Optically connected and reconfigurable GPU architecture for optimized peer-to-peer access

Erik Anderson; Jorge González; Alexander Gazman; Rodolfo Azevedo; Keren Bergman

doi:10.1145/3240302.3240418

Optically connected and reconfigurable GPU architecture for optimized peer-to-peer access

Erik Anderson, Jorge González, Alexander Gazman, Rodolfo Azevedo, Keren Bergman

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

Abstract

Increasing industry interest in the optimization of inter-GPU communication has motivated this work to explore new ways to enable peer-to-peer access. Specifically, this paper investigates how reconfigurable optical links between GPUs in multi-GPU servers can allow for minimized memory transfer latencies for given machine learning applications. Silicon photonics (SiP) is proposed as the enabling technology for such a reconfigurable architecture due to the potential for scalable and cost-efficient production. We evaluated our architecture using traffic obtained from an NVLink-connected 8 GPU server executing a set of machine learning models including AlexNet, DenseNet, NASNet, ResNet, MobileNet, and VGG16. Our results show up to 24.91% reduction of the total relative transmission latency (RTL) between peers.

Original language	English
Title of host publication	MEMSYS 2018 - Proceedings of the International Symposium on Memory Systems
Publisher	Association for Computing Machinery
ISBN (Electronic)	9781450364751
DOIs	https://doi.org/10.1145/3240302.3240418
State	Published - 1 Oct 2018
Externally published	Yes
Event	2018 International Symposium on Memory Systems, MEMSYS 2018 - Alexandria, United States Duration: 1 Oct 2018 → 4 Oct 2018

Publication series

Name	ACM International Conference Proceeding Series

Conference

Conference	2018 International Symposium on Memory Systems, MEMSYS 2018
Country/Territory	United States
City	Alexandria
Period	1/10/18 → 4/10/18

Access to Document

10.1145/3240302.3240418

Cite this

Anderson, E., González, J., Gazman, A., Azevedo, R., & Bergman, K. (2018). Optically connected and reconfigurable GPU architecture for optimized peer-to-peer access. In MEMSYS 2018 - Proceedings of the International Symposium on Memory Systems (ACM International Conference Proceeding Series). Association for Computing Machinery. https://doi.org/10.1145/3240302.3240418

@inproceedings{baea416a548046a3924cba2baa068c9f,

title = "Optically connected and reconfigurable GPU architecture for optimized peer-to-peer access",

abstract = "Increasing industry interest in the optimization of inter-GPU communication has motivated this work to explore new ways to enable peer-to-peer access. Specifically, this paper investigates how reconfigurable optical links between GPUs in multi-GPU servers can allow for minimized memory transfer latencies for given machine learning applications. Silicon photonics (SiP) is proposed as the enabling technology for such a reconfigurable architecture due to the potential for scalable and cost-efficient production. We evaluated our architecture using traffic obtained from an NVLink-connected 8 GPU server executing a set of machine learning models including AlexNet, DenseNet, NASNet, ResNet, MobileNet, and VGG16. Our results show up to 24.91% reduction of the total relative transmission latency (RTL) between peers.",

author = "Erik Anderson and Jorge Gonz{\'a}lez and Alexander Gazman and Rodolfo Azevedo and Keren Bergman",

note = "Publisher Copyright: {\textcopyright} 2018 Copyright held by the owner/author(s).; 2018 International Symposium on Memory Systems, MEMSYS 2018 ; Conference date: 01-10-2018 Through 04-10-2018",

year = "2018",

month = oct,

day = "1",

doi = "10.1145/3240302.3240418",

language = "English",

series = "ACM International Conference Proceeding Series",

publisher = "Association for Computing Machinery",

booktitle = "MEMSYS 2018 - Proceedings of the International Symposium on Memory Systems",

}

Anderson, E, González, J, Gazman, A, Azevedo, R & Bergman, K 2018, Optically connected and reconfigurable GPU architecture for optimized peer-to-peer access. in MEMSYS 2018 - Proceedings of the International Symposium on Memory Systems. ACM International Conference Proceeding Series, Association for Computing Machinery, 2018 International Symposium on Memory Systems, MEMSYS 2018, Alexandria, United States, 1/10/18. https://doi.org/10.1145/3240302.3240418

Optically connected and reconfigurable GPU architecture for optimized peer-to-peer access. / Anderson, Erik; González, Jorge; Gazman, Alexander et al.
MEMSYS 2018 - Proceedings of the International Symposium on Memory Systems. Association for Computing Machinery, 2018. (ACM International Conference Proceeding Series).

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

TY - GEN

T1 - Optically connected and reconfigurable GPU architecture for optimized peer-to-peer access

AU - Anderson, Erik

AU - González, Jorge

AU - Gazman, Alexander

AU - Azevedo, Rodolfo

AU - Bergman, Keren

PY - 2018/10/1

Y1 - 2018/10/1

N2 - Increasing industry interest in the optimization of inter-GPU communication has motivated this work to explore new ways to enable peer-to-peer access. Specifically, this paper investigates how reconfigurable optical links between GPUs in multi-GPU servers can allow for minimized memory transfer latencies for given machine learning applications. Silicon photonics (SiP) is proposed as the enabling technology for such a reconfigurable architecture due to the potential for scalable and cost-efficient production. We evaluated our architecture using traffic obtained from an NVLink-connected 8 GPU server executing a set of machine learning models including AlexNet, DenseNet, NASNet, ResNet, MobileNet, and VGG16. Our results show up to 24.91% reduction of the total relative transmission latency (RTL) between peers.

AB - Increasing industry interest in the optimization of inter-GPU communication has motivated this work to explore new ways to enable peer-to-peer access. Specifically, this paper investigates how reconfigurable optical links between GPUs in multi-GPU servers can allow for minimized memory transfer latencies for given machine learning applications. Silicon photonics (SiP) is proposed as the enabling technology for such a reconfigurable architecture due to the potential for scalable and cost-efficient production. We evaluated our architecture using traffic obtained from an NVLink-connected 8 GPU server executing a set of machine learning models including AlexNet, DenseNet, NASNet, ResNet, MobileNet, and VGG16. Our results show up to 24.91% reduction of the total relative transmission latency (RTL) between peers.

UR - http://www.scopus.com/inward/record.url?scp=85060987011&partnerID=8YFLogxK

U2 - 10.1145/3240302.3240418

DO - 10.1145/3240302.3240418

M3 - Conference contribution

AN - SCOPUS:85060987011

T3 - ACM International Conference Proceeding Series

BT - MEMSYS 2018 - Proceedings of the International Symposium on Memory Systems

PB - Association for Computing Machinery

T2 - 2018 International Symposium on Memory Systems, MEMSYS 2018

Y2 - 1 October 2018 through 4 October 2018

ER -

Optically connected and reconfigurable GPU architecture for optimized peer-to-peer access

Abstract

Publication series

Conference

Access to Document

Other files and links

Fingerprint

Cite this