應用平行化之DSMC法模擬LPCVD腔體之熱流場

Abstract

本研究利用鏈狀分割法（chain partitioner）將直接模擬蒙地卡羅法（Direct Simulation Monte Carlo method, DSMC）予以平行化，然後在一個有九顆處理器的電腦叢集系統（PC cluster）上執行並與之前蕭啟宏[5]論文中使用單一處理器的例子作比較，以凸顯平行化的優點。在平行化的過程中，在做區域切割時必須考慮到以每顆處理器負責相同的粒子數為分割標準，而非相同的區域大小。此外，在分散式記憶體的系統上作平行計算時，計算速度(speedup)並不能與所使用的處理器數目成線性正比，因此當參與計算的處理器數目增加時，儘管速度變快也不能持續地保持好的運算效率（Efficiency）。最後，模擬粒子數及計算腔體的尺寸和計算時步（time step）分別被提出討論是否對平行程式執行時的性能表現有所影響，結果發現在越多的模擬粒子數及越大的計算腔體的案例中，平行程式能得到較好的性能，而由於本研究中粒子數並不會隨時間劇烈改變，所以計算時步並不影響性能。

In this thesis, the parallelization is applied to DSMC method by utilizing chain partitioner. The computation is carried out on a PC cluster system consisting of nine processors. The preliminary results are compared with the Hsiao’s single-processor results [5], to demonstrate the advantage of parallelization. The computational domain is decomposed according to the same number of simulated particles in each processor. When a parallel program is executed on a distributed memory system, the speedup may not be proportional linearly to the number of processor. The computing load in each processor becomes lighter by increasing the processor number, whereas the communication load becomes heavier at the same time. Therefore the efficiency becomes worse with an increment of processor number. The parametric studies are based on the variations of particle number, computing domain size and time step, respectively. The better speedup and efficiency can be achieved when the particle number and computational domain increase. The performance does not depend on time step in present case because the variation of particle number is not severe.