A Deep Dive on How To Leverage the NVIDIA GB200 for Ultra-Fast...
summary
NVIDIA GB200 NVL72 hardware connects 72 GPUs across 18 nodes with NVLink at ~800 GB/s and the DRA-based ComputeDomain CRD exposes IMX-level partitions to Kubernetes. A DRA driver deploys per-node IMX demons, injects a common IMX channel, and blocks pods until the fabric is ready. Demoed features include clean node replacement, sudden node failover recovery, and scale to thousands of GPUs with global convergence near ~10 seconds; the system works in production with managed Kubernetes.
AI-written summary · about
chapters (11)
- 0:00GB200 on Kubernetes intro
- 1:48GB200 hardware and NVLink
- 3:34NVLink partitioning layers
- 5:00Compute domains on Kubernetes
- 7:50Orchestrating IMX demons
- 8:54Compute domain challenges
- 9:16Node failover in domains
- 12:24Scaling compute domains
- 16:20Demo: elasticity and failover
- 25:46Future work and features
- 27:11Key takeaways
related talks
32mLeveraging DRA for Conquering ML Workload Challenges on...
25mHigh Performance Networking for Distributed DL Training in Production K8s
30mRun Your Own AI Cluster on a DGX Spark: Kubernetes, GPUs, and DRA
30mGPU Sharing at CERN: Cutting the Cake Without Losing a Slice!
34mNetworking Optimizations for Multi-Node Deep Learning on Kubernetes
34mSupercharge Kubeflow Performance on GPU Clusters
30mKubernetes WG Device Management - GPUs, TPUs, NICs and More With DRA
32mA Deep Dive on Supporting Multi-Instance GPUs in Containers and Kubernetes