Project Lightning Talk: Spiderpool Unlocks Fine-Grained RDMA Observability for AI Inference
summary
Spiderpool exposes per-pod RDMA by assigning multiple RDMA devices via SR-IOV and fixed pod IPs so switch-level traffic and per-pod RDMA metrics become visible; it collects metrics at pod and node levels, attaches metadata labels, and surfaces them in Grafana dashboards aggregated by node, AI job, and pod. The setup supports MacVLAN and automatic multi-interface routing, finds RDMA hotspots across planes and switches, and claims a >60% troubleshooting improvement.
AI-written summary · about
related talks
09m⚡ Lightning Talk: Advanced GPU-Orchestrated Workflows and HPC Integrations on K8s for Distributed AI/ML at Scale
19mAI/ML Networking Challenges: The Fast and the Finicky!
06mProject Lightning Talk: Sailing Multi-Host Inference with LWS
05mProject Lightning Talk: Scheduling AI Workload Among Multiple Clusters
32mLeveraging DRA for Conquering ML Workload Challenges on...
15mPerformance Management Q&A
09mLightning Talk: GPU-Scanner: Extending CNCF Observability for Multi-GPU AI Workloads
09mLightning Talk: Optimize the Ray Schedule and Autoscaling on Kubernetes Cloud: A Hetero...