標題: | 基於行為之虛擬化環境的記憶體管理系統 BMSS: A Behavior-based Memory Scheduling System in Virtualized Environments |
作者: | 徐合邦 Hsu, Ho-Bang 黃育綸 Huang, Yu-Lun 電控工程研究所 |
關鍵字: | 虛擬化;記憶體;Virtualization;Memory management;Memroy overcommitment |
公開日期: | 2013 |
摘要: | 虛擬化技術是雲端運算重要的核心技術之一。在一個虛擬化的環境裡,運作在同一台實體機器上的虛擬機器們共享了該實體機器的實體資源(包含處理器、記憶體、等資源),因此 Hypervisor 必須負責實體資源的管理與分配。為了提高實體記憶體的使用率,系統管理者會在 Hypervisor 上加裝 Memory overcommitment 框架,以支援 Memory overcommitment 的功能。然而目前的 Memory overcommitment 框架主要依據每個虛擬機器的記憶體消耗量來分配該虛擬環境之實體記憶體。因此在此研究中,我們提出基於行為之記憶體管理系統(Behavior-based Memory Scheduling System,BMSS)。透過分析處理器使用率(CPU utilization)以及分頁錯誤率(Page fault rate),並且從中萃取出所需要的記憶體管理資訊,BMSS 能夠從虛擬機器中選擇出需要更多記憶體的虛擬機器,以及較為合適回收實體記憶體的虛擬機器。在此研究中,我們將虛擬機器的工作負載分成兩大成分:「背景工作負載」以及「需求工作負載」,而「需求工作負載」是指虛擬機器執行所提供的服務所產生的工作負載。BMSS 透過分配更多的實體記憶體給同時執行「背景工作負載」以及「需求工作負載」的虛擬機器,並且藉此提高該虛擬機器的性能。為了分析 BMSS 的能力,我們採用了記憶體負載型工作組成不同的測試組合:「交錯型工作組合」以及「並行型工作組合」。在配置兩個虛擬機器的實驗中,BMSS 提高了「交錯型工作組合」的性能(大約 10 \%),但降低了「並行型工作組合」的性能(低於 2.5\%)。在配置四個虛擬機器的實驗中,BMSS 則可提高 29.5\% 「交錯型工作組合」的性能。 Virtualization technology is one of the key technologies for cloud computing. In a virtualized environment, the virtual machines (VMs) hosted on the same physical machine share the physical resources, so the hypervisor is in charge of resource management including memory management. To achieve a higher memory utilization, a hypervisor is equipped with an memory overcommitment (MOC) framework. However, without concerning the characteristics of the VM workloads, the existing MOC frameworks allocate memory to VMs mainly based on their memory consumption. We propose a novel MOC framework, BMSS, which exploits the characteristics of the CPU utilization and the page fault rate of each VM. According to the characteristics, BMSS can reclaim memory from more appropriate VMs. We divide the VM workload into the background and the demand workload, while the demand workload is generated from the service provided by the VM. Through allocating more memory to the VMs processing both workloads, BMSS improves performance of those VMs. We conduct several experiments with memory-intensive workloads to evaluate the ability of BMSS. In the experiment with two tester VMs, BMSS improves the performance by around 10\% for interleaved workloads, while incurs an overhead less than 2.5\% for concurrent workloads. In the experiment with four tester VMs, the performance improvement for interleaved workloads even reaches 29.5\%. |
URI: | http://140.113.39.130/cdrfb3/record/nctu/#GT070260024 http://hdl.handle.net/11536/75132 |
Appears in Collections: | Thesis |