An asymmetric dstributed shared memory model for heterogeneous parallel systems

Isaac Gelado, Javier Cabezas, Nacho Navarro, John E. Stone, Sanjay Patel, Wen Mei W. Hwu

Research output: Contribution to journalArticle

Abstract

Heterogeneous computing combines general purpose CPUs with accelerators to efficiently execute both sequential control-intensive and data-parallel phases of applications. Existing programming models for heterogeneous computing rely on programmers to explicitly manage data transfers between the CPU system memory and accelerator memory. This paper presents a new programming model for heterogeneous computing, called Asymmetric Distributed Shared Memory (ADSM), that maintains a shared logical memory space for CPUs to access objects in the accelerator physical memory but not vice versa. The asymmetry allows light-weight implementations that avoid common pitfalls of symmetrical distributed shared memory systems. ADSM allows programmers to assign data objects to performance critical methods. When a method is selected for accelerator execution, its associated data objects are allocated within the shared logical memory space, which is hosted in the accelerator physical memory and transparently accessible by the methods executed on CPUs. We argue that ADSM reduces programming efforts for heterogeneous computing systems and enhances application portability. We present a software implementation of ADSM, called GMAC, on top of CUDA in a GNU/Linux environment. We show that applications written in ADSM and running on top of GMAC achieve performance comparable to their counterparts using programmermanaged data transfers. This paper presents the GMAC system and evaluates different design choices.We further suggest additional architectural support that will likely allow GMAC to achieve higher application performance than the current CUDA model.

Original languageEnglish (US)
Pages (from-to)347-358
Number of pages12
JournalACM SIGPLAN Notices
Volume45
Issue number3
StatePublished - Mar 1 2010

Fingerprint

Computer systems
Data storage equipment
Particle accelerators
Program processors
Data transfer
Computer programming

Keywords

  • Asymmetric distributed shared memory
  • Data-centric programming Models
  • Heterogeneous systems

ASJC Scopus subject areas

  • Computer Science(all)

Cite this

An asymmetric dstributed shared memory model for heterogeneous parallel systems. / Gelado, Isaac; Cabezas, Javier; Navarro, Nacho; Stone, John E.; Patel, Sanjay; Hwu, Wen Mei W.

In: ACM SIGPLAN Notices, Vol. 45, No. 3, 01.03.2010, p. 347-358.

Research output: Contribution to journalArticle

Gelado, I, Cabezas, J, Navarro, N, Stone, JE, Patel, S & Hwu, WMW 2010, 'An asymmetric dstributed shared memory model for heterogeneous parallel systems', ACM SIGPLAN Notices, vol. 45, no. 3, pp. 347-358.
Gelado, Isaac ; Cabezas, Javier ; Navarro, Nacho ; Stone, John E. ; Patel, Sanjay ; Hwu, Wen Mei W. / An asymmetric dstributed shared memory model for heterogeneous parallel systems. In: ACM SIGPLAN Notices. 2010 ; Vol. 45, No. 3. pp. 347-358.
@article{90161feafb764b8091d45a8aaf2b0264,
title = "An asymmetric dstributed shared memory model for heterogeneous parallel systems",
abstract = "Heterogeneous computing combines general purpose CPUs with accelerators to efficiently execute both sequential control-intensive and data-parallel phases of applications. Existing programming models for heterogeneous computing rely on programmers to explicitly manage data transfers between the CPU system memory and accelerator memory. This paper presents a new programming model for heterogeneous computing, called Asymmetric Distributed Shared Memory (ADSM), that maintains a shared logical memory space for CPUs to access objects in the accelerator physical memory but not vice versa. The asymmetry allows light-weight implementations that avoid common pitfalls of symmetrical distributed shared memory systems. ADSM allows programmers to assign data objects to performance critical methods. When a method is selected for accelerator execution, its associated data objects are allocated within the shared logical memory space, which is hosted in the accelerator physical memory and transparently accessible by the methods executed on CPUs. We argue that ADSM reduces programming efforts for heterogeneous computing systems and enhances application portability. We present a software implementation of ADSM, called GMAC, on top of CUDA in a GNU/Linux environment. We show that applications written in ADSM and running on top of GMAC achieve performance comparable to their counterparts using programmermanaged data transfers. This paper presents the GMAC system and evaluates different design choices.We further suggest additional architectural support that will likely allow GMAC to achieve higher application performance than the current CUDA model.",
keywords = "Asymmetric distributed shared memory, Data-centric programming Models, Heterogeneous systems",
author = "Isaac Gelado and Javier Cabezas and Nacho Navarro and Stone, {John E.} and Sanjay Patel and Hwu, {Wen Mei W.}",
year = "2010",
month = "3",
day = "1",
language = "English (US)",
volume = "45",
pages = "347--358",
journal = "ACM SIGPLAN Notices",
issn = "1523-2867",
publisher = "Association for Computing Machinery (ACM)",
number = "3",

