應用電腦叢集之流場計算功能探討

Performance Study of Flow Computation in PC Cluster

林智仁(Chih-Jen Lin) ； 歐廣權(Kuang-Chuan Ou) ； 張克勤(Keh-Chin Chang)

《航空太空及民航學刊.系列B》 38卷2期 (2006/10) Pp. 109-119

https://doi.org/10.6124/JAAA.200610_38(2).05

摘要

?了克服計算複雜的流場問題所需要之大量記憶體空間及CPU時間，本研究使用Gentoo Linux作業系統，採用MPI(Messages Passing Interface)函式庫，並使用至多九台一般市面上容易購得的個人電腦，自行架構出可用於流場計算的平行化工具：電腦叢集(PC cluster)。測試問題?二維及三維上蓋牽動空穴流，採用區域分割法進行平行化並搭配不同的矩陣解法來比較其加速性，同時比較應用一維拓撲學(Topology)與二維拓撲學進行區域分割法的差異性。此外，本研究亦探討不同網路頻寬環境所造成之加速效果。綜合本研究各項測試的結果，建議求解複雜流場時可採用區域切割法；而選擇一維拓撲學或是二維拓撲學來分割流場區域則取決於流場區域的複雜程度，其決定準則只要讓各主機之處理器使用率能夠達到高負載及主機間計算量之負載平衡，便能獲得最好的平行效率。另外使用千兆位元(gigabit)等級的網路環境確能得到比一般高速乙太網路環境?理想的加速性。

關鍵字

電腦叢集 ； 平行計算 ； 區域分割法

並列摘要

In order to overcome the large demand of computer memory and CPU time required for the complex flow computations, a homemade nine-node PC cluster which is constructed by nine personal computers bought in the usual market serves as the computational tool for flow simulation. This parallel computational platform is operated under the Gentoo Linux operating system on which the messages passing interface (MPI) libraries are installed. The test problems consist of 2D and 3D lid-driven cavity flows and are parallelized by domain decomposition method. Two domain decomposition methods based on 1D and 2D topologies are investigated and their performances are compared with each other. Comparison of the speedup extents under various network environments is also performed. It is concluded that the choice of 1D or 2D topology depends on the complexity of the physical problems to be solved. It is also found that the promotion of the network environment such as using the gigabit ethernet to replace the fast ethernet leads to better speedup in the parallel computation.

並列關鍵字

PC cluster ； Parallel computation ； Domain decomposition method