Volcano – Cloud Native Batch System for AI, BigData and HPC
Talkat Kubernetes Batch + HPC Day Europe 2022 · on Tuesday, 2022-05-17· at 14:25· Room 4F | Event Center·
summary
Cloud-native batch system that adds job-level lifecycle, multi-pod templates and queue decoupled from namespaces to Kubernetes, with concrete integrations for Spark, Flink, Kubeflow and Cromwell. Implements queue-based fair-share (DRF), reservation, preemption and SLE deadline reservation; supports elastic training and task-topology scheduling plus IO-aware placement for GPUs. Claims 20–30% training speedups in tests, handles a 300k pods/day production workload for a user, has ~350 contributors and ~50 production adopters, and pushes scheduler throughput optimizations to about 5,000 pods/sec.
AI-written summary · about
chapters (20)
- 0:00Intro
- 0:18Volcano - Cloud Native Batch Computing for Al, big data and HPC
- 0:51Trends of HPC, Big Data, and Al Batch Computing
- 2:21Batch on K8s: Challenges
- 4:48Volcano introduction
- 6:38Key Concept
- 8:10Job management
- 9:05Resource management-Queue
- 9:47Dynamic resource sharing between queues
- 11:02Fair share within Queue
- 12:04Scheduling policy in Volcano
- 12:30Scenario: Elastic scheduling
- 14:27Scenario: Topology awareness
- 16:25Scenario: SLA scheduling
- 17:53Scenario: batch scheduler for Spark
- 18:39Scenario: Big scale cluster: 10k nodes, 1 million Pod
- 19:40Use Case: Al platform at Xiaohongshu
- 20:58Use Case: Batch at Ruitian investment
- 22:24Community
- 23:03Release
related talks
32mVolcano: Intro & Deep Dive
24mIntro To Volcano: Cloud Native Batch System
43mRun Your AI Workloads and Microservices on Kubernetes More Easily and Efficiently...
26mIntro: Kubernetes SIG
30mVolcano: A Unified Scheduling Platform for Cloud Native AI
30mCloud Native AI: Harness the Power of Advanced Scheduling for High-Performance AI/ML Training
29mData Processing Efficiency: Optimizing Batch Workloads on Kubernetes With Custom Schedulers
08mNext-Gen AI Orchestration With Volcano On Kubernetes