本站原创撰写论文时该注意些什麽
俞征武订於2002/6/11.
大纲:
前言
论文基本格式编排
论文基本架构
正确的写作方式
常犯的英文文法错误
论文之引用
论文之
零前言
一般科技性质的学术论文有其一定的格式.初写论文的学生因较不熟悉,故常有很大的困扰.所以写了一些提示,希望提醒自己或对我的学生有所帮助.
论文基本格式
写作论文一般每个Journal有自己的规定.但整篇论文必须有统一的格式(format)却是一致遵守的习惯.以下列出常见的规矩:
字体与字型:
font为timesnewroman,字体大小为12点字.
每个section中的title其字体大小为14(小section)或16点字(大section).
Paper的标题(title)的字体大小为18点字,而且需置中.
段落:
文章中每段落左右需切齐.
每一个section前需空一行.
每一个paragraph前需缩排.
空格:
逗号後面接一个空格,如:等inthepaper,hence等.
在两个word中只用一个空白分隔,不可使用两个空白以上.
斜体:
论文中所有变数都要斜体,如:Whenx>,y,thealgorithm等.
论文中自行定义的名词在第一次出现时需斜体,之後就不再用斜体.
论文中所有constant不要斜体,如:Whenx1>,y2,thealgorithm等.其中1,2虽然在下标一样不需斜体.
其他:
缩写要用小括号,如randomaccesemory(RAM).
使用三个小点(等)代表省略,如x1,x2,等,xn,
参考论文之引用需用中括号,如:In[16],wehepresented等.
利用大括号代表集合(set),如:LetS等于{1,2,3,4,5},等.
整篇文章不用粗体(bold).
Figure的说明在正下方,Tables的说明在正上方.
定理证明结束的最後一行的最右方写入"".
行间的距离请用doublespace,虽然不同的刊物有不同的要求.
论文基本架构
以下利用一个例子来说明论文的基本架构.
论文标题(title)
(论文标题范例)需简短并指出论文的特色及贡献
【范例】
IrregularRedistributionSchedulingbyPartitioningMessages
Chun-IChen,ChangWuYu,Ching-HsienHsu,Kun-MingYu,andC.-K.Liang
DepartmentofComputerScienceandInformationEngineering
ChungHuaUniversity,Hsinchu,Taiwan300,R.O.C.
{cwyu,chh,yu,ckliang}@chu.edu.tw
摘 要(Abstract)
撰写摘 要可用一句话点明研究问题的领域及重要性.用两三句话定义研究问题及动机.再用两三句话说明研究具体的成果,最後一句提及主要贡献.
【范例】
Abstract
Dynamicdataredistributionenhancesdatalocalityandimprovesalgorithmperformancefornumerousscientificproblemsondistributedmemorymulti-puterssystems.RegulardatadistributiontypicallyemploysBLOCK,CYCLIC,orBLOCK-CYCLIC(c)tospecifyarraydeposition.Conversely,anirregulardistributionspecifiesanunevenarraydistributionbasedonuser-definedfunctions.Performingdataredistributionconsistsoffourcosts:indexputationalcost,scheduleputationalcost,messagepacking/unpackingcost,anddatatranercost.Previousresultocusonreducingtheformerthreecosts.However,inirregularredistribution,messageswithvaryingsizesaretranittedinthesamemunicationstep.Therefore,thelargestsizedmessagesinthesamemunicationstepdominatethedatatranertimerequiredforthiunicationstep.Thisworkpresentsanefficientalgorithmtopartitionlargemessagesintomultipleallonesandschedulesthembyusingtheminimumnumberofstepswithoutmunicationcontentionand,indoingso,reducingtheoverallredistributiontime.Whenthenumberofprocessorsorthemaximumdegreeoftheredistributiongraphincreasesortheselectedsizeofmessagesiedium,theproposedalgorithmcansignificantlyreducetheoverallredistributiontimeto52%.Moreover,theproposedalgorithmcanbeappliedtoarbitrarydataredistributionwhileslightlyincreasingthemunicationschedulingtime.
Keywords:dataredistribution,scheduling,edgecoloring,bipartitegraphs,multi-graphs
1.简介(Introduction)
1.1首先界定您的研究问题的研究领域并说明此领域在广度上的重要性.
