IEEE International Conference on Data Engineering1084-4627/09 $25.00 - PDF document

IEEE International Conference on Data Engineering1084-4627/09 $25.00 © 2009 IEEEDOI 10.1109/ICDE.2009.79565 566 567 568 569 570 571 572 573 574 575 576 IEEE International Conference on Data Engineering1084-4627/09 $25.00 © 2009 IEEEDOI 10.1109/ICDE.2009.79565 566 567 Weadoptasubsetof XCORE expressionsinTableII,which issufÞcienttocapture XPATH 1.0and XQUERYFLWOR expressions[8].Weusearepresentationof XPATH pathsinour XCORE grammarthatkeepsconsecutivestepstogether,rather thannestingeachstepinaseparate for -loop(whenallowed Ðtheuseof position() precludesthis).Suchanoptimization iscommonin XQUERY engines,andispartof XQUERY normalization,furtherdescribedinSectionIV.Additionally, wedeÞnetwonewrulesfortheXRPCextension[30]: ::= ::= excuteatŽ { ŽExprSingle } ŽfunctionŽXRPCParam { ŽExpr } Ž ()Ž | (Ž $ ŽVar:=ŽVarRef(,ŽXRPCParam) ? )Ž 27: XRPCExpr 28: XRPCParam Rule 27 identiÞesan xrpc:// URI inexpression ExprSingle ,and declaresanew anonymousfunction thatistobeexecuted remotely.Itisnoticeablethat thesegrammarruleslackthe expressivepowertodeÞnerecursivefunctions.Thisdoes notmatterfor XQUERY decomposition,asourdecomposition strategieswillnotgeneraterecursivefunctions.Wealsonote thatthesyntaxdeÞnedbyrules 27 and 28 differsfromthe actual XRPC syntax( executeat { Expr }{ FunApp(ParamList) } ). Thesyntaxusedhereisonlyforpresentationpurpose,to avoidtheneedtodeÞneallrulesconcerningdeclarationof user-deÞnedfunctions.Thus,oursimple XCORE rulewithout explicituser-deÞnedfunctiondeclarationsallowstoexpress allqueriesinasingle Expr ,whichinturncanbemappedtoa querygraph.ThissimpliÞestheformulationofanalysissteps. A.XCoreDependencyGraph Weintroduceadependencygraph( d -graph)foran XCORE query.Considerthe XQUERY query Q 2 inTableIII,which asksforthegradein course42 ofstudentshavinga tutor who isalsoa student ,andits XCORE equivalence Q c 2 . A dependencygraph isadirected,orderedandconnected graph G withvertices V ( G ) andedges E ( G ) .Eachvertex v is denotedas v i :rule[val] ,where v i isauniquevertexidentiÞer, rule isthegrammarrulerepresentedby v i ,and val isan optionalvalueindicatingtheright-hand-sideof rule .There isasingle root vertexwithoutincomingedges. E ( G ) consists ofparseedges E p ( G ) andvarrefedges E v ( G ) .Each parse edge isanorderedvertexpair ( u,v ) ,where u correspondsto aparsingrule r u thatdirectlycausestheuseofanotherparsing rule r v .A varrefedge isanorderedvertexpair ( w,x ) denoting avariableusage.Whena VarRef ruleisused,anadditional edgeiscreatedbetweenthe VarRef vertexandthe Var vertex thatdeÞnesthevariable. Example3.1: Figure2showsthe d -graphof Q c 2 inTableIII. Solidanddashedlinesrepresentparseandvarrefedges, respectively.Forinstance,thevariablebindingintheÞrst let expressioncorrespondstovertices v 2 ,...,v 7 ,andvertices v 8 , ··· ,v 39 depictits return clause.Theedge ( v 4 ,v 5 ) isa parseedge.Theedge ( v 30 ,v 9 ) isavarrefedge,asthevariable usedby v 30 isareferenceofvariable $ c introducedby v 9 . Thus,ad-graphisinessenceaparse-treewithadditional (dashed)edgestoindicatevariableusages. WedeÞnethreetypesofdependencyrelationshipsuponthe reachabilitybetweentwovertices x,y in V ( G ) : (1) x Òparse- v 1 :/grade v 2 :LetExpr v 3 :Var[ $ s] v 4 :/person v 5 :/people v 6 :FunCall[doc] v 7 : Literal [ ··· /students.xml ] v 8 :LetExpr v 9 :Var[ $ c] v 10 :FunCall[doc] v 11 : Literal [ ··· /course 42 .xml ] v 12 :LetExpr v 13 :Var[ $ t] v 14 :ForExpr v 15 :Var[ $ x] v 16 :VarRef[ $ s] v 17 :IfExpr v 18 := v 19 :/tutor v 20 :VarRef[ $ x] v 21 :/name v 22 :VarRef[ $ s] v 23 :ThenElse v 24 :VarRef[ $ x] v 25 :() v 26 :ForExpr v 27 :Var[ $ e] v 28 :/exam v 29 :/enroll v 30 :VarRef[ $ c] v 31 :IfExpr v 32 := v 33 :@id v 34 :VarRef[ $ e] v 35 :/id v 36 :VarRef[ $ t] v 37 :ThenElse v 38 :VarRef[ $ e] v 39 :() Fig.2.D-graph basic XQUERY query (let$s:=doc(“xrpc://A/st udents.xml”)/people/person, $c:=doc(“xrpc://B/course42.xml”), $t:=$s[tutor=$s/name] for$ein$c/enroll/exam Q 2 where$e/@id=$t/id return$e)/grade XCORE variant (let$s:=doc(“xrpc://A/students.xml”)/child::people/child::personreturn let$c:=doc(“xrpc://B/course42.xml”)return let$t:=for$xin$sreturn if($x/child::tutor=$s/child::name)then$xelse() returnfor$ein$c/child::enroll/child::examreturn Q c 2 if($e/attribute::id=$t/child::id)then$eelse())/child::grade normalized XCORE variant (let$t:=(let$s:=doc(“xrpc://A/students.xml”)/child::people/child::person returnfor$xin$sreturn if($x/child::tutor=$s/child::name)then$xelse()) returnfor$ein(let$c:=doc(“xrpc://B/course42.xml” Q n 2 return$c/child::enroll/child::exam) returnif($e/attribute::id=$t/child::id)then$eelse())/child::grade TABLEIII E XAMPLE Q UERY Q 2 dependsonÓ y ,denotedas x p  y ,if y isreachablefrom x via only parseedges; (2) x Òvarref-dependsonÓ y ,denotedas x v  y ,if y isreachablefrom x viaatleastonevarrefedge; and (3) x ÒdependsonÓ y ,denotedas x  y ,ifeither x p  y or x v  y holds.Thecompositionalnatureof XQUERY meansthat x  y conciselycapturesallsemanticdependenciesbetween subexpressions. ConsiderFigure2, v 15 p  v 16 ,since ( v 15 ,v 16 ) isa parseedge; v 15 v  v 3 ,as v 3 isreachablefrom v 15 via ( v 15 ,v 16 ) , ( v 16 ,v 3 ) and ( v 16 ,v 3 ) isavarrefedge. Fora d -graph G andavertex r s  V ( G ) ,weusethe term subgraph tomeanthevertex-inducedsubgraphof r s , including r s andall u  V ( G ) where r s p  u ; r s iscalled the root ofthesubgraph.Forinstance,thesubgraphrootedat vertex v 26 containsvertices v 26 , ··· v 39 ,butdoesnotcontain vertices v 13 ,...,v 25 .Throughoutthispaper,weusetheterms (sub)graphand(sub)queryinterchangeably,asa(sub)queryis representedbytheinducedsubgraphrootedatsomevertex. 569 570 571 Excerptofrequestmessagefor makenodes()  request   projection-paths   used-path /   returned-path  parent::a  / xrpc:returned-path   / projection-paths   fragments /  Excerptofresponsemessagefor makenodes()  env:Envelope...   env:Body   response   fragments   fragment  a  b  c/  /b  /a  /fragment   /fragments   call   sequence  element fragid=“1”nodeid=“2”  /sequence   /call   /request   /env:Body   /env:Envelope  Fig.5.Pass-By-ProjectionMessages usedandreturnednodesplusthedescendantsofthereturned nodes,andisqueriedwith Q . Therearethreereasonswhyprojecting XML isextremely interestingfordistributed XML processing: (i) untilnow,when sendingnodes,wehadtoserializealldescendantsÐwhich potentiallycontainhugesubtreesthatmayremainuntouched ontheotherside.Thisamountstowastednetworkbandwidth aswellasserializationandshreddingeffort. (ii) ifdocuments areprojectedintoleanskelet onsthatonlycontaintherelevant portions,itbecomesfeasibletoserialize XML fragments fromsome lowestcommonancestor on,possiblyeventhe documentroot.Evenwithpass-by-fragment,theexecutionof reverse/horizontal XPATH axesonremotenodesisimpossible. Byextendingprojecting XML withsupportforreverseand