Exploring AI-powered kernel schedulers with Rust
summary
Builds a Rust user‑space scheduler using scx_rustland_core on top of sched_ext BPF to offload scheduling decisions to a PyTorch model (trained from traces of task name, priority and runtime). The system runs the model in user space via a shared address space and torch crate, predicts per‑task time slices (model: 2 inputs, four hidden layers of 32 neurons) and doubles demo FPS from ~30 to ~60 under stress‑ng; blocking and verifier/FP limits remain practical constraints.
AI-written summary · about
chapters (16)
- 0:00AI-powered kernel schedulers with Rust
- 1:20Scheduling in modern systems
- 5:12Linux kernel scheduler limits
- 8:26ML integration challenges
- 11:14BPF for kernel scheduling
- 15:09Rust scheduler with SCX RLAN
- 18:14Neural network for task prediction
- 20:25Live demo: AI scheduler
- 24:59User space scheduling challenges
- 26:51Q&A: Hardware locality
- 32:53Heterogeneous hardware interaction
- 35:23Use cases beyond gaming
- 37:29Model quantization for kernel
- 38:54Neural network architecture
- 42:06CPU mode switch impact
- 44:57Closing remarks
related talks
57mCrafting a Linux kernel scheduler in Rust
01h 00mFrom Tiny Chips to Server Racks: Deploying AI Everywhere with Rust
25m"Rust for AI & Accelerated Computing"
55mScaling out operational research with Rust
19mEnhancing Gen-AI with Rust: Scalable AI Systems
29mRust for ML Infrastructure: A Case Study in Building an ONNX Runtime
28mRewriting space mission analysis in Rust
59mParallel Programming in Rust: Techniques for Blazing Speed