ParallelDynamicLoad-BalancingforAdaptiveDistributive MemoryPDESolvers. NasirTouheed by

Size: px
Start display at page:

Download "ParallelDynamicLoad-BalancingforAdaptiveDistributive MemoryPDESolvers. NasirTouheed by"

Transcription

1 ParallelDynamicLoad-BalancingforAdaptiveDistributive MemoryPDESolvers NasirTouheed by Submittedinaccordancewiththerequirements forthedegreeofdoctorofphilosophy SchoolofComputerStudies TheUniversityofLeeds Thecandidateconrmsthattheworksubmittedishisownandthatappropriate September1998 credithasbeengivenwherereferencehasbeenmadetotheworkofothers.

2 withtheparalleladaptivesolutionofpartialdierentialequations(pdes).weare Thisthesisisconcernedwiththeissueofdynamicload-balancinginconnection Abstract ii schemesonunstructuredgridsandweassumethatgeometricparallelismisused, wherebytheniteelementornitevolumegridsarepartitionedacrosstheavailable parallelprocessors.forparalleleciencyitisnecessarytomaintainawellbalanced partitionandtoattempttokeepcommunicationoverheadsaslowaspossible.when interestedinparallelsolutionsbaseduponeitherniteelementornitevolume anddierentkindsofparallelmachinesarementioned.theniteelementmethodis adaptivityoccurshoweveragivenpartitionmaydeteriorateinqualityandsoitmust alsointroducedanditsparallelimplementationisdiscussedinsomedetail:leading workinthiseld.inchapteroneabriefhistoryofparallelcomputersispresented bemodieddynamically.thisistheproblemthatweconsiderinthiswork. tothederivationofastaticload-balancingproblem.anumberofimportantstatic Chaptersoneandtwooutlinetheprobleminmoredetailandreviewexisting descriptionofsomeerrorindicatorsandcommontechniquesformeshadaptivity. Itisshownhowthisadaptivitymayleadtoaloadimbalanceamongtheavailable loadbalancingalgorithmsarethendiscussed.chaptertwocommenceswithabrief processorsofaparallelmachine.wethendiscusssomewaysinwhichthestatic load-balancingalgorithmsofchapteronecanbemodiedandusedinthecontext ofdynamicload-balancing.theprosandconsofthesestrategiesarediscussedand thennallysomespecicdynamicload-balancingalgorithmsareintroducedand dertaken.inthispreliminary(sequential)versionthedualgraphofanexisting algorithmareoutlinedandanumberofpreliminarynumericalexperimentsareun- discussed. partitionedcomputationalmeshisrepartionedamongthesamenumberofprocessorssothataftertherepartitioningstepeachprocessorhasanapproximateequal anotherarerelativelysmall. tationofthisnewalgorithmandmakingcomparisonwithexistingtechniques.in loadandthenumberofedgesofthisdualgraphwhichcrossfromoneprocessorto Chapterfourthealgorithmisimplementedfora2-dadaptiveniteelementsolver uponanumberofgeneralisationsofexistingalgorithms.thedetailsofthenew InChapterthreeanewdynamicload-balancingalgorithmisproposedbased Theremainderofthethesisisconcernedwiththepracticalparallelimplemen-

3 forsteady-stateproblems,andinchaptervethegeneralityoftheimplementation isenhancedandthealgorithmisappliedinconjunctionwitha3-dadaptivenite volumesolverforunsteadyproblems.inthissituationfrequentrepartitioningof iii whichinvolveverynon-uniformrenement. theworkofthisthesis.thesecomparisonsareveryfavourableforcertainproblems gorithmdetailedhereagainstnewsoftwarethatwasdevelopedsimultaneouslywith themeshisrequired.inthischapterperformancecomparisonsaremadefortheal- andansgiorigin2000.forthepurposesofnumericalcomparisonsalltimings ansgipowerchallenge,dierentworkstationnetworks(sgiindysandsgio2s), codehasbeentestedbymakinguseofavarietyofplatforms,includingacrayt3d, CusingMPIversion1.1(whereapplicable).ThePortabilityoftheload-balancing AllsoftwareimplementationsdescribedinthisthesishavebeencodedinANSI quotedinthisthesisareforthesgiorigin2000unlessotherwisestated.

4 encouragementthroughoutthecourseofthisresearch.notonlywasheveryhelpful IwouldliketothankmysupervisorDr.PeterJimackforhisguidanceand Acknowledgements iv inguidingmethroughoutmystayatleeds,buthewasalsoverypatientwhenit theirhelpfuladviceanddiscussionsduringthistime. ThanksalsotoDr.MartinBerzins,Dr.DavidHodgsonandDr.PaulSelwoodfor cametocorrectingmypoorlydraftedchaptersasregardstotheenglishlanguage. madraqasim,fazilahharon,zahidhussain,jaw-shyongjan,sharifullahkhan, RashidMahmood,SarfrazAhmadNadeem,AllahNawaz,ProfessorMuhammad werealwayshappytohelpme. ThanksarealsoduetomycolleaguesIdreesAhmad,SyedShafaatAli,Muham- MythanksalsogototheGeneralOceandSupportstaoftheSchoolwho UniversityofEdinburghforallowingmetousetheirparallelcomputingfacilities, necessarilyrelatedtotheresearch. Abdul-RaufQuraishi,ShujaMuhammadQuraishiandAlexTsaiformattersnot helpandsupportintheearlydaysofmystayinyork. includingthecrayt3d,andtothankdr.alanwoodofuniversityofyork,forhis IwouldalsoliketoacknowledgetheEdinburghParallelComputingCentreatthe Iespeciallywishtothankmyparentsandsisters,brother-in-laws,mother-in-law andfather-in-lawfortheirencouragement.mytwodaughtersmaryamandsidrah intheukhavebeenagreatencouragementthroughouttheperiodofthisresearch. (whowasamuchneededandwelcomeadditiontoourfamilyinthemiddleofthis MembersofmyextendedfamilyinPakistanandofmyimmediatefamilyhere project)havebeenmostpatientwhileinishedthistask.thisprojectwouldnot formsofstudy-leave,cotsandorsawardsrespectively. PakistanandtheCommitteeofVice-ChancellorsandPrincipalsoftheUniversities oftheunitedkingdomforsupportingmenanciallythroughoutmyresearchinthe havebeencompletedwithouttheconstantloveandsupportofmywife,shagufta. AttheveryendIwouldliketothankTheAlmighty,forthemuchneededcourage Finally,mythanksalsogototheUniversityofKarachi,theGovernmentof theproject. andstrengthwhichhegrantedmeatthisrelativelyoldageofmylifetonalise

5 Contents 1Introduction 1.2ComparisonBetweenSIMDandMIMDComputers:::::::::7 1.1IntroductiontoParallelComputers:::::::::::::::::: GeneralMIMDSystems:::::::::::::::::::: SIMDSystems::::::::::::::::::::::::: FiniteElementMethodsforEllipticPDEs::::::::::::::9 1.4Time-DependentProblems:TheLinearDiusionEquation:::::15 1.5ParallelFiniteElementandLoad-Balancing::::::::::::: PiecewiseLinearFiniteElements:::::::::::::::11 1.6RecursiveGraphPartitioningHeuristics::::::::::::::: TheMethodofLines:::::::::::::::::::::: AlgorithmicDetails::::::::::::::::::::::: RecursiveCoordinateBisection(RCB):::::::::::: ModiedRecursiveGraphBisection(MRGB):::::::: RecursiveInertialBisection(RIB)::::::::::::::: RecursiveGraphBisection(RGB)::::::::::::::: RecursiveSpectralBisection(RSB)::::::::::::::22 1.8OtherGraphPartitioningTechniques:::::::::::::::::25 1.7MultisectionalGraphPartitioningHeuristics::::::::::::: RecursiveNodeClusterBisection(RNCB):::::::::: KernighanandLinTypeAlgorithms::::::::::::: GreedyAlgorithm(GR):::::::::::::::::::: StripwiseMethods::::::::::::::::::::::: MultidimensionalSpectralGraphPartitioning:::::::: StateoftheArtSoftwareToolsforGraphPartitioning:::26 v

6 2AdaptivityandDynamicLoadBalancing CONTENTS 2.1SpatialErrorIndicators::::::::::::::::::::::::30 2.2DierentTypesofRenements::::::::::::::::::::31 29 vi 2.4DiusionAlgorithms::::::::::::::::::::::::::39 2.3RelationBetweenAdaptivityandDynamicLoadBalancing::::: RegenerationSchemes::::::::::::::::::::: BasicDiusionMethod::::::::::::::::::::: LocalMeshAdaptationSchemes:HierarchicalRenement: AMulti-LevelDiusionMethod:::::::::::::::: GeneralisationsofStaticAlgorithms:::::::::::::36 2.6TwoParallelMultilevelAlgorithms::::::::::::::::::45 2.5MinimisingDataMigration:::::::::::::::::::::: ParMETIS:::::::::::::::::::::::::::: ParJOSTLE::::::::::::::::::::::::::: DimensionExchangeMethod:::::::::::::::::42 3ANewDynamicLoadBalancer 3.1MotivationoftheAlgorithm::::::::::::::::::::::51 2.7TwoFurtherParadigms::::::::::::::::::::::::47 3.2DescriptionoftheAlgorithm:::::::::::::::::::::: AlgorithmofVidwansetal.:::::::::::::::::: AlgorithmofOliker&Biswas::::::::::::::::: LocalMigration::::::::::::::::::::::::: GroupBalancing:::::::::::::::::::::::: FurtherRenementoftheAlgorithm:LocallyImprovingthePartitionQuality:::::::::::::::::::::::::::::::56 4ParallelApplicationoftheDynamicLoadBalancerin2-d 3.5Examples::::::::::::::::::::::::::::::::60 3.6Conclusions:::::::::::::::::::::::::::::::73 3.4GlobalLoad-BalancingStrategy:DivideandConquerApproach::58 4.2AParallelDynamicLoad-BalancingAlgorithm::::::::::::80 4.1Introduction::::::::::::::::::::::::::::::: GroupBalancing::::::::::::::::::::::::81 77

7 CONTENTS 4.3DiscussionoftheAlgorithm:::::::::::::::::::::: LocalMigration::::::::::::::::::::::::: DivideandConquerandParallelImplementation::::::84 vii 4.4DescriptionofRelatedDataStructuresAssociatedWiththeRedistributionoftheMesh:::::::::::::::::::::::::: ActivityofType2Processors:UnpackingtheLoad::::: ActivityofType1Processors:PackingtheLoad::::::88 4.5DierentIssuesandRelatedFunctionsUsedintheMainAlgorithm 4.3.3ActivityofType3Processors:ThirdPartyAdjustment::88 ByProcessorsofType1:::::::::::::::::::::::: HandlingofVertices::::::::::::::::::::::93 4.7DierentIssuesWhichareRelatedWithProcessorsofType3::: DierentIssuesWhichareRelatedWithProcessorsofType2::: insertion():::::::::::::::::::::::::::: deletion():::::::::::::::::::::::::::: HandlingofEdges:::::::::::::::::::::::95 4.9SomeExamples::::::::::::::::::::::::::::: UseofMessagePassingInterface(MPI):::::::::::::::: Discussion:::::::::::::::::::::::::::::::: AlternativeAlgorithms::::::::::::::::::::: ComparativeResults::::::::::::::::::::::105 5ParallelApplicationoftheDynamicLoadBalancerin3-d 5.1Introduction::::::::::::::::::::::::::::::: Conclusions::::::::::::::::::::::::::::::: DiscussionII:::::::::::::::::::::::::: DiscussionI::::::::::::::::::::::::::: AParallelAdaptiveFlowSolver:::::::::::::::::::: AParallelFiniteVolumeSolver:::::::::::::::: AParallelAdaptiveAlgorithm:::::::::::::::: ApplicationoftheParallelDynamicLoad-BalancingAlgorithm:: DynamicLoadBalancing:::::::::::::::::::::::: CalculationofWPCG::::::::::::::::::::: UseofTokens::::::::::::::::::::::::::134

8 CONTENTS 5.5ComputationalResults::::::::::::::::::::::::: UseofGlobalCommunication::::::::::::::::: NoColouring::::::::::::::::::::::::::135 viii 5.6Discussion:::::::::::::::::::::::::::::::: Examples::::::::::::::::::::::::::::138 6ConclusionandFutureAreasofResearch 5.7InvestigationintoScalabilityoftheAlgorithm:::::::::::: Conclusions::::::::::::::::::::::::::::::: DiscussionII:::::::::::::::::::::::::: DiscussionI::::::::::::::::::::::::::: SummaryofThesis::::::::::::::::::::::::::: PossibleExtensionstotheResearch::::::::::::::::::

