Building Distributed TensorFlow Using Both GPU and CPU on Kubernetes [I]
summary
Describes a production system that runs distributed TensorFlow on Kubernetes using both CPUs and GPUs: Kubernetes handles scheduling, lifecycle and monitoring while Docker device mapping and kubelet discovery expose NVIDIA GPUs. Current constraint is one GPU per worker with multiple workers placed on GPU-rich nodes; networking is the main bottleneck. Experiments show 3.5x speed on six GPUs versus one. Deliverable is a hosted product TensorFlow as a service that automates distribution, queuing, resource pools and failure handling.
AI-written summary · about
chapters (26)
- 0:00Introduction
- 0:27Agenda
- 1:47Deep Learning
- 4:22ImageNet
- 5:48Deep Learning Technology
- 7:41Open Source TensorFlow
- 8:56Google DeepMind
- 9:48Traditional Machine Learning Algorithms
- 11:18Distributed TensorFlow Flow
- 12:21Distributed TensorFlow GPU
- 13:05Distributed TensorFlow Management
- 13:51Distributed TensorFlow without GPU
- 15:48Leverage GPU power
- 16:09Enable GPU
- 18:56Device
- 20:49Discovery
- 22:45GPU Management
- 23:31GPU Resource
- 24:18CRI Support
- 25:29GPO 12 Volume Support
- 27:13Distribute GPU Resource
- 29:10Problems
- 32:41Product
- 33:32Comparison
- 35:49Interface
- 36:36Company Architecture
related talks
34mBuilding GPU-Accelerated Workflows with TensorFlow and Kubernetes [I]
36mProduction GPU Cluster with K8s for AI and DL Workloads
38mMinimizing GPU Cost for Your Deep Learning on Kubernetes
38mKubernetes for AI Workloads – What Works and What… Doesn’t
34mDemocratizing Machine Learning on Kubernetes [I]
51mDeploying Deep Learning models on Kubernetes clusters with GPU Enabled
38mMinimizing GPU Cost for Your Deep Learning on Kubernetes
35mLarge Scale Distributed Deep Learning on Kubernetes Clusters