Cisco Engineer Builds Emulated Version of Multipath Reliable Connection
Cisco was noticeably absent from the development of Multipath Reliable Connection (MRC), an OpenAI-developed transport protocol for large-scale AI training networks. However, one of its engineers, Bruce McDougall, decided to get hands-on with the new architecture by building an emulated version using Docker containers.
The emulator simulates MRC over a segment routing over IPv6 (SRv6) dataplane and is designed to give engineers and operators a feel for the multiplanar network design and its packet spraying abilities. McDougall wrote in a LinkedIn post that he didn't have a hardware budget to get hands-on with real MRC gear.
MRC was developed by OpenAI alongside Nvidia, AMD, Intel, and Broadcom to speed up AI model training on vast clusters, reducing packet delays and congestion when connecting tens-of-thousands of graphics processing units (GPUs). The protocol has been deployed across Oracle Cloud Infrastructure's supercomputer in Abilene, Texas, which OpenAI uses to train its frontier models.