DeSync: Proactive Congestion Control via Random Delay Offsets for Large-Scale ML Training

Xingbo Feng; Zhuyun Qi; Yi Wang; Ziyao Huang; Yan Liu; Jiashuo Lin; Chenxi Ling; Weichao Li; Jin Zhang; Jianping Wang

doi:10.1109/IWQoS65803.2025.11143372

DeSync: Proactive Congestion Control via Random Delay Offsets for Large-Scale ML Training

Xingbo Feng
, Zhuyun Qi^*
, Yi Wang
, Ziyao Huang
, Yan Liu
, Jiashuo Lin
, Chenxi Ling
, Weichao Li
, Jin Zhang
, Jianping Wang

^*Corresponding author for this work

Department of Computer Science

Research output: Chapters, Conference Papers, Creative and Literary Works › RGC 32 - Refereed conference paper (with host publication) › peer-review

Abstract

Synchronization-induced congestion is a critical performance bottleneck in modern distributed machine learning (ML) training, where simultaneous gradient exchanges create bursty traffic patterns. Existing solutions, both reactive and proactive, struggle to balance throughput and latency in the presence of synchronized flows. We propose DeSync, a proactive traffic shaping scheme that introduces structured random delay to de-synchronize communication rounds. Evaluations with DCQCN, HPCC, DCTCP, and TIMELY demonstrate that DeSync significantly improves FCT, job completion times, and congestion metrics, enhancing existing CC mechanisms without specialized hardware. © 2025 IEEE.

Original language	English
Title of host publication	2025 IEEE/ACM 33rd International Symposium on Quality of Service (IWQoS)
Publisher	IEEE
Number of pages	2
ISBN (Electronic)	9798331549404
ISBN (Print)	9798331549411
DOIs	https://doi.org/10.1109/IWQoS65803.2025.11143372
Publication status	Published - 2025
Event	33rd IEEE/ACM International Symposium on Quality of Service (IWQoS 2025) - Gold Coast, Australia Duration: 2 Jul 2025 → 4 Jul 2025 https://iwqos2025.ieee-iwqos.org/

Publication series

Name	IEEE International Workshop on Quality of Service, IWQoS
ISSN (Print)	1548-615X
ISSN (Electronic)	2766-8568

Conference

Conference	33rd IEEE/ACM International Symposium on Quality of Service (IWQoS 2025)
Abbreviated title	IEEE/ACM IWQoS 2025
Place	Australia
City	Gold Coast
Period	2/07/25 → 4/07/25
Internet address	https://iwqos2025.ieee-iwqos.org/

Funding

This work is supported by the Guangdong High-Level Talents Special Support Program (2021TX05X205), the Guangdong Basic and Applied Basic Research Foundation (2024A0101010001), Pengcheng Laboratory The Major Key Project of PCL (PCL2023A03)

Access Output

10.1109/IWQoS65803.2025.11143372

Other Access Options

Check CityUHK Library

Permanent Link

Right click to copy link

Cite this

Feng, X., Qi, Z., Wang, Y., Huang, Z., Liu, Y., Lin, J., Ling, C., Li, W., Zhang, J., & Wang, J. (2025). DeSync: Proactive Congestion Control via Random Delay Offsets for Large-Scale ML Training. In 2025 IEEE/ACM 33rd International Symposium on Quality of Service (IWQoS) (IEEE International Workshop on Quality of Service, IWQoS). IEEE. https://doi.org/10.1109/IWQoS65803.2025.11143372

@inproceedings{e767fbcf19f043489f3eba5918a01455,

title = "DeSync: Proactive Congestion Control via Random Delay Offsets for Large-Scale ML Training",

abstract = "Synchronization-induced congestion is a critical performance bottleneck in modern distributed machine learning (ML) training, where simultaneous gradient exchanges create bursty traffic patterns. Existing solutions, both reactive and proactive, struggle to balance throughput and latency in the presence of synchronized flows. We propose DeSync, a proactive traffic shaping scheme that introduces structured random delay to de-synchronize communication rounds. Evaluations with DCQCN, HPCC, DCTCP, and TIMELY demonstrate that DeSync significantly improves FCT, job completion times, and congestion metrics, enhancing existing CC mechanisms without specialized hardware. {\textcopyright} 2025 IEEE.",

author = "Xingbo Feng and Zhuyun Qi and Yi Wang and Ziyao Huang and Yan Liu and Jiashuo Lin and Chenxi Ling and Weichao Li and Jin Zhang and Jianping Wang",

