2026-10-11T13:42:19Z https://drops.dagstuhl.de/oai

oai:drops-oai.dagstuhl.de:25604 2026-06-23T10:10:10Z ddc:004 open_access

Simulate Before Sending: Rethinking Transport in Datacenter Networks Straussman, Dan Keslassy, Isaac Shpiner, Alexander Liss, Liran Datacenter networks transport protocols AI training lossless networks Existing transport protocols in commodity datacenter networks struggle to provide low collective completion times (CCTs) to AI training collectives, as packet losses and retransmissions significantly degrade performance. We propose dcSim, an efficient transport that achieves low CCTs and practically lossless performance with commodity switches. In dcSim, each packet first employs a small simulation probe to traverse the network and explore congestion along a candidate path. Only packets whose simulation probes succeed are then transmitted, expecting to succeed as well. Evaluations confirm that dcSim achieves faster CCTs than existing schemes, with small queues and virtually zero packet loss. Finally, dcSim also excels in adverse conditions, including oversubscribed topologies. Schloss Dagstuhl – Leibniz-Zentrum für Informatik Dan Straussman and Isaac Keslassy and Alexander Shpiner and Liran Liss 2026 Is Part Of OASIcs, Volume 139, 1st New Ideas in Networked Systems (NINeS 2026) InProceedings Text doc-type:ResearchArticle publishedVersion application/pdf doi:10.4230/OASIcs.NINeS.2026.19 urn:nbn:de:0030-drops-256044 https://drops.dagstuhl.de/entities/document/10.4230/OASIcs.NINeS.2026.19 eng https://creativecommons.org/licenses/by/4.0/legalcode