格網系統排程機制的製作與模擬

Abstract

格網運算是分散式平行運算中一個新的研究課題.格網運算平台試著去整合網路上大量分散且異質的運算及資料儲存資源,以形成一個虛擬的超級電腦. 因而此超級電腦的運算效能與所整合的資源個別的運算能力及資源間的網路傳輸狀況有很大的關係.既使是同一個程式在不同的時候執行,其使用的資源配置也可能有很大的不同,進而影響到其執行的效能.雖然現今有許多針對各種不同平行運算平台的資源配置方式及演算法被提出.然而,由於各資源的運算能力及網路傳輸狀況的高度動態變化的特性,這些資源配置演算法必須作進一步的修正以符合格網運算平台的需要.在這篇論文中,我們提出運用模擬的方式來幫助資源的管理,並且探討使用這種模擬方法的效率和可行性.為了達到此目的,我們設計並且實作了一個可延伸的計算及排程系統,使一個平行的程式能以相同的排程方法,在此系統中以實際或模擬的模式執行,進而分析其執行效能.此外,新的排程演算法也可以在此系統上發展與研究.為了驗證我們所提方法的實用性,我們使用Grid-NPB中不同的benchmark做為對象,探討這些benchmark在不同的環境下的執行結果.雖然更廣泛的實驗仍需進行,不過初期的實驗結果顯示,在許多情況下我們的方法對平行程式的效能提升能提供有效的幫助.

Grid computing is emerging as a new parallel and distributed computing discipline that attempts to bring world-wide computing resources into a gigantic virtual supercomputing machine. Unlike traditional platforms, the capabilities and power of these computing resources as well as the communication speed between them vary dramatically. As a result, different runs of the same grid program may perform differently, and proper resource allocation and management becomes crucial to make effective use of the resources involved. Today many resource allocation approaches have been proposed and put to use in various grid computing platforms, each employs different kind of heuristics. However, because of the high variation of the performance characteristics of the underlying resources, making proper resource allocation decisions is by no means trivial. In this thesis we propose to use simulation techniques as a means to help resource management, and investigate the effectiveness and feasibility of our proposal. In particular, we design and implement an extensible framework for distributed computing and scheduling, where parallel programs can execute in both real-time and simulation modes using the same scheduling policy. Moreover, new scheduling algorithms can be developed and their performance studied. To assess the usefulness of our approach, we investigate the impact of several scheduling algorithms on different classes of problems using the Grid-NPB benchmarks. Although more extensive experiments are needed, initial experiment results show that our approach can be significant under certain circumstances