Application of CFD parallel computing on NT system

• Main Authors: Fu-Wen Yao, 姚富文
• Other Authors: R. H. Yen
• Format: Others
• Language: zh-TW
• Published: 1999
• Online Access: http://ndltd.ncl.edu.tw/handle/61408820475375984899

碩士 === 國立臺灣大學 === 機械工程學研究所 === 87 === The present study utilizes Message Passing Interface (MPI) to explore the application of CFD parallel computing on NT system. The present study intends to fully utilize the computer facility with two parallel computing environments：the distributed memory mode through high-speed Ethernet and the shared memory mode on dual CPU computers. For the Fourier-spectral element method, the unbalancing job distribution among the Fourier modes would cause a worse speedup. The bank-queue technique can effectively increase the parallel performance. For the spectral element method, the load balancing can be achieved by equal allocation of elements. However, the global communication that is needed after every iteration in the conjugate gradient method might affect the parallel performance by decreasing the ratio of computing time and communication time. If the number of elements is increased, the parallel performance can be increased due to shorter communication time ratio. The WENO method achieves the load balancing by using blocks of same interior points. If the data exchanges of neighbor blocks are employed in a time step, the global communication can be avoided. Hence, high speedup can be obtained by using multiple CPUs.The bank-queue technique can effectively balance the loading and increase the parallel performance when the load distribution of jobs are unbalancing.