【范例】
Parallelputingsystemshebeenextensivelyadoptedtoresolveplexscientificproblemsefficiently.Whenprocessinariousphasesofapplications,parallelsystemsnormallyexploitdatadistributionschemestobalancethesystemloadandyieldabetterperformance.Generally,datadistributionsareeitherregularorirregular.RegulardatadistributiontypicallyemploysBLOCK,CYCLIC,orBLOCK-CYCLIC(c)tospecifyarraydeposition[14,15].Conversely,anirregulardistributionspecifiesanunevenlyarraydistributionbasedonuser-definedfunctions.Forinstance,HighPerformanceFortranversion2(HPF2)providesageneralizedblockdistribution(GEN_BLOCK)[19,20]format,allowingunequallysizedmessages(ordatasegments)ofanarraytobemappedontoprocessors.GEN_BLOCKpesthewayforprocessorswithvaryingputationalabilitiestohandleappropriatelysizeddata.
Arrayredistributioniscrucialforsystemperformancebecauseaspecificarraydistributionmaybeappropriateforthecurrentphase,butinpatibleforthesubsequentone.Manyparallelprogramminglanguagesthussupportrun-timeprimitiveorrearrangingaprogram'sarraydistribution.Thereforedevelopingefficientalgorithmorarrayredistributionisessentialfordesigningdistributedmemorypilerorthoselanguages.Whilearrayredistributionisperformedatruntime,atrade-offoccursbetweentheefficiencyofthenewdatarearrangementfortheingphaseandthecostofarrayredistributingamongprocessors.
1.2在此研究领域上有何研究问题需要解决可和先人的成果比较以突显本论文的创意.
【范例】
Performingdataredistributionconsistsoffourcosts:indexputationalcostTi,scheduleputationalcostTs,messagepacking/unpackingcostTpanddatatranercost.Theindexandscheduleputationsareexecutedinpliertime,withtheremaininginruntime.Thedatatranercostforeachmunicationstepconsistsofstart-upcostTuandtranissioncostTt.Lettheunittranissiontime(denotethecostoftranerringamessageofunitlength.ThetotalnumberofmunicationstepsisdenotedbyC.TotalredistributiontimeequalsTi+Ts+,where等于Max{d1,d2,d3.,dk}anddjrepresentsthesizeofmessagescheduledinithmunicationstepforj等于1tok.
Previousresultocusonreducingtheformerthreecosts(i.e.,Ti,Ts,andTu).Inirregularredistribution,messagesofvaryingsizesarescheduledinthesamemunicationstep.Therefore,thelargestsizeofmessageinthesamemunicationstepdominatesthedatatranertimerequiredforthiunicationstep.
1.3此问题期望如何被解决:如降低演算法的时间复杂度或减少电源的耗费等尽可能利用量化说明.
【范例】
Basedonthefact,thisworkpresentsanefficientalgorithmtopartitionlargemessagesintomultipleallonesandschedulesthembyusingtheminimumnumberofstepswithoutmunicationcontentionand,indoingso,reducingtheoverallredistributiontime.
1.4.说明研究动机:如果此问题没被解决或是充分了解,会有多大的负面的问题明确地直接地说明本研究的目标.
1.6.您的研究将利用何种技巧达成预期的成果.
【范例】
Specifically,theminimumvalueofTs,andCarederived,alongwiththevalueofmireducedbyshorteningtherequiredmunicationtimeforeachmunicationstep.
1.7.说明预期的结果.
1.8预期此研究对整个领域的贡献(广度)
【范例】
Whenthenumberofprocessorsorthemaximumdegreeoftheredistributiongraphincreasesortheselectedsizeofmessagesiedium,theproposedalgorithmcansignificantlyreducestheoverallredistributiontimeto52%.Moreover,theproposedalgorithmcanbeappliedtoarbitrarydataredistributionwhileslightlyincreasingthemunicationschedulingtime.
紧接着是整篇论文的大纲:
【范例】
Therestofthepaperisanizedaollows.Section2presentsnecessarydefinitionsandnotations.Next,Section3describesthebasicgraphmodelalongwithrelatedwork.ThemaincontributionofthepaperisshowninSection4.WealsoconductsimulationsinSection5todemonstratethemeritsofouralgorithm.Finally,Section6concludesthepaper.
2.前人成果介绍(survey)[16,18]Routingisanimprovementtothetable-drivenanddistance-vectorbasedDSDValgorithm.WithDSDV(Destination-SequencedDistance-Vector)Routing[17],everymobilenodemaintainsaroutingtablerecordingallthepossibledestinationsandnumberofhopstoeachdestination.Inordertomaintainroutingtableconsistency,itrequiresnodestoperiodicallybroadcastroutingupdatesthroughoutthework.