9 ListofFigures 1.5Two-dimensionaltorus.:::::::::::::::::::::::::8 1.4(a)Two-dimensionalmesh,(b)three-dimensionalmesh.:::::::8 1.2Aringofprocessors.::::::::::::::::::::::::::6 1.3Hypercubesof(a)dimension1,(b)dimension2and(c)dimension Alineararrayofprocessors.::::::::::::::::::::::6 2.4ThematrixA.::::::::::::::::::::::::::::::44 2.3Dimensionexchangemethod.:::::::::::::::::::::43 2.1Diusionmethod.::::::::::::::::::::::::::::41 2.2Multi-leveldiusionmethod.::::::::::::::::::::::42 1.6EntriesoftheLaplacianmatrix.::::::::::::::::::::23 3.1CalculationofSender,ReceiverandMigtot:::::::::::::::54 3.6Group-balancingalgorithm:versiontwoofloadbalancingofthetwo 3.5Analgorithmforreningthepartitionsbetweenapairofprocessors Thecalculationofgain.:::::::::::::::::::::::::55 3.4Initialversionofloadbalancingofthetwogroups.::::::::::57 3.3Updationofgaindensitiesandedgescutbetweentheprocessors.::56 3.8ThecoarsemeshofExample1.::::::::::::::::::::63 3.7Adivide&conquertypedynamicload-balancingalgorithm.::::61 3.9ThecoarsemeshofExample2.:::::::::::::::::::: Thecoarse\Texas"meshofExample4.:::::::::::::::69 groups.::::::::::::::::::::::::::::::::::59 4.1Updatingthegains.::::::::::::::::::::::::::: Adaptedmeshafter240time-stepsforExample5.:::::::::: Coarsemeshof5184elementsadaptedtoinitialshockconditionfor Example5.:::::::::::::::::::::::::::::::72 ix

10 LISTOFFIGURES 4.3Paralleldynamicload-balancingalgorithm.::::::::::::::86 4.2Loadbalancingofthetwogroups.:::::::::::::::::::85 4.4Thearraynonodessuballwhichcanaccommodateninecoarseelements.92 x 4.5ThefunctionShared().:::::::::::::::::::::::::94 4.8ThefunctionChangenbhd2().::::::::::::::::::::: ThefunctionEdgeChange().::::::::::::::::::::::99 4.6ThefunctionShared2().::::::::::::::::::::::::94 4.7ThefunctionChangenbhd().::::::::::::::::::::::96 4.9ThefunctionChangenbhd3().::::::::::::::::::::: ThefunctionDirichEdgeChange().::::::::::::::::::: ThecoarsemeshofExample3.:::::::::::::::::::: ThepartialviewofthecoarsemeshofExample4.:::::::::: Regularrenementdissectinginteriordiagonal:::::::::::: Meshdata-structuresinTETRAD:::::::::::::::::: Scalabilitycomparisonusingare-balancingtoleranceof5%forExample1(whereTime=RedTime+SolTime).:::::::::::: CalculationofarowoftheweightedLaplacianmatrix.::::::: Greenrenementbytheadditionofaninteriornode:::::::: Calculationofweightsofverticesandedgesoftheweighteddualgraph Scalabilitycomparisonusingare-balancingtoleranceof15%forExample1(whereTime=RedTime+SolTime).:::::::::::: Scalabilitycomparisonusingare-balancingtoleranceof10%forEx- 5.11Scalabilitycomparisonusingare-balancingtoleranceof15%forExample2(whereTime=RedTime+SolTime).::::::::::::15ample2(whereTime=RedTime+SolTime).:::::::::::: Scalabilitycomparisonusingare-balancingtoleranceof5%forEx- 5.10Scalabilitycomparisonusingare-balancingtoleranceof10%forEx- 5.12Scalabilitycomparisonusingare-balancingtoleranceof15%forExample1(whereTime=RedTime+0.2*SolTime).::::::::: Scalabilitycomparisonusingare-balancingtoleranceof15%forExample1(whereTime=RedTime+5*SolTime).::::::::::157

11 LISTOFFIGURES 5.15Scalabilitycomparisonusingare-balancingtoleranceof15%forExample2(whereTime=RedTime+0.2*SolTime).:::::::::15ample2(whereTime=RedTime+25*SolTime).::::::::: Scalabilitycomparisonusingare-balancingtoleranceof15%forExample1(whereTime=RedTime+25*SolTime).:::::::::157 xi 5.17Scalabilitycomparisonusingare-balancingtoleranceof15%forEx- 5.16Scalabilitycomparisonusingare-balancingtoleranceof15%forExample2(whereTime=RedTime+5*SolTime).::::::::::158

12 ListofTables 3.1Partitiongeneratedinparallelon8processorsalongwithournal 3.2SummaryofresultswhentheNew,Vidwansetal.,ChacoandJOS- 3.3Partitiongeneratedinparallelon8processorsalongwithournal partitionsforexample1.::::::::::::::::::::::::64 3.4SummaryofresultswhentheNew,Vidwansetal.,ChacoandJOS- TLEalgorithmsareappliedtotheinitialpartition(seeTable3.1)of Example1.:::::::::::::::::::::::::::::::64 3.5Partitiongeneratedinparallelon8processorsalongwithournal partitionsforexample2.::::::::::::::::::::::::66 3.6SummaryofresultswhentheNew,Vidwansetal.,ChacoandJOS- TLEalgorithmsareappliedtotheinitialpartition(seeTable3.3)of Example2.:::::::::::::::::::::::::::::::66 3.7Partitiongeneratedinparallelon16processorsalongwithournal partitionsforexample3.::::::::::::::::::::::::68 3.8SummaryofresultswhentheNew,Vidwansetal.andJOSTLEalgorithmsareappliedtotheinitialpartition(seeTable3.7)ofExample partitionsforexample4.::::::::::::::::::::::::70 TLEalgorithmsareappliedtotheinitialpartition(seeTable3.5)of Example3.::::::::::::::::::::::::::::::: SummaryofresultswhentheNew,Vidwansetal.,ChacoandJOS- 3.9Initialandnalpartitions(producedbytheNewalgorithm)forExample5.:::::::::::::::::::::::::::::::::73 4.:::::::::::::::::::::::::::::::::::::70 TLEalgorithmsareappliedtotheinitialpartition(seeTable3.9)of Example5.:::::::::::::::::::::::::::::::73 xii

13 LISTOFTABLES 3.12SummaryofresultswhentheNew,Vidwansetal.,ChacoandJOS- 3.11Initialandnalpartitions(producedbytheNewalgorithm)forExample6.:::::::::::::::::::::::::::::::::74 xiii 4.1DataforthepartitionsofExample1(involvingparallelmeshgenerationandrepartitioningon2processors).:::::::::::::::107 TLEalgorithmsareappliedtotheinitialpartition(seeTable3.11) 4.2DataforthepartitionsofExample2(involvingparallelmeshgenerationandrepartitioningon4processors).:::::::::::::::107 ofexample6.::::::::::::::::::::::::::::::74 4.3DataforthepartitionsofExample3(involvingparallelmeshgenerationandrepartitioningon4processors).::::::::::::::: DataforthepartitionsofExample4(involvingparallelmeshgenerationandrepartitioningon4processors).::::::::::::::: DataforthepartitionsofExample5(involvingparallelmeshgenerationandrepartitioningon2processors).::::::::::::::: DataforthepartitionsofExample6(involvingparallelmeshgenerationandrepartitioningon4processors).::::::::::::::: Comparisonofdynamicload-balancingresultsusingfouralgorithms 4.8DataforthepartitionsofExample7(involvingparallelmeshgenerationandrepartitioningon8processors).::::::::::::::: DataforthepartitionsofExample8(involvingparallelmeshgenerationandrepartitioningon16processors).::::::::::::::115 forexamples1to6.:::::::::::::::::::::::::: DataforthepartitionsofExample9(involvingparallelmeshgenerationandrepartitioningon8processors).::::::::::::::: DataforthepartitionsofExample10(involvingparallelmeshgenerationandrepartitioningon16processors).::::::::::::: DataforthepartitionsofExample11(involvingparallelmeshgenerationandrepartitioningon8processors).::::::::::::::11erationandrepartitioningon16processors).::::::::::::: DataforthepartitionsofExample12(involvingparallelmeshgen- 4.14Comparisonofdynamicload-balancingresultsusingfouralgorithms forexamples7to12.::::::::::::::::::::::::::120

14 LISTOFTABLES 5.2Solutiontimes,redistributiontimes,totalmigrationweightsandmi- 5.1Somepartition-qualitymetricsimmediatelybeforeandafterasingle re-balancingstepforexample1.::::::::::::::::::::141 xiv 5.3Solutiontimes,redistributiontimes,totalmigrationweightsandmi- 5.4Solutiontimes,redistributiontimes,totalmigrationweightsandmigrationfrequenciesfor300time-stepsusingare-balancingtolerance of10%forexample1.:::::::::::::::::::::::::143 of5%forexample1.:::::::::::::::::::::::::: Solutiontimes,redistributiontimes,totalmigrationweightsandmi- 5.5Somepartition-qualitymetricsimmediatelybeforeandafterasingle re-balancingstepforexample2.::::::::::::::::::::146 of15%forexample1.::::::::::::::::::::::::: Solutiontimes,redistributiontimes,totalmigrationweightsandmi- 5.8Solutiontimes,redistributiontimes,totalmigrationweightsandmigrationfrequenciesfor300time-stepsusingare-balancingtolerance of10%forexample2.:::::::::::::::::::::::::148 of5%forexample2.::::::::::::::::::::::::::147 of15%forexample2.:::::::::::::::::::::::::149

15 Chapter1 Sincethemiddleofthecurrentcentury,breakthroughsincomputertechnologyhave Introduction madeatremendousimpactonnumericalmethodsingeneralandthenumerical solutionsofpartialdierentialequationsinparticular.duringtheinfancyperiod andhasasmuchmemoryasispossible(oraordable).theprocessoriscommonly ofcomputerstheywereserialinnature.thismeansthattheywerebuiltusing knownasthecentralprocessingunit(cpu)andisfurtherdividedintoacontrol thevonneumannparadigm:withasingleprocessorwhichrunsasfastaspossible unitandanarithmetic-logicunit(alu).thememorystoresbothinstructionsand data.thecontrolunitdirectstheexecutionofprograms,andthealucarries outthecalculationscalledforintheprogram.whentheyarebeingusedbythe program,instructionsanddataarestoredinveryfastmemorylocations,called registers.asfastmemoryisquiteexpensive,therearerelativelyfewregisters. carryingoutthreetrillioncopiesofdatabetweenmemoryandregisterspersecond isthatofthespeedoflight,soinordertobuildacomputerwhichiscapableof say,onehastoteach32-bitwordintoasquarewithsidelengthof10?10meters example,themaximumspeedatwhichthedatacantravelfrommemorytocpu Theperformanceofsuchcomputersisclearlylimitedbyphysicallaws.For (thisisapproximatelyequaltothesizeofarelativelysmallatom).thisissimply theuseofcachememory-whichisimplementedonthesamechipasthecpu.the ideabehindcacheistheobservationthatprogramstendtoaccessbothdataand takenbythedatawhiletravellingfrommemorytoregisters.thisisachievedby notpossible-see[79]fordetails. Inordertospeedupthemachine,onepossibilityistoreducethetransfertime 1

