72Online cycle detection and difference propagation for pointer analysis-第2页
上亿文档资料，等你来发现
72Online cycle detection and difference propagation for pointer analysis-2
antisenforcedwhichstates；isinsertedtherearetwocas；1.-Thetwoverticesarealre；correctlyandwedonothing.；2.；-Vertexispositionedbefor；theordering.Toresolvethi；Thealgorithmachievesanam；co
antisenforcedwhichstatesthatifthencomesafterintheorder().Thus,whenanedgeisinsertedtherearetwocasestoconsider:1.-Thetwoverticesarealreadyorderedcorrectlyandwedonothing.2.-Vertexispositionedbeforeintheordering.Toresolvethisadepth-?rstsearchisper-formedstartingfrom,limitedtonodesbetweenandintheorder.Thisuncoversverticesreachablefromwhichshouldnowcomeafter.Thesearethenshiftedpastintheordering.Ifisreachedduringthesearchthenacyclehasbeendetectedandwebackpropagatethisinformationtouncoverthosenodesinvolved.Thealgorithmachievesanamortisedcostoverin-sertionsof,whichisagoodimprovementuponthecomplexityoftheof?inealgorithm.Aproofofthiscanbefoundin[18].Thereaderisreferredto[20]foramoredetailedexaminationofthisalgorithmanditscom-plexity.3.3DifferencePropagationDifferencepropagationisatechnique?rstsuggestedbyFechtandSeidl[13].Theyproposedageneralframeworkforapplyingittodistributiveconstraintsystems.However,thisisunabletodescribeconstraintswhichhaveaderefer-encedvariableonthelefthandside.Thus,thealgorithmweprovidehereisreallyaninstanceoftheirframeworkex-tendedtocopewiththeconstraintsystemofSection2.1.Theroughaimofthetechniqueistoreducethecostofpropagatingthesolutionforanodetoitssuccessors.Inastandardworklistsolver(seee.g.Figure2),propagationalonganedgeoccursbymerginginto.Thisoperationislikelytobelinearinthesizeofthesmallersetand,therefore,reducingthesizeofsetsinvolvedshouldyieldanimprovement.Thekeyidea,then,isrealisingthateachelementofonlyneedstobepropagatedalonganedgeonce.Figure6attemptstoclarifythis.Clearly,forthistoworkasuf?cientnumberofnodesmustbevisitedmorethanonceandtherearethreelikelyreasonswhythiscanhappen:pooriterationorder,complexconstraintsandcycles.Figures3and4provideexamplesofthe?rsttwo.ThenewsolverisgiveninFigure7.Akeycomponentisthedifferenceset,,whichcontainstheapproximatechangeinsolutionforeachnode.Wecanthinkofasthecollectionpointforelementspropagatedto.Itisapproximateasitmaycontainvalueswhicharealreadyin.Eachtimeisvisitedthealgorithmcomputes,theactualchangeinsolution,bytakingthedifferencebe-tweenand.Thisislikelytobesmallerthanprocedureadddfsshiftproceduredfsmarkasforalldoifthenifunmarkeddodfsifmarkedprocedureshiftunmarkfortodoifmarkedthenifmarkedthenpushelsepushelseallocate//placevisitednodesaftertinorderingfortodoallocate//checkifnewcycledetectedifthencycle(a)(b)(c)Figure6.Illustratingunnecessaryworkper-formedbythestandardworklistsolver.Thediagramsshowpartofaconstraintgraphdur-ingconvergence.Weseeaninitialpropaga-tionfromtooccurring(a).Thesolutionforislaterupdated(b)andthisiseventuallypropagatedto(c).Thepointisthat“a”doesnotneedtoberepropagatedin(c).Infact,onlythechangein’ssolutiondoes,whichinthiscaseis.4ExperimentalStudyInthissectionweprovideempiricaldata,overarangeofbenchmarks,ontheruntimesofworklistsolversusingdifferentcombinationsofthetechniquesdescribedinSec-tion3.Thepurposeistofacilitateanunderstandingofhoweffectivethesemethodscanbe.Table1providesinformationonourbenchmarksuite.AllarepartoftheGNUsystemand,asaresult,theirfullsourcecanbeobtainedfromhttp://www.gnu.org.TheSUIF2.0researchcompilerfromStanford[2]wasdeployedasthefrontendforgeneratingconstraintsets.Inallcases,wewereabletocompilethebenchmarkswithonlysuper?cialmodi?cations,suchasaddingextra“#include”directivesformissingstandardlibraryhead-ers.Theconstraintgeneratoroperatesonthefull‘C’lan-guageandafewpointsmustbemadeaboutthis:Heapmodel-Asingleheapobjectperstaticallocationsitewasused.Structs-Allelementsofastructureweremappedtoasingleconstraintvariable.Arrays-Treatedinasimilarfashiontostructs,byig-noringtheindexexpression.proceduresolveforeachdowhiledoselect//computeactualchangeinsolution//processconstraintsinvolvingforeachdocaseof:foreachdoifdoifchangedthen:foreachdoifdoifchangedthen//propagatetosuccessorsofnforeachdoifchangedthen//endwhileFigure7.TheDifferencepropagatingworklistsolver.andareinitialisedthesameasforFigure2.StringConstants-Thesewerepreservedintactandnotignoredorcombinedintoasingleobject.IndirectCalls-Indirectcallswerehandledusingamechanismsimilartoprocessingthecomplexcon-straints.Exactdetailsare,unfortunately,beyondthescopeofthispaper.ExternalLibraryFunctions-These,almostentirely,camefromtheGNUClibraryandweremodelledus-inghandcraftedsummaryfunctions,whichcapturedonlyaspectsrelevanttopointeranalysis.Theresults,listedinTables2and3,weregeneratedona900MhzAthlonbasedmachinewith1Gbofmainmemory,runningRedhat8.0(Pysche).Theexecutableswerecom-piledusinggcc3.2,withoptimisationlevel“-O2”.TimingBenchmarkboolbcgrepuucpbash0.111.062.01.06.12.05L.O.C.2813ConstraintsSet2365#Cycles26/(190)53/(208)49/(349)46/(250)211/(1203)17/(294)52/(342)45/(414)41/(505)201/(2154)Size7.435.031.Table1.TheBenchmarkSuite.L.O.C.reportsnon-blank,non-commentlinesonly.Constraintsare),simple(),Complex(involving‘’)orAdded[duringconvergence].countedasTrivial(Thenumberofcyclesintheinitialand?nalconstraintgraphsareprovided,alongwiththetotalnumberofvariablesinvolvedinacycle(showninbrackets).Thelastcolumnprovidestheaveragesizeofatargetsetinthesolution.booltimebcchessgrepmakeuucpgawkbashWL0.0..1..0WRWDFWSLWSRWDSFWCLWCRWDCF0....0...5170....0.270.190.480.50.550.610.80.0.40.0.0.0.0.160.22958.641..1.349.417..287.97.0512.7WLWR0543131WDF9682963WCLWDCLwasperformedusingthegettimeofdayfunctionandtheimplementationwasinC++,makingextensiveuseoftheStandardTemplateLibraryandBoostLibrary[1].4.1DiscussionWewillnowpointoutkeyfeaturesofthedata,alongwithsomeremarks.Indoingthis,itishelpfultosplitthebenchmarksintotwocategories:thesmallones(bool,time,bc,grepandchess)andthelargeones(make,uucp,gawkandbash).Also,notethatweregardWL,WFandWRasthebaselinesolvers.SelectionStrategy-LookingatTable2andthebaselinealgorithmsWL,WFandWRweobservethatLRFisneveroptimalforthesmallbenchmarks,butisgenerallybestforthelargeones.ThistrendisrepeatedthroughouttheWD,WSandWDSfamiliesand,intheWCandWDCcategories,weseeLRFgainingasigni?cantadvantageoverLIFOandFIFO.Table3suggeststhatLRFalwayshasalowervisitcountthantheotherstrategies.However,wenotethatthisreductionisnotalwayssigni?cantforthesmallerbenchmarks.Comments:TheLRFschemerequirestheuseofapriorityqueue,givingitahigheroverhead.Hence,