Manage OpenMP GPU data environment under unified address space

Lingda Li; Hal Finkel; Martin Kong; Barbara Chapman

doi:10.1007/978-3-319-98521-3_5

Manage OpenMP GPU data environment under unified address space

Lingda Li
, Hal Finkel
, Martin Kong
, Barbara Chapman

Computer Science

Brookhaven National Laboratory
Argonne National Laboratory

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

6 Scopus citations

Abstract

OpenMP has supported the offload of computations to accelerators such as GPUs since version 4.0. A crucial aspect in OpenMP offloading is to manage the accelerator data environment. Currently, this has to be explicitly programmed by users, which is non-trival and often results in suboptimal performance. The unified memory feature available in recent GPU architectures introduces another option, implicit management. However, our experiments show that it incurs several performance issues, especially under GPU memory oversubscription. In this paper, we propose a compiler and runtime collaborative approach to manage OpenMP GPU data under unified memory. In our framework, the compiler performs data reuse analysis to assist runtime data management. The runtime combines static and dynamic information to make optimized data management decisions. We have implement the proposed technology in the LLVM framework. The evaluation shows our method can achieve significant performance improvement for OpenMP GPU offloading.

Original language	English
Title of host publication	Evolving OpenMP for Evolving Architectures - 14th International Workshop on OpenMP, IWOMP 2018, Proceedings
Editors	Pedro Valero-Lara, Sergi Mateo Bellido, Xavier Martorell, Jesus Labarta, Bronis R. de Supinski
Publisher	Springer Verlag
Pages	69-81
Number of pages	13
ISBN (Print)	9783319985206
DOIs	https://doi.org/10.1007/978-3-319-98521-3_5
State	Published - 2018
Event	14th International Workshop on OpenMP, IWOMP 2018 - Barcelona, Spain Duration: Sep 26 2018 → Sep 28 2018

Publication series

Name	Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume	11128 LNCS
ISSN (Print)	0302-9743
ISSN (Electronic)	1611-3349

Conference

Conference	14th International Workshop on OpenMP, IWOMP 2018
Country/Territory	Spain
City	Barcelona
Period	09/26/18 → 09/28/18

Keywords

Compiler
Data management
LLVM
OpenMP offloading
Runtime
Unified memory

Access to Document

10.1007/978-3-319-98521-3_5

Cite this

Li, L., Finkel, H., Kong, M., & Chapman, B. (2018). Manage OpenMP GPU data environment under unified address space. In P. Valero-Lara, S. M. Bellido, X. Martorell, J. Labarta, & B. R. de Supinski (Eds.), Evolving OpenMP for Evolving Architectures - 14th International Workshop on OpenMP, IWOMP 2018, Proceedings (pp. 69-81). (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Vol. 11128 LNCS). Springer Verlag. https://doi.org/10.1007/978-3-319-98521-3_5

Li, Lingda ; Finkel, Hal ; Kong, Martin et al. / Manage OpenMP GPU data environment under unified address space. Evolving OpenMP for Evolving Architectures - 14th International Workshop on OpenMP, IWOMP 2018, Proceedings. editor / Pedro Valero-Lara ; Sergi Mateo Bellido ; Xavier Martorell ; Jesus Labarta ; Bronis R. de Supinski. Springer Verlag, 2018. pp. 69-81 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)).

@inproceedings{0e68ab2eef59416d85d10aa3d0102328,

title = "Manage OpenMP GPU data environment under unified address space",

abstract = "OpenMP has supported the offload of computations to accelerators such as GPUs since version 4.0. A crucial aspect in OpenMP offloading is to manage the accelerator data environment. Currently, this has to be explicitly programmed by users, which is non-trival and often results in suboptimal performance. The unified memory feature available in recent GPU architectures introduces another option, implicit management. However, our experiments show that it incurs several performance issues, especially under GPU memory oversubscription. In this paper, we propose a compiler and runtime collaborative approach to manage OpenMP GPU data under unified memory. In our framework, the compiler performs data reuse analysis to assist runtime data management. The runtime combines static and dynamic information to make optimized data management decisions. We have implement the proposed technology in the LLVM framework. The evaluation shows our method can achieve significant performance improvement for OpenMP GPU offloading.",

keywords = "Compiler, Data management, LLVM, OpenMP offloading, Runtime, Unified memory",

author = "Lingda Li and Hal Finkel and Martin Kong and Barbara Chapman",

note = "Publisher Copyright: {\textcopyright} Springer Nature Switzerland AG 2018.; 14th International Workshop on OpenMP, IWOMP 2018 ; Conference date: 26-09-2018 Through 28-09-2018",

year = "2018",

doi = "10.1007/978-3-319-98521-3\_5",

language = "English",

isbn = "9783319985206",

series = "Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)",

publisher = "Springer Verlag",

pages = "69--81",

editor = "Pedro Valero-Lara and Bellido, \{Sergi Mateo\} and Xavier Martorell and Jesus Labarta and \{de Supinski\}, \{Bronis R.\}",

booktitle = "Evolving OpenMP for Evolving Architectures - 14th International Workshop on OpenMP, IWOMP 2018, Proceedings",

