Towards accurate and fast evaluation of multi-stage log-structured designs

摘要: Multi-stage log-structured (MSLS) designs, such as LevelDB, RocksDB, HBase, and Cassandra, are a family of storage system designs that exploit the high sequential write speeds hard disks flash drives by using multiple append-only data structures. As first step towards accurate fast evaluation MSLS, we propose new analytic primitives MSLS design models quickly give performance estimates. Our model can almost perfectly estimate cost inserts in whereas conventional worst-case analysis gives 1.8- 3.5× higher estimates than actual cost. A few minutes offline our find optimized parameters decrease LevelDB's insert up to 9.4-26.2%; also suggest changes RocksDB reduce its 32.0%, without reducing query or requiring extra memory.