AI and ML applications are growing at a fast pace in data centers. Ethernet-based networks are gaining interest as an alternative to InfiniBand in AI data center networking. RDMA over Converged Ethernet version 2 (ROCEv2) encapsulates RDMA/RC protocol packets within UDP packets for transport over Ethernet networks. Juniper’s Arun Gandhi and Michal Styszynski discuss the data transfer protocols and congestion considerations for AI workloads.

AI/ML Data Center Networking on Ethernet:
[ Ссылка ]

Part 2 of 3 Video Series: RDMA Over Converged Ethernet Version 2 (ROCEv2):
[ Ссылка ]

AI Data Center Networking:
[ Ссылка ]

0:00 Introduction
1:00 Interview with Michal Styszynski
5:06 How are RDMA V2 components coordinating inside the fabric
7:33 How can an operator be sure that RDMA V2 is working
9:42 Advanced RDMA V2 capabilities
16:42 Advanced load balancing

свернуть