Tasking in accelerators: Performance evaluation

Leonel Toledo; Antonio J. Pena; Sandra Catalan; Pedro Valero-Lara

doi:10.1109/PDCAT46702.2019.00034

Tasking in accelerators: Performance evaluation

Leonel Toledo, Antonio J. Pena, Sandra Catalan, Pedro Valero-Lara

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

7 Scopus citations

Abstract

In this work, we analyze the implications and results of implementing dynamic parallelism, concurrent kernels and CUDA Graphs to solve task-oriented problems. As a benchmark we propose three different methods for solving DGEMM operation on tiled-matrices; which might be the most popular benchmark for performance analysis. For the algorithms that we study, we present significant differences in terms of data dependencies, synchronization and granularity. The main contribution of this work is determining which of the previous approaches work better for having multiple task running concurrently in a single GPU, as well as stating the main limitations and benefits of every technique. Using dynamic parallelism and CUDA Streams we were able to achieve up to 30% speedups and for CUDA Graph API up to 25x acceleration outperforming state of the art results.

Original language	English
Title of host publication	Proceedings - 2019 20th International Conference on Parallel and Distributed Computing, Applications and Technologies, PDCAT 2019
Editors	Hui Tian, Hong Shen, Wee Lum Tan
Publisher	Institute of Electrical and Electronics Engineers Inc.
Pages	127-132
Number of pages	6
ISBN (Electronic)	9781728126166
DOIs	https://doi.org/10.1109/PDCAT46702.2019.00034
State	Published - Dec 2019
Externally published	Yes
Event	20th International Conference on Parallel and Distributed Computing, Applications and Technologies, PDCAT 2019 - Gold Coast, Australia Duration: Dec 5 2019 → Dec 7 2019

Publication series

Name	Proceedings - 2019 20th International Conference on Parallel and Distributed Computing, Applications and Technologies, PDCAT 2019

Conference

Conference	20th International Conference on Parallel and Distributed Computing, Applications and Technologies, PDCAT 2019
Country/Territory	Australia
City	Gold Coast
Period	12/5/19 → 12/7/19

Funding

This project has received funding from the EPEEC project from the European Union's Horizon 2020 research and innovation programme under grant agreement No 801051, from the Spanish Ministry of Economy and Competitiveness under the project Computacion de Altas Prestaciones VII (TIN2015-65316-P ) and the Departament d'Innovacio, Universitats i Empresa de la Generalitat de Catalunya , under project MPEXPAR: Models de Pro-gramacio i Entorns d'Execucio Parallels (2014-SGR-1051). Finally, this project also received funding from the Spanish Ministry of Economy and Competitiveness under the Juan de la Cierva Grant Agreement No IJCI-2017-33511 , and from the European Union's Horizon 2020 research and innovation program under the Marie Sklodowska Curie grant agreement No. 749516 This project has received funding from the EPEEC project from the European Union’s Horizon 2020 research and inno-vationprogramme under grant agreement No 801051, from the Spanish Ministry of Economy and Competitvieness under the project Computación de Altas Prestaciones VII ( TIN2015-65316-P ) and the Departament d’Innovació, Universitats i Empresa de la Generalitat de Catalunya , under project MPEX-PAR:Models de Pro-gramació i Entorns d’Execució Paral · lels ( 2014-SGR-1051 ). Finally, this project also received funding from the Spanish Ministry of Economy and Competitvieness under the Juan de la Cierva Grant Agreement No IJCI-2017-33511 , and from the European Union’s Horizon 2020 research and innovation program under the Marie Sklodowska Curie grant agreement No. 749516 .

