以Globus為基礎的網格計算環境中 工作分配法之設計與實作

Abstract

摘要 Globus是一個探討如何實現網格環境的論壇，目的是將散居在世界各地的資源統合起來提供服務，無論是計算資源、網路資源或記憶資源。根據globus網格架構，我們針對網格分配管理(GRAM)區塊，探討如何管理資源來提供公平的網格計算服務。此篇論文中，研究目標有三點： (1) 如何定義多樣性的計算要求 (2) 當工作的工作量不同時，如何提供公平的服務 (3) 如何將工作分配法以及GRAM實做在globus的網格計算環境 我們研究的網格環境是單一網格計算場所(site)的同質性資源，在當中無論軟體或硬體的能力都是相同的。在globus中的網格分配管理(GRAM)區塊有清楚定義工作的執行程序，但對於如何安排工作到適合的資源提供處卻沒有清楚地描述，因此我們根據GRAM已有的執行程序，推展Mor Harchol-Balte的概念配合工作量的估計，補充設計出在資源管理伺服器中的分散排隊管理架構。在這個架構裡，首先針對使用者的工作描述，藉由可實行確認模組(feasibility check module)判斷是否能夠處理此項要求；接著運用編譯器(compiler)來估計工作的工作量，再根據工作量大小，將工作放到合適的隊伍中等待執行。 藉由猜測工作量的延伸工作指派(ETAGS)演算法，我們設計出分散排隊工作演算法(separating queue policy SQP)。在分散排隊工作法中，因為有每ㄧ工作的工作量估計資訊，我們修改ETAGS中的等待時間分析，並將排隊管理方式從連續工作群改為獨立工作群。根據工作的統計工作量分配、MMC的系統環境與公平性的限制條件，我們計算對工作群分類的服務時間上限。模擬的結果可看出SQP因利用工作量估計資訊而比ETAGS有較短的工作等待時間而且系統的工作也分配地更平均。 最後我們將SQP實做在Globus為基礎的網格計算環境中，這個網格計算環境是將Globus toolkit 4.0安裝在Fedora Core 3 TM上。我們安裝了簡易認證中心(Simple CA)，提供安全的服務，我們安裝了GRAM與GridFTP，讓使用者可以傳輸資料與執行工作。我們架設一個Client-Server程式，這樣即可按照我們的工作分配政策，指派工作到合適的資源提供處執行。

Abstract Globus alliance is a forum discussing how to make Grid environment come true. The purpose is to couple a variety of heterogeneous computational resources, storage systems, databases and other special purpose computing devices and present them as a unified integrated resource. According to the globus architecture, we focus on the Grid Resource Allocation Management (GRAM) Server and research in managing resources and providing fair Grid services. In this thesis, our research objectives are on (1) how to specify diversified computing requirements (2) how to provide fair services among jobs of different workload and (3) how to implement job assignment policy with GRAM on a simple prototype Globus grid computing environment. In this site, the capabilities of hardware and software are the same. The GRAM function in globus defines jobs’ processing procedures patently, but it doesn’t define how to assign jobs to suitable computing nodes clearly. So we base on the processing procedures in GRAM and extend the ideas of Mor Harchol-Balte to the situation with workload estimation. Then the Resource Management Server (RMS) is designed for the architecture of separating queue management. In this architecture, the feasibility check module is first used, and it determines whether the job requests can be done or not on the foundation of jobs descriptions. Second, the compiler is operated and it estimates workload information. Finally, RMS puts jobs to appropriate queue waiting for processing; the step is based on the amount of workload information. With job workload estimation available, we design a separating queue policy (SQP) by extending the Extend Task Assignment based on Guessing Size (ETAGS) algorithm. In SQP, we modify the waiting time analysis of ETAGS from sequential to independent analysis among groups and separate jobs in different groups according to different workload levels. We calculate the upper bound service time (threshold) based on known job workload distribution and M/M/c system standing on fairness definition. Simulation results show SQP outperforms ETAGS in shorter job waiting time and more balanced system load. Finally we implement the SQP in a simple Globus-based GC Environment. The environment is built on Fedora Core 3 TM with Globus toolkit (GT) 4.0 installed on it. We install simple CA function to provide secure and warranting services. We install GRAM and GridFTP functions for transferring data and job processing. We build a client-server program with SQP. According to the SQP, jobs can be handled to suitable computing nodes.