An adaptive performance modeling tool for GPU architectures

Sara S. Baghsorkhi, Matthieu Delahaye, Sanjay J. Patel, William D. Gropp, Wen Mei W. Hwu

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Abstract

This paper presents an analytical model to predict the performance of general-purpose applications on a GPU architecture. The model is designed to provide performance information to an auto-tuning compiler and assist it in narrowing down the search to the more promising implementations. It can also be incorporated into a tool to help programmers better assess the performance bottlenecks in their code. We analyze each GPU kernel and identify how the kernel exercises major GPU microarchitecture features. To identify the performance bottlenecks accurately, we introduce an abstract interpretation of a GPU kernel, work flow graph, based on which we estimate the execution time of a GPU kernel. We validated our performance model on the NVIDIA GPUs using CUDA (Compute Unified Device Architecture). For this purpose, we used data parallel benchmarks that stress different GPU microarchitecture events such as uncoalesced memory accesses, scratch-pad memory bank conflicts, and control flow divergence, which must be accurately modeled but represent challenges to the analytical performance models. The proposed model captures full system complexity and shows high accuracy in predicting the performance trends of different optimized kernel implementations. We also describe our approach to extracting the performance model automatically from a kernel code.

Original languageEnglish (US)
Title of host publicationPPoPP'10 - Proceedings of the 2010 ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming
Pages105-114
Number of pages10
DOIs
StatePublished - Mar 15 2010
Event2010 ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming, PPoPP'10 - Bangalore, India
Duration: Jan 9 2010Jan 14 2010

Publication series

NameProceedings of the ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming, PPOPP

Other

Other2010 ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming, PPoPP'10
CountryIndia
CityBangalore
Period1/9/101/14/10

Fingerprint

Flow graphs
Data storage equipment
Graphics processing unit
Flow control
Analytical models
Tuning

Keywords

  • Analytical model
  • GPU
  • Parallel programming
  • Performance estimation

ASJC Scopus subject areas

  • Software

Cite this

Baghsorkhi, S. S., Delahaye, M., Patel, S. J., Gropp, W. D., & Hwu, W. M. W. (2010). An adaptive performance modeling tool for GPU architectures. In PPoPP'10 - Proceedings of the 2010 ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming (pp. 105-114). (Proceedings of the ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming, PPOPP). https://doi.org/10.1145/1693453.1693470

An adaptive performance modeling tool for GPU architectures. / Baghsorkhi, Sara S.; Delahaye, Matthieu; Patel, Sanjay J.; Gropp, William D.; Hwu, Wen Mei W.

PPoPP'10 - Proceedings of the 2010 ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming. 2010. p. 105-114 (Proceedings of the ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming, PPOPP).

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Baghsorkhi, SS, Delahaye, M, Patel, SJ, Gropp, WD & Hwu, WMW 2010, An adaptive performance modeling tool for GPU architectures. in PPoPP'10 - Proceedings of the 2010 ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming. Proceedings of the ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming, PPOPP, pp. 105-114, 2010 ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming, PPoPP'10, Bangalore, India, 1/9/10. https://doi.org/10.1145/1693453.1693470
Baghsorkhi SS, Delahaye M, Patel SJ, Gropp WD, Hwu WMW. An adaptive performance modeling tool for GPU architectures. In PPoPP'10 - Proceedings of the 2010 ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming. 2010. p. 105-114. (Proceedings of the ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming, PPOPP). https://doi.org/10.1145/1693453.1693470
Baghsorkhi, Sara S. ; Delahaye, Matthieu ; Patel, Sanjay J. ; Gropp, William D. ; Hwu, Wen Mei W. / An adaptive performance modeling tool for GPU architectures. PPoPP'10 - Proceedings of the 2010 ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming. 2010. pp. 105-114 (Proceedings of the ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming, PPOPP).
@inproceedings{4110c091a015451a9eab9f16bb6fa2e5,
title = "An adaptive performance modeling tool for GPU architectures",
abstract = "This paper presents an analytical model to predict the performance of general-purpose applications on a GPU architecture. The model is designed to provide performance information to an auto-tuning compiler and assist it in narrowing down the search to the more promising implementations. It can also be incorporated into a tool to help programmers better assess the performance bottlenecks in their code. We analyze each GPU kernel and identify how the kernel exercises major GPU microarchitecture features. To identify the performance bottlenecks accurately, we introduce an abstract interpretation of a GPU kernel, work flow graph, based on which we estimate the execution time of a GPU kernel. We validated our performance model on the NVIDIA GPUs using CUDA (Compute Unified Device Architecture). For this purpose, we used data parallel benchmarks that stress different GPU microarchitecture events such as uncoalesced memory accesses, scratch-pad memory bank conflicts, and control flow divergence, which must be accurately modeled but represent challenges to the analytical performance models. The proposed model captures full system complexity and shows high accuracy in predicting the performance trends of different optimized kernel implementations. We also describe our approach to extracting the performance model automatically from a kernel code.",
keywords = "Analytical model, GPU, Parallel programming, Performance estimation",
author = "Baghsorkhi, {Sara S.} and Matthieu Delahaye and Patel, {Sanjay J.} and Gropp, {William D.} and Hwu, {Wen Mei W.}",
year = "2010",
month = "3",
day = "15",
doi = "10.1145/1693453.1693470",
language = "English (US)",
isbn = "9781605587080",
series = "Proceedings of the ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming, PPOPP",
pages = "105--114",
booktitle = "PPoPP'10 - Proceedings of the 2010 ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming",

