臺灣博碩士論文加值系統

　　本篇論文說明Provim系統的發展：一個以界限容積階層技術及平行分散式計算來加速光線追蹤法之成像系統。本系統之發展係以異質綱路工作站所構成之分散式算環境為主，系統之架構獨立於綱路工作站的連結方式並藉由呼叫平行虛擬電腦（Parallel Virtual Machine）的程式庫來處理系統中有關綱路傳輸問題。
　　Provim系統使用處理群模型（Processor farm model）並提出集中式與分散式兩種方法；其中差異在於工作（task）產生與分配的不同。集中式方法中由主管（master）負責工作產生與分配，而分散式方法工作產生與分配的負載（load）是儘可能平均分散到可利用的工作者（slaves）以減少在主管處瓶頸的發生。系統的負載平衡（load balancing）是利用需求傾向（demanddriven）的要求工作策略結合適當的工作大小（task granularities）而達到。
　　Provim系統以C語言寫成且實驗結果證實使用28臺異質工作站執行光線追縱法生成影像，時間可從數小時減至數分鐘。實驗結果也說明分散式方法快於集中式方法。
　　本文的結論為在分散式計算環境對於有高計算的應用程式是有明顯的加速優點。再者平行虛擬電腦可簡化分散式計算環境中發展系統的成本與時間。

　　This thesis presents the development of Provim system, a ray tracing sys-tem accelerated by the bounding volume hierarchy developed in distributed computing environments consisting of a network of heterogeneous worksta-tions. The system has been designed to be independent on the network topology, and the the processor farm computational model is utilized to harness the computing. To facilitate the system development Parallel Virtual Ma-chine (PVM) is adopted where the network communications are proceeded by calling routines in the PVM library.
　　The Provim system supports the centralized and decentralized approaches with the difference in the way a task be generated and scheduled. In the centralized approach, the master is responsible for the task generation and scheduling. For the decentralized approach, the load of task generation and scheduling is distributed to available slaves reducing the potential bottleneck possibly encoutered in the master. The system load balancing is achieved by using the demand-driven task request strategy associated with appropriate task granularities.
　　The Provim system has been implemented using C programming lan-guage. The system has demonstrated that ray traced images that would have taken between several hours to generate on a single workstation can now be generated in between several minutes using 28 workstations. Results have also shown that the decentralized approach is faster than the centralized approach.
　　It is concluded that computing in distributed computing environments such as those based on network workstations is potential to bring significant benefits to computationally intensive applications offering advantages of low cost and ready availability. Furthermore, the PVM can facilitate the times and costs spent on developing parallel/distributed systems.
　　This is a ray traced image for the magic-star model consisting of 4132 objects. It can be rendered within 8.22 minutes by teh software developed in this thesis using a network of 28 heterigeneous workstations.