horizontalaxes,however,wegetatooltopreciselyidentify thislowestcommonancestorofan XML documentthatneeds tobeincludedtoallowcorrectremoteexecutionofthoseaxes. (iii) theprojectiontechniquecanevenbeappliedtosupportthe built-infunctions fn:root() and fn:id()/fn:idref() ,i.e.,bytaking thelowestcommonancestorofthose,ifapathcontainsone ofthesefunctions. Forthesereasons,wefurtherreÞnethepass-by-fragment messagepassingsemanticsintotheso-called pass-by- projection semantics. XML projectioncanbeusedinboth directions:toprojectthepar ametersinarequestmessage,and toprojectthefunctionÕsresultsequencebeforeshippingback theresponse. Insertionconditions. Pass-by-projectionremovestheby- fragmentinsertionconditions(inSectionV)iandiv,such thatonlyiiandiii,i.e.,theapplicationofnodecomparison, nodesetoperatorsandaxisstepsontopofmultiplecallsto fn:doc() withthesame URI ,remainsillegal. Messageextension:projectionpaths. Weintroduceanop- tionalelementasasub-elementofa request tag: projection- paths ,whichinturnhaszeroormorechildelements returned- path and used-path .Inthenewpass-by-projectionsemantics, theabsenceorpresenceofthis elementdetermineswhether theresponsemessageshouldbeintheoriginalpass-by-value orthenewpass-by-projectionformat. ::= ::= ::= | | | ::= ProjectionPath SimplePath Axis NodeTest doc (ŽLiteral::ŽLiteral)Ž(  / ŽSimplePath )* AxisNodeTest | SimplePath / ŽAxisNodeTest self::Ž | child::Ž | attribute::Ž descendant::Ž | descendant-or-self::Ž  ancestor ::Ž |  ancestor-or-self ::Ž |  parent ::Ž  root() Ž |  id() Ž |  idref() Ž (( NCName | ):)?( NCName | ) | node()Ž | text()Ž TABLEV G RAMMAR R ULE E XTENSIONOF ProjectionPath ( BOLD ) Example6.1: Toillustrateprojected XRPC messages,the upperpartofFigure5showspartoftherequestmessage forthecallfrom Q 1 (discussedinProblem 4 ): let$bc:=executeat { “example.org” }{ makenodes() } sincetheprojectionpathanalysisdetectsthat $bc willsub- sequentlybeusedascontextnodebyaparentstep: $abc :=$bc/parent::a ,therequestmessagespeciÞes parent::a asa returnedpath.Therefore,theresponsemessagecontainsthe fullfragment  a  b  c/  /b  /a  towhich $abc thengets correctlybound. A.ExtendingProjectedXML Weextendthepathgrammarrules[18]andpathan- notations,tohandlefull-ßedged XQUERY involvingre- verse/horizontal XPATH stepsandbuilt-infunctions.Theex- tendedgrammarrulefor ProjectionPath isgiveninTableV. Wedenotepathannotationsinprojected XML asfollows: Env ( v i )  Expr  Paths 1 using Paths 2 Thenotation Env ( v i ) isusedtoidentifythepathannotation environmentatacertainvertex v i inthe XQUERY d -graph. Suchannotationsareconstructedbottomupby pathanalysis rules thatderivetheused( UPaths )andreturned( Paths ) pathsforeach XCORE expressionintermsofusedand returnedpathsofitssubexpressions.Thebasicpathanalysis ruleshavebeendiscussedin[18],suchas literalvalues , sequences , for and let expressionsand XPATH steps,etc.Our extensiontoincludereverse/horizontal XPATH stepsbringsno changesforthepathanalysisrules,butmustbesupported bytheloadingalgorithm,whichisdescribedinSectionVI- B.Wecomplementtherulesforbuilt-infunctions,which apartfromtheunsolvedcasesmentionedunder Problem 5 in SectionII( fn:root(),fn:id(),fn:idref() )alsoincludes fn:doc() .The descriptionofthebasicprojectiontechniqueassumesasingle document.Asindistributedqueryprocessingtherearealways multipledocuments,ourpathsalwaysstartwith fn:doc( URI ) . Pathanalysisrules. Weprovideonerulefor fn:doc() witha constantparameterandanotherforcomputed URI s: ( DOC 1 )() Env ( v i )  doc ( Literal 1 )  doc ( Literal 1 :: v i ) using  ( DOC 2 ) Env ( v j )  Expr j  Paths j using UPaths j Env ( v i )  doc ( Expr j )  doc (  :: v i ) using Paths j UPaths j AsmentionedinSectionIV,inthedeÞnitionof D ( v x ) 3 ,we useawildcard URI  ifthedocumentnameisanexpression. 