OPPerTune: Post-Deployment Configuration Tuning of Services Made Easy
Talkat 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI '24) · 2024-04-16..04-18·
summary
OpperTune is a deployed service that automatically picks which of hundreds of config knobs to tune, chooses whether to share or split tuning instances across contexts, and applies a novel optimizer that jointly handles numeric and categorical settings. It implements hybrid Bandits for fast joint tuning, a decision tree scoping method called autoscope, and a microbenchmarking filter to limit restarts. Deployed on Azure across two enterprise apps and a 28‑service social benchmark, it reduces P95 latency by more than 50% versus expert defaults.
AI-written summary · about
chapters (6)
- 0:00OPPerTune introduction
- 2:21How OPPerTune works
- 6:39Hybrid parameter spaces
- 11:58Tuning with context
- 16:11Efficient parameter filtering
- 17:14Summary
related talks
23mAutoSys: The Design and Operation of Learning-Augmented Systems
27mThrough the Looking Glass: Leveraging Overton Window Concepts To Redefin...
20mSoftware Performance Optimization and Scaling
40mTurning Uncertainty into Success: How a Modern Approach can Thrive in a Conservative Environment
25mMulti-Tenancy Architecture Lessons Learned
20mBetter Resilience Through Self Tuning
15mCost of Soundness in Mixed-Precision Tuning
53mChange your tools, change your outcome