Load Balancing for AI Training Workloads

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: McClure, Sarah, Cohen, Evyatar, Shpiner, Alex, Silberstein, Mark, Ratnasamy, Sylvia, Shenker, Scott, Keslassy, Isaac
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866915795223707648
author McClure, Sarah
Cohen, Evyatar
Shpiner, Alex
Silberstein, Mark
Ratnasamy, Sylvia
Shenker, Scott
Keslassy, Isaac
author_facet McClure, Sarah
Cohen, Evyatar
Shpiner, Alex
Silberstein, Mark
Ratnasamy, Sylvia
Shenker, Scott
Keslassy, Isaac
contents The extreme bandwidth demands of AI training has made load-balancing a critical component in AI fabrics, and a variety of load-balancing designs have emerged in recent work from both industry and research. However, there is currently little consensus on which design approach dominates or the conditions under which an approach dominates. We also lack an understanding of how far these approaches are from optimal. We provide a technical foundation for answering these questions by systematically evaluating leading load-balancing designs, while decoupling them from specific congestion control and loss recovery stacks. We find that load-balancing based on packet spraying dominates traditional approaches that load balance traffic at flow, flowlet, or subflow granularities. When comparing host- vs switch-based approaches to packet spraying, we find that they perform similarly in failure-free scenarios but that a host-based approach dominates under link failure because of its rapid visibility into end-to-end path conditions. We also identify that no leading approach achieves optimal O(1) queue scaling at maximum utilization. We demonstrate why a destination-based rotation (DR) discipline can reach this optimum and introduce Ofan, a switch-based implementation of DR that we show offers valuable performance gains over other packet spraying approaches.
format Preprint
id arxiv_https___arxiv_org_abs_2507_21372
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Load Balancing for AI Training Workloads
McClure, Sarah
Cohen, Evyatar
Shpiner, Alex
Silberstein, Mark
Ratnasamy, Sylvia
Shenker, Scott
Keslassy, Isaac
Networking and Internet Architecture
Machine Learning
The extreme bandwidth demands of AI training has made load-balancing a critical component in AI fabrics, and a variety of load-balancing designs have emerged in recent work from both industry and research. However, there is currently little consensus on which design approach dominates or the conditions under which an approach dominates. We also lack an understanding of how far these approaches are from optimal. We provide a technical foundation for answering these questions by systematically evaluating leading load-balancing designs, while decoupling them from specific congestion control and loss recovery stacks. We find that load-balancing based on packet spraying dominates traditional approaches that load balance traffic at flow, flowlet, or subflow granularities. When comparing host- vs switch-based approaches to packet spraying, we find that they perform similarly in failure-free scenarios but that a host-based approach dominates under link failure because of its rapid visibility into end-to-end path conditions. We also identify that no leading approach achieves optimal O(1) queue scaling at maximum utilization. We demonstrate why a destination-based rotation (DR) discipline can reach this optimum and introduce Ofan, a switch-based implementation of DR that we show offers valuable performance gains over other packet spraying approaches.
title Load Balancing for AI Training Workloads
topic Networking and Internet Architecture
Machine Learning
url https://arxiv.org/abs/2507.21372