itseemsreasonabletoconcludethatthiscanoutweighanysmallsavinginvisitcount.CycleDetection-FromthedatainTable2,weseethattheWSalgorithmsalmostalwaysbeattheirbaselinecounterparts.WealsoseefromTable2thatonlinedetectiondoesnotperformwellonthesmallbenchmarks.However,WCRisusuallybetterthanWSRonthelargeones.Inparticular,WCRandWDCRarebyfarthefastestsolversforthetwolargestbenchmarks,gawkandbash.Table3indicatesthatonlinecycledetectiondramaticallyreducesthevisitcount,irrespectiveofbenchmarksize.Comments:Onlinecycledetectionisexpensiveandthusahighreturnisneededtoshowanyimprovementinrun-time.Lookingatvisitcountaloneisnotsuf?cienttoexplainwhyitonlypaysoffonthelargerbenchmarks.Wemustalsoconsidertherelativecostofvisitinganodeandthe?nalcol-umnofTable1attemptstomeasurethis.The?guresshowonlytheaveragesetsizeonceconvergenceiscomplete,butwebelievethisgivesanindicationofthesetsizesinvolvedduringconvergence.Thus,inTable1weobservethatthelargerbenchmarkshaveasigni?cantlygreateraveragesetsize.Wefeelthisimpliesthecostpervisittoberelativelybiggerforthelargerbenchmarksand,hence,moreisgainedbyreducingthenumberofvisits.Theperformanceoftheonlinecycledetector,ingeneral,ispleasingasitdemonstratesthatthetechniquecanbeef?cientandbene?cialoverstaticdetection.DifferencePropagation-Forthelargebenchmarks,theresultsinTable2showthatdifferencepropagationisalwaysanimprovementoverthebaseline.Theexactoppositeholdsforthesmallbenchmarks.Wenotewithcuriositythat,forthetwolargestbenchmarks,WDLandWDFgosigni?cantlyfasterthanWCLandWCF,butthatWDRismuchslowerthanWCR.WealsoseethataddingdifferencepropagationtotheWSfamilyalmostalwaysyieldssomethingbetterforthelargebenchmarks,butthatWDCRisworsethanWCRinallbutone.Finally,Table3showsthatdifferencepropagationcausestheexpectedincreaseinvisitcount.Comments:Again,thistechniqueintroducesanover-headwhichappearsunjusti?edonthesmallerbenchmarks.However,itdoesappeartoshowpromise,althoughitre-mainsunclearwhytheWDCRalgorithmdoesnotdobet-ter.Thedataforbashshowsasizeableincreaseinvisitcount,withonlyasmalldropinperformance.Thissug-geststhatvisitcountmaynotberelevantinexplainingtheproblem.Andso,wecanonlyspeculatethatwithcyclesremoved,therearesimplynotenoughnodesbeingvisitedmorethanonce.ThismaybesupportedbythedatafromTa-ble1,whichindicatesthatthe?nalconstraintgraphshaveahighoutdegreeandthuswillpropagatetheirsolutiontomanynodesinasinglevisit.5ConclusionWehaveexploredtheuseofsomeexistingandsomeoriginalmethodsforimprovingtheconvergencetimeofinclusion-constraintsolversinthe?eldofpointeranalysis.Ourresultsindicatethatworklistselectionstrategyisim-portant,thatonlinecycledetectionisfeasibleandeffectiveand,?nally,thatdifferencepropagationhaspotential.Inthefuturewehopetolookatlargerbenchmarks,in-vestigatethedifferencepropagationtechniquefurther,addanimplementationoftheHeintze-Tardieusolverandintro-duce?ow-sensitivity.6AcknowledgementsDavidJ.PearceissupportedbyaPhDstudentshipfromtheU.K.EngineeringandPhysicalSciencesResearchCouncil(EPSRC).References[1]BoostC++libraries,http://www.boost.org.[2]TheSUIF2researchcompiler,StanfordUniversity,http://suif.stanford.edu.[3]A.Aiken.Introductiontosetconstraint-basedprogramanal-ysis.ScienceofComputerProgramming,35(2C3):79C111,Nov.1999.