16 CHAPTER1.INTRODUCTION instructionssequentially.hence,ifwestoreasmallblockofdataandasmallblock ofinstructionsinfastmemory(cache),mostoftheprogram'smemoryaccesseswill usethiscachememoryratherthantheslowermainmemory.thismemorywill 2 outsidethechip. beslowerthanregistersbutitwillbefasterthanthemainmemoryimplemented thistrendcontinueduntilthe1980s.atthistimeanewdesignphilosophycalledthe ofhigh-levellanguagesaswellasthecomplexfunctionsofoperatingsystems,and dressingmodesandothermechanisms.thesenewfeaturesallowedtheexecution usedanincreasedchipareatointroducenewandsophisticatedinstructions,ad- Duringtheinitialstagesinthedevelopmentofmicrochips,designerstypically reducedinstructionsetcomputer(risc)emerged.theriscsupportersarguethat allthesenewinstructionscomplicatethedesignofthecontrolunit,slowingdown theexecutionofbasicoperations.asimpleinstructionsetallows,inprinciple,a byriscsupportersisthatthesimplicationofthecontrolunithelpstosavechip bemorethancompensatedforbytheincreasedspeed.anotheradvantageclaimed areaforthecontrolimplementation.thiscanbeusedtoimplementspecialfeatures simple,fastimplementation,sothelargernumberofinstructionsthatisrequirecan intheoperatingunit,aimedatimprovingtheexecutionspeed.therearemany therewereandstillareimportantclassesofprobleminscienceandengineering variantsofriscprocessors,amongthemareberkeleyrisc,microprocessorswithoutinterlockedpipestages(mips)andtheinmostransputer(seechapter10of [20]). whichpractitionershavenotbeenabletosolvesuccessfully.forexample,toattack the\grandchallenges"([17])monthsorevenyearsareneededbythebestofthese computers.agrandchallengeisafundamentalprobleminscienceorengineering Evenafteralltheseadvancementsinthedevelopmentofthecomputerindustry thathasabroadeconomicandscienticimpact,andwhosesolutioncouldbeadvancedbyapplyinghigh-performancecomputingtechniquesandresources([67]). Manyoftheseproblemsarebasicallylargecomputationaluiddynamicsproblems whichcanbemodelledbyasetofpartialdierentialequations(pdes). asmentionedin[26].thisisthesimulationofathree-dimensional,fullyresolved, icallyweconsiderheretwoexamples.therstoneisstudiedbycaseetal.([16]) turbulentowasmightoccurinthedesignofaportionofashiphull.theprimary Tohaveanideaofthecomputingrequirementstosolvesuchproblemsnumer-

17 CHAPTER1.INTRODUCTION parameterforcharacterisingtheturbulentuidowisthedimensionlessquantity smallwavenumberintheowoneneedsr9=4meshpoints([21]).thatisn=109 berofabout104orgreater.inordertofullyresolveimportantdisturbancesofa knownasthereynoldsnumber(r).suchasimulationwouldhaveareynoldsnum- 3 meshpointsforeachtimestep.eachmeshpointhasonepressuretermandthree velocitytermsforboththecurrentandtheimmediatepasttimestep.thisisa totalof8109scalarvariables.iftemperatureorotherparametersmustalsobe numberofarithmeticoperationsvarieswidely,dependinguponthesolutionmethod maintainedforeachpoint,thenabout1010wordsofdatamemoryarerequired.the employed.oneecientapproachthattakesadvantageoftheproblemgeometry, wanttopredicttheweatheroveranareaof milesfortwo-dayperiod hasbeenestimatedtorequireonlyabout500additionsand300multiplicationsper gridpoint.thisleadstoanoperationscountof1012operationspersingletime step(see[26]fordetails). iftheareaisbeingmodeleduptoaheightof11milesandonewishestopartition andtheparametersneedtocomputedonceeveryhalfhour.asmentionedin[67], this cubicmiledomainintosegmentsofsize Thesecondproblemisthatofmodelingandforecastingofweather.Supposewe memory.itisalsoestimatedin[67]thatforthispredictionthetotalnumberof operationsis1015. thentherewouldbe1011dierentsegments.soweneedatleast1011wordsofdata 108operations/second. wasnotevenclosetohavingenoughcapability([84])toperformthesecalculations. It'sprimarymemorywaslimitedto106words,andtheexecutionratewasabout Hencebythistimeitwasclearthatnew,morepowerfulcomputersystemswould Theworld'smostpowerfulcomputerofthemid70'swastheCRAY-1,which beguntostartapproachingitsphysicallimits,thecommunityhadnochoicebutto beneededtosolvethisclassofproblems.sincethesingleprocessormachineshad Parallelcomputersperformtheircalculationsbyexecutingdierentcomputational consideralternativeparadigmssuchasparallelmachines. tasksonanumberofprocessorsconcurrently.theprocessorswithinaparallel 1.1IntroductiontoParallelComputers

18 CHAPTER1.INTRODUCTION Thisexchangeofinformationoccurseitherintheformofexplicitmessagessentby computergenerallyexchangeinformationduringtheexecutionoftheparallelcode. oneprocessortoanotherordierentparallelprocessorssharingaspeciedcommonmemoryresourcewithintheparallelcomputer.theparallelload-balancing In1966MichaelFlynn([33])classiedsystemsaccordingtothenumberofin- 4 structionstreamsandthenumberofdatastreams.thetwoimportantsystems are:simd-singleinstructionstream,multipledatastream, algorithms,proposedinthisthesis,workverywellontheseparadigms. computingarchitecture SIMDSystems MIMD-MultipleInstructionstream,MultipleDatastream. SuchasystemhasasingleCPUdevotedtoexclusivelytocontrol,andalarge Thissectionprovidesabriefintroductiontotheseimportantclassesofparallel collectionofsubordinateprocessors,eachhavingonlyalus,andtheirown(small amountof)memory.duringeachinstructioncycle,thecontrolprocessorbroadcasts aninstructiontoallofthesubordinateprocessors,andeachofthesubordinate processorseitherexecutestheinstructionorisidle. tionmachinesthatwereproducedbythinkingmachines.thecm-2hadupto 65,3561-bitprocessorsandupto8billionbytesofmemory.Masparalsoproduced SIMDmachines.TheMP-2hasupto16,38432-bitALUsandupto4billionbytes ofmemory. ThemostfamousexamplesofSIMDmachinesaretheCM-1andCM-2connec GeneralMIMDSystems ThekeydierencebetweenMIMDandSIMDsystemsisthatwithMIMDsystems, theprocessorsareautonomous:eachprocessorisafull-edgedcpuwithbotha controlunitandanalu.thuseachprocessorsiscapableofexecutingitsown programatitsownpace.theworldofmimdsystemsisdividedintosharedmemoryanddistributed-memorysystems.

19 CHAPTER1.INTRODUCTION Shared-MemoryMIMD Agenericshared-memorymachineconsistsofacollectionofprocessorsandmemory modulesinterconnectedbyanetwork.eachprocessorhasaccesstotheentire 5 advantageofbeingveryrapid(inprinciple)andisgenerallysimplertoprogram. addressspaceofthememorymodules.sothatanydatastoredintheshared However,itsmaindrawbackisthattherecanbeseriousdelays(contentiontime) memoryiscommonto,andcanbeaccessedby,anyoftheprocessors.thishasthe ofabus,thesearchitecturesdonotscaletolargenumberofprocessors:thelargest time.thesimplestnetworkconnectionisbusbased.duetothelimitedbandwidth ifmorethanoneprocessorwantstousethesamelocationinmemoryatthesame congurationofthecurrentlypopularsgichallengexlhasonly36processors. RecentlySiliconGraphics,Inc.hasdesignedandmanufacturedtheOrigin2000 network.forexamplethebasicunitoftheconvexspp1200isa55crossbar shared-memoryarchitecturesrelyonsometypeofswitch-basedinterconnection utilisesscalableshared-memorymultiprocessing(s2mp)architecture.mostother R10000processorswithapeakperformanceof400Mopeach.Thecomputer computer.thebasicbuildingblockoftheoriginisanodebuiltupontwomips processorsareconnecteddirectlyorindirectlybymeansofcommunicationwires. switch. Distributed-MemoryMIMD Indistributed-memorysystems,eachprocessorhasitsownprivatememory.These Fromtheperformanceandprogrammingpointofviewtheidealinterconnection cost)ofsuchanetworkmakesitimpracticaltoconstructsuchamachinewithmore networkisafullyconnectednetwork,inwhicheachprocessorisdirectlyconnected toeveryotherprocessor.unfortunately,theexponentialgrowthinthesize(and thanafewprocessors.attheoppositeextremefromafullyconnectednetworkis immediatelyadjacentneighbouringprocessors(seefigure1.1).aringisaslightly alineararray:astaticnetworkinwhichallbuttwooftheprocessorshavetwo morepowerfulnetwork.thisisjustalineararrayinwhich\terminal"processors onlyadditionalcostisthecostofp-1orpwiresforanetworkofpprocessors. havebeenjoined(seefigure1.2).thesenetworksarerelativelyinexpensive;the Moreoveritisverycheaptoupgradethenetwork-toaddoneprocessorweonly

20 CHAPTER1.INTRODUCTION 6 Figure1.1:Alineararrayofprocessors. needoneextrawire.therearetwoprincipaldrawbacks: iftwoprocessorsarecommunicating,it'sverylikelythatthiswillprevent otherprocessorswhicharealsoattemptingtocommunicatefromdoingso, Figure1.2:Aringofprocessors. networkthatgivesagoodbalancebetweenthehighcostandhighspeedofthe inalineararraytwoprocessorsthatareattemptingtocommunicatemayhave fullyconnectednetworkandthelowcostbutpoorperformanceofthelineararray Inbetweenthetwoextremesahypercubeisapracticalstaticinterconnection necessarytoforwardthemessagealongasmanyasp/2wires. toforwardthemessagealongasmanyasp-1wires,andinaringitmaybe taketwohypercubesofdimensiond?1andjointhecorrespondingprocessorswith ofasingleprocessor.inordertoconstructahypercubeofdimensiond>0,we communicationwires(seefigure1.3).itisclearthatahypercubeofdimension orring.hypercubesaredenedinductively:adimension0hypercubeconsists

21 CHAPTER1.INTRODUCTION 7 theshortestpaththenthemaximumnumberofwiresamessagehastotravelis d.thisismuchfewthanforthelineararrayofring.theprincipaldrawbackto dwillconsistof2dprocessors.itisalsoclearthatinahypercubeofdimension deachprocessorisdirectlyconnectedtodotherprocessorsandthatifwefollow Figure1.3:Hypercubesof(a)dimension1,(b)dimension2and(c)dimension3. thehypercubeisthatitisnoteasytoupgradethesystem:eachtimewewish toincreasethemachinesize,wemustdoublethenumberofprocessorsandadd hypercube(anncube10with1024processors). anewwiretoeachprocessor.therst\massivelyparallel"mimdsystemwasa fromthen-dimensionalmeshbyadding\wrap-around"wirestotheprocessorson theborder.asfarasupgradingisconcernedmeshesandtoriarebetterthan arraysandrings,respectively.observethatann-dimensionaltoruscanbeobtained (seefigures1.4and1.5),whicharesimplyhigherdimensionalanaloguesoflinear Intermediatebetweenhypercubesandlineararraysarethemeshesandtori hypercubes(althoughnotasgoodaslineararraysandrings).forexample,if onewishestoincreasethesizeofaqqmesh,onesimplyaddsaq1mesh andqwires.meshesandtoriarecurrentlyquitepopular.theintelparagonis tori. 1.2ComparisonBetweenSIMDandMIMDCom- atwo-dimensionalmesh,andthecrayt3dandt3eareboththree-dimensional SIMDcomputersrequirelesshardwareandlessmemorythanMIMDcomputers In[67],Kumaretal.discusstheprosandconsofSIMDandMIMDcomputers.

