摘要： 传统的缓存系统为了追求更高的性能大多是基于内存存储的，数据的持久化功能并不完善，因而系统会受到内存容量的限制，并且在系统宕机时会导致数据全部丢失，无法恢复。为此，在分析传统缓存系统的基础上，针对数据的持久化运用LSM-Tree理论以及Merge-Dump存储引擎进行改进，并参考Google的单机持久化存储系统LevelDB，实现一个分布式的Key-Value持久化缓存系统SSDB，结合传统缓存系统的优点并利用一致性哈希、布隆过滤器等思想对SSDB进行一系列优化。对SSDB性能测试的结果表明，优化后的持久化缓存系统SSDB是纯内存存储的，能有效降低数据的存储成本，且在读写性能上只比Redis下降约 600 QPS。

关键词: LSM-Tree理论, Merge-Dump存储引擎, 缓存系统, 持久化存储, 一致性哈希, 布隆过滤器

Abstract: Most of the traditional caching systems are based on memory storage in order to achieve higher performance, and their data persistence is not perfect. So these systems may be limited to memory capacity. Also they may lose all the data and be impossible to restore when systems break down. After analyzing the traditional caching systems, this paper applies the Log Structured Merge-Tree(LSM-Tree) theory and Merge-Dump storage engine to improve their data persistence, and then implements a distributed Key-Value persistent caching system Sorted Set DB(SSDB) by referencing the stand-alone persistent storage system LevelDB of Google. It combines SSDB with advantages of traditional caching systems, consistent Hashing, Bloom filter and so on to optimize its performance. It tests the performance of SSDB, and the results show that because of pure memory storage, SSDB can effectively reduce the cost of data storage, and it has just a slight decrease of 600 Query Per Second(QPS) in reading and writing performance compared with Redis.

Key words: LSM-Tree theory, Merge-Dump storage engine, caching system, persistent storage, consistent Hashing, Bloom filter

中图分类号: 

• TP311