Research On Parallel I/O Technologies For Structured Grids CFD Applications

Computational Fluid Dynamics(CFD) uses the numerical method to solve the control equations of fluid motion and to obtain the flow field parameters. CFD relates to fluid dynamics, computational mathematics, computer and other disciplines. At present, CFD has been widely applied to the design of aircraft, rocket, spacecraft and other aerospaces and automobile, train and other ground transportation tools. CFD has become the typical numerical high performance computing application. CFD computing involves a lot of data access. In the case of large-scale CFD parallel computing, serial I/O performance does not match the computing performance and hence becomes the performance bottleneck. Parallel I/O is one of the effective ways to solve this problem.According to the demand for I/O performance in CFD parallel computing, we carry out the research of parallel I/O and the programming work of CFD programs in this paper. The main contents are summarized as follows:(1) We study a parallel I/O technology which is on the basis of a common file format(Native mode) in this paper. For a three-dimensional structure grid CFD parallel program LM3D used to solve the slightly compressible low-flow problems, we design a single-zone structure grid parallel I/O algorithm. For a high-precision structure grid CFD program HOSTA, we design two kinds of multi-zone structured grid parallel I/O algorithms:1) all of the processes are involved in reading and writing more than one ordinary file; 2) part of the processes are involved in reading and writing more than one ordinary file. The actual performance testings are carried out in a parallel computer system. The results show that the parallel I/O method has good performance. For a triangular airfoil test case with the grid size of 0.109 billion, parallel I/O achieves a performance speedup of 21.27 compared with the serial I/O and obtains a maximal I/O throughput of 2283 MB/s.(2)We study the data storage and a parallel I/O technology which is on the basis of HDF5 in this paper. For a three-dimensional structure grid CFD parallel program LM3D used to solve the slightly compressible low-flow problems, we design a polymeric parallel I/O algorithm. The results show that the data writing speed of this HDF5 based parallel I/O method outperforms the parallel I/O method (i.e. each process write an ordinary data file independently) for about 6.9x to 16. lx, when 4 to 32 processes are used. The data reading speed of this HDF5 based parallel I/O method is slower than the parallel I/O methods, with a speed of about 0.2x to 0.7x of the latter. However, the cost of data read time is much smaller than the cost of data write time. For a high-precision structure grid CFD program HOSTA, we design three kinds of independent parallel I/O algorithms based on HDF5:1) Each grid partition corresponds to an HDF5 file, all processes are involved in reading and writing; 2) all of the processes are involved in reading and writing an HDF5 file; 3) part of the processes are involved in reading and writing an HDF5 file.For a triangular airfoil test case with the grid size of 0.112 billion, parallel I/O achieves a performance speedup of 21.27 compared with the serial I/O. It achieves a maximal I/O throughput of 5.81 GB/s and improves the application’s performance by over 10%. For a simple airfoil test case with larger grid size, parallel I/O achieves a maximal I/O throughput of 6.72 GB/s.