雲端運算下的節能負載平衡

Abstract

雲端運算是近年來熱門的議題，透過虛擬化的機制讓虛擬機器共享同一台實體機器上的硬體資源進而提高實體機器上的資源利用度。然而，雲端資料中心需要負載平衡來避免系統中因為某些實體機器上的負擔過重，導致執行效率降低、硬體錯誤甚至當機的情形發生。我們在這篇論文中提出一個將雲端系統中的負載平衡問題轉換成一個權重雙分圖，再透過Hungarian演算法求解其上之最小權重的一組配對，依照得到的配對，平行的將虛擬機器遷移到負擔較低的機器上。一次的配對可以處理系統中多個負擔過重的實體機器來減低系統達到平衡的時間，降低虛擬機器間競爭資源的情形，進而提升系統的產能。我們利用CloudSim這個模擬器來測試我們演算法的效能，實驗的結果顯示我們在各種資源負載平衡的狀態之下，不論是系統達到平衡的時間以及各個虛擬機器上應用程式的執行完成時間皆大幅的下降。我們更利用XCP 雲端平台來搭建我們自己的小型雲端環境來證明我們的演算法在實際的雲端平台上也可以使用。此外，雲端的資料中心為了提供大量運算環境而需要許多伺服器，這些伺服器同時運行造成了大量的能源消耗和碳的排放量。所以我們改良我們的負載平衡演算法，當系統中的工作量降低時，去評估目前需要的伺服器數目，並將多餘的伺服器關閉，在負載平衡的同時達到節能的效果。

Recent movement in cloud computing is showing the trend that more and more companies start to deploy their services on the cloud instead of worrying about the troublesome server configuration and administration by themselves. This trend also helps in consolidating hardware resource usage because the same set of machine resources can serve multiple companies. Through a virtualization technique, different companies’ virtual machines (VMs) could share hardware resources on a single physical machine to improve the resource utilization. In order to maximize resource utilization, a load-balancing mechanism for cloud computing is needed to avoid performance degradation, hardware error or failure due to some overloading physical machines. Load balancing for cloud computing can be performed by migrating a VM from an overloaded host to an underutilized host. Since VM migration time are proportional to the amount of physical memory allocated to the VM. In a cloud data center with tens of thousands of physical machines and hundreds of thousands of VMs, one by one VM migration may take long time to reach a system load equilibrium state. In this paper, we propose an algorithm that transforms the load-balancing problem into a minimum weighted matching problem of a weighted bipartite graph. According to the minimum weighted matching obtained from Hungarian method, we concurrently migrate VMs from overloaded hosts to underutilized hosts. We use the CloudSim toolkit to test our algorithm’s performance and the experimental results show that our algorithm not only obtains a good load balance but also reduces the time to reach system load equilibrium state. We also build a cloud development platform via XCP OS to prove that our algorithm could be used in a realistic cloud environment. Furthermore, a major cause of energy inefficiency in a cloud datacenter is the idle power wasted when servers run at low utilization. Therefore, we modify our load-balancing algorithm to migrate all VMs out of low utilized hosts, then to turn off them to save energy during load-balancing process.