虛擬化平台在個人電腦叢集環境效能之研究

Study on the Performance of Virtualization Platform on Personal Computer Cluster Environment

洪榮泰(Rong-Tai Hong)

《中華科技大學學報》 65期 (2016/01) Pp. 77-87

摘要

雲端運算技術近來成為主流，主要可移植需要大量伺服器的應用程式至虛擬機器，藉由調整硬體的彈性特點，整合虛擬機器硬體提升整體運算效率，並降低電腦機房空間及管理成本。本文分別建置虛實體及擬機器叢集平行計算環境，其中虛擬平台安裝XenServer 半虛擬化軟體。藉由和計算節點問題相依之空間分割平行分子動力學模擬演算方法，討論不同計算節點數量時虛擬化的計算效能，並和實體機器比較。研究參數分別為不同計算量下所使用計算節點及網路卡數量。根據本文結果，半虛擬化中需安裝硬體存取權限的機器，其所佔用的計算資源會隨實際運算虛擬機器負載增加而增加，在半虛擬化最大值約為8%。由於間接存取的影響，降低高頻寬輸出/入裝置的效能，使虛擬化叢集平行計算環境其整體效能不如實體機器，但對較高計算量的問題並無明顯差異。在增加網路卡數量，可增加整體網路頻寬，但受到存取延遲影響，增加兩倍網路卡所增加的網路頻寬約20%。此外，亦可解決實體電腦中接點數量不足的問題，維持整體平行效能。

關鍵字

半虛擬化 ； 電腦叢集平行計算 ； 空間分割平行分子動力學模擬 ； MPI

並列摘要

Virtualization technology is a key aspect of the modern cloud computing. Migrating applications which once executed on many physical machines into virtual machines and consolidated those virtual machines onto just a few physical platform dramatically increase computing efficiency and reduce memory space and management costs. The computer cluster parallel computational facilities are constructed by physical and virtual machines. In which Xen Server of para-virtualization virtual machines is installed in this work. To study the parallel performance in virtual machines, the spatial decomposition parallel molecular dynamics used including number of nodes and network cards under various number of atoms as the investigating parameters. By the results, the CPU resource is used by privileged VM and the CPU usages are proportional to the loading of unprivileged VMs, the highest value is near 8% in para-virtualization. Because the virtual machines are unprivileged domains with no direct access to the hardware or device, the overall performance of virtual machine is inferior to that of physical machine but the difference is insignificant in larger number of atoms. As increasing the number of network cards, the overall bandwidth is just improved 20% due to the indirect hardware access.

並列關鍵字

Para-virtualization ； Spatial Decomposition Molecular Dynamic ； Message Passing Interface