進階搜尋


下載電子全文  
系統識別號 U0026-1708201016120300
論文名稱(中文) 基於主從式監控架構之ARMA/SVR階層式虛擬機器資源預測方法
論文名稱(英文) A Hierarchical ARMA/SVR Approach Based on Master-Slave Monitor Model for Resource Utilization Prediction in Virtual Machines
校院名稱 成功大學
系所名稱(中) 資訊工程學系碩博士班
系所名稱(英) Institute of Computer Science and Information Engineering
學年度 98
學期 2
出版年 99
研究生(中文) 楊翔名
研究生(英文) Siang-Ming Yang
學號 p7696101
學位類別 碩士
語文別 英文
論文頁數 52頁
口試委員 指導教授-郭耀煌
口試委員-石維寬
口試委員-陳培殷
口試委員-張大緯
口試委員-洪盟峰
中文關鍵字 虛擬機器  虛擬機器內省器  低負擔在線預測器  資源預測器  時間序列  支持向量回歸 
英文關鍵字 Virtual Machine  Virtual Machine Introspection  low overhead on-line prediction  resource utilization prediction  time series  support vector regression 
學科別分類
中文摘要 由於虛擬化技術的快速增長,建構於實體機器中的虛擬機器性能是十分備受關注的。虛擬機器排程是決定虛擬機器性能的關鍵。在過去,虛擬機器排程器的運作方式,採用先到先得的機制。這些虛擬機器排程器,是基於測量系統的狀態而發展的。這些排程器會根據測量到的結果,作為資源分配的依據。然而,現今有些排程器有考量到未來的變化進而排程,這些排程器所做排程的面向較廣,因此通常會有較好的效果。換句話說,如果我們能夠分析過去事件(資源使用率)來預測未來的事件,我們可以提供給許多虛擬機器排程器一個支持。這是本論文的主要動機。一個好的虛擬機資源利用率預測過程有兩個步驟:1)觀察、記錄、並分析虛擬機器的資源使用率,2)在系統運行時,預測模型執行應為低系統負擔的運作。因此,在這篇論文中,我們提出主從式監控架構(MASM)來解決傳統虛擬機器內省器的高運作負擔,並且提出一個低負擔在線預測器,雙階層式預測模型(TLPM)來當作我們的資源預測器,用來預測VM的資源利用率。
此TLPM由L1-子模型和L2-子模型所組成,兩個預測的子模型是用來預測虛擬機器的資源使用率。如果使用適當的預測步驟,可降低預測所伴隨的負擔量。TLPM預測有三個步驟。第一步驟是使用L1-子模型,L1-子模型為時間序列預測模型,ARMA模型。利用此模型的季節特性,進而預測下一個季節期間的資源利用率。第二步驟是要找到關鍵區域,此區域為資源使用率有大變化的期間。L2-子模型僅在關鍵區域中執行預測機制,以減少預測所伴隨的系統負擔。第三步是使用L2-子模型,L2-子模型為另一種預測模型,支持向量回歸,在關鍵領域中執行更精確的在線預測機制。
最後,TLPM的準確度運算是採用MSE衡量準則以及錯誤容忍區間的計算方式。在實驗中所採用的資源使用率的資料來自收集一個台灣的大學的Web伺服器。實驗結果顯示,使用TLPM的運算複雜度為傳統在線預測的37.5%,而MSE僅上升14%,而在錯誤容忍區間的衡量中,設定容忍區間大於資源變化最大值的1%條件下,TLPM和在線的誤差率相差不到3%。
英文摘要 Due to the fast growth of virtualization technology, the performance of virtual machines (VM) on specific platforms is greatly concerned. VM scheduler is the key to determine the performance of VMs. In past, the scheduler of VMs is operated in a manner of first-come-first-serve (FCFS). These VM schedulers benefit from the measurement of the system status ahead. Based on the measurements, the resource request from VMs is predicted for the resource allocation in future. However, proactive scheduler executes schedule base on considering the future changes. The proactive scheduler considers more events, so it usually gets the better performance than traditional scheduler. In other words, if we can analyze past events (resource utilization) to predict future events, we can give a support to many VM schedulers. This is the major motivation of this thesis. The flow of a good VM resource utilization prediction has two steps: 1) the preprocessing includes: observing, recording, and analyzing resource utilization of VMs; 2) the prediction execution, prediction model should have low execution overhead in running system. Therefore, in this thesis, a Master-Slave Monitor Model(MASM) was proposed to solve the high computational observation by Virtual Machine Introspection, a Two-Level Prediction Model (TLPM) which is a low overhead on-line prediction to execute resource utilization prediction and to predict VM resource utilization.
The TLPM is composed of L1-submodel and L2-submodel, the two prediction submodels are used to predict the resource utilization of VMs. Prediction overhead can be properly reduced if a proper prediction scheme is used. The TLPM has three steps. The first step is to use a L1-submodel, time series prediction model, ARMA, which is used to predict the resource utilization of the next seasonal period. The second step is to find Key areas, where the resource utilization varies largely. The L1-submodel prediction is only used in the found key areas in order to reduce the online prediction overhead. The third step is to use L2-submodel, SVR, for a more precise prediction in the Key area.
Finally, the accuracy of the TLPM is evaluated by MSE and Tolerant error threshold between the actual and predicted resource utilization. The actual resource utilization dataset is collected from a public Web server of the University in Taiwan. In the experiments, the TLPM outperforms the traditional online prediction by 37.5% in terms of overhead evaluation, and the MSE only rises 14%. In the presence of Tolerant error threshold, the difference of TLPM error rate is less 3% than traditional online prediction when the threshold is more than 1% of maximal resource utilization.
論文目次 Chapter 1 Introduction 1
1.1 Motivation 2
1.2 Contribution of This Thesis 4
1.3 Organization 6
Chapter 2 Background and Relate Work 7
2.1 The Background of Virtual Machine System 7
2.2 Virtual Machine Introspection 12
2.3 Prediction Model 14
Chapter 3 A Master-Slave Monitor and A Two-Level Prediction Model 16
3.1 A Master-Slave Monitor 16
3.2 A Two-Level prediction model 21
3.2.1. Design Concept 20
3.2.2. The Flow of the TLPM 21
3.3 Level-1 Prediction Model 26
3.4 Discovery of Key Area for L2 Predictor 29
3.5 Level-2 prediction model 32
Chapter 4 Experimental Results 35
4.1 Experimental Environment and Dataset 35
4.2 Numeric Results 37
Chapter 5 Conclusions and Future Work 49
5.1 Conclusions 49
5.2 Future Work 49
References 50
參考文獻 [Xia08] Xianghua Xu1, Peipei Shan2, Jian Wan, Yucheng Jiang “Performance Evaluation of the CPU Scheduler in XEN” 2008 International Symposium on Information Science and Engieering