}

TY - GEN

T1 - An adaptive performance modeling tool for GPU architectures

AU - Baghsorkhi, Sara S.

AU - Delahaye, Matthieu

AU - Patel, Sanjay J.

AU - Gropp, William D.

AU - Hwu, Wen Mei W.

PY - 2010/3/15

Y1 - 2010/3/15

N2 - This paper presents an analytical model to predict the performance of general-purpose applications on a GPU architecture. The model is designed to provide performance information to an auto-tuning compiler and assist it in narrowing down the search to the more promising implementations. It can also be incorporated into a tool to help programmers better assess the performance bottlenecks in their code. We analyze each GPU kernel and identify how the kernel exercises major GPU microarchitecture features. To identify the performance bottlenecks accurately, we introduce an abstract interpretation of a GPU kernel, work flow graph, based on which we estimate the execution time of a GPU kernel. We validated our performance model on the NVIDIA GPUs using CUDA (Compute Unified Device Architecture). For this purpose, we used data parallel benchmarks that stress different GPU microarchitecture events such as uncoalesced memory accesses, scratch-pad memory bank conflicts, and control flow divergence, which must be accurately modeled but represent challenges to the analytical performance models. The proposed model captures full system complexity and shows high accuracy in predicting the performance trends of different optimized kernel implementations. We also describe our approach to extracting the performance model automatically from a kernel code.

AB - This paper presents an analytical model to predict the performance of general-purpose applications on a GPU architecture. The model is designed to provide performance information to an auto-tuning compiler and assist it in narrowing down the search to the more promising implementations. It can also be incorporated into a tool to help programmers better assess the performance bottlenecks in their code. We analyze each GPU kernel and identify how the kernel exercises major GPU microarchitecture features. To identify the performance bottlenecks accurately, we introduce an abstract interpretation of a GPU kernel, work flow graph, based on which we estimate the execution time of a GPU kernel. We validated our performance model on the NVIDIA GPUs using CUDA (Compute Unified Device Architecture). For this purpose, we used data parallel benchmarks that stress different GPU microarchitecture events such as uncoalesced memory accesses, scratch-pad memory bank conflicts, and control flow divergence, which must be accurately modeled but represent challenges to the analytical performance models. The proposed model captures full system complexity and shows high accuracy in predicting the performance trends of different optimized kernel implementations. We also describe our approach to extracting the performance model automatically from a kernel code.

KW - Analytical model

KW - GPU

KW - Parallel programming

KW - Performance estimation

UR - http://www.scopus.com/inward/record.url?scp=77749337497&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=77749337497&partnerID=8YFLogxK

U2 - 10.1145/1693453.1693470

DO - 10.1145/1693453.1693470

M3 - Conference contribution

AN - SCOPUS:77749337497

SN - 9781605587080

T3 - Proceedings of the ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming, PPOPP

SP - 105

EP - 114

BT - PPoPP'10 - Proceedings of the 2010 ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming

ER -