year = "2025",

doi = "10.1109/IWQoS65803.2025.11143372",

language = "English",

isbn = "9798331549411",

series = "IEEE International Workshop on Quality of Service, IWQoS",

publisher = "IEEE",

booktitle = "2025 IEEE/ACM 33rd International Symposium on Quality of Service (IWQoS)",

address = "United States",

note = "33rd IEEE/ACM International Symposium on Quality of Service (IWQoS 2025), IEEE/ACM IWQoS 2025 ; Conference date: 02-07-2025 Through 04-07-2025",

url = "https://iwqos2025.ieee-iwqos.org/",

}

Feng, X, Qi, Z, Wang, Y, Huang, Z, Liu, Y, Lin, J, Ling, C, Li, W, Zhang, J & Wang, J 2025, DeSync: Proactive Congestion Control via Random Delay Offsets for Large-Scale ML Training. in 2025 IEEE/ACM 33rd International Symposium on Quality of Service (IWQoS). IEEE International Workshop on Quality of Service, IWQoS, IEEE, 33rd IEEE/ACM International Symposium on Quality of Service (IWQoS 2025), Gold Coast, Australia, 2/07/25. https://doi.org/10.1109/IWQoS65803.2025.11143372

DeSync: Proactive Congestion Control via Random Delay Offsets for Large-Scale ML Training. / Feng, Xingbo; Qi, Zhuyun; Wang, Yi et al.
2025 IEEE/ACM 33rd International Symposium on Quality of Service (IWQoS). IEEE, 2025. (IEEE International Workshop on Quality of Service, IWQoS).

Research output: Chapters, Conference Papers, Creative and Literary Works › RGC 32 - Refereed conference paper (with host publication) › peer-review

TY - GEN

T1 - DeSync

T2 - 33rd IEEE/ACM International Symposium on Quality of Service (IWQoS 2025)

AU - Feng, Xingbo

AU - Qi, Zhuyun

AU - Wang, Yi

AU - Huang, Ziyao

AU - Liu, Yan

AU - Lin, Jiashuo

AU - Ling, Chenxi

AU - Li, Weichao

AU - Zhang, Jin

AU - Wang, Jianping

PY - 2025

Y1 - 2025

N2 - Synchronization-induced congestion is a critical performance bottleneck in modern distributed machine learning (ML) training, where simultaneous gradient exchanges create bursty traffic patterns. Existing solutions, both reactive and proactive, struggle to balance throughput and latency in the presence of synchronized flows. We propose DeSync, a proactive traffic shaping scheme that introduces structured random delay to de-synchronize communication rounds. Evaluations with DCQCN, HPCC, DCTCP, and TIMELY demonstrate that DeSync significantly improves FCT, job completion times, and congestion metrics, enhancing existing CC mechanisms without specialized hardware. © 2025 IEEE.

AB - Synchronization-induced congestion is a critical performance bottleneck in modern distributed machine learning (ML) training, where simultaneous gradient exchanges create bursty traffic patterns. Existing solutions, both reactive and proactive, struggle to balance throughput and latency in the presence of synchronized flows. We propose DeSync, a proactive traffic shaping scheme that introduces structured random delay to de-synchronize communication rounds. Evaluations with DCQCN, HPCC, DCTCP, and TIMELY demonstrate that DeSync significantly improves FCT, job completion times, and congestion metrics, enhancing existing CC mechanisms without specialized hardware. © 2025 IEEE.

UR - https://www.scopus.com/pages/publications/105017008608

UR - https://www.scopus.com/record/pubmetrics.uri?eid=2-s2.0-105017008608&origin=recordpage

U2 - 10.1109/IWQoS65803.2025.11143372

DO - 10.1109/IWQoS65803.2025.11143372

M3 - RGC 32 - Refereed conference paper (with host publication)

SN - 9798331549411

T3 - IEEE International Workshop on Quality of Service, IWQoS

BT - 2025 IEEE/ACM 33rd International Symposium on Quality of Service (IWQoS)

PB - IEEE

Y2 - 2 July 2025 through 4 July 2025

ER -

DeSync: Proactive Congestion Control via Random Delay Offsets for Large-Scale ML Training

Abstract

Publication series

Conference

Funding

Access Output

Other Access Options

Permanent Link

Other Files and Links

Fingerprint

Cite this