Preprint
Sep 2026
Sharing a Fabric with Collective Communication: Two Storage Penalties in Deep Learning Training
This work shows that node-local NVMe staging via DYAD eliminates both effects of I/O stall latency and traffic-class contention on collective communication by keeping storage I/O off that path.
Chen Wang, Wen-Zhao Wu, Hyojin Kim et al.
· 0 citations