Compute-based Fault Tolerance for DNN

Adi Molkho; Amit Golander; Oded Schwartz

doi:10.1145/3757347.3759145

Compute-based Fault Tolerance for DNN

Adi Molkho
, Amit Golander
, Oded Schwartz

The Rachel and Selim Benin School of Engineering and Computer Science

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › peer-review

Abstract

Deep neural network (DNN) systems use many GPUs, which can fail—making fault tolerance (FT) essential to avoid cluster restarts. Traditional FT relies on frequent checkpointing, incurring high bandwidth and memory costs. We propose an alternative strategy using GPU redundancy, introducing uniform and heterogeneous encoding approaches.

Original language	English
Title of host publication	Proceedings of the 18th ACM International Systems and Storage Conference, SYSTOR 2025
Publisher	Association for Computing Machinery, Inc
Pages	216
Number of pages	1
ISBN (Electronic)	9798400721199
DOIs	https://doi.org/10.1145/3757347.3759145
State	Published - 8 Sep 2025
Event	18th ACM International Systems and Storage Conference, SYSTOR 2025 - Virtual, Online, Israel Duration: 8 Sep 2025 → 9 Sep 2025

Publication series

Name	Proceedings of the 18th ACM International Systems and Storage Conference, SYSTOR 2025

Conference

Conference	18th ACM International Systems and Storage Conference, SYSTOR 2025
Country/Territory	Israel
City	Virtual, Online
Period	8/09/25 → 9/09/25

Bibliographical note

Publisher Copyright:
© 2025 Copyright held by the owner/author(s).

Keywords

DNN
Distributed Systems
Fault Tolerance
GPU

Access to Document

10.1145/3757347.3759145

Cite this

@inproceedings{6fac7074df5d48899646520544388b00,

title = "Compute-based Fault Tolerance for DNN",

abstract = "Deep neural network (DNN) systems use many GPUs, which can fail—making fault tolerance (FT) essential to avoid cluster restarts. Traditional FT relies on frequent checkpointing, incurring high bandwidth and memory costs. We propose an alternative strategy using GPU redundancy, introducing uniform and heterogeneous encoding approaches.",

keywords = "DNN, Distributed Systems, Fault Tolerance, GPU",

author = "Adi Molkho and Amit Golander and Oded Schwartz",

note = "Publisher Copyright: {\textcopyright} 2025 Copyright held by the owner/author(s).; 18th ACM International Systems and Storage Conference, SYSTOR 2025 ; Conference date: 08-09-2025 Through 09-09-2025",

year = "2025",

month = sep,

day = "8",

doi = "10.1145/3757347.3759145",

language = "אנגלית",

series = "Proceedings of the 18th ACM International Systems and Storage Conference, SYSTOR 2025",

publisher = "Association for Computing Machinery, Inc",

pages = "216",

booktitle = "Proceedings of the 18th ACM International Systems and Storage Conference, SYSTOR 2025",

}

Molkho, A, Golander, A & Schwartz, O 2025, Compute-based Fault Tolerance for DNN. in Proceedings of the 18th ACM International Systems and Storage Conference, SYSTOR 2025. Proceedings of the 18th ACM International Systems and Storage Conference, SYSTOR 2025, Association for Computing Machinery, Inc, pp. 216, 18th ACM International Systems and Storage Conference, SYSTOR 2025, Virtual, Online, Israel, 8/09/25. https://doi.org/10.1145/3757347.3759145

Compute-based Fault Tolerance for DNN. / Molkho, Adi; Golander, Amit; Schwartz, Oded.
Proceedings of the 18th ACM International Systems and Storage Conference, SYSTOR 2025. Association for Computing Machinery, Inc, 2025. p. 216 (Proceedings of the 18th ACM International Systems and Storage Conference, SYSTOR 2025).

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › peer-review

TY - GEN

T1 - Compute-based Fault Tolerance for DNN

AU - Molkho, Adi

AU - Golander, Amit

AU - Schwartz, Oded

PY - 2025/9/8

Y1 - 2025/9/8

N2 - Deep neural network (DNN) systems use many GPUs, which can fail—making fault tolerance (FT) essential to avoid cluster restarts. Traditional FT relies on frequent checkpointing, incurring high bandwidth and memory costs. We propose an alternative strategy using GPU redundancy, introducing uniform and heterogeneous encoding approaches.

AB - Deep neural network (DNN) systems use many GPUs, which can fail—making fault tolerance (FT) essential to avoid cluster restarts. Traditional FT relies on frequent checkpointing, incurring high bandwidth and memory costs. We propose an alternative strategy using GPU redundancy, introducing uniform and heterogeneous encoding approaches.

KW - DNN

KW - Distributed Systems

KW - Fault Tolerance

KW - GPU

UR - https://www.scopus.com/pages/publications/105017692446

U2 - 10.1145/3757347.3759145

DO - 10.1145/3757347.3759145

M3 - ???researchoutput.researchoutputtypes.contributiontobookanthology.conference???

AN - SCOPUS:105017692446

T3 - Proceedings of the 18th ACM International Systems and Storage Conference, SYSTOR 2025

SP - 216

BT - Proceedings of the 18th ACM International Systems and Storage Conference, SYSTOR 2025

PB - Association for Computing Machinery, Inc

T2 - 18th ACM International Systems and Storage Conference, SYSTOR 2025

Y2 - 8 September 2025 through 9 September 2025

ER -

Compute-based Fault Tolerance for DNN

Abstract

Publication series

Conference

Bibliographical note

Keywords

Access to Document

Other files and links

Fingerprint

Cite this