非侵入式虛擬機記憶體效能瓶頸之觀測

Abstract

虛擬化技術已經廣泛的應用於雲端運算的環境中。在虛擬化的環境中一些硬體資源可以在虛擬機器中彼此分享利用像是CPU和網路卡，但像是記憶體在大部分的情況下是在配置虛擬機器時就已決定好且不能共享於虛擬機器間。再者在雲端環境中有著很多記憶體需求的應用程式，本身都會使用到大量的記憶體。所以虛擬機的記憶體估測是一個很重要的議題。在這篇研究中我們提出一個嶄新的系統NIMBLE去偵測虛擬機中的記憶體效能瓶頸名。NIMBLE是利用作業系統中的分頁機制去評估虛擬機中是否有記憶體效能瓶頸，並回報需要多少額外記憶體去化解此瓶頸。另外我們也可以依據不同的記憶體增加量去估測系統可以縮短多少由記憶體效能瓶頸所造成的額外執行時間。從實驗中顯示NIMBLE對虛擬機所造成的額外效能開銷最多為1.1%。除此之外，為了可以將我們的系統實際的應用於現有的雲端環境中，我們要求系統對虛擬機是非侵入式的且不需修改虛擬機任何內部的環境即可使用。另外因為雲端環境中可能有各種不同的作業系統，所以我們的系統設計成可以支援大部分常見的作業系統像是Windows和Linux。

Virtualization technology has been widely adapted in cloud environment. In a virtualized environment, some hardware resources such as processors and network bandwidth can be directly shared across virtual machines (VMs), but the memory resource is statically bound to a VM and cannot be shared. On the other hand, a lot of applications are memory-intensive. Efficient utilization of memory resource is a key issue in the use of virtualization technology in cloud environment. In this research, we present NIMBLE, a novel system to detect memory performance bottleneck for VMs in a cloud datacenter. NIMBLE monitors the paging activities of VMs to detect memory performance bottleneck. It will estimate the amount of additional memory needed for removing memory performance bottleneck. NIMBLE can also predict the reduction of the VM execution time due to the additional memory size. The experimental results indicate that the maximum runtime overhead of NIMBLE is about 1.1% on average. For NIMBLE to be applicable in a wide-range of cloud environments, we design it to be non-invasive. NIMBLE does not require modification or manual access to the guest systems is mostly agnostic to guest operating system type.