Userspace Assisted Scheduling via Sc...
summary
Proposes a userspace-to-kernel QoS stack that maps four classes (user interactive, user initiated, utility, background) through a middleware library using zero-API config-based tagging and netlink events, not forcing apps to opt in. Kernel changes include per-task push callbacks for wakeup/loadbalancer parity, EEVDF mapping of runtime/period hints, and a DVFS ramp-up multiplier to speed frequency response. Claims default should be utility, adoption takes ~12–18 months, and experiments become easy because configs let testing without widespread app rewrites.
AI-written summary · about
chapters (13)
- 0:00Introduction and roadmap
- 2:07Linux scheduler challenges
- 4:32QS plumbing framework approach
- 7:21QS classes and mapping
- 13:17Real-time and deadline discussion
- 19:00Access control and abuse prevention
- 20:04Wake-up path and load balancer
- 22:07Kernel improvements for QS
- 27:49Adoption problem and zero API
- 31:48Developer and user motivation
- 37:49Default scheduler values
- 41:00Performance inheritance and access control
- 42:52Summary and Q&A
related talks
31mSIG Scheduling Update: Transition From Pod To Workload Scheduling
39mOptimizing CephFS with Combining MDS QoS Scheduling and Static-Dynamic Subtree Partitioning
05mCombat Chaos-Driven Delivery By Thinking Like An OS
42mInter Process Confusion - Navigating Zephyr's IPC APIs
39mPractical Container Scheduling: Juggling Optimizations, Guarantees, and Trade-Offs at Netflix
35mLeveraging the Kubernetes Enqueue Scheduling Plugin for Smarter Workload Placement
34mSIG-Scheduling Intro and Deep Dive
10mLightning Talk: Resource Fairness and Utilization for Heterogeneous Batch/M...