[4]A.AikenandE.L.Wimmers.Solvingsystemsofsetcon-straints.InProc.7thAnnualIEEESymposiumonLogicinComputerScience,pages329C340,1992.[5]B.Alpern,R.Hoover,B.K.Rosen,P.F.Sweeney,andF.K.Zadeck.Incrementalevaluationofcomputationalcircuits.InProc.1stAnnualACM-SIAMSymposiumonDiscreteAl-gorithms,pages32C42,1990.[6]L.O.Andersen.ProgramAnalysisandSpecializationfortheCProgrammingLanguage.PhDthesis,DIKU,Univer-sityofCopenhagen,1994.[7]F.Bourdoncle.Ef?cientchaoticiterationstrategieswithwidenings.InProc.ConferenceonFormalMethodsinPro-grammingandtheirApplications,pages128C141,1993.[8]L.-L.ChenandW.L.Harrison.Anef?cientapproachtocomputing?xpointsforcomplexprogramanalysis.InProc.ACMConferenceonSupercomputing,pages98C106,1994.[9]M.Das.Uni?cation-basedpointeranalysiswithdirectionalassignments.InProc.ACMConferenceonProgrammingLanguageDesignandImplementation,pages35C46,2000.[10]M.F¨ahndrich,J.S.Foster,Z.Su,andA.Aiken.Partialonlinecycleeliminationininclusionconstraintgraphs.InProc.ACMConferenceonProgrammingLanguageDesignandImplementation,pages85C96,1998.[11]M.F¨ahndrich,J.Rehof,andM.Das.Scalablecontext-sensitive?owanalysisusinginstantiationconstraints.InProc.ACMConferenceonProgrammingLanguageDesignandImplementation,pages253C263,2000.[12]C.FechtandH.Seidl.Anevenfastersolverforgeneralsystemsofequations.InProc.StaticAnalysisSymposium,pages189C204,1996.[13]C.FechtandH.Seidl.Propagatingdifferences:Anef?cientnew?xpointalgorithmfordistributiveconstraintsystems.InProc.EuropeanSymposiumonProgramming,pages90C104,1998.[14]N.HeintzeandO.Tardieu.Ultra-fastaliasinganalysisusingCLA:AmillionlinesofCcodeinasecond.InProc.ACMConferenceonProgrammingLanguageDesignandImple-mentation,pages254C263,2001.[15]S.Horwitz,A.J.Demers,andT.Teitelbaum.Anef?cientgeneraliterativealgorithmfordata?owanalysis.ActaInfor-matica,24(6):679C694,1987.[16]L.G.Jones.Ef?cientevaluationofcircularattributegram-mars.ACMTransactionsonProgrammingLanguagesandSystems,12(3):429C462,1990.[17]A.KanamoriandD.Weise.Worklistmanagementstrate-gies.TechnicalReportMSR-TR-94-12,MicrosoftRe-search,1994.[18]A.Marchetti-Spaccamela,U.Nanni,andH.Rohnert.rma-tionProcessingLetters,59(1):53C58,1996.[19]F.Nielson,H.R.Nielson,andC.L.Hankin.PrinciplesofProgramAnalysis.Springer-Verlag,1999.[20]D.J.PearceandP.H.J.Kelly.Onlinealgorithmsformain-tainingthetopologicalorderofadirectedacyclicgraph(workinprogess,availableuponrequest).Technicalreport,ImperialCollege,2003.[21]A.RouhtevandS.Chandra.Off-linevariablesubstitutionforscalingpoints-toanalysis.InProc.ACMConferenceonProgrammingLanguageDesignandImplementation,pages47C56,2000.[22]B.Steensgaard.Points-toanalysisbytypeinferenceofpro-gramswithstructuresandunions.InProc.6thInterna-tionalConferenceonCompilerConstruction,pages136C150,1996.[23]B.Steensgaard.Points-toanalysisinalmostlineartime.InProc.ACMSymposiumonPrinciplesofProgrammingLan-guages,pages32C41,1996.包含各类专业文献、应用写作文书、高等教育、外语学习资料、幼儿教育、小学教育、中学教育、72Online cycle detection and difference propagation for pointer analysis等内容。　Posts - 250,
Articles - 0,
Comments - 575
18:36 by DanielWise, ... 阅读,
当在分布式系统中工作时，事务有时必须要跨越服务边界。例如，如果一个服务管理客户信息而另一个服务管理订单，一个客户提交一个订单并想产品可以发送到一个新的地址，系统将需要调用每个服务上的操作。如果事务完成，用户将会期待两个系统上的信息都被合适的更新。
& 如果基础架构支持一个原子事务协议，服务可以像刚才描述的那样被组合到一个复合事务中。WS-AT(网络服务原子事务)提供在参与的服务间共享信息的平台来实现ACID事务必须的两步语义提交。在WCF中，在服务边界间的流事务信息被称作事务流。
