Too Many Knobs to Tune? Towards Faster Database Tuning by Pre-selecting Important
Talkat 12th USENIX Workshop on Hot Topics in Storage and File Systems (HotStorage '20) · 2020-07-13..07-14·
summary
Experimental study shows that tuning five knobs out of 155 on Cassandra with YCSB-A reaches about 99% of the best throughput, identified by a random forest feature-importance ranking after Latin hypercube sampling. Ground-truth collection generates 25,000 samples (later reduced toward ~400), measured on real machines, and applying pre-selection cuts the tuner search space so an existing autotuner hits equivalent performance with roughly 5× fewer iterations. Results generalize to other workloads and PostgreSQL.
AI-written summary · about
chapters (9)
- 0:00Intro
- 0:20Database tuning is important!
- 3:03Motivation
- 4:13Experimental study
- 5:37Methodology
- 8:21(II) Identify Important Knobs
- 10:21How many knobs matter?
- 11:57What about a different workload?
- 14:12Discussion
related talks
33mCassandra Performance Tuning Like You've Been Doing It for Ten Years
04mPostgres Tuning is saved by Sovenath Shaw
20mDebate: Autotuning
28mAutotuning PostgreSQL on Azure Flexible Server
53mMake Your Database Dream of Electric Sheep: Designing for Autonomous Operation
18mWhy Postgres Is Better Than MySQL for Automated Knob Tuning
22mOPTIMUSCLOUD: Heterogeneous Configuration Optimization for Distributed Databases
41mOptimizing your app by understanding your PostgreSQL database