Navigating the AI/ML Networking Maze in Kubernetes: Lessons From the Trenches#

Antonio Ojea, Google — Core

• Why: HPC networking primitives — RDMA, MPI, collective-communication patterns — are being dragged into Kubernetes by AI workloads. Bridging traditional networking experience into this is exactly the PM's transition, told by a SIG-Network tech lead.
• What: Practical lessons from building networking solutions for AI/ML workloads, including a DRA-based network driver. (from abstract only)
• How: Discusses integrating specialized hardware, managing out-of-band RDMA, and the collective-communication patterns of distributed training, with Dynamic Resource Allocation (DRA) exposing and managing these resources. (from abstract only)
• Where applicable: Kubernetes clusters running distributed AI/ML training/inference with RDMA-class hardware; the DRA network-driver approach is general but still maturing.