& 为了在服务边界间十万事务流转的语义，下面的5步必须实现:
& 1. (服务契约) SessionMode.Required.& 服务契约必须要求会话，因为这是信息如何在合作者和服务组成部分间共享消息的方式。
&& 2. (操作行为) TransactionScopeRequired=true. 操作行为必须要求一个事务范围。如果没有事务存在，那么将会按照要求创建一个新的事务。
&& 3.(操作契约) TransactionFlowOption.Allowed. 操作契约必须允许事务信息在消息头中流转。
&& 4.(绑定定义) TransactionFlow=true. 绑定必须使能事务流以便于信道可以将事务信息加到SOAP消息头中。也要注意绑定必须支持会话因为wsHttpBinding支持但是basicHttpBinding不支持。
& 5.(客户端)TransactionScope. 这部分初始化事务，一般对客户端来说，当调用服务操作时必须使用一个事务范围。它也必须调用TransactionScope.Close() 来执行改变。
图片5.12 扩展服务边界的事务
& 关于TransactionScopeRequired 属性的.NET 3.5 文档包含了下面的表来描述这些元素间的关系。为了方便我们在这里重述一遍。
TransactionScopeRequired
允许事务流的绑定
调用事务流
方法不在事务内执行。
方法在一个新的事务中创建执行。
True or False
对这个事务头会返回一个SOAP错误。
方法不在事务内执行。
方法在事务内执行。
& 列表5.18 描述了如何使用这些元素。代码与列表5.15 中显示的类似，5.15 中的代码是确定一个服务操作的事务实现，而5.18代码使用TransactionFlowOption 属性显示跨服务的事务实现。注意几个点。
首先，ServiceContract被标记为需要会话。为了实现这个需求，必须使用一个支持会话的协议，比如wsHttpBinding或者netTcpBinding。
其次，为了例证的目的，TransactionAutoComplete设置成false 同时方法的最后一行是SetTransactionComplete.如果执行达不到SetTransactionComplete,事务将自动回滚。
第三，TransactionFlowOption.Allowed 在每一个OperationContract 上设置来允许跨服务的事务调用。
列表5.18 跨边界的事务流上下文
[ServiceContract(SessionMode=SessionMode.Required)]
public interface IBankService
[OperationContract]
double GetBalance(string accountName);
[OperationContract]
void Transfer(string from, string to, double amount);
public class BankService : IBankService
[OperationBehavior(TransactionScopeRequired = false)]
public double GetBalance(string accountName)
DBAccess dbAccess = new DBAccess();
double amount = dbAccess.GetBalance(accountName);
dbAccess.Audit(accountName, "Query", amount);
[OperationBehavior(TransactionScopeRequired = true, TransactionAutoComplete=true)]
public void Transfer(string from, string to, double amount)
Withdraw(from, amount);
Deposit(to, amount);
catch(Exception ex)
[OperationBehavior(TransactionAutoComplete = false, TransactionScopeRequired = true)]
[TransactionFlow(TransactionFlowOption.Allowed)]
private void Withdraw(string accountName, double amount)
DBAccess dbAccess = new DBAccess();
dbAccess.Withdraw(accountName, amount);
dbAccess.Audit(accountName, "Withdraw", amount);
OperationContext.Current.SetTransactionComplete();
[OperationBehavior(TransactionAutoComplete=false, TransactionScopeRequired=true)]
private void Deposit(string accountName, double amount)
DBAccess dbAccess = new DBAccess();
dbAccess.Deposit(accountName, amount);
dbAccess.Audit(accountName, "Deposit", amount);
OperationContext.Current.SetTransactionComplete();
class DBAccess
private SqlC
public DBAccess()
string cs = ConfigurationManager.ConnectionStrings["sampleDB"].ConnectionS
conn = new SqlConnection(cs);
conn.Open();
public void Deposit(string accountName, double amount)
string sql = string.Format("Deposit {0}, {1}", accountName, amount);
SqlCommand cmd = new SqlCommand(sql, conn);
cmd.ExecuteNonQuery();
public void Withdraw(string accountName, double amount)
string sql = string.Format("Withdraw {0}, {1}", accountName, amount);
SqlCommand cmd = new SqlCommand(sql, conn);
cmd.ExecuteNonQuery();
public double GetBalance(string accountName)
SqlParameter[] paras = new SqlParameter[2];
paras[0] = new SqlParameter("@accountName", accountName);
