Swing: Short-cutting Rings for Higher Bandwidth Allreduce

arXiv:2401.09356v2 Announce Type: replace-cross
Abstract: The allreduce collective operation accounts for a significant fraction of the runtime of workloads running on distributed systems. One factor determining its performance is the distance between communicating nodes, especially on networks like torus, where a higher distance implies multiple messages being forwarded on the same link, thus reducing the allreduce bandwidth. Torus networks are widely used on systems optimized for machine learning workloads (e.g., Google TPUs and Amazon Trainium devices), as well as on …

abstract arxiv bandwidth collective cs.dc cs.lg cs.ni cs.pf distributed distributed systems messages multiple networks nodes performance running systems type workloads