Scholarlink

Database systems need well-tuned parameters for effective performance. Traditional tuning methods relying on expert knowledge or manual rules often face slow convergence and unstable results in high-dimensional parameter spaces or complex workloads. In automated tuning, invalid or extreme parameter values may cause severe performance swings or system crashes, disrupting the process and reducing reliability. Many tools lack parameter validity verification, allowing configurations that violate transactional consistency and system correctness. To address these issues, FasterTune is introduced. It is a robust tuning system combining domain knowledge with Bayesian Optimization (BO). FasterTune uses a two-stage, LLM-guided search to accelerate exploration during cold start, incorporates a high-robustness checkpoint recovery feature to instantly roll back to the best historical configuration after failures, and embeds a parameter validity-checking module built on resource-constraint modeling and SQL rule verification to block illegal settings. Experiments on TPC-C and TPC-H benchmarks show FasterTune delivers up to 25.4% higher throughput, 19.8% lower transaction latency, and 41.6% shorter tuning time compared to GPTuner and LlamaTune.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

FasterTune: A New Paradigm for Database Tuning with Large Language Models and Bayesian Optimization

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Xiaoqiang Chen,
  • Xinyuan Sun,
  • Bohan Li,
  • Yaofeng Tu,
  • Xiugang Dong

摘要

Database systems need well-tuned parameters for effective performance. Traditional tuning methods relying on expert knowledge or manual rules often face slow convergence and unstable results in high-dimensional parameter spaces or complex workloads. In automated tuning, invalid or extreme parameter values may cause severe performance swings or system crashes, disrupting the process and reducing reliability. Many tools lack parameter validity verification, allowing configurations that violate transactional consistency and system correctness. To address these issues, FasterTune is introduced. It is a robust tuning system combining domain knowledge with Bayesian Optimization (BO). FasterTune uses a two-stage, LLM-guided search to accelerate exploration during cold start, incorporates a high-robustness checkpoint recovery feature to instantly roll back to the best historical configuration after failures, and embeds a parameter validity-checking module built on resource-constraint modeling and SQL rule verification to block illegal settings. Experiments on TPC-C and TPC-H benchmarks show FasterTune delivers up to 25.4% higher throughput, 19.8% lower transaction latency, and 41.6% shorter tuning time compared to GPTuner and LlamaTune.