3 Weusethe doc(..) preÞxesofthe returnedpaths annotationson v asa morepreciseformofthe D ( v ) property.Documentsthatwereonlyusedbut notreturnedwerealsobe partoftheoriginal D ( v ) ,butthesewillnotcause semanticproblems. Alsonotethatallpathsstartwith doc( URI :: v i ) ,thusidentifying bothdocument URI aswellasthevertex v i whereitisloaded. ThisnotationfacilitatestheidentiÞcationofsituationswhere thesame URI isloadedtwice(thefunction hasMatchingDoc() ). Asimilarrulecanbeformulatedfor XML elementconstruc- tion,producingareturnpath doc ( v i :: v i )withanartiÞcialunique URI .Therulefor fn:root() is: ( ROOT ) Env ( v j )  Expr j  Paths j using UPaths j Env ( v i )  fn : root ( Expr j )  p  Paths j p /root()using UPaths j Thebuilt-infunction fn:root() withasingleparameteristreated inthepathannotationsmuchlike XPATH axissteps,wherethe parameterhasbecomethepathpreÞx.Inthispathnotation, functionsremaineasilyrecognizablebytheparentheses.The rulesforthebuilt-infunctions fn:id()/fn:idref() ,arehighly similar(only fn:id() provided): Env ( v j )  Expr j  Paths j using UPaths j ( ID ) Env ( v k )  Expr k  Paths k using UPaths k Env ( v i )  fn : id ( Expr j ,Expr k )  p  Paths k p /id() using Paths j UPaths j UPaths k TheÞrstparameterof fn:id() isignoredbytheannotations, asitcontainsstringvalues,andtheprojectionannotation frameworkonlyallowsfortheestimationofnodesets.This hastheconsequencethatourloadingalgorithmwillconserve all elementswithan ID/IDREF attribute. B.RuntimeXMLProjection Theextensionswemadeto XML projection,namelysupport forreverse/horizontal XPATH axesand fn:root(),fn:id()/fn:idref() , couldnotbetriviallyintegratedintheloadingalgorithm of[18].However,inthiscasewearenotreallylookingfor aloadingalgorithmthatefÞcientlyreads(shreds)an XML Þleintoaprojectedrepresentation.Rather,thedocumentsare alreadypresent(andindexed)inthe XQUERY engine,and runtimemessageprojectionisa serialization task.Therefore, weproposeanew runtime approachforprojection,targeted atserialization,ratherthanat shredding.Whereastheoriginal loadingalgorithmstartsatthedocumentroot,andevaluates absolute usedandreturnedpaths,ourruntimeprojectionalgo- rithmstartsinarun-timestate,thatis,withareal, materialized contextsequence (e.g.,theparametervaluesthatareaboutto beserializedina SOAP message),andexecutesonly relative pathsonthem.Becausethenodesequenceboundatrun- timetoafunctionparameterisonlyasubsetofthenode setcharacterizedbyitscompile-timepathannotation(e.g.,its contentsmaywellhavebeenre ducedbyapplyingaselection predicate),thisruntimeproj ectiontechniquecanbemuchmore precisethantheoriginalprojectionalgorithm. Forthesereasons,our runtime approachforprojection simplyreliesonthenormal XPATH evaluationcapabilitiesof the XQUERY engineforfullyevaluatingallusedandreturned pathannotationsone-by-one(andunitingthemwith union() ). Doingso,itproducesa usednodesetU anda returnednode setR .Thesetwosetsaretheinputfortheruntimeprojection algorithmlistedinAlgorithm1. Algorithm1 :R UNTIME XMLP ROJECTION ( U,R, D ) input : U -usednodes,sortedondocumentorder R -returnednodes,sortedondocumentorder D -theoriginal XML document output : D  -theprojectionof