摘要
近年來，由於使用者對於服務的需求量隨者網路服務的盛行也日漸劇 增，雲端運算的架構也因此被提出，雲端運算與以往的網格運算並沒有顯著的不同。兩者都是分散式運算的延伸，網格運算是著重於整合眾多異構平台，而雲端運算則強調在本地端資源有限的情況下，利用網路方式取得遠方的運算資源，提供使用者便利，相對的需考量的問題也隨之變多，例如資料的遷移性、資訊的安全性和系統的容錯性等。 Chubby是2007年Google提出來的一個分散式容錯系統實現於雲端運算架構中；也是Yahoo釋出的hadoop軟體中的zookeeper，實現方式是通過對檔案的建立操作來實現“加鎖”，系統主要是能支持著多個google的系統，如Google File System(GFS)和bigtable等系統。Chubby利用備份的方式來達到容錯的功能，其架構本身主要參考paxos演算法來實現，主要是利用訊息傳遞的方式，來維護各個伺服器在雲端運算下的各個資料一致性。 本論文基於paxos的方法來提出一個新的機制，希望可以應用於雲端運算的架構上。我們利用劃分區域的方式來分配各個伺服器需要處理的任務，並且採用輪替的方式來平衡處理的伺服器的負載，如此方法可以有效的提升方法的效能，並且利用consisten hashing方式來維護伺服器配置，讓我們的方法更可以實現於雲端架構上。 實驗證明，我們機制提出的方式，能更比其他機制更有效率的處理訊息並且可以維持著低延遲的方式處理；對於伺服器的擴充方面，相較於其他機制更可以維持著高效能，有效的利用各個伺服器擴充的資源。
並列摘要
The Internet usage growth rapidly, for Internet service provider, they have to handle big data, due to Cloud Computing has been widely discussed in recent years. Therefore, how to maintain high availability in Cloud Computing is a main challenge. Cloud Computing maintain high availability to serve large number of users through replication. Therefore, we have to replicate a service on multiple servers. They make sure that if some replicas fail, the service can still be performed by the operational nodes. However, once a service is replicated, ensuring consistency of all the replicas is not easy. Conflicting requests submitted simultaneously to different nodes can lead to inconsistent states in the replicas and meaningless replies to clients. Atomic broadcast can ensure service transmission consistency. libPaxos is one of atomic broadcast algorithm and it is an algorithm based on message-passing model of consistency. In this paper, we compared libPaxos, Mencius and RingPaxos. libPaxos use two-phase commit protocol(2PC) to ensure consistency. Mencius improves libPaxos method to distribute every server’s loading by rotating mechanism. RingPaxos use ring topology to improve libPaxos’ tree topology. The proposed scheme is improving libPaxos method, and first, we divide region that can let messages distributed processing, and use rotating mechanism to balance server’s loading, These features can improve protocol’s performance effectively. The experiment shows that our scheme can sustain high throughput under high client load. For scalability of protocols, our scheme also can sustain high throughput and low latency.