Cloud Native AI: Harness the Power of Advanced Scheduling for High-Performance AI/ML Training
Talkat KubeCon + CloudNativeCon Europe 2025 · on Friday, 2025-04-04· at 11:45· Level 3 | ICC Capital Suite 14-16·
summary
Volcano adds a hyper node API that models nested, switch-aware groups of Kubernetes nodes so schedulers can place tightly communicating LLM training pods on one-hop domains, with auto-discovery plugins translating vendor networks into topology trees and health checks. New lifecycle controls let jobs reset only failed tasks instead of full restarts. GPU virtualization exposes fractional allocation via WGPU memory and WGPU number. Project is CNCF-incubating and runs in production at over 60 orgs.
AI-written summary · about
chapters (13)
- 0:00Cloud Native AI scheduling intro
- 1:50Cloud Native AI trends
- 4:44Volcano Project deep dive
- 7:15Hyper Node abstraction
- 13:07Mapping workloads to Hyper Nodes
- 16:47Network topology benefits
- 19:40Job lifecycle management
- 21:10GPU virtualization for AI
- 21:51Multi-cluster scheduling
- 24:30Resource management with queues
- 25:14Future developments
- 26:42Community contributions
- 27:31Q&A
related talks
30mVolcano: A Unified Scheduling Platform for Cloud Native AI
08mNext-Gen AI Orchestration With Volcano On Kubernetes
43mRun Your AI Workloads and Microservices on Kubernetes More Easily and Efficiently...
32mVolcano: Intro & Deep Dive
24mVolcano – Cloud Native Batch System for AI, BigData and HPC
26mIntro: Kubernetes SIG
24mIntro To Volcano: Cloud Native Batch System
08mKeynote: Scaling Model Training with Volcano: iFlytek’s Kubernetes Breakthrough