Intro: Scheduling SIG
summary
Describes a Flow Network Graph scheduler (Poseidon + Foreman) that models pod placement as min‑cost max‑flow with equivalence classes to shrink graph size, uses a CS2 solver via gRPC, and assigns costs per machine to optimize spread. Claims mock tests at 100,000 nodes and 300–400k pods reach about 30k pods/sec throughput. Notes limitations: no dynamic volume provisioning, no global rescheduling of already bound pods, and affinity/anti‑affinity handled with added policies.
AI-written summary · about
chapters (7)
- 0:00Start
- 9:41A Novel Flow Network Graph Based Scheduling Approach
- 10:31Firmament Kubernetes integration Design Overview
- 12:18Poseidon Design Overview
- 15:08Flow Graph Before Solver
- 21:36Scheduling Throughput Comparisons
- 23:47Roadmap - what lies ahead
related talks
39mSIG-Scheduling Deep Dive
39mIntro: Scheduling SIG
39mPractical Container Scheduling: Juggling Optimizations, Guarantees, and Trade-Offs at Netflix
30mSched.net: A Network-Aware Kubernetes Scheduler [I]
35mSIG Scheduling Deep Dive
23mOptimizing Storage Assignment via Pod Scheduling Under Disturbance Factors
36mNetwork-aware Scheduling in Kubernetes
36mDeep Dive: SIG Scheduling