}

Li, L, Finkel, H, Kong, M & Chapman, B 2018, Manage OpenMP GPU data environment under unified address space. in P Valero-Lara, SM Bellido, X Martorell, J Labarta & BR de Supinski (eds), Evolving OpenMP for Evolving Architectures - 14th International Workshop on OpenMP, IWOMP 2018, Proceedings. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), vol. 11128 LNCS, Springer Verlag, pp. 69-81, 14th International Workshop on OpenMP, IWOMP 2018, Barcelona, Spain, 09/26/18. https://doi.org/10.1007/978-3-319-98521-3_5

Manage OpenMP GPU data environment under unified address space. / Li, Lingda; Finkel, Hal; Kong, Martin et al.
Evolving OpenMP for Evolving Architectures - 14th International Workshop on OpenMP, IWOMP 2018, Proceedings. ed. / Pedro Valero-Lara; Sergi Mateo Bellido; Xavier Martorell; Jesus Labarta; Bronis R. de Supinski. Springer Verlag, 2018. p. 69-81 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Vol. 11128 LNCS).

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

TY - GEN

T1 - Manage OpenMP GPU data environment under unified address space

AU - Li, Lingda

AU - Finkel, Hal

AU - Kong, Martin

AU - Chapman, Barbara

N1 - Publisher Copyright: © Springer Nature Switzerland AG 2018.

PY - 2018

Y1 - 2018

N2 - OpenMP has supported the offload of computations to accelerators such as GPUs since version 4.0. A crucial aspect in OpenMP offloading is to manage the accelerator data environment. Currently, this has to be explicitly programmed by users, which is non-trival and often results in suboptimal performance. The unified memory feature available in recent GPU architectures introduces another option, implicit management. However, our experiments show that it incurs several performance issues, especially under GPU memory oversubscription. In this paper, we propose a compiler and runtime collaborative approach to manage OpenMP GPU data under unified memory. In our framework, the compiler performs data reuse analysis to assist runtime data management. The runtime combines static and dynamic information to make optimized data management decisions. We have implement the proposed technology in the LLVM framework. The evaluation shows our method can achieve significant performance improvement for OpenMP GPU offloading.

AB - OpenMP has supported the offload of computations to accelerators such as GPUs since version 4.0. A crucial aspect in OpenMP offloading is to manage the accelerator data environment. Currently, this has to be explicitly programmed by users, which is non-trival and often results in suboptimal performance. The unified memory feature available in recent GPU architectures introduces another option, implicit management. However, our experiments show that it incurs several performance issues, especially under GPU memory oversubscription. In this paper, we propose a compiler and runtime collaborative approach to manage OpenMP GPU data under unified memory. In our framework, the compiler performs data reuse analysis to assist runtime data management. The runtime combines static and dynamic information to make optimized data management decisions. We have implement the proposed technology in the LLVM framework. The evaluation shows our method can achieve significant performance improvement for OpenMP GPU offloading.

KW - Compiler

KW - Data management

KW - LLVM

KW - OpenMP offloading

KW - Runtime

KW - Unified memory

UR - https://www.scopus.com/pages/publications/85057258853

U2 - 10.1007/978-3-319-98521-3_5

DO - 10.1007/978-3-319-98521-3_5

M3 - Conference contribution

AN - SCOPUS:85057258853

SN - 9783319985206

T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

SP - 69

EP - 81

BT - Evolving OpenMP for Evolving Architectures - 14th International Workshop on OpenMP, IWOMP 2018, Proceedings

A2 - Valero-Lara, Pedro

A2 - Bellido, Sergi Mateo

A2 - Martorell, Xavier

A2 - Labarta, Jesus

A2 - de Supinski, Bronis R.

PB - Springer Verlag

T2 - 14th International Workshop on OpenMP, IWOMP 2018

Y2 - 26 September 2018 through 28 September 2018

ER -

Li L, Finkel H, Kong M, Chapman B. Manage OpenMP GPU data environment under unified address space. In Valero-Lara P, Bellido SM, Martorell X, Labarta J, de Supinski BR, editors, Evolving OpenMP for Evolving Architectures - 14th International Workshop on OpenMP, IWOMP 2018, Proceedings. Springer Verlag. 2018. p. 69-81. (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)). doi: 10.1007/978-3-319-98521-3_5

Manage OpenMP GPU data environment under unified address space

Abstract

Publication series

Conference

Keywords

Access to Document

Other files and links

Fingerprint

Cite this