【范例二】
Techniqueorregulararrayredistributioncanbeclassifiedintotwogroups:themunicationsetsidentificationandmunicationoptimizations.TheformerincludesthePITFALLS[17]andtheScaLAPACK[16]methodorindexsetsgeneration.Parketal.[14]devisedalgorithmorBLOCK-CYCLICdataredistributionbetweenprocessorsets.Dongarraetal.[15]proposedalgorithmicredistributionmethodorBLOCK-CYCLICdepositions.Zapataetal.[1]designedparallelsparseredistributioncodeforBLOCK-CYCLICdataredistributionbasedonCRSstructure.Also,theGeneralizedBasic-CycleCalculationmethodwaspresentedin[3].Techniqueormunicationoptimizationsprovidedifferentapproachestoreducethemunicationoverheadsinaredistributionoperation.
3.定义或背景(notationsanddefinitions)
【范例】
AnydataredistributioncanberepresentedbyabipartitegraphG等于(S,T,E),calledaredistributiongraph.WhereSdenotessourceprocessorset,Tdenotesdestinationprocessorset,andeachedgedenotesamessagerequiredtobesent.Forexample,aBlock-Cyclic(x)toBlock-Cyclic(y)dataredistributionfromPprocessorstoQprocessors(denotedbyBC(x,y,P,Q))canbemodeledbyabipartitegraphGBC(x,y,P,Q)等于(S,T,E)whereS等于{s0,s1,等,s(s(-1}(T等于{t0,t1,等,t(t(-1})denotesthesourceprocessorset{p0,p1,等,p(s(-1}(destinationprocessorset{p0,p1,等,p(t(-1})andwehe(si,tj)(Ewithweightwifsourceprocessorpihastosendtheamountofwdataelementstodestinationprocessorpj.Forsimplicity,weuseBC(x,y,P)todenoteBC(x,y,P,P).
4.主要成果(mainresults)ccordingly,foragivendataredistributionproblem,aconflict-freeschedulingwiththeminimumnumberofmunicationstepscanbeobtainedbycoloringtheedgesofthecorrespondingredistributiongraphG.WhenGisbipartite,itiswellknownthat(((G)等于((G)[22].Asaresult,theminimumnumberofrequiredmunicationstepsequalsthemaximumdegree(ofthegivendistributiongraphG.
Previousworkisequivalenttofindingoutanedgecolorings{E1,E2,E3,等,E(}ofGsothat(i.e.,thedatatranertime)canbedecreased.Tothebestofourknowledge,itisstillopentodeviseanefficientalgorithmtominimizebothoftheoverallredistributiontimeandmunicationsteps.
Unlikeexistingalgorithms,themainideabehindourworkistopartitionlargedatasegmentsintomultiplealldatasegmentsandproperlyschedulethemindifferentmunicationstepswithoutincreasingthenumberoftotalmunicationsteps.
4.2Anexample
【范例】
Forexample,Figure6depictsaredistributiongraphwiththemaximumdegree(等于4.
Figure6.Aredistributiongraphwith(等于4.
Weneedfourmunicationsteporthisdataredistributionsince(((G)等于((G)等于4.Inaddition,theoverallcostofthecorrespondingschedulingis38(SeeTable1).
Table1.TheschedulingcorrespondstotheedgecoloringinFigure4.
Step1(red)2(yellow)3(green)4(purple)TotalCost18631138
NotethatthetimecostofStep1(coloredinred)isdominatedbythedatasegment(with18dataelements)fromP0toQ0.Supposethatweevenlypartitionthesegmentintotwodatasegments(with9and9dataelementsrespectively)andtranitthemindifferentsteps,thenthetimerequiredforStep1isreducedto10(dominatedbythedatasegmentfromP3toQ3).Notethatthedatapartitionaddsanedge(P0,Q0)intheoriginalredistributiongraph.Similarly,wecanpartitionanylargedatasegmentintomultiplealldatasegmentifthemaximumdegreeoftheresultingredistributiongraphremainsunchanging.Afterseveraldatapartitions,theoverallmunicationcostcanbereducedto29andthenumberofrequiredmunicationstepisstillminimized(seeFigure7andTable2).
Figure7.Theresultingredistributiongraphafterpartitioninglongdatasegments.
Table2.Theschedulingafterpartitionlongdatamunications.
Step1(red)2(yellow)3(green)4(purple)TotalCost995629
较正式的成果说明.
