Application-aware Virtual Machine Provision Strategy

在雲端計算中，資料中心的能源消耗是一個非常重要的議題，因此，開發省電演 算法來替資料中心節約能源就變得很重要。大部分的省電演算法透過減少低使用 率的實體機器機器，並且關掉它們來達到省電的效果。 省電演算法必須要處理兩個問題，第一個問題是虛擬機器image存放的問題，如果 儲存虛擬機器image的實體機器被關掉之後，那台虛擬機器便無法在繼續運行，第 二個問題是虛擬機器的特性，我們發現頻繁執行I/O運作的虛擬機器跑在沒有存放 image的實體機器上的效能只有跑在存放image的實體機器的40\%，相反地，頻繁 執行CPU運作的虛擬機器卻不會受到有沒有跑在image存放的實體機器上而受到影 響。 這篇論文描述了我們在加強版的Roystonea上執行的省電演算法，Roystonea是一 套雲端作業的雛型，我們提出近似演算法以及動態規劃演算法來解決我們省點演 算法中要如何有效集中image，以及找到需要存放image的實體機器數目，我們跑 模擬來比較我們近似演算法和動態規劃演算法所算出需要儲存image的實體機器 數目的差距，模擬的結果顯示我們的近似演算法在比動態規劃演算法花更少的執 行時間下，可以找到一個和動態規劃不會相差太遠的結果。Energy consumption of data centers is an important issue in cloud computing, therefore it is important to develop energy conservation algorithms to reduce energy consumption of data centers. Most energy conservation algorithms achieve its goal by consolidating workload, reducing the number of underutilized servers, and shutting idle servers. An energy conservation algorithm must address two issues. The first issue is the placement of virtual machine image. If the physical machine that stores the image is off-line, then we cannot run this virtual machine. The second issue is the characteristics of virtual machines. We observe that the performance of an I/O-intensive virtual machine running on the physical machine without its image is only 40\% of the performance while running on the physical machine {em with} its image. On the other hand, the performance of a CPU-intensive virtual machine is almost the same whether the physical machine has the image or not. This paper describes our energy conservation algorithms as enhancement of Roystonea, which is a cloud computing system that manages crucial resources in a cloud system. We propose an approximation algorithm and two dynamic programmings to consolidate images of virtual machines, and find the number of physical machines that need to be stored images. We conduct simulations and compare the number of physical machines that need to be stored images in approximation algorithm with the optimal number of physical machines found by the dynamic programming. The simulation results indicate that our approximation algorithm can find good solutions in much less time than the dynamic programming