OpenFOAM Parallel Acceleration Research For DC

OpenFOAM is an open-source and portable computational fluid dynamics software.To cope with the situation where industrial software is sanctioned,China’s exploration of open-source industrial software is progressing rapidly.With the development of Chinese technology,domestic computing power is gradually increasing,and a new "Eastern" supercomputing system loaded with DCU coprocessors is gradually taking shape.Utilizing OpenFOAM’s advantages in fluid mechanics on domestic supercomputing systems is not only a requirement for OpenFOAM’s powerful computing capabilities,but also a requirement for expanding the adaptability of supercomputing software.Based on previous work,this paper investigates and analyzes the porting and optimization of OpenFOAM,mainly focusing on the following three aspects.1.Using OpenFOAM as the research object,by analyzing the structural performance of "Eastern" and the functional architecture of OpenFOAM,a solver suitable for DCU was developed for OpenFOAM,which implements multiple sparse matrix storage formats such as CSR,ELL,and COO,as well as diagonal matrix pre-processing.2.Building on the first point,further analysis and research were conducted on the performance of matrix storage formats.Combined with the hardware characteristics of the "Eastern" node,parallel matrix-vector multiplication(SpMV)was implemented on a single-node,multi-DCU platform.3.Through analysis of various data types in OpenFOAM,it is found that there is a large performance difference between single-precision and double-precision calculations.To improve performance,this paper implements a mixed-precision data processing method on OpenFOAM,which combines single-precision floating-point numbers and doubleprecision floating-point numbers to ensure computational accuracy and improve computational efficiency.Through testing,this mixed-precision data processing method can achieve approximately 1.5 times performance improvement.The pitzDaily test case provided by OpenFOAM is used to verify the algorithms in this paper,and the test results are analyzed using various performance acceleration methods.The test results are proven to be accurate and reliable,with a maximum performance acceleration of 17.9 times for DCU matrix multiplication compared to CPU matrix multiplication,and a performance acceleration ratio of up to 3 times for the entire DCU solver compared to the original OpenFOAM.