【范例】
Thealgorithmoftheselectionstepisshownaollows.
AlgorithmSelection()
Input:AredistributiongraphG等于(S,T,E)withmaximumdegree.
Output:AredistributiongraphG等于(S,T,E(D)withmaximumdegree,whereDrepresentsthosedummyedgesaddedinthealgorithm.
Step1.Selecttheedgeek等于(si,tj)fromEsuchthatthevaluewk/(1+vk)isthelargestanddG(si)<,anddG(tj)<,,wherevkdenotesthenumberofaddeddummyedgewiththesameendpointsofek.Ifnosuchedgeexists,terminatethisalgorithm.
Step2.Addadummyedgeek'等于(si,tj)toDandsetvk等于vk+1.
Step3.GotoStep1.
ThetimeplexityofSelectionisO(mlogm),wheremisthesizeofedgesetoftheinputredistributiongraph.
4.4Complexity
5.理论上的证明(simulationresults).实验结果GEN_BLOCKdistribution.GivenanirregulararrayredistributiononA[1:N]overPprocessors,theeragesizeofdatablocksisN/P.LetTb(Ta)denotethetotalredistributioncostwithout(with)applyingouralgorithm.ThereductionratioRequals(Tb-Ta)/Tb.Moreover,let{E1,E2,E3,等,E(}ofGdenotetheoutputofSchedulingstep.WealsodefineCi等于.Asaresult,theoverallredistributiontimeisboundedbyB等于sincetheproposedalgorithmdoesnotselectmaximum-degreeedgeorfurtherpartition.Otherwise,therequiredmunicationstepwillbeincreased.
Tothoroughlyevaluatehowouralgorithmaffectsthedatatranercost,oursimulationsconsiderdifferentscenarios.Eachdatapointinthefollowingfiguresrepresentsanerageofatleast10000runsineachdifferentscenario.
Thefirstscenarioassumesthatthesizeofdataarrayiixed,i.e.,N等于100,thenumberofprocessorsrangefrom4,8,16,32,64,to128,thesizeofdatablocksisrandomlyselectedbetween1and50.InFigure12,thevalueofTbdrasticallyraisesasthenumberofprocessorsincreases.However,afterapplyingouralgorithm,theoveralldistributiontimeTaoothlyraisesasthenumberofprocessorsincreases.NotethattheBvaluedropsasthenumberofprocessorincreaseduetothedecreaseoftheeragevaluesofdataelementsinasinglemunication.Inshort,whenthenumberofprocessorsincreases,thereductionratioRraisesifapplyingourpartitionalgorithm.
Figure12.SimulationresultsofScenarioI.
Thesecondscenarioassumesthatthenumberofprocessorsiixed,i.e.,P等于32,thesizeofdataarrayNequals1600,3200,6400,9600,or12800,andthesizeofdatablocksisrandomlyselectedbetween1and2((N/P).AsshowninFigure13,thevaluesofTa,Tb,andBraisesasthesizeofdataarrayNincreasesduetotheincreaseoftheeragenumberofdataelementsinasinglemunication.However,thereductionratiostaysabout52%byapplyingourpartitionalgorithm,evenwiththelargesizeofdataarray.
Figure13.SimulationresultsofScenarioII.
结论
简单摘 要整篇论文的贡献及使用的技巧
此研究对该领域有何启发
贡献有何处可扩展
WereyourresultsexpectedIfnot,whynot
Whatgeneralizationsorclaimsareyoumakingaboutyourresults
Doyourresultscontradictorsupportotherexperimentalresults
Dotheysuggestotherobservationsorexperimentswhichcouldbedonetoconfirm,refute,orextendyourresults
Doyourresultssupportorcontradictexistingtheory
DoyourresultssuggestthatmodificationsorextensionsneedtobemadetoexistingtheoryWhatarethey
Couldyourresultsleadtoanypracticalapplications
Stresshowtheresultsinthisstudyconfirmyourengineering/Scientificmotivations(specificandgeneral)and,ultimately,yourreader'sinterests(i.e.Engineering/Scientificneed).
【范例】
Wehepresentedanefficientalgorithmtoreducetheoverallredistributiontimebyapplyingdatapartition.Simulationresultsindicatesthatwhenthenumberofprocessorsorthemaximumdegreeoftheredistributiongraphincreasesortheselectedsizeofdatablocksisappropriate,ouralgorithmeffectivelyreducetheoverallredistributiontime.Infuture,wetrytoestimatethereductionratioprecisely.Wealsobelievethatthetechniquesdevelopedinthestudycanbeappliedtoresolveotherschedulingproblemsindistributionsystems.
