Abstract: Distributed training (DT) in shared clusters usually deploys a scheduler for resource allocation to multiple concurrent jobs. Meanwhile, a recent acceleration primitive, In-Network ...