Question

A data center team is deploying a GPU cluster for AI/ML training workloads across Nexus 9000 switches using RoCEv2 for RDMA transport. During large distributed training jobs, they observe throughput collapse and retransmissions caused by packet drops in the fabric. The team wants to build a lossless Ethernet fabric that also reacts to congestion before buffers overflow. Which combination of technologies should they implement to meet these requirements?

Accepted Answer

Priority Flow Control (PFC) for the RoCEv2 traffic class combined with Explicit Congestion Notification (ECN) and DCQCN-based rate control

Answer

Enhanced Transmission Selection (ETS) alone to guarantee bandwidth for the RoCEv2 class with tail-drop queuing

Answer

TCP with selective acknowledgment (SACK) and increased switch buffer allocation on all interfaces

Answer

Standard 802.3x global pause frames applied to all ports to prevent any packet loss