Scalable Multi-Node AI Workloads in Multi-Tenant AI Clouds U...
summary
Claim: multi-node GPU training runs at near line rate on Ethernet by combining RoCE/RDMA with OVS+OVN+OVN-Kubernetes CNI and careful topology. Implementation uses per-GPU PFs mapped to leaf switches as per-GPU rails via PFs, eight logical OVN switches for east‑west rails, and enforces isolation and fairness with ACLs and QoS. Key performance lever is hardware offload of TC/OpenFlow into NICs or DPUs so Open vSwitch flow rules install in silicon and subsequent packets stay wire speed.
AI-written summary · about
chapters (7)
- 0:00Scalable multi-tenant AI clouds
- 3:30Hardware acceleration for AI
- 9:38OSS stack and GPU interconnects
- 15:45Pod networking with OVS/OVN
- 24:42Hardware offloading in Open vSwitch
- 31:29Network topology design
- 37:34AI data centers with open source
related talks
39mScalable Multi-Node AI Workloads in Multi-Tenant AI Clouds U...
34mNetworking Optimizations for Multi-Node Deep Learning on Kubernetes
01h 15mTutorial: Build, Operate, and Use a Multi-Tenant AI Cluster Based Entirely on Open Source
34mNetworking Optimizations for Multi-Node Deep Learning on Kubernetes
25mHigh Performance Networking for Distributed DL Training in Production K8s
21mScaling RoCE Networks for AI Training
40mOptimize Your AI Cloud Infrastructure: A Hardware Perspective
17mMeta’s Network Journey to Enable AI