Using the GPU to accelerate distributed computing at the forecast of air temperature extreme values

Problem statement. The climate change giving rise to the emergence of extreme values of meteorological parameters, in particular temperature anomalies, threaten the sustainable development of the Russian Federation. The modern hydrodynamic models allow us to calculate these anomalies, but require high computing power. So, the service of Roshydromet cluster system consists of 96 compute nodes, each of which contains two of the most high-performance Intel Xeon 5680 with a clock frequency of 3.33 GHz (total of 192 processors) performance of 15.33 TFLOPS (trillions of operations per second). Existing compact hydrodynamic models, such as WRF, allows you to do fewer computational resources, but the computing speed is significantly reduced. We propose to accelerate the computation through the use of distributed computing on multiple computers, and using the opportunity of parallel computing through GPUs of those computers. It is shown that the joint use of the power of the computer’s CPU and its graphics processor is the time required for the calculations of extreme values of temperature when using the model WRF, the core of which is installed on the server side of the system is reduced 8 times. Proposed possible configuration of such a system, which may consist of two computers: high-performance servers, such as ASUS ESC4000 G3 c processor Intel Xeon E5-2600 and client workstations to display the results of calculations and input data into the system. It is proposed to use on the server four powerful graphics card company AMD Radeon HD 7970 DDR5 memory to 3 GB and a capacity of up to 3 TFLOPS each. The proposed software allows you to organize distributed computing in the client-server and parallel computing on GPUs. The proposed technical solution characteristics of the construction of such a system requires a robust power supply and cooling. The scheme of construction of distributed computing extreme values of meteorological parameters can be useful in operational forecasting for aviation, MES, MOD, etc. services.