分散式物件導向系統之行為模式與配置策略之研究分析

Abstract

在平行計算機系統（parallel computer systems）設計的領域中，分散 式物件導向計算（object-oriented computation）是一種高效率的設計 趨勢。在一個分散式環境下，物件（objects）分置於各個處理點（ processing nodes）中，在評估一個系統的整體效應時，處理點間物件程 序呼叫（object invocations）所產生的負荷，即是評估的重點。 根據 我們的分析， 物件程序呼叫包含了兩大部分：一為處理點間訊息傳送所 需的負荷（communication overhead）； 另一則是物件內部訊息常式查 詢（method lookup）所耗費的負荷。傳統的配置策略均未考慮物件程序 呼叫時程序查詢所產生的負荷。另外， 在評估一個配置策略的效能時， 我們必需針對在某一配置策略下物件程序呼叫的行為模式，予以分析。 為求詳盡分析在物件程序呼叫時， 狀態及資源同步化（ resource syn- chronization ）的行為，本論文建立一個 Colored generalized stochastic Petri net （ CGSPN ）模型，並配合馬可夫鏈（ Markov chains ）轉換技術，以評估一個配置策略之效能，另外並提出設計一個 高效能配置策略的準則。針對以上的敘述，本論文的研究成果可分為三大 部分：（一）提出在分散式物件導向程式環境下物件程序呼叫行為模式。 （二）建立評估物件程序呼叫行為的解析模型。 並提出提升整體效能的 基本準則，以供設計物件配置策略的參考。（三）設計高效能物件配置策 略。 根據前兩項的研究成果，我們分析出影響分散式物件導向計算效能 的兩大因素， 並進而設計靜態（ static）以及動態（ dynamic ）物件 配置策略。在靜態配置策略方面， 我們採用 clustering 的技術，將具 有高度聚合力的物件群（ groups of objects ）置於同一個處理點中， 以減少負荷的產生； 對於動態配置策略的設計， 我們採用 class information bit string (CIBS) 訊息，以提升在動態環境下，決定物件 配置能力的精確度。除此之外，在本論文中，佐以實驗結果，應證我們的 研究成果。 Distributed object-oriented computation is a cost-effective approach to the design of parallel computer systems. In a distributed object-oriented system,objects are normally distributed across the processing nodes.The cost of object invocations includes the communication overhead and the time spent on method lookup.Therefore,our study is to develop effective object assignment strategies which optimize the above cost. To evaluate the effectiveness of object assignment strategies in the quantitative and qualitative viewpoints,it is necessary to give an analytical model for the behavior of object invocations. In this thesis, we asopt the colored generalized stochastic Petri net (CGSPN), since it clearly describes the behavior of different types of resources (objects) in a distributed system. Moreover, we use the Markov chain transformation to measure the effectiveness of object invocations when an object assignment strategy is applied. Based on the above descriptions, the goal of our study is shown as follows: (1)We derive an operational model of method invocations for contemporary message-passing systems. (2)With the guidelines stated in (1), we use the CGSPN model to construct an analytical mode and develop the criteria to design an effective object assignment strategy. (3)Based on (1) and (2), we propose static and dynamic assignment strategies. In the design of static object assignment strategy, we use the clustering technique which groups related objects in nodes. Besides, in the design of dynamic object assignment strategy, we accommodate the Class Information Bit String (CIBS) messages with traditional load information to determine the target node for object assignment in run-time. We also perform simulation to prove the effectiveness of our object assignment strategies compared with other strategies.