RDMA over Ethernet for Distributed AI Training at Meta Scale (SIGCOMM'24, Paper 246)

Meta has shared insights into its extensive deployment of RDMA over Ethernet for AI training. Their strategy consists of three key components: analyzing AI training workloads, translating those insights into effective network design, and striking a balance between performance and ease of operation. They discovered that different AI tasks, such as ranking and generative AI, exhibit distinct scales and traffic patterns, which require tailored routing and congestion control solutions. To address this, Meta implemented a multi-layered network topology with separate front-end and back-end systems. They utilized traffic engineering and Equal-Cost Multi-Path (ECMP) for load balancing, opting for a collective communication library for congestion control instead of DCQCN. This approach enabled them to achieve optimal performance by co-tuning the network and the communication library, underscoring the significance of understanding and optimizing for the unique characteristics of various AI workloads.