軟體定義網路下多控制器之快速一致性演算法

Abstract

軟體定義網路(SDNs)是一個能增進網路彈性並優化網路管理的新興網路架構。它最主要的概念是將資料層和控制層分離，將網路管理事務交由一個中央控制器處理。軟體定義網路多控制器的概念被提出以解決，單一控制器無法負荷網路的快速成長及單一控制器單點故障的問題。而如何達到網路狀態的一致性對SDN多控制器架構而言是一項重要議題。為了解決上述議題並確保所有控制器有相同一致的網路狀態，故在控制器間建立一套有效的一致性機制是必要的。Raft 是一強一致性並用於OpenDaylight (ODL) Clustering的一致性演算法。但因Raft演算法會定期選出一主控制器，故有主控制器重選超時問題，且系統負載會集中於主控制器。因此，我們提出一基於Paxos協定並具強一致性之快速一致性機制 (FPC)，並實作此演算法於ODL Clustering。FPC演算法可以保證當多個控制器同時提出資料變更時，必定會選出一個控制器負責執行並發布正確的資料變更順序給所有控制器，使多個控制器間的網路狀態達到一致。我們所提出的FPC在三台及四台SDN多控制器環境下時，比OpenDaylight Clustering Raft擁有較低且穩定的同步時間，分別減少35.3%及33.4%的同步時間，並在三台SDN多控制器環境下擁有較低的資料存取時間，在讀取資料及寫入資料時分別減少26.0 %和59.7%資料存取時間。

The software defined network (SDN) has been proposed as a promising network technology to improve network flexibility and to simplify network management. It decouples the control plane and the data plane, which enables centralized control of network management and improves network efficiency. Multiple SDN controllers architecture has been proposed to improve the scalability problem and to avoid a single point of failure. One major issue in the SDN multiple controllers architecture is how to reach a consistent network state among SDN controllers. In order to resolve the above issue and to make sure that all controllers have the same network state and the whole system becomes consistent, an efficient consensus mechanism to synchronize the control state of each controller is required. The Raft is a consensus algorithm used in the OpenDaylight (ODL) Clustering, which can provide strong consistency. However, the Raft consensus algorithm has the leader election timeout issue and its loading may fall on a certain controller. In view of these problems, we propose a Fast Paxox-based Consensus algorithm (FPC) which also provides strong consistency. We implemented it on the ODL Clustering to maintain a consistent global network state. The proposed FPC uses a controller priority mechanism to guarantee a proposal must be elected in each round and no additional round is needed even more than two proposers get the same votes. The proposed FPC has lower average consensus time (35.3% lower with 3-node ODL Clustering and 33.4% lower with 4-node ODL Clustering) than Raft. With a low consensus time, the proposed FPC can improve the data store access performance (26.0% faster at retrieving data and 59.7% faster at storing data via the REST API comparing with the Raft). Therefore, the proposed FPC is feasible for multiple SDN controller networks.