Democratizing Machine Learning on Kubernetes [I]
summary
Open-source repo provides a lightweight Kubernetes setup and a GUI deep learning workspace to launch Jupyter and distributed TensorFlow jobs; benchmarks use ImageNet on two worker nodes (each with four Tesla K80 GPUs) and one CPU parameter server over gRPC. Single-node training scales near linearly from 1→4 GPUs; distributed scaling varies by model (GoogleNet nearly linear, ResNet50 ~1.6× on two nodes) because the parameter-server design creates a network bottleneck. Recommended tooling: Horovod ring all-reduce, NCCL, FreeFlow CNI, and deep gradient compression to reduce bandwidth needs.
AI-written summary · about
chapters (27)
- 0:00Introduction
- 0:44Joy Qiao Introduction
- 3:27Documentation
- 4:20democratizing ML
- 4:57distributed deep learning
- 7:40model parallelism
- 9:47in the weeds
- 12:16Key learnings
- 12:47Training environment
- 13:48Linear scalability
- 14:45Distributed training
- 16:19Distributed results
- 18:07Compute over communication ratio
- 20:46Observations
- 23:52Ring and Reduced
- 25:51Deep Gradient Compression
- 27:02Open Source Kubernetes
- 27:28Deep Learning Workspace
- 28:06Deep Learning GUI
- 29:32Enable TensorFlow Board
- 30:09Notebook Job
- 30:37Distributed Training Job
- 31:08Pencil Bar
- 31:31Graph
- 31:49FreeFlow CNI
- 33:39Community Engagement
- 33:53Conclusion
related talks
49mDistributed Deep Learning
46mMachine Learning with Kubernetes
28mTraining Large Language Models on Kubernetes
38mKubernetes for AI Workloads – What Works and What… Doesn’t
37mBuilding Distributed TensorFlow Using Both GPU and CPU on Kubernetes [I]
25mMaking the Most Out of Your Hardware Accelerators in a Kubernetes Cluster
31mKubernetes For GPU Powered Machine Learning Workloads In Academia
36mAccelerating High-Performance Machine Learning at Scale i...