22 CHAPTER1.INTRODUCTION 8 Figure1.4:(a)Two-dimensionalmesh,(b)three-dimensionalmesh. Figure1.5:Two-dimensionaltorus.

23 CHAPTER1.INTRODUCTION operatingsystemateachprocessor.simdcomputersarenaturallysuitedfordata- becausetheyhaveonlyoneglobalcontrolunitandonlyonecopyoftheprogram needstobestored.ontheotherhand,mimdcomputersstoretheprogramand9 ecutedonalargedataset(whichisthecaseintheeldofimageprocessingfor example). parallelprograms;thatis,programsinwhichthesamesetofinstructionsareex- oftime.data-parallelprogramsinwhichsignicantpartsofthecomputationare executedierentinstructionsinthesameclockcycle,soifaprogramhasmany tionals,itisentirelypossiblethatmanyprocessorswillremainidleforlongperiods conditionalbranchesorlongsegmentsofcodewhoseexecutiondependsoncondi- AcleardisadvantageofSIMDcomputersisthatdierentprocessorscannot containedinconditionalstatementsarethereforebettersuitedtomimdcomputers thantosimdcomputers. processorshasitsowncontrolunit.itmayseemthatthecostofeachprocessor mustbehigherthanthecostofasimdprocessor.however,itispossibleto usegeneral-purposemicroprocessorsasprocessingunitsinmimdcomputers.in contrast,thecpuusedinsimdcomputershastobespeciallydesigned.hence, IndividualprocessorsinanMIMDcomputeraremorecomplex,becauseeach duetoeconomiesofscale,processorsinmimdcomputersmaybebothcheaperand morepowerfulthanprocessorsinsimdcomputers. 1.3FiniteElementMethodsforEllipticPDEs Taylorseriesandinvolvethevaluesofthesolutionatneighbouringpointsinthe replacedbydierencequotients.thedierenceoperatorsareusuallyderivedfrom Inthenitedierenceapproximation,thederivativesinadierentialequationare equationsarethenitedierence,theniteelementandthenitevolumemethods. Probablythethreemostpopularnumericaltechniquesforsolvingpartialdierential domain.aftertakingtheboundaryconditionsintoaccount,a(sparse)system ofalgebraicsimultaneousequationsisobtainedandcanbesolvedforthenodal unknowns. toimplementonregulardomains.unfortunatelythismethodisdiculttoapply forsystemswithirregulargeometriesand/orunusualboundaryconditions. Thenitedierencesmethod(FDM)iseasytounderstandandstraightforward

24 CHAPTER1.INTRODUCTION forsuchsystems.incontrasttonitedierencetechniques,theniteelement methoddividesthesolutiondomainintosimplyshapedregionsor\elements".an Theniteelementmethod(FEM)providesanalternativethatisbettersuited 10 approximatesolutionforthepdecanbedevelopedforeachoftheseelements.the totalsolutionisthengeneratedbylinkingtogetheror\assembling"theindividual solutionstakingcaretoensurecontinuityattheinterelementboundaries.thusthe Inthisschemethesolutionisrepresentedasaseriesofpiecewiseconstantelements. elements(controlvolumes).foreachcontrolvolumetheareaintegralisconverted PDEisapproximatelysatisedinapiecewisefashion(seebelow). intoalineintegraloveritsedgesandthenumericaluxattheboundariesalso ThediscretisedformofthePDEisfoundbyintegratingtheequationoverthe Thenitevolumemethod(FVM)mayalsobeappliedonunstructuredmeshes. calculated. thisthesis.theinterestedreadercanconsultthebooksofjohnson([58])andstrang &Fix([95]).Howeverwedescribethemethodbrieyinthecaseofaparticular PDE;Poisson'sequationin2dimensions: Acomprehensivedescriptionoftheniteelementmethodisbeyondthescopeof Forclarityweassumethefollowingboundaryconditionsareimposed: Notethatthisequationisalinearsecondorderpartialdierentialequationwhich forx2<2: (1.1) condition. arisesinalargenumberofphysicalsituations(e.g.owofanidealuid). equation(1.1)byatestfunctionwandintegrateovertoget, Werstderivetheweakformoftheequation(1.1).Todothiswemultiplythe Byusingthedivergencetheoremweget, -Rwr2udx=Rwfdx.

25 rstderivativesaresquareintegrableinandwhicharezeroeverywhereon?1) CHAPTER1.INTRODUCTION 11 Forsimplicitywewillassumeg0inwhichcasetheexpressionsimpliesstill further: thenaboveintegralformreducesto, Rru:rwdx?R?2gwds=Rwfdx.?1.TheaboveintegralformthenleadstothefollowingweakformofthePoisson's NowletH1E()bethespaceofallthosefunctionswhoserstderivativesaresquare integrableinandwhichsatisfythedirichletboundaryconditioneverywhereon Rru:rwdx=Rwfdx. PDE. forallw2h10(). Therestofthissectionconsiderstheniteelementapproximationtothesolution Findu2H1E()suchthat ofthisweakform. Zru:rwdx=Zwfdx; (1.2) 1.3.1PiecewiseLinearFiniteElements Theveryrststepintheapproximationofubytheniteelementmethodisto thissectionwewillassumethesearetriangles).thisisalwayspossibleprovided dividethedomainintoalargenumberofsmallnon-overlappingsubdomains(in thedomainusingisoparametricniteelements(see[19]). approachtothecurveboundary(see[95]fordetails).anotherwayistotriangulate thatisitselfapolygon(i.e.therearenocurvedboundaries).therearemethods meansofasetoflinesegmentsinsuchawaythatinthelimittheselinesegments tohandlecurvedboundaries.onewayistoapproximatethecurveboundaryby from1ton=nb+ne(wherenbisthenumberofverticesintheinteriorofthe domainorontheneumannboundary,?2,andneisthenumberofverticeson Letussupposethatthevertices(nodes)ofthetriangleshavebeennumbered

26 CHAPTER1.INTRODUCTION choosetherstdegreepolynomialshere(polynomialsofdegreezerocannotbe thedirichletboundary,?1).oneachtriangleuisapproximatedbyalowdegree polynomial.althoughanydegreepolynomialscanbeselectedforsimplicitywe 12 functionsarelinearoneachtriangleandsatisfypj(x)=1ifxisthepositionvector Nowwecanwriteu(anapproximationstou)intermsofthesebasisfunctionsas, ofthenodejandpj(x)=0ifxisthepositionvectorofanyoftheothernodes. usedsincewerequirethederivativestobesquareintegrable). Wenextdenesimple\basis"functionsPi(x)forallnodesifrom1toN.These Dirichletboundarycondition,u=uE,fori=nB+1,...,nB+nE.(Notethat, whereaiareunknown(tobedetermined)fori=1,...,nb,andaregivenbythe duetoourchoiceofbasisfunctions,aiisthevalueofuwhenevaluatedattheith nodeofthemesh).ifwesubstitutethevalueofufromequation(1.3)foruand u=nxi=1aipi(x); (1.3) elementequationsisgivenby: equationsfortheunknownsa1;:::;anb.thissystem,knownasthegalerkinnite replacewbypj(x),forj=1,...,nb,inequation(1.2)wethengetasystemofnb Typicallythisiswritteninmatrixformas nbxi=1aizrpi(x):rpj(x)dx=zpj(x)f(x)dx?nx i=nb+1aizrpi(x):rpj(x)dx; Ka=f; forj=1,...,nb. (1.5) (1.4) Kji=RrPj(x):rPi(x)dx)andaisavectoroftheunknownsa1;:::;anB. wherekisreferredtoasthe\globalstinessmatrix"(whoseentriesaregivenby isthattheentrykjiofthematrixkwillalwaysbezeroiftheverticesnumberedj 1.3.2AlgorithmicDetails Havingderivedtheniteelementequations(1.5)wenowdiscusshowthematrixK andthevectorfcanbeobtainedsystematically.themostimportantpointtonote andiarenotconnectedbyanedgeofthemesh.thisisbecausethedotproductof rpj(x)andrpi(x)willbezerooneverytriangularelementinsuchacase.since

27 CHAPTER1.INTRODUCTION \sparsematrix". thismeansthatmostoftheentriesofkwillalwaysbezero,werefertothisasa SupposethattheniteelementmeshconsistsofEtriangularelementse(e= 13 for(j=1;jn;j++) Hencewemayusethefollowingpseudo-codetocalculateK: 1,...,E).TheneachentryofKmaybeobtainedfromthefollowingformula: for(i=1;in;i++)f Kji=RrPj(x):rPi(x)dx==PEe=1RerPj(x):rPi(x)dx: K(j,i)=0 for(j=1;jn;j++) Theorderoftheloopscaneasilybere-arranged: g. for(e=1;ee;e++) for(i=1;in;i++) K(j,i)=K(j,i)+RerPj(x):rPi(x)dx for(e=1;ee;e++) for(j=1;jn;j++) K(j,i)=0 NowwecanmakeuseofthesparsitycausedbythelocalnatureofP1;:::;PN: for(j=1;jn;j++) for(i=1;in;i++) for(i=1;in;i++) K(j,i)=0 K(j,i)=K(j,i)+RerPj(x):rPi(x)dx: for(e=1;ee;e++) for(j=1;j3;j++)f j=numberofnodewhichisj-thvertexofelemente g. for(i=1;i3;i++)f gk(j,i)=k(j,i)+rerpj(x):rpi(x)dx i=numberofnodewhichisi-thvertexofelemente Atthispointwecanmakethefollowingobservations.

28 CHAPTER1.INTRODUCTION Itisnecessarytonumbertheverticesofeachelement,e,ofthetriangulation of;1,2,3.also,anintegerarray,\icon"say,needstobesetupwhichstores thenodenumberofeachvertexofeachelement. 14 Itisalsonecessarytostoreanarrayofthepositionvectors,sjsay,ofthe mentpseudo-codeshouldnowlooksomethinglike: Now,ifweassumethatthenEnodeson?1arenumberedlast,thentheniteele- Asimilararrangementcanbemadeinordertocalculatef,where fj=rf(x)pj(x)dx=pee=1ref(x)pj(x)dx: verticesofthemesh. for(j=1;jnb;j++)f f(j)=0 for(i=1;inb;i++) for(e=1;ee;e++) gfor(j=1;jne;j++) a(nb+j)=ue(s(nb+j;1);s(nb+j;2)) K(j,i)=0 for(j=1;j3;j++)f j=icon(e,j) if(jnb)f f(j)=f(j)+ref(x)pj(x)dx for(i=1;i3;i++)f i=icon(e,i) if(inb) ggelsef(j)=f(j)-a(i)rerpj(x):rpi(x)dx K(j,i)=K(j,i)+RerPj(x):rPi(x)dx Solvethesystem:Ka=f: gtheparallelgenerationandsolutionofthissystemwillbediscussedinx1.5.

29 CHAPTER1.INTRODUCTION 1.4Time-DependentProblems:TheLinearDiffusionEquation 15 Wenowmoveontoconsiderhowwemaygeneralisetheabovetheorytodealwitha lineartime-dependentdierentialequation.thesimplestparabolictime-dependent dierentialequationisthelineardiusionequation: andsomeboundaryconditions,suchas u(x;0)=u0(x)forallx2; (1.6) andthelaplacianoperator,r2,isassumedtoapplyonlytothesespatialvariables Notethatinaboveallthespatialvariableshavebeengroupedtogetherasx notthesame. variabletshouldnotbethoughtofasbeing\justanotherindependentvariable", likexandysay,becausetheboundaryconditionsassociatedwiththisvariableare wouldliketocomputethesolutionforarbitraryvaluesoftwhicharelessthant theothervariableswherewegenerallyknowaboutthebehaviourofthesolution (wehavenoideaaboutthebehaviourofthesolutionattimet).thisdiersfrom throughouttheboundaryofthespatialdomain. Asfaras`t'isconcernedweonlyknowthesolutionattheboundaryt=0and whichtreatsthespatialvariablesandtimevariableindependently.fortunatelythe methodoflinesexactlydoesthesame TheMethodofLines Keepinginmindthespecialnatureofthevariable`t'weneedapracticalmethod dierentialequations(odes),byonlydiscretisinginspaceintherstinstance. ThisisageneralmethodwhichreducesasystemofPDEstoasystemofordinary

30 CHAPTER1.INTRODUCTION ofspatialdiscretisations(e.g.niteelementornitevolume)tobeusedwithany standardodesolver(see[9],forexample).weattempttofollowthisapproach Thespatialandtemporaldiscretisationarethusindependent,allowingavariety 16 Thisyieldsthefollowingsystemofequations, toobtaintransientsolutionsusingtheniteelementmethodpresentedinprevious section. multiplyequation(1.6)byatestfunction,pj(x),whichhasnotimedependence. Thismeansweonlytriangulatethespatialpartofthedomain,andthenwe andmakinguseofthedivergencetheoremasbeforethisbecomes, Intwodimensionswemayagaindividethedomain,,intotrianglesandnumber theverticesofthesetrianglesfrom1ton=nb+newherenbandneareas ofthemesh.sinceweareinterestedinatime-dependentniteelementsolutionwe (1.7) seekanapproximation,u(x;t),tothetruesolution,u(x,t),oftheform ofnbequationsfornbunknowns(inthiscasea1(t);:::;anb(t)).thissystemis Dirichletboundarycondition,u=uE,fori=nB+1,...,nB+nE. whereai(t)areunknown(tobedetermined)fori=1,...,nb,andaregivenbythe Now,replacingubyuinequations(1.7)forj=1,...,nBweagainobtainasystem u=nxi=1ai(t)pi(x); givenbynbxi=1dai Asbeforewemayexpressthisinmatrixnotation,inwhichcaseitbecomes, dtzpipjdx=?nbxi=1aizrpi:rpjdx+z?2gpjds+zfpjdx? i=nb+1aizrpi:rpjdx?nb+ne nb+ne X Mda dt=?ka+f(t): i=nb+1dai XdtZPiPjdx: (1.8)

