Learning the Linux Kernel Configuration Space: Results and Challenges
Talkat Open Source Summit Europe + Embedded Linux Conference Europe 2019 · on Wednesday, 2019-10-30· at 15:15· Forum 2·
summary
An open toolchain called TuxML using Docker/Python samples and builds hundreds of thousands of kernels (200K+ across versions) to measure build status, sizes and warnings; a sample of ~95K x86_64 builds shows ~6% fail (~3,000 failures) but most map to a handful of root bugs and three infrastructure-caused errors. Classification trees identify problematic option combinations; regression predicts uncompressed kernel size with ~7% error and highlights DEBUG_INFO and the count of yes options as top influencers. Continuous CI integration and smarter sampling (runconfig limits) remain key challenges.
AI-written summary · about
chapters (10)
- 0:00Start
- 11:13Answers in the rest of the talk
- 12:07TUXML: Sampling, Measuring, Learning
- 14:35Data: version 4.13.3 and 4.15 (x86_64)
- 15:54Application 1: "Smart" build infrastructure
- 17:45Classification problem: predict the class BUILD/ FAILURE Out of options values
- 18:29Classification tree
- 19:23Application 1: Smart build infrastructure
- 25:28Application 2: Kernel Size Prediction
- 32:01Challenges
related talks
46mLinux Kernel Autotuning
41mML applications in Linux kernel
48mUsing Machine Learning to find Linux bugs
11mEnabling Composition of Kernels and Compilers
27mLeverage ML to automatically tune K8s apps
20mUnlocking Performance: Harnessing LLMs To Streamline GPU Kernel Development in...
16mReducing Operational Barriers to Model Training
19mAI Kernel Generation: What's working, what's not, what's next