Evaluation of data distribution patterns in distributed-memory machines

Edgar T. Kalns; Hong Xu; Lionel M. Ni

doi:10.1109/ICPP.1993.165

Evaluation of data distribution patterns in distributed-memory machines

Edgar T. Kalns, Hong Xu, Lionel M. Ni

Research output: Contribution to journal › Conference article published in journal › peer-review

6 Citations (Scopus)

Abstract

Determining an appropriate data distribution among different memories is critical to the performance of data-parallel programs on distributedmemory machines. By analyzing the computational load of data arrays and the communication complexity of various data movement operations in a program, this paper suggests a first-order cost model for determining a small set of appropriate data distribution patterns among many possible choices. A new data distribution specification, namely CYBLOCK, is proposed to enhance the expressiveness of data distribution specifications being proposed in High Performance Fortran. Cost analysis of two case studies: a linear system solver and a Purdue-set benchmark loop, are used to illustrate the proposed evaluation method. The model correctly predicts the relative performance of the case studies when implemented with various regular data distributions on an nCUBE-2 multicomputer.

Original language	English
Article number	4134204
Pages (from-to)	175-183
Number of pages	9
Journal	Proceedings of the International Conference on Parallel Processing
Volume	2
DOIs	https://doi.org/10.1109/ICPP.1993.165
Publication status	Published - 1993
Externally published	Yes
Event	1993 International Conference on Parallel Processing, ICPP 1993 - Syracuse, United States Duration: 16 Aug 1993 → 20 Aug 1993

Bibliographical note

Publisher Copyright:
© 1993 IEEE.

Access to Document

10.1109/ICPP.1993.165

Cite this

@article{7cdf8b2801ac40598f5685c41f71a691,

title = "Evaluation of data distribution patterns in distributed-memory machines",

abstract = "Determining an appropriate data distribution among different memories is critical to the performance of data-parallel programs on distributedmemory machines. By analyzing the computational load of data arrays and the communication complexity of various data movement operations in a program, this paper suggests a first-order cost model for determining a small set of appropriate data distribution patterns among many possible choices. A new data distribution specification, namely CYBLOCK, is proposed to enhance the expressiveness of data distribution specifications being proposed in High Performance Fortran. Cost analysis of two case studies: a linear system solver and a Purdue-set benchmark loop, are used to illustrate the proposed evaluation method. The model correctly predicts the relative performance of the case studies when implemented with various regular data distributions on an nCUBE-2 multicomputer.",

author = "Kalns, \{Edgar T.\} and Hong Xu and Ni, \{Lionel M.\}",

note = "Publisher Copyright: {\textcopyright} 1993 IEEE.; 1993 International Conference on Parallel Processing, ICPP 1993 ; Conference date: 16-08-1993 Through 20-08-1993",

year = "1993",

doi = "10.1109/ICPP.1993.165",

language = "English",

volume = "2",

pages = "175--183",

journal = "Proceedings of the International Conference on Parallel Processing",

issn = "0190-3918",

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

}

TY - JOUR

T1 - Evaluation of data distribution patterns in distributed-memory machines

AU - Kalns, Edgar T.

AU - Xu, Hong

AU - Ni, Lionel M.

PY - 1993

Y1 - 1993

N2 - Determining an appropriate data distribution among different memories is critical to the performance of data-parallel programs on distributedmemory machines. By analyzing the computational load of data arrays and the communication complexity of various data movement operations in a program, this paper suggests a first-order cost model for determining a small set of appropriate data distribution patterns among many possible choices. A new data distribution specification, namely CYBLOCK, is proposed to enhance the expressiveness of data distribution specifications being proposed in High Performance Fortran. Cost analysis of two case studies: a linear system solver and a Purdue-set benchmark loop, are used to illustrate the proposed evaluation method. The model correctly predicts the relative performance of the case studies when implemented with various regular data distributions on an nCUBE-2 multicomputer.

AB - Determining an appropriate data distribution among different memories is critical to the performance of data-parallel programs on distributedmemory machines. By analyzing the computational load of data arrays and the communication complexity of various data movement operations in a program, this paper suggests a first-order cost model for determining a small set of appropriate data distribution patterns among many possible choices. A new data distribution specification, namely CYBLOCK, is proposed to enhance the expressiveness of data distribution specifications being proposed in High Performance Fortran. Cost analysis of two case studies: a linear system solver and a Purdue-set benchmark loop, are used to illustrate the proposed evaluation method. The model correctly predicts the relative performance of the case studies when implemented with various regular data distributions on an nCUBE-2 multicomputer.

UR - http://www.scopus.com/inward/record.url?scp=33747682320&partnerID=8YFLogxK

U2 - 10.1109/ICPP.1993.165

DO - 10.1109/ICPP.1993.165

M3 - Conference article published in journal

AN - SCOPUS:33747682320

SN - 0190-3918

VL - 2

SP - 175

EP - 183

JO - Proceedings of the International Conference on Parallel Processing

JF - Proceedings of the International Conference on Parallel Processing

M1 - 4134204

T2 - 1993 International Conference on Parallel Processing, ICPP 1993

Y2 - 16 August 1993 through 20 August 1993

ER -

Evaluation of data distribution patterns in distributed-memory machines

Abstract

Bibliographical note

Access to Document

Other files and links

Fingerprint

Cite this