31 Mji=RPiPjdx).Inthiscasethevectorfdependsupontthroughthepossible CHAPTER1.INTRODUCTION andthematrixmisknownasthe\galerkinmassmatrix"(withentriesgivenby Again,Kisthe\globalstinessmatrix"(whoseentriesaregivenbyKji=RrPi:rPjdx) 17 dependenceofthefunctionfinequation(1.6)upont,orthepossibledependence ofthedirichletboundaryconditionupont(throughthefunctionue). in[9])fordealingwithequationssuchastheseinanecientmanner(i.e.using system,itisasystemofnbordinarydierentialequationsforwhichwecaneasily manystandardtechniques(e.g.thesoftwarepackagesprintwhichisdescribed obtaininitialvaluesfortheunknownsai(t)(fromthefunctionu0(x)).thereare Itshouldbenoticedthatthesystemofequations,(1.8),isnotanalgebraic localerrorthroughadaptivetime-stepping).nevertheless,ateachtimestepanite elementcalculationssimilartothatdescribedinx1.3mustbeundertaken. memoryandspeedofaserialmachinestarttobecomeaserialbottleneck.also thesystemofequations(1.5)canbeeasilyandquicklysolvedonaserialmachine. Butwhenthenumberofdegreesoffreedomisinexcessofamillionorsothenthe Forasmallproblemwherethenumberofdegreesoffreedomisjustafewthousand 1.5ParallelFiniteElementandLoad-Balancing Byusingsuchamachinenotonlycanwehopetosolvelargerproblems(e.g.in forsomeapplicationswherethesizeoftheproblemisnotsobigthetimetaken structuralmechanics)butwecanalsohopetosolvethemmorequickly. byaserialmachinemaystillbeverylarge(fornon-linearproblemsforexample, wheretheiterativemethodsforsolvingthecorrespondingsystem(1.5)arequite sparsesystemofequations(1.5)inparallel.letussupposethedomainhasbeen expensive).inthesecasesapromisingwayforwardistouseaparallelarchitecture. dividedintonsubdomains1,2,...,nandtheithsubdomainihasbeen ontheinterfacebetweenthesubdomainsarelabelledaandtheunknownsinside assignedtotheithprocessorofaparallelmachine.letusassumethattheunknowns Intherestofthissectionwediscussamethodforassemblingandsolvingthe writtenintheform, a1thenina2,a3,...,anandlastlyinathenthesystemofequations(1.5)canbe eachsubdomainsarelabelleda1,a2,...,an.ifwerstnumbertheunknownsin

32 CHAPTER1.INTRODUCTION 264 A 1A2: C2 C1 18 B1B2::BnA: AnCn : a a2 a : = 2 64fn f2 f f : 1 whereai,bi,ciandaarethemselvesusuallysparse.itisclearfromthedenition ofthebasisfunctionspjthatfi,ai,biandciaretotallyhousedbytheithprocess 3 75; (1.9) distributedacrossdierentprocessors.eachprocessorcancomputeandassemble itsowncontributiontothem,independently,storingthemintheblocksfiandai say(sothatf=f1+f2+...+fnanda=a1+a2+...+an). andhencecanbeassembledindependentofeachotherinparallel.butfandaare form: Inordertosolvethesystemofequations(1.9)werstwriteitincomponent Ifwesubstitutethevalueofaifromequation(1.10)in(1.11)wegetthefollowing equation: Aiai+Cia=fi;i=1;2;:::;n; XiBiai+Aa=f: i(fi?cia)+aa=f; (1.11) IfwedeneAsbytheequation, Onsimplicationthisreducesto, (A?XiBiA?1 ici)a=f?xibia?1 ifi: (1.13) (1.12) thentheequation(1.13)cansimplybewrittenas, Asa=f?XiBiA?1 As=A?XiBiA?1 ici; ifi: (1.14) parallelmachinesbecauseeachsysteminequation(1.10)isentirelyindependent Ifequation(1.15)isthensolvedforathenthiscanbesubstitutedintoequation (1.10)andsolvedforaiforalli.Thisapproachisidealfordistributedmemory

33 CHAPTER1.INTRODUCTION tosolveequation(1.15)thenitisnotnecessarytoexplicitlyformthematrixas aniterativemethod,suchastheconjugategradient(cg)algorithm([40]),isused andmaythereforebesolvedinparallelwiththeotherswhenrequired.moreover,if 19 sowehave wherepisthedirectionvectorobtainedfromtheresidualofthekthiteratesofa, of(1.14).themainstepinvolvedisthematrixvectormultiplicationofw=asp multiplicationandsubdomainsolves(somelocalcommunicationisalsorequired betweenprocessorssharinginterpartitionboundaryvertices). Fromequation(1.16)itisclearthatwcanbeobtainedusingonlymatrix-vector Fromabovediscussionitisclearthatthecommunicationoverheadisproportionaltothenumberofverticesontheinterpartitionboundary,henceoneshould w=ap?xibi(a?1 i(cip)): processorswillbeidlewhileothersarestillbusysolvingtheirsystems). eachsubdomainwilltrytosolvetheequation(1.10)inparallel,henceitisdesirable thatthenumberofunknownsineachofaiisapproximatelysame(otherwisesome trytokeepthisboundaryassmallaspossible.alsooncethevectoraisknown havetwomainfeatures, eachprocessorshouldstoreapproximatelythesamenumberofverticesor Hencethedecompositionoftheelementsofthemeshintosubdomainsshould numberofverticeswhichlieontheboundarybetweentheprocessorsshould elements(toensureequalload), andthatapairofnodesisconnectedbyanedgeonlyifthecorrespondingelements areneighboursofeachother,thenaboveproblembecomesaspecialcaseofamore Thedualgraphofagivenmeshisobtainedbyreplacingeachelementbyanode, Inordertoachievetheabovewerstdenethedualgraphofagivenmesh. bekeptlow. beanundirectedgraphwherenisthesetofnodeswithknknodesandeisthe setofedgeswithkekedges,partitionnintonsubsets,n1,n2,...,nnsuchthat generalproblem,namelythegraphpartitioningproblem. Ni\Nj=;fori6=j,kNik=kNk/nandSiNi=N,andthenumberofedges Then-waygraphpartitioningproblemisdenedasfollows:LetG=G(N,E)

34 obtaina2-waypartitionofn,andthenwefurthersubdivideeachpartusing2- partitionproblemismostfrequentlysolvedbyrecursivebisection.thatis,werst CHAPTER1.INTRODUCTION ofewhoseincidentverticesbelongtodierentsubsetsisminimised.then-way 20 of2-waypartitionsorbisections. problemofperformingan-waypartitionisreducedtothatofperformingasequence waypartitions.afterlognphases,graphgispartitionedintonparts.thus,the,isnotsolvableinpolynomialtime.itisinfactannp-hardproblem([22,36,68]). Neverthelessthereareheuristicapproacheswhichperformwellinmostcases.In thenextfewsectionswereviewsomeofthemoreimportantoftheseheuristics. Unfortunatelythisproblem,whichiswell-knowninthegraphtheoryliterature 1.6RecursiveGraphPartitioningHeuristics concentrateonbisectingthegraphsubjecttotheloadbalancingandcut-weight Forthesakeofsimplicity(asmentionedabove),manygraphpartitioningheuristics nature,butthecorrespondingdisadvantageisthatthetotalnumberofsubdomains approachisthatitiseasytoimplementinparallelbecauseofthedivideandconquer canbeappliedrecursivelyontherecentsubdomains.themainadvantageofthis imisationconstraints.whenmorethantwosubdomainsarerequired,theprocedure (thenumberofedgesontheinter-partitionboundaryiscalledthecut-weight)min- thusproducedmustbeapowerof RecursiveCoordinateBisection(RCB) twoorthree-dimensionalcoordinatesavailableforthenodes.asimplebisection LetG=G(N,E)beagivenundirectedgraph.Wemustalsoassumethatthereare strategy,duetosimon([90]),whichisaslightgeneralisationofanearliermethod maininghalfareassignedtotheothersubdomain. usedbywilliamsin[114],forthegraphgistodeterminethecoordinatedirection ofthelongestexpansionofthedomain.withoutanylossofgenerality,assumethat Halfofthenodeswithsmallx-coordinateareassignedtoonesubdomain,there- thisisthex-direction.thenallnodesaresortedwithrespecttotheirx-coordinate. doesnottakeadvantageoftheconnectivityinformationgivenbythegraph.itis Althougheasytoprogram,theprincipaldrawbackofRCBisthatthemethod

35 CHAPTER1.INTRODUCTION methodisnotgenerallysuitableforourpurpose. thereforeunlikelythattheresultingpartitionwillhavealowcut-weightandsothis RecursiveInertialBisection(RIB) ThismethodisageneralisationofRCBtechniqueswhichisdescribedin[28,75] eithersideofit).thisprocedureisthenrepeatedrecursivelyforeachsubdomain. makingacutwhichisorthogonaltothisaxis(withapproximatelyequalweightson ofinertiaforthesepointmassesisthencalculatedandthedomainisbisectedby locatedatthecentroidoftheircorrespondinginitialelement.theprincipalaxis forexample.here,theverticesofthedualgraphareconsideredaspointmasses HeretheideaistousethegraphdistanceasopposedtoEuclideandistanceusedin arelativelyhighcut-weight([28]) RecursiveGraphBisection(RGB) Thismethodisextremelyfast,butliketheRCBitalsoproducespartitionswith d(ni,nj)=numberofedgesintheshortestpathconnectingniandnj. x1.6.1.recallthatthegraphdistancebetweenthetwonodesniandnjisgivenby whichareclosetothisextremenodeareplacedinonesubdomainandtheremaining halfareplacedintheothersubdomain. nodesaccordingtotheirdistancefromoneoftheextremenodes.halfthevertices thepseudo-diameter)ofthegraph(seegeorgeandliu([37]))andthensortthe Herethestartingpointistondthediameter(or,sincethisisexpensivetond, thatatleastoneofthetwosubdomainsisconnected.butitisstillpossiblethat theothersubdomainmaynotbeconnected.hencewemayendupwithasituation inwhichnotallofthesubdomainsareconnected. Ifwestartoutwithaconnectedgraphthenbyconstructionitisguaranteed 1.6.4ModiedRecursiveGraphBisection(MRGB) In[50]HodgsonandJimackpresenttheirowngraphbisectionmethodMRGB.This byattemptingtoproducesubdomainswhichareallsimplyconnected. methodisamodicationofthergbmethod,whichtriestoimproveontheoriginal InMRGBeachbisectionbeginsbyndingtwoapproximatelyextremalnodes

Schneps, Leila; Colmez, Coralie. Math on Trial : How Numbers Get Used and Abused in the Courtroom. New York, NY, USA: Basic Books, 2013. p i.

Schneps, Leila; Colmez, Coralie. Math on Trial : How Numbers Get Used and Abused in the Courtroom. New York, NY, USA: Basic Books, 2013. p i. New York, NY, USA: Basic Books, 2013. p i. http://site.ebrary.com/lib/mcgill/doc?id=10665296&ppg=2 New York, NY, USA: Basic Books, 2013. p ii. http://site.ebrary.com/lib/mcgill/doc?id=10665296&ppg=3 New

More information

San$Diego$Imperial$Counties$Region$of$Narcotics$Anonymous$ Western$Service$Learning$Days$$ XXX$Host$Committee!Guidelines$ $$

San$Diego$Imperial$Counties$Region$of$Narcotics$Anonymous$ Western$Service$Learning$Days$$ XXX$Host$Committee!Guidelines$ $$ SanDiegoImperialCountiesRegionofNarcoticsAnonymous WesternServiceLearningDays XXXHostCommitteeGuidelines I. Purpose ThepurposeoftheWesternServiceLearningDays(WSLD)XXXHostCommittee(HostCommittee)isto organize,coordinateandproducethewsldxxxeventwithinthe6weekperiodof3weekspriortotheendof

More information

programsitproduces.finally,weshowhowtoproduceecient,optimizingprogramgeneratorsby

programsitproduces.finally,weshowhowtoproduceecient,optimizingprogramgeneratorsby TopicsinOnlinePartialEvaluation TechnicalReport:CSL-TR-93-563 (alsofusememo93-14) March,1993 ErikRuf DepartmentsofElectricalEngineering&ComputerScience ComputerSystemsLaboratory Partialevaluationisaperformanceoptimizationtechniqueforcomputerprograms.Whenaprogram

More information

TABLE OF CONTENTS CHAPTER TITLE PAGE

TABLE OF CONTENTS CHAPTER TITLE PAGE viii TABLE OF CONTENTS CHAPTER TITLE PAGE TITLE PAGE DECLARATION DEDICATION ACKNOWLEDGEMENT ABSTRACT ABSTRAK TABLE OF CONTENTS LIST OF TABLES LIST OF FIGURES LIST OF APPENDICES I II III IV VI VII VIII

