Network-Offloaded Bandwidth-Optimal Broadcast and Allgather for Distributed AI

In the Fully Sharded Data Parallel (FSDP) training pipeline, collective operations can be interleaved to maximize the communication/computation overlap. In this scenario, outstanding operations such as Allgather and Reduce-Scatter can compete for the injection bandwidth and create pipeline bubbles....

Full description

Saved in:

Bibliographic Details
Published in:	SC24: International Conference for High Performance Computing, Networking, Storage and Analysis pp. 1 - 17
Main Authors:	Khalilov, Mikhail, Girolamo, Salvatore Di, Chrapek, Marcin, Nudelman, Rami, Bloch, Gil, Hoefler, Torsten
Format:	Conference Proceeding
Language:	English
Published:	IEEE 17.11.2024
Subjects:	AI accelerators Artificial intelligence Clusters Engines Multicast algorithms Networking Next generation networking Pipelines Protocols Substrates Supercomputers Training
Online Access:	Get full text
Tags:	Add Tag No Tags, Be the first to tag this record!

Description
Summary:	In the Fully Sharded Data Parallel (FSDP) training pipeline, collective operations can be interleaved to maximize the communication/computation overlap. In this scenario, outstanding operations such as Allgather and Reduce-Scatter can compete for the injection bandwidth and create pipeline bubbles. To address this problem, we propose a novel bandwidth-optimal Allgather collective algorithm that leverages hardware multicast. We use multicast to build a constant-time reliable Broadcast protocol, a building block for constructing an optimal Allgather schedule. Our Allgather algorithm achieves 2 \times traffic reduction on a 188 -node testbed. To free the host side from running the protocol, we employ SmartNIC offloading. We extract the parallelism in our Allgather algorithm and map it to a SmartNIC specialized for hiding the cost of data movement. We show that our SmartNIC-offloaded collective progress engine can scale to the next generation of 1.6 Tbit/s links.
DOI:	10.1109/SC41406.2024.00109