The Multipath Reliable Connection (MRC) Transport
For AI/ML practitioners and network engineers, MRC addresses the bottleneck of reliable, high-throughput transport over commodity Ethernet, offering a practical solution for scaling training clusters.
MRC is a transport protocol extending RoCEv2 to improve large-scale AI/ML training over Ethernet, providing per-packet multipath, sender-based congestion control, and enhanced loss recovery and resilience. It achieves production-grade performance without requiring specialized hardware.
MRC is an open, production-grade transport designed for large-scale AI/ML training over best-effort Ethernet. It extends RoCEv2 with explicit, composable primitives for per-packet multipath and sender-based congestion control, decouples packet delivery from semantic processing, adds multiple new capabilities for accelerated packet-loss recovery and adds resilience against port and path failures. This paper presents MRC and details its core capabilities and mechanisms.