Research On GPU-based Parallel Computing For Supersonic Complex Flow Applications

With the increasing maturity of computational fluid dynamics(CFD)and the rapid development of computer technology,CFD plays an increasingly prominent role in aerodynamic performance analysis,optimization design and complex flow mechanism analysis of aircraft.In recent years,the geometric shape,flow mechanism and physical model of CFD numerical simulation have become increasingly complex,and the computational scale has increased unprecedentedly.Therefore,it is urgent to improve the computational efficiency by realizing efficient parallel computing.Graphic processing units(GPU)have powerful floating-point computing capabilities and storage bandwidth in data parallelism.In recent years,with the improvement of GPU programmability,GPU has been widely used in general computing fields such as molecular dynamics,direct simulation Monte Carlo,CFD,weather prediction,and deep learning.The realization of CFD efficient parallel computing on GPU is a research hotspot in the field of general computing.In this paper,large-scale parallel computing research on supersonic complex flows are carried out on GPU,and the characteristics of the GPU and the applicability of the physical model and calculation method to the GPU are considered.According to the different characteristics of CPU and GPU architecture,a GPU-based CFD parallel compressible flow solver is established.The solver is based on cell-centered finite volume method with unstructured grids,which configures the thread blocks and thread grids into a one-dimensional form.The linear reconstruction method is used to obtain second-order accuracy,and the AUSM+UP/AUSMPW+ scheme and center scheme are used to calculate the convective fluxes and viscous fluxes,respectively.In addition,the turbulent solution is solved by the K-ω SST two-equation model.The DP-LUR implicit time-marching method,which is beneficial to parallel computing without data dependence,is implemented on GPU,and the diagonal sweep in the LU-SGS method is replaced by several point-by-point relaxation iteration processes.Numerical results show that,compared with the explicit Runge-Kutta method,the DP-LUR method can improve the convergence speed of the GPU-based program.Moreover,numerical experiments such as NACA0012 airfoil,laminar plate boundary layer,compression corner,supersonic inlet and double ellipsoid,etc.are carried out to verify the GPU-based parallel program.According to the numerical simulation of supersonic complex flows,a hybrid parallel algorithm of message passing interface(MPI)and compute unified device architecture(CUDA)for CFD applications on multi-GPU HPC clusters is established.The 1D domain decomposition method is used to load each GPU with approximately the same number of computational grids.In addition,the exchange of data is done by the CPUs controlling the GPUs.In this paper,three data communication modes are designed,namely blocking communication mode,non-blocking communication mode,asynchronous concurrent communication mode.By using the asynchronous concurrent communication mode,the overlaps among GPU computing,CPU＿CPU communication and CPU＿GPU data transfer can be realized,and the computational efficiency has increased by about 10%.Moreover,numerical experiments such as aerospace plane and supersonic transverse jet interaction,etc.are carried out to verify the multi-GPU parallel algorithm.This paper tests the speedup and scalability of the multi-GPU parallel algorithm on the self-built parallel system based on PC clusters and GPU queue of Cngrid12,and the factors that affect parallel performance are studied.For the numerical cases of turbulent plate boundary layer and supersonic transverse jet interaction problems,the calculation grids reach 369 and 333 million,respectively.Numerical results show that,compared to CPU parallel computing,GPU parallel computing can achieve speedups of one to two orders of magnitude.At the same time,the multi-GPU parallel system has good strong scalability and weak scalability,and maintains high parallel efficiency in large-scale parallel computing.Numerical results show that,compared with CPU-based parallel computing,GPU-based parallel computing can achieve speedups of one to two orders of magnitude.Meanwhile,the multi-GPU parallel system has good strong scalability and weak scalability,and maintains high parallel efficiency in large-scale parallel computing.For the supersonic combustion chemical reaction problems,a numerical simulation method is established on NVIDIA Tesla V100 GPU parallel system based on PC clusters.In this paper,the hydrocarbon fuel/air mixture combustion chemical non-equilibrium flow problems are considered,and the Dryer’s two-step reaction models are used.Numerical simulations are carried out on a supersonic model combustor to verify the numerical simulation algorithm of multi-GPU supersonic combustion chemical reaction.