Funders	Funder number
Computacion de Altas Prestaciones VII
Computación de Altas Prestaciones VII	TIN2015-65316-P
European Union’s Horizon 2020 Research and Innovation Program
Spanish Ministry of Economy and Competitvieness
Horizon 2020 Framework Programme
Generalitat de Catalunya	IJCI-2017-33511, 2014-SGR-1051
Departament d'Innovació, Universitats i Empresa, Generalitat de Catalunya
Ministerio de Economía y Competitividad
Horizon 2020	801051, 749516

Keywords

CUDA
Cuda graph
Cuda stream
Dynamic parallelism
GPU

Access to Document

10.1109/PDCAT46702.2019.00034

Cite this

Toledo, L., Pena, A. J., Catalan, S., & Valero-Lara, P. (2019). Tasking in accelerators: Performance evaluation. In H. Tian, H. Shen, & W. L. Tan (Eds.), Proceedings - 2019 20th International Conference on Parallel and Distributed Computing, Applications and Technologies, PDCAT 2019 (pp. 127-132). Article 9029123 (Proceedings - 2019 20th International Conference on Parallel and Distributed Computing, Applications and Technologies, PDCAT 2019). Institute of Electrical and Electronics Engineers Inc.. https://doi.org/10.1109/PDCAT46702.2019.00034

Toledo, Leonel ; Pena, Antonio J. ; Catalan, Sandra et al. / Tasking in accelerators : Performance evaluation. Proceedings - 2019 20th International Conference on Parallel and Distributed Computing, Applications and Technologies, PDCAT 2019. editor / Hui Tian ; Hong Shen ; Wee Lum Tan. Institute of Electrical and Electronics Engineers Inc., 2019. pp. 127-132 (Proceedings - 2019 20th International Conference on Parallel and Distributed Computing, Applications and Technologies, PDCAT 2019).

@inproceedings{2b1122b986db48cd84a9fd07731aca45,

title = "Tasking in accelerators: Performance evaluation",

abstract = "In this work, we analyze the implications and results of implementing dynamic parallelism, concurrent kernels and CUDA Graphs to solve task-oriented problems. As a benchmark we propose three different methods for solving DGEMM operation on tiled-matrices; which might be the most popular benchmark for performance analysis. For the algorithms that we study, we present significant differences in terms of data dependencies, synchronization and granularity. The main contribution of this work is determining which of the previous approaches work better for having multiple task running concurrently in a single GPU, as well as stating the main limitations and benefits of every technique. Using dynamic parallelism and CUDA Streams we were able to achieve up to 30% speedups and for CUDA Graph API up to 25x acceleration outperforming state of the art results.",

keywords = "CUDA, Cuda graph, Cuda stream, Dynamic parallelism, GPU",

author = "Leonel Toledo and Pena, {Antonio J.} and Sandra Catalan and Pedro Valero-Lara",

note = "Publisher Copyright: {\textcopyright} 2019 IEEE.; 20th International Conference on Parallel and Distributed Computing, Applications and Technologies, PDCAT 2019 ; Conference date: 05-12-2019 Through 07-12-2019",

year = "2019",

month = dec,

doi = "10.1109/PDCAT46702.2019.00034",

language = "English",

series = "Proceedings - 2019 20th International Conference on Parallel and Distributed Computing, Applications and Technologies, PDCAT 2019",

publisher = "Institute of Electrical and Electronics Engineers Inc.",

pages = "127--132",

editor = "Hui Tian and Hong Shen and Tan, {Wee Lum}",

booktitle = "Proceedings - 2019 20th International Conference on Parallel and Distributed Computing, Applications and Technologies, PDCAT 2019",

}

Toledo, L, Pena, AJ, Catalan, S & Valero-Lara, P 2019, Tasking in accelerators: Performance evaluation. in H Tian, H Shen & WL Tan (eds), Proceedings - 2019 20th International Conference on Parallel and Distributed Computing, Applications and Technologies, PDCAT 2019., 9029123, Proceedings - 2019 20th International Conference on Parallel and Distributed Computing, Applications and Technologies, PDCAT 2019, Institute of Electrical and Electronics Engineers Inc., pp. 127-132, 20th International Conference on Parallel and Distributed Computing, Applications and Technologies, PDCAT 2019, Gold Coast, Australia, 12/5/19. https://doi.org/10.1109/PDCAT46702.2019.00034

