A Better and More Efficient ML Experience for CERN Users
summary
CERN runs a centralized Kubeflow service that pools GPUs, FPGAs and cloud TPUs to replace many siloed group clusters, improving utilization and enabling bursts to public cloud; it enforces image provenance via Harbor and Trivy, policy with OPA and runtime protection with Falco. Users get notebook-to-pipeline flow via Kale notebook-to-pipeline, an API for CI-triggered pipelines, and a GPU-availability checker. Credentials use mutating webhooks plus a daemon that handles Kerberos and OAuth2 renewal. Migrating to NVIDIA A100/MIG adds physical partitioning for multi-tenant GPU sharing.
AI-written summary · about
chapters (15)
- 0:00Intro
- 1:32Large Hadron Collider - LHC
- 2:41ML in LHC Data Acquisition
- 3:57ML to Discover New Physics
- 4:53Motivation for Kubeflow
- 6:10Kubeflow at CERN
- 8:50Improved Resource Usage
- 10:47Resources
- 15:06Integration
- 16:35Credential Management
- 18:06Namespace Management
- 19:02Scans and Runtime Checks
- 20:55User Feedback
- 23:18Conclusions
- 24:23Questions?
related talks
29mStreamlining Competitive Data Science at CERN: Running ML...
29mA Scalable Platform for Training and Inference Using Kubeflow
31mBuilding and Managing a Centralized ML Platform with Kubeflow at...
32mMachine Learning Using Various GPU Technology With Kubeflow.
30mGPU Sharing at CERN: Cutting the Cake Without Losing a Slice!
31mAI, CERN, and the Quest for GPU Custody: How CERN Leverages DRA for Efficient GPU Sharing
46mMachine Learning with Kubernetes
38mAll You Need to Know to Build Your GPU Machine Learning Cloud [B]