非侵入式虛擬機處理器效能瓶頸之觀測

Abstract

虛擬化技術已廣泛地運用於建立雲端運算中心。在虛擬化環境中，我們可以動態調整系統上所擁有的資源量。有些硬體資源，像是處理器，會直接影響著應用程式的運行效能。為了使應用程式能獲得運行的最佳效能，同時兼顧最少的花費，我們必須在資源配置上尋求平衡點。現今已有許多不同的資源監視器可以觀看作業系統使用資源的情況。然而，在一部處理器資源使用率達100%的系統中，仍無法有效地辨別出其資源是否足夠，亦或是需要再增加多少的資源已達其最佳執行效能。在這篇研究中我們提出了一套名為NICBLE的系統，可以有效地監測雲端中心內虛擬機器的處理器效能瓶頸。NICBLE從虛擬機器產生的一些事件中推得虛擬機所需要的處理器資源數目。為了量化處理器資源數目對應用程式所造成的效能影響，NICBLE估測增加處理器數量後所能減少應用程式的執行時間。從實驗中顯示NICBLE對於虛擬機所造成的額外效能影響平均為1.2%。為了使得NICBLE能夠輕易地部屬在雲端中心環境中，我們在設計上兼具非侵入式及不針對特定作業系統的優點。

Virtualization technology has been widely used in today’s datacenter construction. In a virtualized environment, we can provision flexible amounts of hardware resources to a system. Some hardware resource, such as CPU, is important and may directly influence the application performance. However, CPU resource is limited, and we have to seek a good balance between maximizing the application performance and minimizing the cost. In other words, we need a tool that can determine the amount of CPU resource as needed by a guest VM. Unfortunately, existing CPU resource monitoring tool has many deficiencies. They cannot precisely indicate if a VM is short of CPU resource when the VM’s CPU utilization is at 100%. They also fail to provide straightforward cost-benefit information. Finally, most of them are designed to operate within the guest VM and cannot be transparently applied to virtualization platform. In this paper, we present NICBLE, a system to detect VM CPU performance bottleneck in virtualized environment. NICBLE monitors the scheduling activities at the hypervisor layer, and based on the activities, NICBLE will infer the CPU demand of each VM running on the hypervisor. NICBLE will also indicate the application execution time for each VM as a result of a change in the CPU resource provisioning. The experiment result shows that the averaged overhead incurred by NICBLE is about 1.2%. For practical deployment in virtualized datacenter environments, NICBLE follows a non-invasive and OS agnostic design for the guest VMs.