• 検索結果がありません。

Japan Advanced Institute of Science and Technology

N/A
N/A
Protected

Academic year: 2021

シェア "Japan Advanced Institute of Science and Technology"

Copied!
4
0
0

読み込み中.... (全文を見る)

全文

(1)

Japan Advanced Institute of Science and Technology

JAIST Repository

https://dspace.jaist.ac.jp/

Title

ワークステーションクラスタによる高速並列処理に関

する研究

Author(s)

奥野, 弘之

Citation

Issue Date

1998‑03

Type

Thesis or Dissertation

Text version

author

URL

http://hdl.handle.net/10119/1123

Rights

Description

Supervisor:堀口 進, 情報科学研究科, 修士

(2)

Hiroyuki Okuno

Scho ol of Information Science,

Japan Advanced Institute of Science and Technology

February 13, 1998

Keywords: WorkstationCluster,Performance Evaluation,Message PassingLibrary,

ParallelProcessing.

1 Introduction

Workstation cluster systems are very attractive for distributed and parallel processing,

sinceitiseasytoobtainparallelprocessingenvironment. Butsuchasworkstationclusters

using slownetworklayerhave aproblem that can not obtainperformances inprop ortion

to numb erof workstations.

In this paper, wediscuss ab out performance of messagepassing library such asPVM

and MPI, and propose hierarchical broadcast, a technique to to improvement PVM's

broadcast. Next we examine some applications on workstation cluster and discuss the

performance of workstation cluster. From the result, we discuss ab out applications run-

ning onworkstation cluster. Finally, we propose eectiveparallel processing method on

workstation clustersuchas sharedvalue system onPVM.

2 Workstation cluster and Massage passing libraly

Recently,manystudiesand approachesareexaminedtoimproveperformance ofworksta-

tion cluster. From hardware improvment,high-sp eed networklayers, suchas 100Base-T,

FDDI, ATM and Myrinet are deveropp ed. These hardwares achievesfaster workstation

cluster'smessagepassing. But,theircostofintro duction ishigh. Sowecannotuseeasily.

On the other hand, software approaches are performed to extend existed message

passinglibrary,suchasPVMand MPItoreducetheiroverheadonmessagepassingsetup

time. TPVM and LPVM are proposed this approach. pardi-PVM achieved fast message

passing using Adsmith proposed to provide new interfaces to PVM such as distributed

shared memory-systems.

Copyright c

1998byHiroyukiOkuno

(3)

interface.

3 Message passing performance on workstation clus-

ter

We measured message passing function's performance such as barrier synchronization,

latency, communication speed, and broadcast speed on workstation cluster, to propose

eectiveprocessingmethodusingworkstationclusterconstructedstandard PVMorMPI,

and 10Base-Tnetworks. This measurementisimp ortant toanalyze b ottlenecksonwork-

station clustr'smessage passing performance.

As the result, we found that PVM's message passing p erformance are good in case

of using TCP connection. MPI is better than PVM about all messagepassing functions.

From these message passing performance, we propose hierarchical broadcast to improve

PVM's broadcast speed, and conrmed their eectiveness. And we propose selectable

broadcast that improveshierachicalbroadcast.

4 Performance of Application programs on Work-

station Cluster

Inthis chapter,weexamineperformance ofworkstationcluster byrunningsome applica-

tions,suchassortingprogram, N-queen problemand TSP(Traveling SalesmanProblem).

We found sorting program is unsuitable on workstation cluster because sort program

communicatewith large data due to increasing message passing time. N-queen problem

indicated that if non-communication parallel processing model could load tasks equally,

performance will increase linearly. In TSP,weexamined and discussed some parallel ex-

ecutionmethod,suchas distributed mo del, master-slavemodel, and co-operationmodel.

We gained co-operation model achieves linear performance increase. And wefound that

TSP is not aected by message passing,so suitablefor workstationclusters. As aresult,

we proposedshared value system tomake availableshared valueusing PVM.

5 Conclusion

Thispaperevaluatedworkstationclusterperformanceanddiscussedtheircharacters. We

found workstationcluster usingslownetworklayersuchas 10Base-T,arenot appropriate

tocommunicatlargedata. Fromthat'sresult,weproposedhierarchicalbroadcastonPVM

and conrmedimprovementofbroad castperformance. Wediscussedabout performance

of applicationsrunningon workstationcluster. As the result,wegaind thatless message

passing model and less message data passing model is suitable for workstation cluster.

And we proposed and discussed about shared value system which makes shared value

system on PVM.

(4)

tem prop osed inthis pap er and evaluation their performance.

参照

関連したドキュメント

As you are well aware, the technology has advanced rapidly for TIG welding machine, not only the basic performance and quality but also the welding quality have improved greatly

utilized for constructing integration rules for the evaluation of weakly and strongly singular integrals also defined in the Hadamard finite part sense, in one or two dimen- sions

More pre- cisely, the dual variants of Differentiation VII and Completion for corepresen- tations are described and (following the scheme of [12] for ordinary posets) the

redex search token passing reduction diagram rewriting. computation

Extended cubical sets (with connections and interchanges) are presheaves on a ground category, the extended cubical site K, corresponding to the (augmented) simplicial site,

In Section 3, we show that the clique- width is unbounded in any superfactorial class of graphs, and in Section 4, we prove that the clique-width is bounded in any hereditary

Arnold This paper deals with recent applications of fractional calculus to dynamical sys- tems in control theory, electrical circuits with fractance, generalized voltage di-

Arnold This paper deals with recent applications of fractional calculus to dynamical sys- tems in control theory, electrical circuits with fractance, generalized voltage di-