Swing - Short-cutting Rings for Higher Bandwidth Allreduce

Explore a 19-minute conference talk from NSDI '24 that introduces Swing, a novel algorithm designed to enhance allreduce performance on torus networks. Learn how this innovative approach reduces the number of hops between communicating nodes by swinging between torus directions, resulting in up to 3x performance improvement over existing allreduce algorithms. Discover the algorithm's effectiveness across various vector sizes and torus-like topologies, regardless of shape and size. Gain insights into the significance of allreduce operations in distributed systems and their impact on workload runtime, particularly in machine learning-optimized systems like Google TPUs and Amazon Trainium devices, as well as Top500 supercomputers. Understand the challenges posed by torus networks and how Swing addresses them to achieve higher bandwidth allreduce operations.

NSDI '24 - Swing: Short-cutting Rings for Higher Bandwidth Allreduce