Tasking in accelerators: Performance evaluation. / Toledo, Leonel; Pena, Antonio J.; Catalan, Sandra et al.
Proceedings - 2019 20th International Conference on Parallel and Distributed Computing, Applications and Technologies, PDCAT 2019. ed. / Hui Tian; Hong Shen; Wee Lum Tan. Institute of Electrical and Electronics Engineers Inc., 2019. p. 127-132 9029123 (Proceedings - 2019 20th International Conference on Parallel and Distributed Computing, Applications and Technologies, PDCAT 2019).

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

TY - GEN

T1 - Tasking in accelerators

T2 - 20th International Conference on Parallel and Distributed Computing, Applications and Technologies, PDCAT 2019

AU - Toledo, Leonel

AU - Pena, Antonio J.

AU - Catalan, Sandra

AU - Valero-Lara, Pedro

PY - 2019/12

Y1 - 2019/12

N2 - In this work, we analyze the implications and results of implementing dynamic parallelism, concurrent kernels and CUDA Graphs to solve task-oriented problems. As a benchmark we propose three different methods for solving DGEMM operation on tiled-matrices; which might be the most popular benchmark for performance analysis. For the algorithms that we study, we present significant differences in terms of data dependencies, synchronization and granularity. The main contribution of this work is determining which of the previous approaches work better for having multiple task running concurrently in a single GPU, as well as stating the main limitations and benefits of every technique. Using dynamic parallelism and CUDA Streams we were able to achieve up to 30% speedups and for CUDA Graph API up to 25x acceleration outperforming state of the art results.

AB - In this work, we analyze the implications and results of implementing dynamic parallelism, concurrent kernels and CUDA Graphs to solve task-oriented problems. As a benchmark we propose three different methods for solving DGEMM operation on tiled-matrices; which might be the most popular benchmark for performance analysis. For the algorithms that we study, we present significant differences in terms of data dependencies, synchronization and granularity. The main contribution of this work is determining which of the previous approaches work better for having multiple task running concurrently in a single GPU, as well as stating the main limitations and benefits of every technique. Using dynamic parallelism and CUDA Streams we were able to achieve up to 30% speedups and for CUDA Graph API up to 25x acceleration outperforming state of the art results.

KW - CUDA

KW - Cuda graph

KW - Cuda stream

KW - Dynamic parallelism

KW - GPU

UR - http://www.scopus.com/inward/record.url?scp=85083156573&partnerID=8YFLogxK

U2 - 10.1109/PDCAT46702.2019.00034

DO - 10.1109/PDCAT46702.2019.00034

M3 - Conference contribution

AN - SCOPUS:85083156573

T3 - Proceedings - 2019 20th International Conference on Parallel and Distributed Computing, Applications and Technologies, PDCAT 2019

SP - 127

EP - 132

BT - Proceedings - 2019 20th International Conference on Parallel and Distributed Computing, Applications and Technologies, PDCAT 2019

A2 - Tian, Hui

A2 - Shen, Hong

A2 - Tan, Wee Lum

PB - Institute of Electrical and Electronics Engineers Inc.

Y2 - 5 December 2019 through 7 December 2019

ER -

Toledo L, Pena AJ, Catalan S, Valero-Lara P. Tasking in accelerators: Performance evaluation. In Tian H, Shen H, Tan WL, editors, Proceedings - 2019 20th International Conference on Parallel and Distributed Computing, Applications and Technologies, PDCAT 2019. Institute of Electrical and Electronics Engineers Inc. 2019. p. 127-132. 9029123. (Proceedings - 2019 20th International Conference on Parallel and Distributed Computing, Applications and Technologies, PDCAT 2019). doi: 10.1109/PDCAT46702.2019.00034

Tasking in accelerators: Performance evaluation

Abstract

Publication series

Conference

Funding

Keywords

Access to Document

Other files and links

Fingerprint

Cite this