一個滿足QoS需求的自動延展IaaS雲端服務

Abstract

資源分配在服務規範中是非常重要的一環，考慮並分配雲端服務所需的資源數目。如果缺乏仔細的控管，雲端服務可能會在尖峰時間無法全力提供服務或是在離峰的時間平白浪費資源。自動延展是基礎設施即服務中常用的一種解決方法，服務提供者可以透過在需要時自動增加或減少資源來維護並減少資源浪費。儘管大部分的雲端服務都有提供自動延展的功能，但是預測因運算資源缺乏而造成效率下降的用戶端經驗仍然是困難的。 有別於以往將實體資源映射到虛擬資源的資源利用優化方式，本論文將專注於軟體層的資源配置最佳化，我們提出一個新穎的雲端資源管理框架。此系統除了可以動態地自動延展，還可以藉由偵測客戶端的回應時間、由Qos訂定的閾值，給予使用者較佳的使用感受。 實驗顯示以回應時間為閾值，能改善系統的資源利用效率，和Husaini[8]的結果相比，CPI是他的兩倍，而且只花費了60%的資源。此系統在雲端應用中以工作量為基礎，動態地配置虛擬資源，改善資源使用率，避免資源匱乏以及減少使用者的支出。

Resource allocation, the process of assigning the amount of resources needed by cloud applications, is very crucial concern in the service provision. Without precise management, the cloud services may starve during the peak load time or waste the resources during the off-peak time. Auto scaling mechanism is one approach used in IaaS in which service providers can maintain the resources and reduce waste resources by automatically increase or decrease them when needed. Most of cloud services support auto-scaling functionality; however, it is still difficult to predict the client side experience which will result in decreasing performance due to lacking of computing instances. This paper focused on allocating resources at the application level rather than mapping the physical resources to virtual resources for the efficiency of resource utilization. We present a novel cloud resource management framework which supports dynamic auto-scaling. The proposed system monitors the end-user’s response time directly from client-side, and several thresholds defined with Quality of Services (QoS) considerations which include response time and error rates sampling. The result showed that using response time as a threshold could improve system in the efficiency of resource utilization which has twice better CPI and cost only 60% of the result shown in Husaini’s report [8]. Thus the system has dynamically allocated the virtual resource among the cloud applications based on their workload changes that improved resource utilization, prevented from resource starvation and reduced the user usage cost.