Numerical Simulation Of Droplet Motion Deformation Based On DCU Acceleration Device

Computational Fluid Dynamics(Computational Fluid Dynamics,CFD)is an interdisciplinary subject that uses computer technology to study mathematics and fluid mechanics,numerical simulation of droplet motion and deformation is an important application field of computational fluid dynamics simulation.Under the interaction of different flow fields,the description and tracking of droplet morphology in the computer has high time complexity,and it is urgent to use accelerators to achieve efficient parallel simulation calculations.The DCU(Deep Compute Unit)accelerator is a GPU-like(Graphic Processing Unit)acceleration device independently developed by China.The numerous internal computing units make it more suitable for processing large-scale data-intensive computing,which has been deployed on many heterogeneous supercomputing platforms.In view of the lack of heterogeneous implementation of CFD algorithms on domestic DCU platforms,and the lack of optimization of algorithm communication and memory access methods for specific accelerator hardware characteristics in related literatures on heterogeneous droplet simulation,this thesis focuses on the CPU+DCU heterogeneous architecture.Drop deformation simulation parallel technology to carry out research.The main work and contributions of the thesis are as follows:(1)Designed and implemented a serial C version of the droplet motion deformation numerical simulation on the CPU,used Immersed Boundary Methods for the calculation of two-phase flow controlled by incompressible and immiscible Navier-Stokes equations,Using the finite difference method to discretize the control equations,the most time-consuming part of the whole program is the calculation of the pressure Poisson equation through static analysis,and the parallelism of the program is determined,which provides a foundation for the migration of the simulation on the DCU platform.(2)Realize and optimize heterogeneous droplet deformation simulation on a single DCU,follow the HIP(Heterogeneous-compute Interface for Portability)heterogeneous programming model,combine the algorithm principle and the characteristics of the DCU accelerator hardware architecture,and process the Poisson equation into a linear equation system through finite difference,and then this The equation system is used as a numerical linear algebra problem for heterogeneous calculation,and the heterogeneous droplet deformation simulation of CPU+single DCU is preliminarily realized;(3)Realize high-precision heterogeneous droplet deformation simulation on multiple DCUs.Since the data scale is doubled after the simulation precision is improved,the peak performance of a single DCU cannot meet the computing requirements of high bandwidth and high computing power.In order to solve the computational efficiency problem of droplet deformation simulation under high precision,the mixed programming mode of MPI+HIP-C is used to realize the concurrency of data transmission and task execution from the host side to the multi-device side,and complete the high-precision heterogeneous liquid on multiple DCUs.The drop deformation simulation makes full use of the hardware characteristics of multiple DCUs in the platform node to further improve the efficiency.The experimental data show that the parallel program on a single DCU after the basic transplantation achieves the highest speedup ratio of 1.96 compared to the CPU serial time consumption,and the optimized single DCU achieves a speedup ratio of2.34 compared to the CPU serial efficiency.Compared with the single DCU,the optimized result of DCU achieves a speedup of 2.41 times,which proves the feasibility and efficiency of the CPU+DCU heterogeneous algorithm for droplet deformation simulation.It expands the application ecology of the domestic DCU platform,and provides a reference for the realization and optimization of other CFD droplet simulation related work on this platform.