軟體定義網路下具低錯誤偵測率之控制器故障偵測機制

Abstract

軟體定義網路(SDN)是一個能讓研究人員透過程式來控制他們的網路以滿足不同應用需求之新興網路架構。在SDN多控制器架構下，有控制器故障偵測的重要議題。現有的多控制器故障偵測機制存在高錯誤偵測率的問題。為了解決此一問題，我們提出一低錯誤偵測率故障偵測機制 (FDLM)。FDLM利用修改過的gossip協定來傳遞heartbeat訊息及使用多台控制器來做控制器故障的判定。即所謂的故障協議，以達到低錯誤偵測率。我們修改了gossip的順序，讓我們的gossip協定可以減少將控制器heartbeat訊息內的資訊傳遞給所有控制器所需要的時間。另一方面，我們的故障協議演算法定義了一新的控制器狀態，錯誤偵測，來回報給其他控制器以降低錯誤偵測率。實驗結果顯示，在設置四台控制器且具有訊息遺失率的環境下，我們所提出的FDLM與Akka相比能有較低的錯誤偵測率 (降低了95.5%) 及較少的訊息傳遞量。但FDLM比起Akka有較高的故障偵測時間 (升高10.0%)。擁有低錯誤偵測率的好處是可以減少不必要的控制器故障回復工作，此包括重選領導控制器、重啟控制器，以及讓重啟的控制器與領導控制器做資料同步。

The software defined network (SDN), an efficient network technology, allows researchers to program their networks to meet different applications need. In this thesis, we focus on failure detection services in multiple SDN controllers. We review some classical failure detection services and discuss how these services work. These classical failure detection services have the high mistake rate problem. To deal with this problem, we propose a failure detection service with low mistake rates (FDLM) mechanism. We use a modified gossip protocol to distribute heartbeat messages and have multiple controllers to make a failure decision in order to achieve low mistake rates. We modify the gossiping order in the gossip protocol. With the modified gossip protocol, the time to distribute a controller’s information in a heartbeat message to all other controllers can be reduced. We also use a modified failure consensus algorithm, where an extra state is defined to report false detection of a specific controller, to reduce the mistake rate. Simulation results show that the proposed FDLM has better performance in terms of average mistake rate, 95.5% (79.8%) lower, and maximum message size per round, 16.0% lower (20% higher), compared to Akka under message loss for the case of four (three) controllers. However, the proposed FDLM has slightly higher (lower) detection time, 10.0% higher (8.7% lower), compared to Akka for the case of four (three) controllers. The significance of achieving low mistake rates in the proposed FDLM is decreasing unnecessary SDN controller failover, which causes overhead of reelecting a leader controller, restarting a controller, and synchronizing the restarted controller with the leader controller.