Analysis of the Impact of Migration on Grid Environments – An Experience on Sun Grid Engine

於分散式系統中，關於遷移這項技術已經使用模擬、實作等方法來研究，並且在網格計算中有許多的排班程式採取了遷移的技術。在效率上而言，遷移能夠改善分散式系統中工作量平衡、可信任度、及服務保證等。然而，遷移也有可能造成副作用，例如資源的消耗及遷移錯誤等。 在這篇論文中，我們利用不同的策略來決定未完成之工作需要搬遷到哪些適合的電腦上繼續執行，這裡的策略包含了一些數學統計的方法，例如馬可夫鏈、波瓦松處理等。而經過決策所選出來的電腦，必須具備了在未來能夠提供一個高可得性之環境給這些未完成的工作，以避免持續遷移的現象。In distributed systems, the techniques of migration have been extensively studied by simulation, implementation, and successful results were accepted in most schedulers for grid computing. In effect, migration is capable of improving the load balancing, reliability, and QoS of distributed systems. However, migration can also be accompanied by negative effects. It may not only consume a considerable amount of resource, but also come with migration failure. In this paper, we present some migration decision strategies and experimental results on job migration. Through different policies, it makes decision to select and migrate to the next available node. Here we use some probability and statistic methods, like Markov chain and Poisson process, to improve the performance of the decision-making. The target node is selected such that it may provide a high availability environment and thus may finish the execution with high probability.List of Tables 6 List of Figures 7 Chapter 1 Introduction 9 1.1 Model of experiment environment 10 1.2 Motivation 11 1.3 Estimation 12 1.4 Paper organization 13 Chapter 2 Overview 14 2.1 The Sun Grid Engine 14 2.2 Experiment architecture and procedure 16 2.2.1 Communication 16 2.2.2 Grid simulation 17 2.2.3 Jobs 22 2.3 Three different models 23 2.3.1 Chaos environment model 23 2.3.2 High availability environment model 24 2.3.3 Low availability environment model 26 2.4 Migration mechanism 28 2.5 Migration strategies 31 2.5.1 Most available strategy 31 2.5.2 Markov chain strategy 32 2.5.3 Poisson process strategy 34 Chapter 3 Experiments 37 3.1 Experiments on Chaos environmental model 37 3.1.1 Most available strategy v.s. SGE default strategy 37 3.1.2 Markov chain strategy v.s. SGE default strategy 41 3.1.3 Poisson process strategy v.s. SGE default strategy 45 3.2 Experiments on high available environment model 47 3.2.1 The most suitable one for HA environment model 47 3.2.2 Adopting Markov chain strategy on HA environment 49 3.3 Experiments on low available environment model 51 3.3.1 The most suitable one for LA environment model 51 3.3.2 Adopting Poisson process strategy on LA environment 53 Chapter 4 Related work 54 Chapter 5 Conclusion 56 5.1 configuration trap 56 5.2 Future work 58 Appendix A References 5