資料中心應用程式網路效能瓶頸偵測暨網路最高流速之估算

Abstract

為了要在資料中心環境探測是否存在應用程式網路頻寬上的瓶頸，我們提出了一個新的網路頻寬瓶頸偵測機制用來測量網路應用程式的預期所需最大網路頻寬。並且此機制不需要在客戶端虛擬機器中安裝額外的驅動程式。除此之外，我們進一步提出網路流扭曲機制來改變網路吞吐量而達到精確測量網路瓶頸的程度。而在測量網路瓶頸程度之後，藉由速度拉回機制將連線回到測量前的狀態。最後我們實作了一個網路瓶頸偵測的原型，實驗結果顯示系統對於預期的最大網路吞吐量以及效能瓶頸有高度的預測準確度，系統額外負擔約佔了總傳輸檔案時間的百分之三十。

In order to detect the presence of network performance bottleneck for applications in datacenter environment, we propose the NBD tool for determining the maximum network throughput of an application in datacenter environment. NBD is fully transparent and does not require modification to the applications. NBD employs a novel technique called network flow distortion for bottleneck level estimation. The network connection states are all maintained throughout the operation of NBD. The NBD prototype has been shown to give accurate detection of network performance bottleneck and incur a mild performance overhead.

Key word: datacenter, cloud computing, network throughput, bottleneck, virtual machine