More information

Data Security at the KOKU

Data Security at the KOKU I. After we proposed our project to the central registration office of the city of Hamburg, they accepted our request for transferring information from their birth records. Transfer of all contact details

More information

DataIntegrationwithXMLandSemanticWeb Technologies

DataIntegrationwithXMLandSemanticWeb Technologies DataIntegrationwithXMLandSemanticWeb Technologies Athesispresented by RubénTous Submittedinpartialfullmentoftherequerimentsfor DoctorateinComputerScienceandDigitalCommunication thedegreeofdoctorofphilosophy

More information

E-MAIL - Spam Filtering

E-MAIL - Spam Filtering E-MAIL - Spam Filtering SPAM FILTERING OVERVIEW... USER S GUIDE...... Common Login Errors......... Lists Tab... 5 Profile Tab... 7 Quarantine Tab... 0 LOGGING IN MAIN SCREEN SPAMBLOCKER TABS (LISTS, PROFILE

More information

PG DIPLOMA IN GLOBAL STRATEGIC MANAGEMENT LIST OF BOOKS*

PG DIPLOMA IN GLOBAL STRATEGIC MANAGEMENT LIST OF BOOKS* PG DIPLOMA IN GLOBAL STRATEGIC MANAGEMENT LIST OF BOOKS* Paper I: INTERNATIONAL BUSINESS ENVIRONMENT Global Business Environment (ICMR Publication Textbook) [Ref. No: GBE 11 2K4 23] [ISBN: 81-7881-693-8]

More information

TITLE 9. HEALTH SERVICES CHAPTER 1. DEPARTMENT OF HEALTH SERVICES ADMINISTRATION ARTICLE 4. CODES AND STANDARDS REFERENCED

TITLE 9. HEALTH SERVICES CHAPTER 1. DEPARTMENT OF HEALTH SERVICES ADMINISTRATION ARTICLE 4. CODES AND STANDARDS REFERENCED TITLE 9. HEALTH SERVICES CHAPTER 1. DEPARTMENT OF HEALTH SERVICES ADMINISTRATION ARTICLE 4. CODES AND STANDARDS REFERENCED R9-1-412. Physical Plant Health and Safety Codes and Standards A. The following

More information

A STUDY OF THE IMPACT OF CONSTRUCTION ACCIDENTS ON THE PROJECT CONTINUITY

A STUDY OF THE IMPACT OF CONSTRUCTION ACCIDENTS ON THE PROJECT CONTINUITY A STUDY OF THE IMPACT OF CONSTRUCTION ACCIDENTS ON THE PROJECT CONTINUITY Final Project Report as one of requirement to obtain S1 degree of Universitas Atma Jaya Yogyakarta By: KARTIKA IRIANTHY ZEBUA NPM.

More information

2010 MegaPath Inc. All rights reserved. Confidential and Proprietary 2

2010 MegaPath Inc. All rights reserved. Confidential and Proprietary 2 Polycom SIP Phones Base Configuration for MegaPath Duet Hosted Voice Service Contents Summary... 3 Configuration Settings Overview... 3 Restarting the Phone... 5 Rebooting the Phone... 6 Uploading Log

More information

Learn AX: A Beginner s Guide to Microsoft Dynamics AX. Managing Users and Role Based Security in Microsoft Dynamics AX 2012. Dynamics101 ACADEMY

Learn AX: A Beginner s Guide to Microsoft Dynamics AX. Managing Users and Role Based Security in Microsoft Dynamics AX 2012. Dynamics101 ACADEMY Learn AX: A Beginner s Guide to Microsoft Dynamics AX Managing Users and Role Based Security in Microsoft Dynamics AX 2012 About.com is a Rand Group Knowledge Center intended to provide our clients, and

More information

"Charting the Course... MOC 20409 B Server Virtualization with Windows Hyper-V and System Center. Course Summary

Charting the Course... MOC 20409 B Server Virtualization with Windows Hyper-V and System Center. Course Summary Description Course Summary This five day course will provide you with the knowledge and skills required to design and implement Microsoft Server solutions using Hyper-V and System. Objectives At the end

More information

TERMS OF REFERENCE FINANCIAL CONSULTING FIRM 6 MONTHS, NATIONAL

TERMS OF REFERENCE FINANCIAL CONSULTING FIRM 6 MONTHS, NATIONAL TERMS OF REFERENCE FINANCIAL CONSULTING FIRM 6 MONTHS, NATIONAL 1. Background Financial statement of the social insurance offices consists of 8 main financial statements and these financial statements

More information

Computer Studies/Information and Communications Technology (ICT)

Computer Studies/Information and Communications Technology (ICT) Page 1 of 10 Professional Master of Education Subject Declaration Form IMPORTANT This declaration form should be returned to the PME provider(s) to which you have applied or the Postgraduate Applications

More information

THE FIRST SCHEDULE (See rule 7) Table I - FEES PAYABLE

THE FIRST SCHEDULE (See rule 7) Table I - FEES PAYABLE Number of entry On what payable Number of the relevant Form THE FIRST SCHEDULE (See rule 7) Table I - FEES PAYABLE Natural For e-filing Small entity, alone or with natural Others, alone or with natural

More information

CalREDIE Browser Requirements

CalREDIE Browser Requirements CalREDIE Browser Requirements Table of Contents Section 1: Browser Settings... 2 Section 2: Windows Requirements... 11 Section 3: Troubleshooting... 12 1 Section 1: Browser Settings The following browser

More information

Online submission of account of receipts and utilization of Foreign Contribution for the year in FC-6 Form

Online submission of account of receipts and utilization of Foreign Contribution for the year in FC-6 Form Online submission of account of receipts and utilization of Foreign Contribution for the year in FC-6 Form Instructions for online filing of annual account in FC-6 Form 1. For online filing of annual accounts,

More information

Continuity Plan Template for Non-Federal Governments

Continuity Plan Template for Non-Federal Governments Continuity Plan Template for Non-Federal Governments [Department/Agency/Organization Name] [Month Day, Year] [Department/Agency/Organization Name] [Street Address] [City, State Zip Code] [Department/Agency/Organization

More information

Software Monthly Maintenance (Non Accounting Use) Quick Reference Guide

Software Monthly Maintenance (Non Accounting Use) Quick Reference Guide Software Monthly Maintenance (Non Accounting Use) Quick Reference Guide When not using the accounting within the software the system will build up information that will affect the performance and speed

More information

Performance Objective Identification Worksheet

Performance Objective Identification Worksheet Performance Objective Identification Worksheet INSTRUCTIONS For each of the performance objective identified, you must indicate by page and paragraph number the equivalent performance objective in your

More information

EIGHTH DIRECTIVE OF DEPUTY RECEIVER (IMPLEMENTING THE HOW/HWC PLAN OE LIQUIDATION)

EIGHTH DIRECTIVE OF DEPUTY RECEIVER (IMPLEMENTING THE HOW/HWC PLAN OE LIQUIDATION) COMMONWEALTH OF VIRGINIA STATE CORPORATION COMMISSION COMMONWEALTH OF VIRGINIA at the Relation of the STATE CORPORATION COMMISSION, Applicant, v. CASE NO. INS-1994-00218 HOW INSURANCE COMPANY, A RISK RETENTION

More information

1 of 7 31/10/2012 18:34

1 of 7 31/10/2012 18:34 Regulatory Story Go to market news section Company TIDM Headline Released Number Ironveld PLC IRON Holding(s) in Company 18:01 31-Oct-2012 0348Q18 RNS Number : 0348Q Ironveld PLC 31 October 2012 TR-1:

More information

1. Who can use Agent Portal? 2. What is the definition of an active agent? 3. How to access Agent portal? 4. How to login?

1. Who can use Agent Portal? 2. What is the definition of an active agent? 3. How to access Agent portal? 4. How to login? 1. Who can use Agent Portal? Any active agent who is associated with Future Generali Life Insurance Company Limited can logon to Agent Portal 2. What is the definition of an active agent? An agent, whose

More information

Declaration to be submitted by directors in the Applicant Company 1

Declaration to be submitted by directors in the Applicant Company 1 Form SNBFI/D1 Name of the Applicant Company: Declaration to be submitted by directors in the Applicant Company 1 1. Personal Details 1.1 Full name: 1.2 National Identity Card number: 1.3 Passport number:

More information

THE PSYCHOLOGY CLUB EASTERN CONNECTICUT STATE UNIVERSITY CONSTITUTION. Article I: Name. Article II: Purpose

THE PSYCHOLOGY CLUB EASTERN CONNECTICUT STATE UNIVERSITY CONSTITUTION. Article I: Name. Article II: Purpose THE PSYCHOLOGY CLUB EASTERN CONNECTICUT STATE UNIVERSITY CONSTITUTION Article I: Name The name of this organization shall be called the Psychology Club of Eastern Connecticut State College. Article II:

More information

How To Get A Financial Aid Award In Athena

How To Get A Financial Aid Award In Athena Athena Self-Service Walkthrough By the UGA Office of Student Financial Aid 220 Holmes/Hunter Academic Building Athens, GA 30602-6114 Phone: (706) 542-6147 Section 1 How to Find Out What is Needed by the

More information

B1 Project Management 100

B1 Project Management 100 Assignment of points B1 Project Management 100 Requirements for Design Presentation Meetings and Proposal Submissions for Key to Project Management Design Presentation Meeting and Proposal Submissions

More information

Workflow Administration of Windchill 10.2

Workflow Administration of Windchill 10.2 Workflow Administration of Windchill 10.2 Overview Course Code Course Length TRN-4339-T 2 Days In this course, you will learn about Windchill workflow features and how to design, configure, and test workflow

More information

ACCUPLACER Arithmetic & Elementary Algebra Study Guide

ACCUPLACER Arithmetic & Elementary Algebra Study Guide ACCUPLACER Arithmetic & Elementary Algebra Study Guide Acknowledgments We would like to thank Aims Community College for allowing us to use their ACCUPLACER Study Guides as well as Aims Community College

More information

TABLE OF CONTENTS CHAPTER DESCRIPTION PAGE

TABLE OF CONTENTS CHAPTER DESCRIPTION PAGE vii TABLE OF CONTENTS CHAPTER DESCRIPTION PAGE TITLE DECLARATION DEDICATION ACKNOWLEDGEMENT ABSTRACT ABSTRAK LIST OF CONTENTS LIST OF FIGURES LIST OF TABLES LIST OF APPENDICES i ii iii iv v vi vii xii

More information

Contents. 1 Introduction. 2 Feature List. 3 Feature Interaction Matrix. 4 Feature Interactions

Contents. 1 Introduction. 2 Feature List. 3 Feature Interaction Matrix. 4 Feature Interactions 1 Introduction 1.1 Purpose and Scope................................. 1 1 1.2 Organization..................................... 1 2 1.3 Requirements Notation............................... 1 2 1.4 Requirements

More information

ADSL Router Quick Installation Guide Revised, edited and illustrated by Neo

ADSL Router Quick Installation Guide Revised, edited and illustrated by Neo ADSL Router Quick Installation Guide Revised, edited and illustrated by Neo A typical set up for a router PCs can be connected to the router via USB or Ethernet. If you wish to use a telephone with the

More information

Financial Health and Funding in Colleges

Financial Health and Funding in Colleges Report by the Comptroller and Auditor General Managing Finances in English Further Education Colleges Ordered by the House of Commons to be printed 2 May 2000 LONDON: The Stationery Office 0.00 HC 454

More information

RadiativeHeatingandCooling incircumstellarenvelopes Dipl.-Phys.PeterWoitke ausberlin von zurerlangungdesakademischengrades dertechnischenuniversitatberlin VomFachbereich04(Physik) DoktorderNaturwissenschaften(Dr.rer.nat.)

