High Performance Networking for Distributed DL Training in Production K8s

Explore the intricacies of high-performance networking for distributed deep learning training in production Kubernetes environments in this 25-minute conference talk. Delve into the design and architecture of an 800 GPU cluster interconnected over RoCE fabric, achieving line rate performance between communicating containers in multi-node jobs. Learn about scalable cookie-cutter POD design for data centers, low latency one-hop network design enabling NCCL rings to avoid output port congestion, and Kubernetes integration with multi-homed networks for optimal GPU utilization. Gain insights into performance numbers for training workloads from production clusters, and discover how to overcome bottlenecks at NIC and switching fabric acting as interconnects between nodes.