U and R on D projectionnodes P U R andsortedondocumentorder; 1 proj Þrstnodein P ; 2 cur Þrstnodeof D ,i.e.,rootnode; 3 while ¬ P.end () do 4 if proj isadescendantof cur then 5 add cur to D  ; 6 cur nextnodein D ; 7 elseif proj = cur then 8 if proj isareturnednode then 9 add cur andalldescendantsof cur to D  ; 10 cur nextfollowingnodeof cur in D ; 11 while proj.next isadescendantof proj do 12 proj proj.next pruneprojectionnodes; 13 end 14 else 15 add cur to D  ; 16 cur nextnodein D ; 17 end 18 proj proj.next nextprojectionnode; 19 else 20 cur nextfollowingnodeof cur in D ; 21 end 22 end 23 cur rootnodeof D  ; 24 while cur hasonlyonechildnode  cur/ { U R } do 25 cur Þrstchildof cur ; 26 end 27 Theruntimeprojectionalgorithm identiÞesallprojection nodesinthe XML treerepresentationoftheoriginaldocument, bytraversingthetreetop-downdepth-Þrst.Duringtraversal,if thecurrentnode cur ofthe XML documentisanancestorofthe currentprojectionnode proj (line5), cur isaddedtooutput D  andmovedtothe nextnode indocumentorder.Ifa proj is found(line8), proj isaddedto D  ;ifthis proj isareturned node,allitsdescendantsarealsoappended.Then cur ismoved toits nextfollowingnode inthedocument.Otherwise,ifthe currentprojectionnode proj isnotadescendantof cur ,the subtreeof cur canbeskipped(line21).Thoughthisalgorithm isformulatedonanabstractlevelthatisindependentofthe particular XML storageschemeusedinan XQUERY engine,it issafetoassumethatskippingasubtreeisfast(either O (1) or O ( log ( |D| )) ).Attheendofthealgorithm(lines24-27), post-processingisperformedtoremoveunnecessarynodes,as weareonlyinterestedinthe lowestcommonancestor ofall inputnodesintheprojecteddocument D  . Example6.2: Consideran XML document D inFigure6(a). Assumethattheusednodeset U is { i } ,andthereturnednode set R is { d,k } .Figure6(b)showstheprojecteddocument D  ofapplyingAlgorithm1on U , R and D . Thealgorithmstartswith P { d,i,k } , proj  d and cur  a .Wetraversethetreeusing cur from a to d .Nodes a , b and c areaddedto D  ,sincetheyareancestorsofthe currentcontextnode d .Nodes d,e and f arealsoaddedto D  ,as d isareturnednode.Then, cur isadvancedto g ( d Õs nextfollowingnode).Becausethenextcontextnode i isnotin a b c d e f g h i j k l m n o (a) Original XML tree D b c d e f i k l m (b) Projectedtree D  Fig.6.Runtime XML ProjectionExample thesubtreeof g ,thesubtreeisskippedbyadvancing cur to i . Recallthat i isausednode,thusonly i isaddedto D  .Thelast contextnodeis k .Ourcurrentdocumentnode cur traverses from i to j ,andthento k ,wherewecanaddnodes k , l and m to D  .Thetraversalcanbeterminated,becausethereis nomorecontextnodestoprocess.However,theintermediate result D  contains all commonancestorsof { d,i,k } .Thepost- processingremovesnode a from D  ,whichproducestheÞnal projecteddocument D  asshowninFigure6(b). Relativeprojectionpaths. At compiletime ,the XQUERY compilerbuildsaquerygraph( d -graph)withroot v root , normalizesitfollowedbydecompositionandcodemotion. ForeachinsertedXRPCExpr v xrpc ,andforeach XRPCParam parametervertex v param ,itthenextractstherelativepaths: U rel ( v param ) = allSufxes ( R ( v param ) ,U ( v xrpc )) R rel ( v param ) = allSufxes ( R ( v param ) ,R ( v xrpc )) U rel ( v xrpc ) = allSufxes ( R ( v xrpc ) ,U ( v root )) R rel ( v xrpc ) = allSufxes ( R ( v xrpc ) ,R ( v root )) ,with: allSufxes ( Paths i ,Paths j ) = { s j | p i /s j  Paths j :  p i  Paths i } At runtime ,   v param U rel ( v param ) and   