More information

CHAPTER 42A. Case management of certain personal injuries actions. 42A.1. (1) Subject to paragraph (3), this Chapter applies to actions

CHAPTER 42A. Case management of certain personal injuries actions. 42A.1. (1) Subject to paragraph (3), this Chapter applies to actions CHAPTER 42A Case management of certain personal injuries actions Application and interpretation of this Chapter 42A.1. (1) Subject to paragraph (3), this Chapter applies to actions proceeding as ordinary

More information

This Version Not For Distribution EMR/EHR

This Version Not For Distribution EMR/EHR This Version Not For Distribution EMR/EHR Cheng B Saw, Ph.D. Chair, Asian-Oceanic Affairs of AAPM Director - Physics, Northeast Radiation Oncology Centers President, CBSaw Publishing, LLC Harrisburg, PA,

More information

BSTP SRF 81 Medical Emergency and Work Place Injury

BSTP SRF 81 Medical Emergency and Work Place Injury Page 1 of 5 Table of Contents I. PRINCIPLE (Purpose):...2 II. ROLE:...2 III. GLOSSARY, ABBREVIATIONS OR DEFINITIONS:...2 IV. INDICATIONS (Policy):...2 V. SPECIMENS (Samples):...2 VI. MATERIALS, REAGENTS,

More information

ISSUING THE AIR OPERATOR CERTIFICATE, OPERATIONS SPECIFICATIONS, AND COMPLETING THE CERTIFICATION REPORT

ISSUING THE AIR OPERATOR CERTIFICATE, OPERATIONS SPECIFICATIONS, AND COMPLETING THE CERTIFICATION REPORT ISSUING THE AIR OPERATOR CERTIFICATE, OPERATIONS SPECIFICATIONS, AND COMPLETING THE CERTIFICATION REPORT GUIDANCE MATERIAL FOR INSPECTORS CA AOC-017 AIR OPERATOR CERTIFICATION RECORD OF AMENDMENTS Amendment

More information

Encrypting with BitLocker for disk volumes under Windows 7

Encrypting with BitLocker for disk volumes under Windows 7 Encrypting with BitLocker for disk volumes under Windows 7 Summary of the contents 1 Introduction 2 Hardware requirements for BitLocker Driver Encryption 3 Encrypting drive 3.1 Operating System Drive 3.1.1

More information

FIRST QUESTIONNAIRE ON OTHER CRA PRODUCTS

FIRST QUESTIONNAIRE ON OTHER CRA PRODUCTS FIRST QUESTIONNAIRE ON OTHER CRA PRODUCTS The Board of the International Organization of Securities Commissions (IOSCO) has approved a project specification for its Committee 6 on Credit Rating Agencies

More information

NATIONAL UNIVERSITY OF SCIENCE AND TECHNOLOGY

NATIONAL UNIVERSITY OF SCIENCE AND TECHNOLOGY NATIONAL UNIVERSITY OF SCIENCE AND TECHNOLOGY FACULTY OF COMMERCE GENERAL MASTERS IN BUSINESS ADMINISTRATION MANAGERIAL ACCOUNTING GMB 562 FINAL EXAMINATION 11 DECEMBER 2003 TIME ALLOWED: 3 HOURS + 30

More information

STATE UNIVERSITY OF NEW YORK COLLEGE OF TECHNOLOGY CANTON, NEW YORK COURSE OUTLINE JUST 201 CRITICAL ISSUES IN CRIMINAL JUSTICE

STATE UNIVERSITY OF NEW YORK COLLEGE OF TECHNOLOGY CANTON, NEW YORK COURSE OUTLINE JUST 201 CRITICAL ISSUES IN CRIMINAL JUSTICE STATE UNIVERSITY OF NEW YORK COLLEGE OF TECHNOLOGY CANTON, NEW YORK COURSE OUTLINE JUST 201 CRITICAL ISSUES IN CRIMINAL JUSTICE Prepared by: Dr. Brian K. Harte SCHOOL OF BUSINESS AND CRIMINAL JUSTICE Department

More information

TABLE OF CONTENT CHAPTER TITLE PAGE TITLE DECLARATION DEDICATION ACKNOWLEDGEMENTS ABSTRACT ABSTRAK

TABLE OF CONTENT CHAPTER TITLE PAGE TITLE DECLARATION DEDICATION ACKNOWLEDGEMENTS ABSTRACT ABSTRAK TABLE OF CONTENT CHAPTER TITLE PAGE TITLE DECLARATION DEDICATION ACKNOWLEDGEMENTS ABSTRACT ABSTRAK TABLE OF CONTENT LIST OF TABLES LIST OF FIGURES LIST OF ABBREVIATIONS LIST OF APPENDICES i ii iii iv v

More information

Installing Globodox Web Client on Windows Server 2012

Installing Globodox Web Client on Windows Server 2012 Installing Globodox Web Client on Windows Server 2012 Make sure that the Globodox Desktop Client is installed. Make sure it is not running. Note: Please click on Allow or Continue for all required UAC

More information

The following report presents financial data only. The full and binding version is available in Polish. K2 INTERNET S.A.

The following report presents financial data only. The full and binding version is available in Polish. K2 INTERNET S.A. The following report presents financial data only. The full and binding version is available in Polish. K2 INTERNET S.A. Annual Financial Statement of K2 Internet S.A. for the twelve-month period ended

More information

E-DISCOVERY IN CANADA

E-DISCOVERY IN CANADA 0 E-DISCOVERY IN CANADA Todd J. Burke Kelly Friedman Andrew J. McCreary James Morton Susan Nickle Vincenzo Rondinelli Glenn Smith James Swanson Susan Wortzman Foreword by The Honourable Colin L. Campbell

More information

FINAL JOINT PRETRIAL ORDER. This matter is before the Court on a Final Pretrial Conference pursuant to R. 4:25-1.

FINAL JOINT PRETRIAL ORDER. This matter is before the Court on a Final Pretrial Conference pursuant to R. 4:25-1. SUPERIOR COURT OF NEW JERSEY MIDDLESEX COUNTY:LAW DIVISION Docket No. Plaintiff(s), v. Defendant(s). FINAL JOINT PRETRIAL ORDER This matter is before the Court on a Final Pretrial Conference pursuant to

More information

Financial Services (Investment and Fiduciary Services) FINANCIAL SERVICES (FEES) REGULATIONS 2011 FINANCIAL SERVICES (FEES) REGULATIONS 2011

Financial Services (Investment and Fiduciary Services) FINANCIAL SERVICES (FEES) REGULATIONS 2011 FINANCIAL SERVICES (FEES) REGULATIONS 2011 Financial Services (Investment and Fiduciary Services) Legislation made under s. 53 and 56. 1989-47 (LN. ) Commencement 1.4.2011 Amending enactments Relevant current provisions Commencement date LN. 2011/036

More information

DavidWhalley ProfessorDirectingThesis TheodoreBaker. GregoryRiccardi CommitteeMember

DavidWhalley ProfessorDirectingThesis TheodoreBaker. GregoryRiccardi CommitteeMember COLLEGEOFARTSANDSCIENCES THEFLORIDASTATEUNIVERSITY GRAPHICALVISUALIZATION COMPILEROPTIMIZATIONS OF AThesissubmittedtothe MICKEYBOYD By DepartmentofComputerScience requirementsforthedegreeof inpartialfulllmentofthe

More information

To define and explain different learning styles and learning strategies.

To define and explain different learning styles and learning strategies. Medical Office Assistant Program Overview The Medical Office Assistant program prepares students for entry-level employment as a medical office assistant. It discusses the fundamentals of medical terminology,

More information

PMP Certification Exam Prep Bootcamp

PMP Certification Exam Prep Bootcamp Commitment / Vision / Results SM Toll Free (US): (800) 877-8129 Office: (919) 495-7371 Fax: (919) 556-0283 Email: info@cvr-it.com Web site: www.cvr-it.com Course Overview PMP Certification Exam Prep Bootcamp

More information

Department of International Trade at Feng Chia University Master s Program Requirements Policy

Department of International Trade at Feng Chia University Master s Program Requirements Policy Department of International Trade at Feng Chia University Master s Program Requirements Policy Revised and approved by the Department Affairs Committee on June 9 th, 2005 Revised and approved by the Department

More information

Regulatory Story. RNS Number : 8343I. DCD Media PLC. 08 July 2013. TR-1: NOTIFICATION OF MAJOR INTEREST IN SHARES i

Regulatory Story. RNS Number : 8343I. DCD Media PLC. 08 July 2013. TR-1: NOTIFICATION OF MAJOR INTEREST IN SHARES i 1 of 7 25/11/2013 11:51 Regulatory Story Go to market news section Company TIDM Headline Released DCD Media PLC DCD Holding(s) in Company 15:19 08-Jul-2013 8343I15 RNS : 8343I DCD Media PLC 08 July 2013

More information

Computer Apps for Careers (#5000)

Computer Apps for Careers (#5000) AASD BUSINESS CURRICULUM Computer Apps for Careers (#5000) Description Students will use a variety of software to create personal as well as business documents. The course will emphasize touch-keyboarding

More information

LIST OF TABLES. 2.4 Variables Related to CRM Implementation 57. 2.11 Variables in Customer Retention at Commercial Banks 74

LIST OF TABLES. 2.4 Variables Related to CRM Implementation 57. 2.11 Variables in Customer Retention at Commercial Banks 74 LIST OF TABLES Table 2.1 Steps in CRM Development at 51 2.2 CRM system at Commercial 53 2.3 Stages in CRM implementation 55 2.4 Variables Related to CRM Implementation 57 2.5 CRM Acceptance Among the Employees

More information

Microsoft Word 2013 Basics

Microsoft Word 2013 Basics Microsoft Word 2013 Basics 1. From Start, look for the Word tile and click it. 2. The Ribbon- seen across the top of Microsoft Word. The ribbon contains Tabs, Groups, and Commands a. Tabs sit across the

More information

AUSTIN INDEPENDENT SCHOOL DISTRICT INTERNAL AUDIT DEPARTMENT HUMAN RESOURCE AUDIT PROGRAM

AUSTIN INDEPENDENT SCHOOL DISTRICT INTERNAL AUDIT DEPARTMENT HUMAN RESOURCE AUDIT PROGRAM GENERAL: The Human Resources (HR) Department deals with the management of people within the organization. This includes responsibility for the process of hiring staff, recruiting and advertising positions

More information

TechAdvisor & Account Coordinator Salesforce.com Training

TechAdvisor & Account Coordinator Salesforce.com Training TechAdvisor & Account Coordinator Salesforce.com Training 1 TABLE OF CONTENTS I. First-Time User... 3 Log Into Salesforce.com... 3 Setup Salesforce.com... 3 II. Salesforce.com Tabs... 6 Home... 6 Contacts...

More information

Introduction to Windchill Projectlink 10.2

Introduction to Windchill Projectlink 10.2 Introduction to Windchill Projectlink 10.2 Overview Course Code Course Length TRN-4270 1 Day In this course, you will learn how to participate in and manage projects using Windchill ProjectLink 10.2. Emphasis

More information

Cyber security standard

Cyber security standard Cyber security standard Brief description This *Standard specifies security standards that protect *ICT systems and data from unintended or unauthorized access, damage or destruction. Related policies

More information

Dealing with digital Information richness in supply chain Management - A review and a Big Data Analytics approach

Dealing with digital Information richness in supply chain Management - A review and a Big Data Analytics approach Florian Kache Dealing with digital Information richness in supply chain Management - A review and a Big Data Analytics approach kassel IH university press Contents Acknowledgements Preface Glossary Figures

More information

NATIONAL UNIVERSITY OF SCIENCE AND TECHNOLOGY FACULTY OF COMMERCE DEPARTMENT OF INSURANCE AND ACTUARIAL SCIENCE

NATIONAL UNIVERSITY OF SCIENCE AND TECHNOLOGY FACULTY OF COMMERCE DEPARTMENT OF INSURANCE AND ACTUARIAL SCIENCE NATIONAL UNIVERSITY OF SCIENCE AND TECHNOLOGY FACULTY OF COMMERCE DEPARTMENT OF INSURANCE AND ACTUARIAL SCIENCE B.COMM (HONS) DEGREE IN RISK MANAGEMENT AND INSURANCE STRATEGIC MARKETING OF INSURANCE [CIN

More information

GOVERNMENT OF BERMUDA Ministry of Education. Public School Reorganization: A Consultation

GOVERNMENT OF BERMUDA Ministry of Education. Public School Reorganization: A Consultation GOVERNMENT OF BERMUDA Ministry of Education Public School Reorganization: A Consultation March 2015 GOVERNMENT OF BERMUDA Ministry of Education 14 Waller s Point Road St. David s DD 03 Bermuda P.O. Box

More information

Microsoft Word 2010 Basics

Microsoft Word 2010 Basics Microsoft Word 2010 Basics 1. Start Word if the Word 2007 icon is not on the desktop: a. Click Start>Programs>Microsoft Office>Microsoft Word 2007 b. The Ribbon- seen across the top of Microsoft Word.

