TATune: A RocksDB Knob Tuning System Based on Transformer

Abstract

RocksDB is a powerful database engine that offers a wide range of adjustable knobs, which greatly influence its performance. However, configuring RocksDB manually for optimal performance is challenging due to the large number of available knobs and their complex settings. To address this issue, we propose Transformer Adaptive Gentic Algorithm Tune(TATune), an auto-tuning system for RocksDB knobs. In TATune, knob configuration files for RocksDB are randomly generated and executed at different preset workloads first. Subsequently, the correlation between the knob and RocksDB performance is learned by the prediction model based on Transformer. Finally, an adaptive genetic algorithm that utilizes the prediction model as a fitness function to recommend the RocksDB knob setting. Additionally, a novel optimization metric is also proposed to evaluate the performance of the auto-tuning RocksDB knob system. TATune is compared with other approaches to configure RockDB knobs on six distinct workloads. The results indicate that TATune is effective and achieves significant performance improvement across various target workloads. The final average optimization performance is 26% better than K2vTune and 72% better than RTune.

Keywords

• RocksDB
• auto-tuning
• knob configuration
• transformer
• adaptive genetic algorithm