- 學位論文
在資源確保的前提來擺放虛擬機器以最佳化資料中心的能源消耗
Power Optimization for Data Center Through Resource-guaranteed VM Placement
摘要
在現今社會，雲端產業蓬勃發展，因此需要越來越多的資料中心 來當做雲端服務的基礎設施。而在之前的研究中不難發現，資料中心 的規模龐大，因此需要大量的電力來支持整個資料中心的運作。所以 如果我們能在資料中心省去一點點比例的電，最後帶來的獲益卻會是 很龐大的。而在資料中心裡，伺服器和交換機就耗掉了60% 的電，同 時，這兩項設備的能源消耗大多有一個特性就是，單純開機沒有額外 的工作量時，就會耗去滿載工作量時不少比例的電能。所以，能否再 滿足資源需求的情況下，關去不必打開的機器就成了資料中心能否更 省電的關鍵。在之前的研究中，對於資料中心的省電大多都只單純考 慮關去伺服器或是交換機，而我們是第一個同時考慮兩者的研究。我 們把這個問題建構出一個問題模型，但要解出這個模型的最佳解，需 要大量的計算時間，因此我們設計了一個多項式時間內可以解決的方 法。在實驗結果中我們可以發現借著同時考慮兩方面的需求，我們設 演算法相較於之前的方法大約有了30%的進步。同時，在考慮當前最 佳情況來做決定的這種設計方法，在有些時候可能會找不到滿足限制 的結果。我們所設計的演算法在我們跑的實驗中，得到不滿足限制的 結果的機率只有不到3%。
並列摘要
In a data center, virtual machines (VMs) not only require the server resource(e.g., CPU, memory) but also might generate a lot of network traffic among themselves. Existing research works on VMs placement are focusing on (1) how to place VMs to minimize server energy consumption without considering consumption of network resources; or (2) how to minimize inter-VM network traffic without considering energy consumption. In this work,we tackle the power saving issue of data centers from the point of view of VM placement as well; but, our idea is to address VM placement problem by joint consideration of both server and network energy consumption for minimizing total power consumption of a data center. First, we validate that the network power consumption plays an important part on the total power consumption of a data center by analyzing the data center architectures and the power benchmark of IT equipments. Next, we propose an energy-aware VM placement algorithm, called GreenGlue, which jointly optimizes the power consumed by servers and switches. GreenGlue is able to address (1) initial VM placement issue, in which VM requests are given and no running VM exists in a data center, and (2) adaptive VM placement issue, in which VM request arrivals and resource demands are dynamic over time. Furthermore, an optimal (initial or adaptive) VM placement model is derived via the mixed integer linear programming. Finally, we present a comprehensive performance study via simulation with different scenarios generated by synthetic and real traffic traces. The results show that GreenGlue can perform better than existing VM placement schemes in terms of the power saving and the rate of finding feasible placement.