Dressing-up Your Cluster for AI in Minutes with a Portable Network CR
Talkat Cloud Native + Kubernetes AI Day NA 2024 · on Tuesday, 2024-11-12· at 13:30· Salt Palace | Level 1 | Grand Ballroom A·
summary
Claim: a single portable Network CR automates per-node secondary NIC discovery, IP range allocation and VPC registration so clusters get pod-to-pod PCIe/DPU-class performance without per-host CRs. Implementation: Multi-NIC CNI controller + node daemon that generates NetworkAttachmentDefinitions, runs on OpenShift via OperatorHub and is open source on GitHub. Tested on an IBM Research VPC with 1,800 GPUs and microbenchmarks (OSU) showing clear bandwidth gains versus primary-only networking. Integrations target Multus, NRI, DRA, KNI.
AI-written summary · about
chapters (13)
- 0:00AI cluster networking intro
- 2:34High-performance networking with MCNA
- 4:22Network custom resources basics
- 6:54Direct port communication importance
- 8:23Limitations of secondary networking
- 10:09Portable network custom resource
- 11:54Portable MCNI implementation
- 13:40Deployment and setup process
- 14:20Live performance benchmarks
- 17:35Alternative implementation approaches
- 19:02Future directions and integrations
- 19:46Open source project resources
- 20:52Q&A
related talks
27mMake Kubernetes Networking Ready for World Class AI and HPC ...
09mLightning Talk: Cloud Native Networking for AI : Strengthen CNI for RDMA
39mScalable Multi-Node AI Workloads in Multi-Tenant AI Clouds U...
46mMulti - Networking Kubernetes Containers with CNI
39mScalable Multi-Node AI Workloads in Multi-Tenant AI Clouds U...
09mMulti-Cluster Networking: A Technology-Agnostic Network Operator for K8 Federation
34mHeterogeneous multi-cluster full mesh communication practice
38mVPCs in Kubernetes
