ML applications in Linux kernel
summary
Argues for a lightweight ML infrastructure that links kernel subsystems to user space via sysfs + eBPF bridge and netlink so models train off‑host and deliver compact inference or rules back to kernel for decisions. Demonstrates concrete use cases: automated log parsing and crash isolation using MCP servers to fetch precise source traces, memory access pattern analysis with prior work using SVM embedded in C for tiny in‑kernel models, and fleet‑scale off‑host failure prediction. Notes costs: no FPUs in kernel, training stays user space, inference must be tiny to avoid measurable latency.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 1:51Why run ML in kernel
- 5:50User space interaction approach
- 10:49Improving testing with ML
- 14:22Memory access pattern analysis
- 26:36Need for ML infrastructure
- 31:12Kernel interfaces for ML
- 38:46Challenges in data extraction
- 40:55Alternative approaches
- 41:16Q&A and closing
related talks
11mEnabling Composition of Kernels and Compilers
25mMachine Learning on Microcontrollers; Tiny, Useful and Fun
01h 53mTinyML: Machine Learning’s Future is Tiny and Bright
52mMachine Learning in IoT Applications
44mIntel Math Kernel Library Perspectives & Latest Advances
34mOptimizing AI Inference with ML Compilers & Hardware
06mML for Systems and Systems for ML
36mTinyML at the Edge: Deploying and Optimizing AI Workloads on Zephyr RTOS