Research And Implementation Of An Efficient KV Store With Matrix Tables Based On A Heterogeneous System

With the advent of the era of big data,the demand of high-performance storage device and system is growing significantly.Key-value stores have been widely used in data centers due to their excellent performance and unlimited horizontal scalability in theory.However,enormously-growing data requires better performance on key-value stores.The emerging of NVM(Non-Volatile Memory)make it a reality.NVM has both the non-volatile feature of HDD and the high-performance feature of DRAM.It can bring favourable performance improvement in conventional storage system.In this paper,we propose MatrixKV,an LSM-tree based persistent KV store on the heterogeneous storage architecture,to mitigate system stalls and write amplification in LSM-tree.We propose a matrix table in NVM to manage L0's data,which adopts a table-stacking structure and index tables in different levels to accelerate query in matrix table.We design a fine granularity column compaction between L0 and L1 based on matrix table to reduce involved data in each compaction,resulting in less performance fluctuation and higher performance.We increase the size of L0 in matrix table to reduce LSM-tree's levels and thus mitigate write amplification without penalties.We implement MartixKV based on RocksDB and evaluate it on a real 3D Xpoint NVM device—Apache Pass.Evaluation results show that MatrixKV gives more stable random write performance without system stalls and reduces both LSM-tree levels and write amplification.For random write,MatrixKV can improve throughput 3.91 x to 4.99 x compared to RocksDB and 1.66 x to 2.27 x compared to NoveLSM.For random read,MatrixKV can improve throughput 2.7x to 8.65 x over RocksDB and 2.14 x to 3.99 x over NoveLSM.