[Kar08] Kara Nance and Brian Hay “Virtual Machine Introspection Observation or Interference” Security & Privacy, IEEE 2008

[Hof07] A. Hoffmann, Kishor S. Trivedi, “A Best Practice Guide to Resource Forecasting for Computing Systems Guenther” IEEE TRANSACTIONS ON RELIABILITY, DECEMBER 2007

[Zhi07] Zhiyuan Luo “Online Network Resource Consumption Prediction with Confidence” Communications and Networking in China, 2007.

[Wei07] Weigend, A. S, Time Series Prediction: Forecasting the Future and Understanding the Past. (Publish in 1994)

[Fra02] Francisco J. Nogales, Javier Contreras, “Forecasting Next-Day Electricity Prices Time Series Models” IEEE TRANSACTIONS ON POWER SYSTEMS MAY 2002

[Kyo03] Kyoung-jae Kim, “Financial time series forecasting using support vector machines” NEUROCOMPUTING, SEP 2003

[Mar08] Marta Beltra´ n, Antonio Guzma´ n, and Jose Luis Bosque, “A New CPU Availability Prediction Model for Time-Shared Systems” IEEE TRANSECTION ON COMPUTER, JULY 2008

[Nic09] Nicholas I. Sapankevych, “Time series prediction using support vector machines”, IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE, MAY 2009

[Bry01] Bryan D. Payne “XenAccess: An Introspection Library for Xen” Initial version: 2001

[Jun09] Jupyung Lee, Student Member, IEEE, and Kyu Ho Park, Member, IEEE “Prediction-Based Micro-Scheduler: Toward Responsive Scheduling of General-Purpose Operating Systems”, IEEE TRANSACTIONS ON COMPUTERS, MAY 2009

[Che10] Cheng-Ping Lai a,b, Pau-Choo Chung a, Vincent S. Tseng c, “A novel two-level clustering method for time series data analysis” Expert Systems with Applications, 2010

[Gad09] Gad Abraham, Graham B. Byrnes, and Christopher A. Bain “Short-Term Forecasting of Emergency Inpatient Flow ” IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, MAY 2009

[Ram09] Ramon Nou “Autonomic QoS control in enterprise Grid environments using online simulation“ Journal of Systems and Software, March 2009
論文全文使用權限
  • 同意授權校內瀏覽/列印電子全文服務,於2011-08-30起公開。
  • 同意授權校外瀏覽/列印電子全文服務,於2011-08-30起公開。


  • 如您有疑問,請聯絡圖書館
    聯絡電話:(06)2757575#65773
    聯絡E-mail:etds@email.ncku.edu.tw