Abstract
在分散式系統中(Distributed System)，效能評估 (Performance Evaluation)變得較為困難，主要是因為異質性的環境(Heterogeneous Environment)，以及不是中央管理式(Centralized)的因素．另一方面， 由於高速網路 (High Speed Network)的便利性，較便宜的工作站 (Workstation)以及方便的程式設計環境(Programming Environment)，使 得群組計算(Cluster Computing)，變得愈來愈流行．本論文提出一套方 法來分析分散式系統的效能．我們比較了不同平台(Platform)之間的差異 ．傳輸排列方式(Communication Pattern)，傳輸時間 (Communication Time)和計算時間(Computation)的比值，傳輸量的大小(Communication Volume)，傳輸時間 (Communication Time)以及傳輸次數( Communication Count)等，這些都有可能影響程式的執行效率．我們使 用 NAS平行指標程式(NAS Parallel Benchmark)做為基本的研究對象．我 們以工作站群(Workstation Cluster)做為基本的平台．他們以PVM這套軟 體連接起來．當我們實際執行 NAS指標程式後，他們會產生追蹤檔( Trace File)．這些追蹤檔可以經由視窗介面(Window Interface)的工具 ，顯視在螢光幕前．經過後續處理將追蹤檔轉換成另外一套工具的輸入之 後，我們可以更清楚了解程式的行為．此外分析執行後的一些數據，我們 可以了解那些是影響分散式系統的最重要因素．根據這些數據，我們可以 找出程式的瓶頸在那裡，藉以改善程式的效能．我們發現負載平衡(Load Balance)，處理機的個數(Number of CPU)，是否同一機型 (Model)，網 路的架構(Network Configuration)等，皆是影響程式執行效率的主要因 素． Cluster computing are integrated into a single logical system to solve a large computational problem, is becoming very popular. Cluster computing is feasible due to the availability of high speed networks, inexpensive workstations, and convenient programming environments. However the heterogeneity nature of the environment, the cost of communication between processors, and the balance of workload all affect the performance of cluster computing greatly. We discussed a methodology to observe the behavior of parallel programs. In the methodology, visualization tools are used to observe system utilization and communication traffic, which in turn give hints about program behaviors. To illustrate the methodology, NAS parallel benchmarks are used to evaluate the performance of workstation clusters of different architectures. We have experimented using NAS benchmarks on clusters of HP and IBM workstations. Communication/computation ratio and execution time, by varying the problem size and communication volume are measured and the performance of these clusters are evaluated.