Machine Learning with Kubernetes
summary
Covers running ML on Kubernetes using EKS with concrete tools and numbers: distributed training uses MPIJob + NVIDIA NCCL for distributed training across p3 instances (multi-node clusters of up to dozens of 8‑GPU nodes), Horovod offers better performance than TFJob for many customers, and FSx for Lustre can cache S3 and avoid repeated copies for large datasets via the CSI driver. Also describes Kubeflow components — Notebooks, Pipelines, KFServing, Fairing — and claims AWS can assemble ~top‑10 supercomputer scale (≈20 p3dn nodes ≈1 petaflop each).
AI-written summary · about
chapters (15)
- 0:00Introduction and speaker background
- 1:04Amazon's ML stack overview
- 3:54Why Kubernetes for ML workloads
- 6:22Large-scale ML training use case
- 8:11Distributed training with Horovod and TF job
- 12:25Data staging with FSx Lustre
- 16:01Simulation workloads for autonomous vehicles
- 21:39AWS compute capacity for ML
- 21:59Introduction to KubeFlow
- 24:10Jupyter notebooks in KubeFlow
- 28:07KF Serving for inference endpoints
- 31:00KubeFlow pipelines for ML workflows
- 33:52KubeFlow 1.0 enterprise features
- 38:09Resources and next steps
- 39:15Q&A session
related talks
51mMachine Learning using Kubeflow and Kubernetes
31mKubernetes for Machine Learning: Managing and Scaling AI Workloads
46mDeploying scalable Machine Learning models with Kubernetes
30mFrom High Performance Computing To AI Workloads on Kubernetes: MPI Runtime in Kubeflow TrainJob
42mKubernetes for Machine Learning: Managing and Scaling AI Workloads
26mMachine Learning Meets K8s
31mBuilding and Managing a Centralized ML Platform with Kubeflow at...
28mMachine Learning on Kubernetes at Shell: A Kubeflow Journey