paras[1] = new SqlParameter("@sum", System.Data.SqlDbType.Float);
paras[1].Direction = System.Data.ParameterDirection.O
SqlCommand cmd = new SqlCommand();
cmd.Connection =
mandType = mandType.StoredP
mandText = "GetBalance";
for (int i = 0; i & paras.L i++)
cmd.Parameters.Add(paras[i]);
int n = cmd.ExecuteNonQuery();
object o = cmd.Parameters["@sum"].V
return Convert.ToDouble(o);
public void Audit(string accountName, string action, double amount)
Transaction txn = Transaction.C
if (txn != null)
Console.WriteLine("{0} | {1} Audit:{2}",
txn.TransactionInformation.DistributedIdentifier,
txn.TransactionInformation.LocalIdentifier, action);
Console.WriteLine("&no transaction& Audit:{0}", action);
string sql = string.Format("Audit {0}, {1}, {2}",
accountName, action, amount);
SqlCommand cmd = new SqlCommand(sql, conn);
cmd.ExecuteNonQuery();
&& 列表5.19 显示了配置文件。注意绑定是支持会话的wsHttpBinding。因为代码在服务契约中声明了SessionMode.Required 所以这是必须的。也要注意transactionFlow=&true&在绑定配置部分定义。
列表5.19 在配置文件中使能事务流
&?xml version="1.0" encoding="utf-8" ?&
&configuration&
&connectionStrings&
&!--Change connectionString refer to your environment--&
&add connectionString="Data Source=SQL2K8CLUSTER\SQL2K8CLUSTER;Initial Catalog=BankSIntegrated Security=True" name="sampleDB"/&
&/connectionStrings&
&system.serviceModel&
&bindings&
&wsHttpBinding&
&binding name="transactions" transactionFlow="true"&
&security&
&transport&
&extendedProtectionPolicy policyEnforcement="Never" /&
&/transport&
&/security&
&/binding&
&/wsHttpBinding&
&/bindings&
&behaviors&
&serviceBehaviors&
&behavior name="metadata"&
&serviceMetadata httpGetEnabled="true" /&
&/behavior&
&/serviceBehaviors&
&/behaviors&
&services&
&service behaviorConfiguration="metadata" name="Services.BankService"&
&endpoint address="" binding="wsHttpBinding" bindingConfiguration="transactions"
contract="Services.IBankService" /&
&baseAddresses&
&add baseAddress="http://localhost:8000/EssentialWCF" /&
&/baseAddresses&
&/service&
&/services&
&/system.serviceModel&
&/configuration&
& 列表5.20 显示了将两个服务的工作合并到一个单独事务的客户端代码。创建了三个代理，两个指向一个服务，第三个指向另外一个服务。两个查询操作和一个Withdraw 操作在proxy1上调用，然后在proxy2上调用Deposit。如果在那些服务操作内所有事情都很顺利，它们每个都会执行自己的SetTransactionComplete().在两个操作都返回后，客户端调用plete()来完成事务。只有事务中所有部分都执行它们的SetTransactionComplete()方法，事务才会被提交；如果它们中有没有成功的，整个事务将会被回滚。最后，proxy3会调用两个查询操作来确定经过事务处理以后改变是一致的。
列表5.20 在一个客户端合作完成一个分布式事务
using (TransactionScope scope = new TransactionScope(TransactionScopeOption.RequiresNew))
BankServiceClient proxy1 = new BankServiceClient();
BankServiceClient proxy2 = new BankServiceClient();
Console.WriteLine("{0}: Before - savings:{1}, checking {2}",
DateTime.Now,
proxy1.GetBalance("savings"),
proxy2.GetBalance("checking"));
proxy1.Withdraw("savings", 100);
proxy2.Deposit("checking", 100);
proxy1.Close();
proxy2.Close();
BankServiceClient proxy3 = new BankServiceClient();
Console.WriteLine("{0}: After - savings:{1}, checking {2}",
DateTime.Now,
proxy3.GetBalance("savings"),
proxy3.GetBalance("checking"));
proxy3.Close();
& 图片5.13 显示了一个客户端和两个服务端的输出。左边的客户端打印了savings账户的总额并在转账前后检查。两个服务端在右边。最上面的服务被Proxy1和Proxy3访问；底下的被Proxy2访问。上面的服务执行了两个查询操作，一个Withdraw操作和另外两个查询操作。下面的服务执行了一个Deposit操作。注意分布式事务身份id 在两个服务端都是一致的，意味着它们都是同一个事务的一部分。
图片5.13 一个单一食物中的两个合作的事务服务的输出