Optimization Of CPU-GPU Heterogeneous Data Flow And Its Application In Aerodynamic Numerical Algorithms

With the increasing of the complexity and the date scale in numerical computation of aerodynamics, the traditional multi-core processor and high-performance CPU can not meet the demand of scientific computing. It is urgent to use the parallel computing technology to realize the numerical calculation in more efficient way, and as a result of improving the efficiency of the program.At present, the development of GPU parallel computing technology is more and more matured. GPU has a strong floating-point computing capability and good programmability. Using the CPU-GPU heterogeneous architecture on high-performance computer systems has become one of the development trends. The paper focus on the basis of CPU-GPU heterogeneous computing environment, especially thekey technologies about optimizing the data stream in heterogeneous computing; uses the CUDA platform provided by NVIDIA todevelop our program.Condcting the research about the data streamon the processing model framework.We introduce the execution of the application, the source of the data stream and the specific processing of the data stream. In order to improving the efficiency of program, the paper uses the feature of Hyper in the Kepler architecture to optimize the processing of the data stream and texture memory on the device to reduce the access latency, and then analysis the effect of the number of threads in the block on the processing efficiency. We describe the data dependency appeared in the application, and offer the corresponding solutions. The experiments show that the overall performance of the simulation program using the CPU-GPU architecture to calculate has been improved about 10 times than that on CPU.For the large-scale data, the paper further analyzesthe MPI-GPU hybrid parallel programming architecture. In this architecture, the host structure adopts the master-slave mode to partition the tasks equally on the CPU and distribute to other compute nodes. At last, using the GPU to accelerate the process, while using MPI to communication between compute nodes. At the structural level, the MPI-GPU realize the coarse-grained parallelism on the CPU and the fine parallelism on the GPU. We researchthe initialization and switching of CUDA multi-context when multiple CUDA accesses the same GPU, and analyzes the overhead of context switching. The MPS technology is applied to the MPI-GPU architecture, so that multiple CUDA processes share the same context, eliminating the switching overhead and further improving the performance.In this paper, the aerodynamic simulation application is applied to the heterogeneous architecture of CPU-GPU and MPI-GPU respectively.Utilizing the GPU to computing large-scale data in parallelism way and optimize the data stream in the process to improve the computational efficiency. A certain acceleration effect has been achieving the expected goal,and the paper has a certain significance for the research about GPU on the other field.