More information

Hi iv. Declaration Certificate Acknowledgement Preface. List o f Table. List o f Figures. viii xvi xvii. 1.1 Introduction 1

Hi iv. Declaration Certificate Acknowledgement Preface. List o f Table. List o f Figures. viii xvi xvii. 1.1 Introduction 1 Declaration Certificate Acknowledgement Preface. List o f Table. List o f Figures Hi iv V viii xvi xvii Chapter 1 Introduction 1-13 1.1 Introduction 1 1.2 Statement of the Problem 3 1.2.1 Terms used in

More information

Children s Council of the International Technology and Engineering Educators Association

Children s Council of the International Technology and Engineering Educators Association Children s Council of the International Technology and Engineering Educators Association (Children s Council of ITEEA) Original bylaws approved March, 1998 Bylaws updated 2007 Bylaws revised (April 25,

More information

Managing Successful Projects

Managing Successful Projects 2008 AGI-Information Management Consultants May be used for personal purporses only or by libraries associated to dandelon.com network. Managing Successful Projects with PRINCE2 Office of Government Commerce

More information

Life Insurance. Basic Life Insurance. Optional Life Insurance

Life Insurance. Basic Life Insurance. Optional Life Insurance Life Insurance Basic Life Insurance PEIA fers active employees under age 65 a basic $10,000 decreasing term life insurance policy with accidental death and dismemberment (AD&D) benefits. The value this

More information

ZTE Australia Help Guides MF91

ZTE Australia Help Guides MF91 MF91 Gaming & IP Services Configuration Guides Introduction The purpose of this document is to walk you through the setup of your MF91 for Gaming or any other usage that requires transparency of data traffic

More information

1. Access your account Log in to your online account at http://service.ringcentral.com using your main Ring Central phone number and password.

1. Access your account Log in to your online account at http://service.ringcentral.com using your main Ring Central phone number and password. Ring Central Quick Set Up settings for GA Hotline For assistance call National Hotline Committee Chair hotlinechair@trusteewebsite.com Ring Central Support (888) 898-4591 The purpose of this document is

More information

SITE PHOTOGRAPHS (GOOGLE EARTH): ROAD REHABILITATION Refer to Appendix B for a map of the viewpoint locations.

SITE PHOTOGRAPHS (GOOGLE EARTH): ROAD REHABILITATION Refer to Appendix B for a map of the viewpoint locations. 469335: N7 Rehab BA Report_Appendix B5 Page i SITE PHOTOGRAPHS (GOOGLE EARTH): ROAD REHABILITATION Refer to Appendix B for a map of the viewpoint locations. R1 R1 R2 R2 Looking north with Okiep east of

More information

3 BUSINESS ACCOUNTING STANDARD,,INCOME STATEMENT I. GENERAL PROVISIONS

3 BUSINESS ACCOUNTING STANDARD,,INCOME STATEMENT I. GENERAL PROVISIONS APPROVED by Resolution No. 1 of 18 December 2003 of the Standards Board of the Public Establishment the Institute of Accounting of the Republic of Lithuania 3 BUSINESS ACCOUNTING STANDARD,,INCOME STATEMENT

More information

SECTION II. A sample LIFE return can be found on OSFI s website.

SECTION II. A sample LIFE return can be found on OSFI s website. SECTION II General The LIFE return is generally self-explanatory, and accordingly, the text on each page of the Return is considered to be part of the instructions. Additional information to assist in

More information

Worksheet to Review Vector and Scalar Properties

Worksheet to Review Vector and Scalar Properties Worksheet to Review Vector and Scalar Properties 1. Differentiate between vectors and scalar quantities 2. Know what is being requested when the question asks for the magnitude of a quantity 3. Define

More information

Administrative Services

Administrative Services Policy Title: Administrative Services De-identification of Client Information and Use of Limited Data Sets Policy Number: DHS-100-007 Version: 2.0 Effective Date: Upon Approval Signature on File in the

More information

The following report presents financial data only. The full and binding version is available in Polish. K2 INTERNET S.A.

The following report presents financial data only. The full and binding version is available in Polish. K2 INTERNET S.A. The following report presents financial data only. The full and binding version is available in Polish. K2 INTERNET S.A. Annual Financial Statement of K2 Internet S.A. for the twelve-month period ended

More information

CONSTITUTION. with a view to speaking with a single voice on matters of importance, declare their intent to constitute themselves into an association.

CONSTITUTION. with a view to speaking with a single voice on matters of importance, declare their intent to constitute themselves into an association. CONSTITUTION The Graduate Students of the Department of Law of the School of Graduate Studies of the University of Toronto, with a view to speaking with a single voice on matters of importance, declare

More information

THE PERFORMANCE MANAGEMENT GROUP LLC

THE PERFORMANCE MANAGEMENT GROUP LLC THE PERFORMANCE MANAGEMENT GROUP LLC ON-CAMPUS TRAINING: LEAN SIX SIGMA EXCELLENCE IN HEALTHCARE DELIVERY BLACK BELT CERTIFICATION Performance Improvement Training for the Healthcare Industry ABOUT THE

More information

PROJECT MANAGEMENT PROFESSIONAL CERTIFIED ASSOCIATE IN PROJECT MANAGEMENT (PMP & CAPM) EXAM PREPARATION WORKSHOP

PROJECT MANAGEMENT PROFESSIONAL CERTIFIED ASSOCIATE IN PROJECT MANAGEMENT (PMP & CAPM) EXAM PREPARATION WORKSHOP TSE015 PROJECT MANAGEMENT PROFESSIONAL CERTIFIED ASSOCIATE IN PROJECT MANAGEMENT (PMP & CAPM) EXAM PREPARATION WORKSHOP Course Outline I. Introduction and Course Objectives A. About PMI B. PMP and CAPM

More information

STATE UNIVERSITY OF NEW YORK COLLEGE OF TECHNOLOGY CANTON, NEW YORK COURSE OUTLINE EADM 400 INCIDENT COMMAND: SYSTEM COORDINATION AND ASSESSMENT

STATE UNIVERSITY OF NEW YORK COLLEGE OF TECHNOLOGY CANTON, NEW YORK COURSE OUTLINE EADM 400 INCIDENT COMMAND: SYSTEM COORDINATION AND ASSESSMENT STATE UNIVERSITY OF NEW YORK COLLEGE OF TECHNOLOGY CANTON, NEW YORK COURSE OUTLINE EADM 400 INCIDENT COMMAND: SYSTEM COORDINATION AND ASSESSMENT Prepared By: Dr. Michael J. O Connor Jr. SCHOOL OF BUSINSS

More information

How To Build A House

How To Build A House CONTRACT TO BUILD A HOUSE AT I. Contract Parties Now comes (builder s name) (address) and (buyer s name) (address) to hereby agree to build a house on property located at (common address) and legally described

More information

Business Administration of Windchill PDMLink 10.0

Business Administration of Windchill PDMLink 10.0 Business Administration of Windchill PDMLink 10.0 Overview Course Code Course Length TRN-3160-T 3 Days After completing this course, you will be well prepared to set up and manage a basic Windchill PDMLink

More information

elink Guidance Document

elink Guidance Document elink Guidance Document Managing Activities Updated April 2014 Purpose: This document provides guidance on how to manage activities in elink. What is an Activity? An activity in elink is a project, task,

More information

TABLE OF CONTENTS. List of Figures iii. List of Tables v. I. Introduction 1. A. New Technology & EHR Survey 1. B. Types of Electronic Systems 1

TABLE OF CONTENTS. List of Figures iii. List of Tables v. I. Introduction 1. A. New Technology & EHR Survey 1. B. Types of Electronic Systems 1 TABLE OF CONTENTS List of Figures iii List of Tables v Page I. Introduction 1 A. New Technology & EHR Survey 1 B. Types of Electronic Systems 1 C. Types of Optometrists 1 II. Characteristics of Responding

More information

Class Climate Online Evaluation Instructions

Class Climate Online Evaluation Instructions I: GETTING STARTED II: IMPORTING MULTIPLE INSTRUCTORS AND COURSES III: CREATING and/or EDITING INDIVIDUAL USERS IV: ADDING and/or EDITING INDIVIDUAL COURSES V: IMPORTING E-MAIL ADDRESSES (BULK) VI: ADDING

More information

Dealer Tutorial. Uplink Customer Service 1-888-9UPLINK sales@uplink.com. 2010 Uplink Security, LLC. All rights reserved.

Dealer Tutorial. Uplink Customer Service 1-888-9UPLINK sales@uplink.com. 2010 Uplink Security, LLC. All rights reserved. Welcome to the u-traq Dealer Tutorial Uplink Customer Service 1-888-9UPLINK sales@uplink.com 2010 Uplink Security, LLC. All rights reserved. Table of Contents I. Device Overview Introduction to u-traq

More information

PRESS RELEASE. End of press release

PRESS RELEASE. End of press release Unofficial Translation This is an unofficial translation of the press release made below and it has been prepared for information purposes only. In the case of any discrepancy between this translation

More information

REDUCTION IN FORCE, LAYOFF, AND RECALL ADMINISTRATORS, TEACHERS, AND SUPPORT STAFF

REDUCTION IN FORCE, LAYOFF, AND RECALL ADMINISTRATORS, TEACHERS, AND SUPPORT STAFF REDUCTION IN FORCE, LAYOFF, AND RECALL ADMINISTRATORS, TEACHERS, AND SUPPORT STAFF I. Purpose of regulation; definitions A. The purpose of this regulation is to establish a procedure for reduction in force

More information

(Japanese Note) 1. With reference to subparagraph (m) of paragraph 1 of Article 3 of the Convention:

(Japanese Note) 1. With reference to subparagraph (m) of paragraph 1 of Article 3 of the Convention: (Japanese Note) Translation London, February 2,2006 Excellency: I have the honour to refer to the Convention between Japan and the United Kingdom of Great Britain and Northern Ireland for the Avoidance

More information

Testing and Accommodation Enhancements

Testing and Accommodation Enhancements Testing and Accommodation Enhancements CECAS users will notice that a set of changes to the CECAS production (live) site have occurred. The changes are: Child Demographics screen has been updated to contain

More information

Form I: AGENCY'S IDENTIFICATION DATA

Form I: AGENCY'S IDENTIFICATION DATA Form I: AGENCY'S IDENTIFICATION DATA 1 Name 2 Head office (a) Address (b) (c) (d) (e) (f) (g) City Pin E-mail address Phone Nos. with STD code Fax No Contact person's name and designation with mobile number

More information

BREVARD COUNTY MERIT SYSTEM PROCEDURES PROCEDURE II

BREVARD COUNTY MERIT SYSTEM PROCEDURES PROCEDURE II BREVARD COUNTY MERIT SYSTEM PROCEDURES PROCEDURE II Title: PAY PLAN I. PURPOSE AND SCOPE To implement the Pay Plan Policy. II. DEFINITIONS AND REFERENCES Merit System Policy II, Pay Plan. III. AMENDMENTS

More information

Apple Pro Training Series. OS X Server. Essentials. Arek Dreyer. and Ben Greisler

Apple Pro Training Series. OS X Server. Essentials. Arek Dreyer. and Ben Greisler Apple Pro Training Series OS X Server Essentials Arek Dreyer and Ben Greisler Table of Contents Configuring and Monitoring OS X Server Lesson 1 About This Guide 3 Learning Methodology 4 Lesson Structure

More information

Core Training Quick Reference Guide Version 2.0

Core Training Quick Reference Guide Version 2.0 Core Training Quick Reference Guide Version 2.0 Page 1 of 34 Contents Changes from Previous Version... 3 Introduction... 5 Guidance for Professional Users based in Colleges/ Schools/ Departments... 5 Logging

More information

THE PERFORMANCE MANAGEMENT GROUP LLC

THE PERFORMANCE MANAGEMENT GROUP LLC THE PERFORMANCE MANAGEMENT GROUP LLC ONLINE TRAINING: LEAN SIX SIGMA SERVICE EXCELLENCE BLACK BELT CERTIFICATION Performance Improvement Training for Service Industries: Financial Services Telecommunications

More information

TCS Scheduler Copyright 2008

TCS Scheduler Copyright 2008 TCS Scheduler Copyright 2008 I TCS Scheduler Table of Contents Foreword 0 Part I Welcome & Introduction 2 Part II Setup/Configuration 2 Part III Browse Event Lists 4 Part IV View Calendars 5 Part V Calendar

More information

How To Compete At Las Vegas All Stars

How To Compete At Las Vegas All Stars Registration Procedures 1) Submit an e-mail to allstarslasvegas@gmail.com after completing the attached registration form with all necessary information that follows in this packet. Please feel free to

More information

INVESTMENT POLICY. Financial institutions, brokers and similar institutions that desire to do investment business with the College.

INVESTMENT POLICY. Financial institutions, brokers and similar institutions that desire to do investment business with the College. INVESTMENT POLICY I. Policy Section 11.0 Business Functions II. Policy Subsection 11.8 Investment III. Policy Statement The Vice President for Finance and Administration, or designee, shall be responsible

More information