Managed Kubernetes — Next Gen Academic Infrastructure?
summary
Managed Kubernetes runs a production academic platform with 2500 CPU cores, 60–100 TB flash, 50 GPUs supporting ~200 users and ~1,000 pods/day, integrated with HPC (32k cores, 15 PB, 360 GPUs) and workflow engines Nextflow and Snakemake. Built components include JupyterHub, BinderHub, AlphaFold on demand; patched storage drivers (special XFS/CSI driver) and a modified NVIDIA GPU operator enable shared GPUs and enforced allocation. Benchmarks show Kubernetes with NUMA tuning matches or beats PBS for genomics (short tasks much faster); remaining problems are DNS/certificate multi-tenant conflicts and isolation for sensitive data.
AI-written summary · about
chapters (13)
- 0:00Intro
- 0:40Overview
- 2:27Managed Kubernetes
- 6:28Beyond Just an Infrastructure
- 9:44Examples
- 10:53Implementation Details
- 13:39Sensitive Data
- 14:38HPC Jobs on Kubernetes
- 15:57Research
- 17:00Efficient Resource Allocation II
- 22:52Infrastructure Comparisons II
- 27:46Green Computing I
- 29:37Hybrid Cloud I
related talks
31mKubernetes For GPU Powered Machine Learning Workloads In Academia
23mImplementation Challenges: From HPC to Containers in the Academy
09mKeynote: Kubernetes for GenAI: The New Operating System of the AI-First World
32mScaling Kubernetes Clusters for Generative Models: Managing GPU Resources for AI App...
07mKeynote: High Performance Computing on Google Kubernetes Engine
27mHarnessing Kubernetes for Scalable AI/ML Workloads
29mTowards Sustainable Multi-Cluster Management in Real-World Academic Infrastructure
41mBuilding & Operating a Large-scale HPC AI Cluster on Kubernetes