Abstract: “This article considers results of performance test of experimental model of cluster hybrid computing system based on GPU accelerators in Linpack. It was created six variants of configuration of cluster system. These variants differ from each other by count of nodes, GPU accelerators in each node and amount of random access memory. It was carried out qualitative analysis of test results using approximation methods. As a result it was created simulation model for determination of dependence of real performance (Tflops) of cluster system on count GPU accelerators in the node. The simulation model allows to estimate optimal value of cost-performance ratio in dependence on characteristics of electronic components of cluster computing system.”

Keywords: parallel computing, high-performance computing, cluster computing system, GPU accelerator, CUDA technology.