Volcano: Intro & Deep Dive
summary
Volcano provides Kubernetes-native batch job CRDs with multiple pod templates, gang (co‑scheduling) semantics, preemption/backfill, queue-based fair share and hierarchical quotas, and integrations for MPI/TensorFlow/Spark/Flink. It adds elastic training to scale GPU replicas, a task-topology and IO-aware scheduling mode plus a NUMA/node-aware plugin, and reports up to 1.5k pods/sec throughput after scheduler and kube-apiserver optimizations. Production adopters include Huawei Cloud and others, one user runs 300k pods/day in production and about 50 enterprises report active use.
AI-written summary · about
chapters (22)
- 0:00Intro
- 0:39Cloud Native for Intelligent Workload
- 1:48Batch on K8s: Challenges
- 3:01Volcano Overview
- 4:55Key Concept
- 6:38Job management
- 8:19Resource management-Queue
- 9:04Dynamic resource sharing between queues
- 10:13Fair share within Queue
- 11:45Case: hierarchical queue
- 12:51Scheduling policy in Volcano
- 13:25Scenario: Elastic scheduling
- 15:53Scenario: Topology awareness
- 18:19Scenario: SLA scheduling
- 20:08Scenario: CPU Topology awareness
- 22:07Scenario: batch scheduler for Spark
- 23:12Scenario: Co-location and oversubscription
- 25:30Scenario: Big scale cluster: 10k nodes, 1 million Pod
- 26:57Use Case: Al platform at Xiaohongshu
- 28:25Use Case: Batch at Ruitian investment
- 30:06Community
- 30:42Release
related talks
24mVolcano – Cloud Native Batch System for AI, BigData and HPC
24mIntro To Volcano: Cloud Native Batch System
26mIntro: Kubernetes SIG
43mRun Your AI Workloads and Microservices on Kubernetes More Easily and Efficiently...
37mImproving Performance of Deep Learning Workloads With Volcano
30mVolcano: A Unified Scheduling Platform for Cloud Native AI
28mEfficient Scheduling Of High Performance Batch Computing For Analytics Workloads With Volcano
30mCloud Native AI: Harness the Power of Advanced Scheduling for High-Performance AI/ML Training