiments that allow to compare the energy efficiency of CPU and GPU, as well as identify its dependence on a number of parameters of program implementation.
Analysis of characteristics of software implementation of parallelization of process of constructing of discrete dynamical models was conducted in this paper. SIMD-architecture was used for the task of parallelization. Technology CUDA and GPU NVIDIA was used for this software implementation.