Kubernetes For GPU Powered Machine Learning Workloads In Academia
summary
Project builds a bare‑metal Kubernetes cluster tuned for ML: three control‑plane nodes, Juju models for scaling, and MAAS for automated provisioning. It deploys four NVIDIA DGX A100 nodes (three whole‑GPU, one configured with 7‑slice vGPU profiles to yield 56 vGPU instances) using the GPU operator; storage is an external Ceph cluster, authentication via Keystone→SAML, GitLab CI runners and registry integration, MetalLB layer‑2 load balancing, Prometheus/Grafana logging, and Vault HA. Cluster runs ML training and CI; fair GPU queuing is still unresolved.
AI-written summary · about
chapters (15)
- 0:00Kubernetes GPU project
- 1:03Bare metal architecture
- 3:35Security and observability setup
- 5:21Nvidia GPU integration
- 7:50Networking with MetalLB
- 9:18SAML via Keystone
- 10:23GitLab CI/CD overview
- 11:07University of Alabama research
- 12:32Kubernetes research use cases
- 15:17HPC batch compute
- 18:29OpenStack for development
- 20:15Future Kubernetes research
- 21:20Next-gen data workflows
- 23:04Reproducible HPC workloads
- 25:12Q&A
related talks
06mKubernetes Love Machine Learning, Even on Private Cloud
46mMachine Learning with Kubernetes
32mScaling Kubernetes Clusters for Generative Models: Managing GPU Resources for AI App...
31mManaged Kubernetes — Next Gen Academic Infrastructure?
34mDeep Customized Kubernetes for Machine Learning in Tencent
36mProduction GPU Cluster with K8s for AI and DL Workloads
25mEmpowering AI-Driven Drug Discovery: Overcoming Challenges in Building a ML Platform on Kubernetes
38mAll You Need to Know to Build Your GPU Machine Learning Cloud [B]