Performance models for MPI collective communications with network contention

Hyacinthe Nzigou Mamadou, Takeshi Nanri, Kazuaki Murakami

研究成果: ジャーナルへの寄稿記事

抄録

The paper presents a novel approach to estimate the performance of MPI collective communications. Our objective is to help researchers to make appropriate decisions on their message-passing applications. For each collective communication, we attempt to apply LogGP and P-LogP standard point-to-point models. The resulted models are compared with the empirical data in order to identify the most suitable for performance characterization of collective operations. For the communications on large clusters with large size messages, the network contention problem can significantly affect the performance. Hence, to reduce the relative gap between the prediction and the measured runtime, the contention issue is also modeled, by a queuing theory analysis method, and taken in account with the total performance estimation. The experiments performed on a cluster which consists of 64 processors interconnected by Gigabit Ethernet network show encouraging results. For any collective operation, given a number of processors and a range of message sizes, there is at least one model that predicts the performance precisely. We could achieve a gap between the predicted and the measured run-time around 15%. Thus, by handling the contention problem, we could reduce around 80% of the relative gap.

元の言語英語
ページ(範囲)1015-1024
ページ数10
ジャーナルIEICE Transactions on Communications
E91-B
発行部数4
DOI
出版物ステータス出版済み - 1 1 2008

Fingerprint

Communication
Message passing
Ethernet
Experiments

All Science Journal Classification (ASJC) codes

  • Software
  • Computer Networks and Communications
  • Electrical and Electronic Engineering

これを引用

Performance models for MPI collective communications with network contention. / Mamadou, Hyacinthe Nzigou; Nanri, Takeshi; Murakami, Kazuaki.

:: IEICE Transactions on Communications, 巻 E91-B, 番号 4, 01.01.2008, p. 1015-1024.

研究成果: ジャーナルへの寄稿記事

Mamadou, Hyacinthe Nzigou ; Nanri, Takeshi ; Murakami, Kazuaki. / Performance models for MPI collective communications with network contention. :: IEICE Transactions on Communications. 2008 ; 巻 E91-B, 番号 4. pp. 1015-1024.
@article{d2a6457945d1441588a048730f175185,
title = "Performance models for MPI collective communications with network contention",
abstract = "The paper presents a novel approach to estimate the performance of MPI collective communications. Our objective is to help researchers to make appropriate decisions on their message-passing applications. For each collective communication, we attempt to apply LogGP and P-LogP standard point-to-point models. The resulted models are compared with the empirical data in order to identify the most suitable for performance characterization of collective operations. For the communications on large clusters with large size messages, the network contention problem can significantly affect the performance. Hence, to reduce the relative gap between the prediction and the measured runtime, the contention issue is also modeled, by a queuing theory analysis method, and taken in account with the total performance estimation. The experiments performed on a cluster which consists of 64 processors interconnected by Gigabit Ethernet network show encouraging results. For any collective operation, given a number of processors and a range of message sizes, there is at least one model that predicts the performance precisely. We could achieve a gap between the predicted and the measured run-time around 15{\%}. Thus, by handling the contention problem, we could reduce around 80{\%} of the relative gap.",
author = "Mamadou, {Hyacinthe Nzigou} and Takeshi Nanri and Kazuaki Murakami",
year = "2008",
month = "1",
day = "1",
doi = "10.1093/ietcom/e91-b.4.1015",
language = "English",
volume = "E91-B",
pages = "1015--1024",
journal = "IEICE Transactions on Communications",
issn = "0916-8516",
publisher = "Maruzen Co., Ltd/Maruzen Kabushikikaisha",
number = "4",

}

TY - JOUR

T1 - Performance models for MPI collective communications with network contention

AU - Mamadou, Hyacinthe Nzigou

AU - Nanri, Takeshi

AU - Murakami, Kazuaki

PY - 2008/1/1

Y1 - 2008/1/1

N2 - The paper presents a novel approach to estimate the performance of MPI collective communications. Our objective is to help researchers to make appropriate decisions on their message-passing applications. For each collective communication, we attempt to apply LogGP and P-LogP standard point-to-point models. The resulted models are compared with the empirical data in order to identify the most suitable for performance characterization of collective operations. For the communications on large clusters with large size messages, the network contention problem can significantly affect the performance. Hence, to reduce the relative gap between the prediction and the measured runtime, the contention issue is also modeled, by a queuing theory analysis method, and taken in account with the total performance estimation. The experiments performed on a cluster which consists of 64 processors interconnected by Gigabit Ethernet network show encouraging results. For any collective operation, given a number of processors and a range of message sizes, there is at least one model that predicts the performance precisely. We could achieve a gap between the predicted and the measured run-time around 15%. Thus, by handling the contention problem, we could reduce around 80% of the relative gap.

AB - The paper presents a novel approach to estimate the performance of MPI collective communications. Our objective is to help researchers to make appropriate decisions on their message-passing applications. For each collective communication, we attempt to apply LogGP and P-LogP standard point-to-point models. The resulted models are compared with the empirical data in order to identify the most suitable for performance characterization of collective operations. For the communications on large clusters with large size messages, the network contention problem can significantly affect the performance. Hence, to reduce the relative gap between the prediction and the measured runtime, the contention issue is also modeled, by a queuing theory analysis method, and taken in account with the total performance estimation. The experiments performed on a cluster which consists of 64 processors interconnected by Gigabit Ethernet network show encouraging results. For any collective operation, given a number of processors and a range of message sizes, there is at least one model that predicts the performance precisely. We could achieve a gap between the predicted and the measured run-time around 15%. Thus, by handling the contention problem, we could reduce around 80% of the relative gap.

UR - http://www.scopus.com/inward/record.url?scp=67651047265&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=67651047265&partnerID=8YFLogxK

U2 - 10.1093/ietcom/e91-b.4.1015

DO - 10.1093/ietcom/e91-b.4.1015

M3 - Article

AN - SCOPUS:67651047265

VL - E91-B

SP - 1015

EP - 1024

JO - IEICE Transactions on Communications

JF - IEICE Transactions on Communications

SN - 0916-8516

IS - 4

ER -