Linux Kernel Autotuning
summary
Automated tuning of Linux kernel parameters uses Bayesian optimization and simple adapters to search configuration space and it produces concrete wins: a Diamond/ZRAM memory case yields about 30% memory reduction and an nginx experiment tunes 16 kernel parameters to outperform a kernel engineer and converges in roughly 17 minutes. Implementation runs experiments, applies statistical tests to compare distributions, and uses acquisition functions; results work for isolated, mostly static workloads but face limits on mixed/dynamic servers and on kernel-space vs user-space tradeoffs.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 1:09Need for autotuning
- 3:05Machine learning basics
- 6:03Kernel heuristics challenges
- 10:16Evolving kernel tuning
- 12:37ML optimization methods
- 19:15Practical case studies
- 23:05Kernel-space ML integration
- 28:24Discussion and Q&A
related talks
16mXkernel: Principled Performance Tunability of Operating System Kernels
41mAutomated Performance Tuning with Bayesian Optimization
27mLeverage ML to automatically tune K8s apps
22mAutomating Performance Tuning with Machine Learning
29mContinuous tuning of Microservices via Bayesian Optimization
18mLet the machines optimize the machines
34mPerformance tuning Twitter services with Graal and Machine Learning
17mWorkload-Based MySQL Tuning