The Advanced Performance of Improved LevelDB on NUMA Systems: Three Practical Approaches

The ARM-based server processors have been gaining momentum in many scenarios. Key-Value (KV) stores also support many crucial applications and services. The LevelDB(Key-Value store, as a kind of NoSQL database) run in ARM(Advanced RISC Machine) server, which resulting in that while it performs fast in handling requests in memory, it is limited by issues such as I/O(Input/Output) and the NUMA(Non-Uniform Memory Access) architecture. It does not fully utilize the potential of the hardware and software.This paper considers three practical aspects, which have improved the performance of LevelDB, including multiple MemTable write mechanism, FilterTable that narrows the lookup scope and NUMA node load balance. Our evaluations show that improved LevelDB outperforms the origin LevelDB by up to 1.15∼2.04× in fillseq and by 1.74∼4.61× in fillrandom. In read performance, improved LevelDB can bring 20%∼30% improvement for readreq and 20%∼60% for readrandom.