Enabling Heterogeneity-Aware Raft Consensus for Distributed Key-Value Storage

In this thesis, we focus on optimizing the performance of Raft in heterogeneous environments, with a distributed key-value store as an example. Raft is a widely adopted consensus protocol for building fault-tolerant distributed systems given its understandability and simplicity. However, its design does not consider the performance heterogeneity of servers caused by, for example, varying hardware or dynamic performance fluctuations due to I/O stress. This degrades the overall I/O performance of a Raft cluster when a poorly performed server is elected as the leader. We present HRaftKV, a distributed key-value store that incorporates heterogeneity awareness into Raft consensus, so as to achieve high performance while preserving the core Raft properties, such as leader completeness. HRaftKV builds on prioritized leader election (PLE), which prioritizes key-value servers by performance and elects a server with the highest priority as the leader. To adapt to the varying server performance, PLE further dynamically adjusts the priority of a server at runtime periodically and transfers the leadership when a server with higher priority than the current leader is available. We prove the safety and correctness of PLE, and implement PLE in HRaftKV. Our evaluation on Amazon EC2 demonstrates the high efficacy of HRaftKV in heterogeneous environments.這篇論文圍繞於改善 Raft 共識演算法在不均一環境下的性能，並以分布式鍵-值資料庫作為應用例子。因為 Raft 共識演算法易於理解和簡潔，它被廣泛採用於建構具容錯能力的分佈式系統。不過，此演算法的設計並未考慮因為不同的硬體配置或運行其間的性能波動所引致的性能不均一性。當一些性能較差的伺服器被選為主伺服器（Leader）時，整個 Raft 集群的讀寫性能會因而下降。因此，我們提出了一個名為 HRaftKV 的分布式鍵-值資料庫，它把對不均一性的處理和 Raft 共識演算法結合起來，在保留 Raft 的特性的同時達致高性能。HRaftKV建基於區分優次的主伺服器選舉（Prioritized Leader Election，PLE），它按伺服器的性能排列優次，並從中選出最優先的為主伺服器。為顧及變動的伺服器性能，PLE 更會於運行其間調整伺服器優次及視乎情況轉移主伺服器。我們證明了 PLE 的正確性並將其實現於 HRaftKV。我們在 Amazon EC2 上的實驗展示了 HRaftKV 在不均一環境能達致高性能。WONG, Tsz Fung.M.Phil. Chinese University of Hong Kong 2021.Includes bibliographical references (leaves )Abstracts in English and Chinese.Title from PDF title page (viewed on ...