Kubernetes for AI Workloads – What Works and What… Doesn’t
summary
Demo shows Kubernetes running a training job from a Docker image that uses NVIDIA CUDA TensorFlow Keras and a baked dataset, creating a Job that spawns a Pod and prints epochs to logs. Cluster is 1 machine 8 CPU 30GB RAM 2 GPUs. A namespace resourceQuota limits to 1 GPU, so a 2‑GPU request is rejected. Kubernetes handles retries and completion but lacks gang scheduling, elasticity, preemption and topology‑aware packing for large distributed training.
AI-written summary · about
chapters (15)
- 0:00Introduction
- 3:05AI Ecosystem
- 4:23What is Kubernetes
- 6:27What is AI
- 13:43Docker Image
- 15:55Kubernetes Job
- 19:07AI Job
- 22:54AI Job Delete
- 25:43Whats Missing
- 27:04Cube Scheduler
- 29:06Challenges
- 32:15Scheduling for batch jobs
- 33:59Topology awareness
- 35:17Run AI
- 36:41Outro
related talks
27mHarnessing Kubernetes for Scalable AI/ML Workloads
16mAI + Kubernetes: What Beginners...
36mProduction GPU Cluster with K8s for AI and DL Workloads
42mKubernetes for Machine Learning: Managing and Scaling AI Workloads
38mAI in Kubernetes: How to Get Started?
37mBuilding Distributed TensorFlow Using Both GPU and CPU on Kubernetes [I]
31mKubernetes For GPU Powered Machine Learning Workloads In Academia
34mBuilding GPU-Accelerated Workflows with TensorFlow and Kubernetes [I]