MOSCOW, November 24. /TASS/. Researchers from the Siberian Federal University (SFU) together with colleagues from Moscow State University and the Siberian Division of RAS have come up with an idea to use a system of several graphics processors instead of one machine for calculating fluid dynamics. The article with the results of the study has been published in the journal - Communications in Computer and Information Science.
Computational fluid dynamics is essential in solving and analyzing problems, which arise when the flows of liquids and gases interact with any surface. For example, when air friction drags on an aircraft wing or water with the screw-propellers. In particular, computational fluid dynamics is used when designing aircraft, cars, drilling wells, and research into blood circulation.
When compared to central processing units (CPU), a graphics processing unit (GPU), which is responsible for treating images, makes it possible to carry out computational fluid dynamics faster. According to the scientists, in recent years, the efficiency of GPUs has increased to such a level that it is possible to use them in scientific and engineering simulations, for example, for modeling molecular behavior, computations for aerodynamics or flying vehicles, and fluid dynamics.
"In many areas, including in computational fluid dynamics, computational algorithms are being developed that use graphic accelerators to perform labor-intensive operations. At the same time, the principles of parallel calculations with GPUs and CPUs are notably different. Hence, only part of the algorithms can be easily adapted for GPU usage," commented Andrey Sentyabov, a study coauthor from the Institute of Engineering Physics and Radioelectronics at SFU.
GPU’s memory capacities are not sufficient to store all data needed for these computations. The researchers suggested using several GPUs for computations instead of just one unit to facilitate simultaneous parallel calculations, which mitigated the problem of insufficient memory. The scientists tested the new technique in the modeling of non-condensable laminar and turbulent currents. When compared with using a modern six-core CPU processor, the speed of calculations with the multi-GPU-system appeared to be 2-3 times faster.
The study’s authors presume that in the future, the program code needs to be further developed to get a powerful tool for modeling flows.