v param R rel ( v param ) areusedtoprojecttheparametersintheoutgoing XRPC requestmessage.The U rel ( v xrpc ) and R rel ( v xrpc ) arepassed inthe projection-paths elementsuchthattheremotepeer canappropriatelyapplythesepathstoprojecttheresponse message. ProjectingadocumentwithAlgorithm1requirespre- calculatedusedandreturnednodesets.Thesesetsaresimply computedusingthe XPATH evaluationinfrastructureofthe underlying XQUERY engine,byfeedingtheintermediateresult $ctx param correspondingto v param ascontextsequenceinto allsufÞxpaths s i  U rel ( v param ) (resp. R rel ( v param ) ): union($ctx param /s 1 ,union($ctx param /s 2 , .... union($ctx param /s n Š 1 ,$ctx param /s n )...)) Paths $ctx /path i / root() /path j withfunction fn:root() areexe- cutedas root($ctx) /path j .Similarly, $ctx /path i / id() /path j is executedas root($ctx) // attribute() ::( a 1 | .. | a n ) /../path j ,where a 1 ,..,a n areall ID attributes(resp. IDREF incaseof idref() ). Therequesthandlerontheremotesideusesthe samemethodtoevaluatethesufÞxpaths U rel ( v xrpc ) and R rel ( v xrpc ) usingtheresultsequenceofthefunctionas $ctx xrpc duringserializationoftheresponsemessage. Incaseof XML datawithauser-deÞned XMLSCHEMA ,the defaultprojectionalgorithmislikelytothrowawaymandatory elementsandattributes.Forthisreason,theruntimeprojection algorithmshouldbemadeschema-aware.Asimplesolution istoensurethatonlyelementswitha minoccurs declaration ofzero(i.e.,optionalelements)areremoved.Onecanalso envisionmoreadvancedvariantsthatfurtherreducethesize ofatyped XML document. VII.E VALUATIONIN M ONET DB/XQ UERY Wehaveimplementedtheproposedalgorithmsin Mon- etDB/XQuery [4],apurelyrelational XML databasesystemthat usesthe PathÞnder [10] XQUERY compiler.Weusethe XRPC extensionforremotefunctionevaluation.Thetestplatform consistedofthree 2 GHzAthlon64Linuxmachinesconnected via 1 Gb/sEthernet.Eachwasequippedwitha 2 GBRAM. ThebenchmarkdatausedisXMark[23],apopular XML benchmarkforevaluating XQUERY efÞciencyandscalability. Thedatasetwasgeneratedusingscalarfactors 0.1,0.2, 0.4,0.8 and 1.6 .Adatasetisstoredoneachremotepeer. Weconductedthreegroupsofexperiments:bandwidthusage, queryexecutiontimeandruntimeprojectionprecision.Note that,astherearenoothercompar ativeresultsexist,themain goalofourexperimentsistoshowtheimpactoftheproposed techniquesinastep-by-stepfashion. WeslightlymodiÞedthequery Q n 2 (inTableIII)sothatit conformstotheXMarkschemaasthefollowing: (let$t:=let$s:=doc(“xrpc://peer1/xmk nn MB.xml”) /child::site/child::people/child::person returnfor$xin$sreturnif($x/descendant::age 40)then$xelse() returnfor$ein(let$c:=doc(“xrpc://peer2/xmk nn MB.auctions.xml”) return$c/descendant::open auction) returnif($c/child::seller/attrib ute::person=$t/attribute::id) then$c/child::annotationelse())/child::author Alltechniquesdiscussedinthispaperareappliedontheabove query: (i) underthepass-by-valuesemantics,onlytheexpres- sion doc(“xrpc://peer1/xmk nn MB.xml”)/.../child::person can bedecomposedandexecutedon peer1 ; (ii) underthepass- by-fragmentsemantics,wecandecomposeboththesecond let clause( “let$s:=...” )andthesecond for -loop( “for$ein ...” ),andexecutethemon peer1 and peer2 respectively.The variable $t becomestheparameterofthegeneratedfunction containingthesecond for -loop(seealsoTableIV); (iii) under thepass-by-projectionsemantics,thequeryisdecomposed inthesamewayasusingpass-by-fragment,however,when serializingtherequestmessages,aprojectionof $t/attribute::id (parameterprojection)and $c/child::annotation/child:author (re- sultprojection)iscalculated.Thetestsetthuscontainsfour queriesintotal,andeachofthemisexecutedon 2 documents ofsizes 10,20,40,80 and 160 MB. Bandwidthusage. Figure7showsthebandwidthusedby eachbenchmarkqueryondifferentsetofdocuments,i.e., thetotalsizeof XML documentsplustotalsizeof XML messagestransferredamongpeers,initsy-axis.Thex-axis isthetotalsizeofthe XML documentsusedbyeachquery. Thepuredata-shipping XQUERY query(theleftmostbar) coststhelargestbandwidthusage,asbothdocumentshave tobeshipped.By-valuedecompositioncanpushthe XPATH step doc(“xrpc://peer1/xmk nn MB.xml”)/.../child::person tobe evaluatedon peer1 ,whichreducestheamountofdatasent from peer1 tothelocalpeer.However,theseconddocument 575 protocol,whichlackspropersupportfor XML elementsand sequences;aproblemaddressedby XRPC usingaspeciÞc literal SOAP messageencoding.ActiveXML(AXML)[1],[2] isadeclarativeframeworkthatharnesseswebservicesfordata integrationinapeer-to- peerarchitecture.Like XRPC ,italso useda(document/literalencoding) SOAP protocoltorepresent XML subtreevalues.However,thefocusin AXML hasbeen inadaptivecallmaterializationstrategies,notonautomatic querydecompositionandthesemanticchallengesthisbrings in XQUERY ,suchasdistributednodeidentity.XQueryD[22], like XRPC supportsfunctionshippingin XQUERY ,butitdoes notdeÞneanopennetworkprotocol. Decomposingqueriestoaddressmultipledatasourcesisby nowawell-studiedprobleminrelationaldatabases[28]and object-orienteddatabases[13],[17].Manyoftheseideasand methodscanbeappliedto XQUERY ,yetwehaveshownhere thattheissueofefÞcientlymanagingdistributednodeidentity anddocumentorderaddinterestingchallenges.[24],[25] discussthedecompositionofunstructuredquerylanguages onlyonasemi-structureddatab ase(arooted,labeledgraph). In XML data-bases,previousapproachesrequirestructural informationaboutpeersfors upervisingdecomposition[27]. Otherworks[6],[7],[26]onlyfocusonarestrictedsetof XQUERY queries. XML projection[18]drasticallyreducesthesizeofthe datamodelrepresentationusingcompile-timequerycharac- terization.[5]introducesaprecise XML pruningtechnique forasubsetof XQUERYFLWOR expressions,basedonthe apriori knowledgeofadataguideforunderlying XML data. However,itdoesnothandle XPATH predicates,backwardaxes and XQUERY -likelanguages.Atype-based XML projection technique[3]isstudiedtoimprovecurrentsolutionswith comparableorhigherprecisionandlesspruningoverhead,as wellassupportingbackward XPATH axes.However,a DTD is required.[15]discussesruntime XML projectiontechniques. Basedonthestaticcompilationofruntimelookup-tablesanda runtime-automatonfromprojectionpathsanda DTD ,theycan Þltertheinput XML documentefÞcientlyusingstringmatching algorithms.Thistechnique,however,stilllackssupportfor reverse XPATH axesand XQUERY built-infunctions. IX.C ONCLUSION Wehavedescribedaframeworkfordistributedexecution offull-ßedged XQUERY ,focusingontheissueofproviding equivalentquerydecompositions,inthefaceofsemantic differenceswhen(partsof)nodesareshippedacrossthe networkin XML messages.WeÞrstcarefullycharacterized theproblemsthatmayoccu rregardingnodeidentityand structural XPATH relationshipsinsuchadistributedsetting. Then,weproposedaseriesoftechniquessuchaspass-by- fragmentandtheuseofanovelruntime XML projection methodforserializing XML messages,thatremovevirtually allsemanticproblemsandstronglyimproveperformance,as shownbyexperimentsontheopen-sourceMonetDB/XQuery