Problem
Research questions and friction points this paper is trying to address.
Multipath Reliable Connection
AI/ML training
RoCEv2
congestion control
packet-loss recovery
The Multipath Reliable Connection (MRC) Transport
🤖 AI Summary
This work addresses the challenges of packet loss, congestion, and path failures in best-effort Ethernet networks during large-scale AI/ML training by proposing a highly reliable transport architecture built upon an extended RoCEv2 protocol. The proposed system introduces, for the first time in a production-grade protocol, per-packet multipath transmission, sender-driven congestion control, and a clean decoupling of packet delivery from semantic processing. It further integrates accelerated loss recovery and fault-tolerant mechanisms for path and port failures. Experimental results demonstrate that the architecture substantially enhances the reliability, resilience, and performance of AI training communications, exhibiting strong robustness against real-world network anomalies.
📝 Abstract
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.
Innovation
Methods, ideas, or system contributions that make the work stand out.
Multipath Reliable Connection
RoCEv2
sender-based congestion control
packet-loss recovery
fault resilience
🔎 Similar Papers
No similar papers found.
