Optical networks keep data flowing fast despite failures
OMRC: Multipath Reliability for Direct-Connect Optical Interconnects
Networking and Internet Architecture
Summary
Network failures can slow down or stop data tasks in direct optical connections used in data centers. The authors introduce Optical MRC (OMRC), a method that reroutes these connections when problems occur. This approach helps keep machine learning training fast and restarts halted tasks. It also greatly increases the network's bandwidth compared to previous methods.
What this means in practice
- •For data center network engineers: Maintain high-speed communication for AI training workloads in optical direct-connect networks despite link failures using OMRC’s remapping techniques.
- •For machine learning infrastructure teams: Restore stalled AI inference operations in direct optical networks by applying OMRC to reroute network traffic automatically after failures.
Authors
Eric Ding, Keshav Pradeep, Rachee Singh
Abstract
Network failures degrade Multipath Reliable Connection (MRC) in direct-connect optical datacenters. We propose Optical MRC (OMRC), which remaps collectives and circuits, preserving 95.7% of training throughput, restoring stalled inference, and achieving 4.7x MRC's simulated bandwidth.