Improving Performance of Deep Learning Workloads With Volcano
Talkat KubeCon + CloudNativeCon North America 2019 · on Thursday, 2019-11-21· at 10:55· Room 1AB - San Diego Convention Center, Upper Level·
summary
Volcano is a Kubernetes-oriented project that provides CRD-based resource models and unified capabilities such as group scheduling, fair share, priority queues, and job suspend/resume to better manage large-scale deep learning training workloads. The talk explains how Baidu migrated an existing job-based training operator to use Volcano, and presents measured results showing about 15% higher resource utilization and 10% faster training completion after the switch.
AI-written summary · about
chapters (21)
- 0:00Start
- 6:13What's Changed for DL Cluster
- 7:19PServer on K8s
- 8:45Paddle Operator Fault Toleranto
- 11:31Volcano Introduction
- 12:44How DLP leveraging Volcano
- 14:03Volcano Architecture
- 15:36Scenarios: MPI
- 17:25Scenarios: Fair Share
- 19:07Scenarios: Job Priority
- 20:15Gang Scheduling
- 22:59Task Topology & Binpack
- 24:56Job minResource (Spark)
- 25:39What get improved - Summarize
- 26:54Refactoring for Volcano
- 28:34Elastic CTR Estimation Demo
- 29:38CTR with Volcano
- 30:31CTR yami
- 31:01CTR Performance Comparison
- 31:48Pipeline
- 32:27Contact Info
related talks
32mVolcano: Intro & Deep Dive
26mIntro: Kubernetes SIG
08mKeynote: Scaling Model Training with Volcano: iFlytek’s Kubernetes Breakthrough
30mVolcano: A Unified Scheduling Platform for Cloud Native AI
43mRun Your AI Workloads and Microservices on Kubernetes More Easily and Efficiently...
24mIntro To Volcano: Cloud Native Batch System
28mEfficient Scheduling Of High Performance Batch Computing For Analytics Workloads With Volcano
08mNext-Gen AI Orchestration With Volcano On Kubernetes