Zapata,"SparseMatrixBlock-CyclicRedistribution,"ProceedingofIEEEInt'l.ParallelProcessingSymposium(IPPS'99),SanJuan,PuertoRico,April1999.FredericDesprez,JackDongarraandAntoinePetitet,"SchedulingBlock-CyclicDataredistribution,"IEEETrans.onPDS,vol.9,no.2,pp.192-205,Feb.1998.
C.-HHsu,S.-WBai,Y.-CChungandC.-SYang,"AGeneralizedBasic-CycleCalculationMethodforEfficientArrayRedistribution,"IEEETPDS,vol.11,no.12,pp.1201-1216,Dec.2000.
C.-HHsu,Dong-LinYang,Yeh-ChingChungandChyi-RenDow,"AGeneralizedProcessorMappingTechniqueforArrayRedistribution,"IEEETransactionsonParallelandDistributedSystems,vol.12,vol.7,pp.743-757,July2001.
MinyiGuo,"CommunicationGenerationforIrregularCodes,"TheJournalofSuperputing,vol.25,no.3,pp.199-214,2003.
MinyiGuoandI.Nakata,"AFrameworkforEfficientArrayRedistributiononDistributedMemoryMultiputers,"TheJournalofSuperputing,vol.20,no.3,pp.243-265,2001.
MinyiGuo,I.NakataandY.Yamashita,"Contention-FreeCommunicationSchedulingforArrayRedistribution,"ParallelComputing,vol.26,no.8,pp.1325-1343,2000.
MinyiGuo,I.NakataandY.Yamashita,"AnEfficientDataDistributionTechniqueforDistributedMemoryParallelComputers,"JSPP'97,pp.189-196,1997.
MinyiGuo,YiPanandZhenLiu,"SymbolicCommunicationSetGenerationforIrregularParallelApplications,"TheJournalofSuperputing,vol.25,pp.199-214,2003.
EdgarT.Kalns,andLionelM.Ni,"ProcessorMappingTechniqueTowardEfficientDataRedistribution,"IEEETrans.onPDS,vol.6,no.12,December1995.
第三章正确的写作方式
论文写作着实不易,市面上也有相当多着作.我个人认为:如何将您的研究成果,用最容易被接受的方式准确地表示出来,就是论文写作的目的之一.
3.1突显您的贡献的重要程度
背景(background)说明须完整,尤其是前人的弱点.
突显动机(motivation)常可用来突显论文的重要.
表达需准确简洁
一篇好文章一般需要以下特色:
创意
贡献
分析
仔细地实验或严谨的证明
3.2表达的逻辑需连贯一致
上一个句子和下一个句子之间,逻辑上必须因果相互连贯的.写一句文章後,必须确认是否已经充分将您的概念表达出来了.写一段文章後,必须重新改写得使只更简洁有力.
3.3可读性
写一句文章後,必须确认是否读者可以轻松地就可了解.
勿用太多符号干扰读者.
多利用图解及实例.
3.4阅读自己的文章
先说明你的创见的直觉概念
再用例子解释ideas
最後用algorithm来准确描述方法.
於适当时机提出证明或实验
第四章常犯的英文文法错误
您可能常犯的英文文法错误罗列如下:
even,therefore,then,是adverb,不是conjunction.
第五章论文之引用
在本章中我们介绍IEEE对引用论文时的规定
排列顺序要按照Lastname的顺序来排.
2.Journal,Conference,书的名称需斜体
引用专书时之写法如下:
T.CoverandJ.Thomas,ElementsofInformationTheory.NewYork:Wiley,1991.
引用Journal之写法如下:
MichaelLuby,"AsimpleparallelalgorithmforthemaximalindependentsetProblem,"SIAMJournalonComputing,vol.15,pp.1036-1053,1986.
引用Conference之写法如下:
L.Hu,"DistributedcodeassignmentorCDMApacketradioworks,"inProc.INFOCOM,1991,pp.1500-1509.
引用之写法如下:
第六章论文之
请将下列电子档分不同目录烧成一份光碟给我.
论文(请用英文撰写)
投稿会议论文(请用英文撰写)
口试投影片(请用英文撰写)
实验程式码需加注解
实验数据及图档(需用mathlab划图)
论文前人相关文献.
参考书目:
1.柯泰德线上英文论文编修owl.cmgt.nctu.edu.tw/