Supercomputer simulation of gas flow properties in metallic microchannel

In our work we consider gas flows into metallic microchannels of technical systems. The gas consists of two components: nitrogen and hydrogen. The material which is often used for microchannels production is nickel and its alloys. The supersonic expansion of gas mixtures is accompanied by several simultaneous non equilibrium processes: mixing of layers, shock waves, heat exchange with walls and etc. We will take into account real geometry for metal (including microtopography) and thermal heating. To investigate such complex problems it is essential to develop adequate numerical technique that implies using stable and robust numerical algorithms and calculations on fine grids by means of high performance computer systems. Parallel computer realization of numerical approach is oriented to using of supercomputers with hybrid architecture. Each node of such supercomputer has several multi-core central processors (CPU) and a few graphics processors (GPU). We propose multi-scale method for computation of mixture gas flow in metal microchannel. This method integrates macroscopic approach with molecular description of processes near metal walls. Macroscopic approach is based on the quasi-gasdynamic equations for mixture (QGDM) [2]. Microscopic approach is the correction of the flow parameters in Knudsen's layer that is performed by methods of molecular dynamics (MMD). The QGDM system is a generalization of the quasi-gasdynamic equations for one component gas. The quasi-gasdynamic equations are solved numerically on both structured or unstructured grids by finite volumes method. The system of molecular dynamic equations is used as a sub-grid algorithm. The MMD give a chance to get some information about processes on molecular scale, and on times of a few nanoseconds. The parallel algorithm is based on "domain decomposition" technique and a splitting on physical processes. Gas dynamics is calculated on CPUs units. The computational area is split into sub domains. Parallelization is realized on a grid of the processing units. Molecular dynamic calculations are distributed on GPUs. For MMD parallelization an equal number of particles is sent to each computing unit. According to the algorithm the molecular dynamic calculations do not require exchanges. Splitting on physical processes means that QGD-equations are processed by CPUs and MD is processed by GPUs. The results of calculations obtained in this paper show that the algorithms give high efficiency when it is implemented on modern hybrid systems with a large number of cores in use