BitalosDB: Partitioned Key-Value Storage Engine for Lower Write Amplification
摘要
Modern storage systems often depend on two core data structures: the B+ tree and the Log-Structured Merge Tree (LSM-tree). The B+ tree excels in read performance but suffers from significant write amplification. On the other hand, the LSM-tree experiences higher write amplification because of multi-level static sorted table compaction. The paper introduces BitalosDB, an innovative storage architecture designed to reduce write amplification. The architecture partitions all data into multiple pages called Bitpages. Data partitioning localizes compaction to a single Bitpage, significantly narrowing the merging scope during this process and consequently reducing write amplification. Evaluations show that BitalosDB’s random write performance is 1.1x to 2.6x that of RocksDB in most cases, and its random read performance is 2.5x to 8.5x that of RocksDB, and write amplification is reduced by 16% to 55%.