Who's Using That GPU? Identity-Aware Access Control for Kubernetes GPU…
summary
An open‑source validating admission webhook intercepts DRA ResourceClaim creation and enforces identity‑aware rules: OIDC via Dex for human auth, group‑based device‑class entitlements, a regex justification field, and expiry limits so claims can be shortened but not extended. SPIRE issues SVIDs to bind workload identity to the GPU, a TTL reclaim controller automatically returns devices when leases expire, and the live demo shows five denial cases plus a full approval‑to‑cleanup lifecycle.
AI-written summary · about
chapters (16)
- 0:00GPU access control challenges
- 1:04Limitations of GPU claims
- 2:30Demo environment overview
- 3:36Anonymous GPU claims gaps
- 4:39Admission controllers setup
- 5:42Team-based access policies
- 7:29Testing denied GPU requests
- 10:41Validating approved GPU claims
- 11:44Preventing policy circumvention
- 13:37Running GPU workloads with tracking
- 15:04Tracking GPU leaseholders with SPIFFE
- 16:53Preventing identity spoofing on leases
- 17:58Automated GPU lease reclamation
- 18:42Summary of identity-aware control
- 19:44Implementation complexity
- 22:40Q&A on idle GPU detection
related talks
25mAgentic Zero Trust: at Rest, in Transit, and at Runtime
31mAI, CERN, and the Quest for GPU Custody: How CERN Leverages DRA for Efficient GPU Sharing
34mDeclarative Device Virtualization: Orchestrating GPUs & Hardware in Cloud Native Environments
32mLeveraging DRA for Conquering ML Workload Challenges on...
17mTrustworthy, Sovereign AI: Creating Resilient Spaces
28mGPU sharing done right: Secrets, security, and scaling with Vault and vCluster
30mRunning AI Workloads in Containers and Kubernetes
32mProduction Workload Identity with SPIRE