Spiderpool Updates For AI Workloads: DRA, NRI, and RDMA Obser...
Lightningat KubeCon + CloudNativeCon Europe 2025 · on Tuesday, 2025-04-01· at 09:45· Platinum Suite | Level 3·
summary
Spiderpool adds a DRA resource claim to declare RDMA devices and network interfaces so Pod YAML no longer needs manual PCIe names, and uses NRI dynamic insertion at Pod startup to attach RDMA devices with GPU affinity for GPU Direct RDMA (example: when pods request fewer than eight GPUs Spiderpool allocates matching RDMA devices on demand). It also exports Prometheus RDMA metrics per Pod and node (SR‑IOV aware) and ships Grafana dashboards.
AI-written summary · about
related talks
32mLeveraging DRA for Conquering ML Workload Challenges on...
29mWhat Can I Get You? An Introduction to Dynamic Resource Allocation
24mFit-to-Serve: How a New DRA Capability for Dynamic Device...
50mDynamic Resource Allocation - the path towards GA
31mAI, CERN, and the Quest for GPU Custody: How CERN Leverages DRA for Efficient GPU Sharing
34mRDMA Enabled Kubernetes for High Performance Computing
09mLightning Talk: Cloud Native Networking for AI : Strengthen CNI for RDMA
30mRun Your Own AI Cluster on a DGX Spark: Kubernetes, GPUs, and DRA