LSM Tree Compaction Offloading Using NVMe-oF

Sungho Moon, Hera Koo, Hyeongjun Jeon, Beomseok Nam

http://doi.org/10.5626/JOK.2022.49.7.569

NVMe-over-Fabrics (NVMe-oF) is drawing attraction in the industry as an alternative to disaggregated storage by providing fast access to remote NVMe SSDs through NVMe commands. In this paper, we propose RocksDB-oF, an LSM-Tree-based key-value store optimized for disaggregated storage using NVMe-oF. RocksDB-oF alleviated the Write Stall problem by offloading compaction from the computing node onto the storage node in consideration of the characteristics of NVMe-oF. In addition, a file system that uses Storage Performance Development Kit (SPDK) effectively solves the file system consistency problem of two nodes accessing the same NVMe SSD at the same time. Experimentally, in a disaggregated storage environment with NVMe-oF, RocksDB-oF showed higher write throughput than legacy RocksDB.