XML databasesystem( monetdb.cwi.nl ). Ourmainfutureworkisanissueleftout-of-scopehere: decidingondistributedqueryplacementafterdecomposition. Inthisarea,wealsocontemplateusingruntimemethodstoim- proveoptimizationquality.Anotherdirectionisdecomposition ofqueriescontaining XQUF updateexpressions.Thechallenge hereisthatupdatesarenecessarilytiedtoexecutionontheir sourcepeer,whichrestrictsdecompositiontocaseswhereat compile-timeasingleaffect edpeercanbeidentiÞed. R EFERENCES [1]S.Abiteboul,O.Benjelloun,B.Cautis,I.Manolescu,T.Milo,and N.Preda.LazyQueryEvaluationforActiveXML.In SIGMOD ,2004. [2]S.Abitebouletal.AFrameworkf orDistributedX MLDataManage- ment.In EDBT ,2006. [3]V.Benzakenetal.Type-BasedXMLProjection.In VLDB ,2006. [4]P.Bonczetal.MonetDB/XQuery:AFastXQueryProcessorPowered byaRelationalEngine.In SIGMOD ,2006. [5]S.Bressanetal.AcceleratingqueriesbypruningXMLdocuments. Data Knowl.Eng. ,54(2),2005. [6]P.Bunemanetal.UsingPartialEvaluationinDistributedQuery Evaluation.In VLDB ,2006. [7]G.Congetal.Distributedqueryevaluationwithperformanceguarantees. In SIGMOD ,2007. [8]D.Draperetal.XQuery1.0andXPath2.0FormalSemantics.W3C CandidateRecommendation8June2006. [9]M.Fern« andezetal.HighlyDistributedXQuerywithDXQ.In SIGMOD , 2007. [10]T.Grustetal.XQueryonSQLHosts.In VLDB ,2004. [11]M.Gudgin,M.Hadley,N.Mendels ohn,J.-J.Moreau,andH.F.Nielsen. SOAPVersion1.2Part1:MessagingFramework.W3CRecommen- dation24June2003.http://www.w3.or g/TR/2003/REC-soap12-part1- 20030624. [12]M.Gudgin,M.Hadley,N.Mendels ohn,J.-J.Moreau,andH.F.Nielsen. SOAPVersion1.2Part2:Adjuncts.W3CRecommendation24June 2003.http://www.w3.org/TR/2003/REC-soap12-part2-20030624. [13]V.JosifovskiandT.Risch.Querydecompositionforadistributed object-orientedmediatorsystem. DistributedandParallelDatabases , 11(3):307Ð336,2002. [14]J.KnoopandB.Steffen.Codemotionforexplicitlyparallelprograms. SIGPLANNot. ,34(8):13Ð24,1999. [15]C.Kochetal.XMLPreÞlteringasaStringMatchingProblem.In ICDE ,2008. [16]D.Kossmann.Thestateoftheart indistributedqueryprocessing. ACM ComputingSurveys ,32(4),2000. [17]H.Kozankiewicz,K.Stencel,andK.Subieta.Distributedquery optimizationinthestack-basedapproach.In HPCC ,2005. [18]A.Marianetal.ProjectingXMLDocuments.In VLDB ,2003. [19]N.Mitra.SOAPVersion1.2Part0:Primer.W3CRecommendation24 June2003.http://www.w3.org/TR/2003/REC-soap12-part0-20030624. [20]N.OnoseandJ.Sim« eon.XQueryatYourWebService.In WWW ,2004. [21]M.T. ¬ OzsuandP.Valduriez. Principlesofdistributeddatabasesystems (2nded.) .Prentice-Hall,Inc.,NJ,USA,1999. [22]C.Reetal.DistributedXQuery.In IIWeb ,September2004. [23]A.Schmidtetal.XMark:ABen chmarkforXMLDataManagement. In VLDB ,2002. [24]D.Suciu.Querydecompositionandviewmaintenanceforquery languagesforunstructureddata.In VLDB ,1996. [25]D.Suciu.Distributedqueryev aluationonsemistructureddata. ACM Trans.DatabaseSyst. ,27(1),2002. [26]K.TajimaandY.Fukui.Answeri ngXPathqueriesovernetworksby sendingminimalviews.In VLDB ,2004. [27]L.T.T.Thuy,D.D.Duong,V.C.Bhavsar,andH.Boley.Abottom-up strategyforquerydecomposition.In ICDIM ,2006. [28]E.WongandK.YousseÞ.Decomposition-astrategyforquery processing. ACMTrans.DatabaseSyst. ,1(3):223Ð241,1976. [29]C.YuandC.Chang.Distributedqueryprocessing. ACMComputing Surveys ,16(4),1984. [30]Y.ZhangandP.Boncz.XRPC:Inter operableandEfÞcie ntDistributed XQuery.In VLDB ,2007. [31]Y.ZhangandP.Boncz.DistributedXQueryandupdatesprocessing withheterogeneousXQueryengines.In SIGMOD ,2008.