}

TY - JOUR

T1 - An asymmetric dstributed shared memory model for heterogeneous parallel systems

AU - Gelado, Isaac

AU - Cabezas, Javier

AU - Navarro, Nacho

AU - Stone, John E.

AU - Patel, Sanjay

AU - Hwu, Wen Mei W.

PY - 2010/3/1

Y1 - 2010/3/1

N2 - Heterogeneous computing combines general purpose CPUs with accelerators to efficiently execute both sequential control-intensive and data-parallel phases of applications. Existing programming models for heterogeneous computing rely on programmers to explicitly manage data transfers between the CPU system memory and accelerator memory. This paper presents a new programming model for heterogeneous computing, called Asymmetric Distributed Shared Memory (ADSM), that maintains a shared logical memory space for CPUs to access objects in the accelerator physical memory but not vice versa. The asymmetry allows light-weight implementations that avoid common pitfalls of symmetrical distributed shared memory systems. ADSM allows programmers to assign data objects to performance critical methods. When a method is selected for accelerator execution, its associated data objects are allocated within the shared logical memory space, which is hosted in the accelerator physical memory and transparently accessible by the methods executed on CPUs. We argue that ADSM reduces programming efforts for heterogeneous computing systems and enhances application portability. We present a software implementation of ADSM, called GMAC, on top of CUDA in a GNU/Linux environment. We show that applications written in ADSM and running on top of GMAC achieve performance comparable to their counterparts using programmermanaged data transfers. This paper presents the GMAC system and evaluates different design choices.We further suggest additional architectural support that will likely allow GMAC to achieve higher application performance than the current CUDA model.

AB - Heterogeneous computing combines general purpose CPUs with accelerators to efficiently execute both sequential control-intensive and data-parallel phases of applications. Existing programming models for heterogeneous computing rely on programmers to explicitly manage data transfers between the CPU system memory and accelerator memory. This paper presents a new programming model for heterogeneous computing, called Asymmetric Distributed Shared Memory (ADSM), that maintains a shared logical memory space for CPUs to access objects in the accelerator physical memory but not vice versa. The asymmetry allows light-weight implementations that avoid common pitfalls of symmetrical distributed shared memory systems. ADSM allows programmers to assign data objects to performance critical methods. When a method is selected for accelerator execution, its associated data objects are allocated within the shared logical memory space, which is hosted in the accelerator physical memory and transparently accessible by the methods executed on CPUs. We argue that ADSM reduces programming efforts for heterogeneous computing systems and enhances application portability. We present a software implementation of ADSM, called GMAC, on top of CUDA in a GNU/Linux environment. We show that applications written in ADSM and running on top of GMAC achieve performance comparable to their counterparts using programmermanaged data transfers. This paper presents the GMAC system and evaluates different design choices.We further suggest additional architectural support that will likely allow GMAC to achieve higher application performance than the current CUDA model.

KW - Asymmetric distributed shared memory

KW - Data-centric programming Models

KW - Heterogeneous systems

UR - http://www.scopus.com/inward/record.url?scp=77949851572&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=77949851572&partnerID=8YFLogxK

M3 - Article

AN - SCOPUS:77949851572

VL - 45

SP - 347

EP - 358

JO - ACM SIGPLAN Notices

JF - ACM SIGPLAN Notices

SN - 1523-2867

IS - 3

ER -