Research On Fast Implementation Of Large-scale Parallel Turbulent Flow Simulation Based On LW-ACM

Computational fluid dynamics(CFD)has been developing relatively slowly in recent decades.One of the most important reasons is because the computational time of numerical simulation is excessive,if not impractical.Proposed several years ago,Lattice Boltzmann method(LBM)has good locality and parallelism.Compared with the traditional CFD method,LBM achieves many times improvement of performance.However,due to the large amount of memory required in each iteration,it becomes the bottleneck of LBM for further performance improvement.Link-Wise Artificial Compressibility Method(LW-ACM)is a derivative of LBM,but it requires much less memory than LBM.Thus,LW-ACM has a better performance than LBM.However,due to the short development time,previous work only use LW-ACM to simulate some simple fluid dynamics in an enclosed space,and there is no research related to inlet/outlet boundary conditions,nor any work released on complex flows such as three-dimensional turbulent flow.Consequently,this thesis conducts the researches from four aspects: turbulent flow simulation method,standardized software platform design,performance optimization technology and application validation.It aims to expand the functions and application scenes of Link-Wise Artificial Compressibility Method and provides basic support for engineering applications based on LW-ACM.Specifically,the main contributions and innovations of this thesis are as follows:(1)Proposing a fast implementation of turbulent flow simulation based on Link-Wise Artificial Compressibility MethodBefore this thesis,there was no work to adopt Link-Wise Artificial Compressibility Method to simulate three-dimensional turbulent flow and no complex boundary conditions proposed for LW-ACM.Therefore,this thesis first proposes inlet/outlet boundary conditions for Link-Wise Artificial Compressibility Method.Then,combining Synthetic Eddy Method and Synthetic Turbulent Inflow Generator,this thesis introduces Synthetic Eddy Method as the source term into Synthetic Turbulent Inflow Generator,and we add an additional step in Link-Wise Artificial Compressibility Method for Synthetic Turbulent Inflow Generator as the external force.A turbulent channel flow is performed to validate the implementation.This thesis also analyzes the program behaviour of Synthetic Eddy Method and then leverage Open MP to accelerate the turbulent inflow generation.Therefore,this thesis expands the application range of LW-ACM and provide the fundamental numerical method of LW-ACM to deal with the turbulent flow simulation on complex engineering application.(2)Designing and implementing Fast LW,a software platform for large-scale fast turbulent flow simulationBased on the above numerical method,this thesis design and implement Fast LW,a software platform for large-scale fast turbulent flow simulation from scratch.This platform includes pre-processing,numerical simulation,post-processing,turbulent inflow generation,parallel optimization,acceleration and so on.this thesis adopts a macro-based approach to achieve selecting operation of each functional module for user.Unneeded modules can be removed during pre-compilation,preventing code redundancy.Fast LW runs on the heterogeneous architecture of CPU + GPU,and makes full use of global memory,shared memory and constant memory,which greatly reduces data communication and improves communication efficiency.(3)Proposing a machine learning accelerated turbulent flow simulation method for Fast LWSynthetic Eddy Method is effective in introducing turbulent inflow,but we find that this method consumes a lot of CPU computational time,which greatly increases the overall simulation time.Therefore,a CNN+LSTM based machine learning method is used to predict and generate turbulent inflow instead of Synthetic Eddy Method.This thesis uses CNN to capture the features of images and LSTM to process the temporal image features.The machine learning model adopted in this thesis can predict the fluctuating velocity with correct turbulence characteristics even after a long time.Experiments show that the CNN+LSTM based turbulent inflow generator is 15 times faster than that by Synthetic Eddy Method,which greatly improves the performance of Fast LW.(4)Implementing a topology-aware multi-GPU method for Fast LWBased on the heterogeneous CPU+GPU architecture,this thesis designs and implements a multi-GPU parallel algorithm that increases the available global memory size and the number of GPU computing elements.This thesis leverages the combination of MPI and NCCL,which makes full use of NVLink and the advantages of Infini Band to achieve the topology-aware multi-GPU simulation.In this way,we can make the best use of NVLink and the shortest network path in a cluster for GPU communication.We also analyze the efficiency of cross-GPU computing and communication and scalability in a typical case and design the corresponding optimization method to achieve efficient scalable heterogeneous parallel computing for Fast LW,meeting the requirements of computing performance for super-large scale and high-resolution mesh applications.(5)Validating Fast LW by simulating several complex cases of turbulent flowIn this thesis,Fast LW is simulated and validated by three cases: three-dimensional turbulent channel flow,three-dimensional turbulent square-duct flow and three-dimensional turbulent square-duct flow with obstacles.The correctness of Fast LW and the correctness of multi-GPU implementation are validated by three-dimensional turbulent channel flow simulation.From the three-dimensional turbulent channel flow simulation,we can clearly see the development of the fluid,which gradually develops from laminar flow to fully developed turbulent flow in the channel after the area of influence of Synthetic Turbulent Inflow Generator.Besides,the eight-vortex secondary flow and the hairpin structure of the vortex near the wall are observed in the three-dimensional turbulent square-duct flow simulation.Finally,a rib-roughened square-duct turbulent flow is simulated,and the numerical simulation results are validated by experiments.