An Agent-based Network Control Architecture

Architecture

J.Nicklisch,J.Quittek,A.Kind,S.Arao

inca@ccrle.nec.deC&CResearchLaboratories,Berlin

NECEuropeLtd.

Abstract

ThispaperdescribesthedesignandimplementationofINCA,anopenarchi-tectureforthedistributedmanagementofmulti-servicenetworksandsystemsap-plications.TheIntelligentNetworkControlArchitectureispopulatedbystation-aryandmobileintelligentagents.Theseagentsperformmonitoringandcontrolofnetworkandsystemscomponents,therebysupportingtheintegratedmanage-mentofnetworksandservices.

Thearchitectureprovidestransactioncapabilitiestocontroltransportandmo-bilityofagents,agentprioritization,andmultipleagentcodetransferschemes.Managedobjectsusedtoaccessresourcesonnetworkelementsandnewsystemfunctionalitiescanbecreated,distributed,andreplaceddynamically.AnexampleINCAapplicationdemonstratesthatprioritizedagentsarenecessarytosupportthetimelyexecutionofcriticaltasks.

Thedesignofouragentbasednetworkmanagementplatformisnotboundtoaparticularprogramminglanguageorcomputingenvironment;thecurrentim-plementationhowever,isbasedonJavaandRMI.

Keywords:MobileAgents,NetworkManagementPlatform

1Introduction

Thecentralizedorhierarchicalapproachestonetworkmanagement,beingtightlycou-pledtotheclient-serverparadigm,showwellknownlimitations.Facingthegrowingcomplexityandextensionofcomputerandtelecommunicationnetworks,itiscom-monlyagreedthatthereisneedtostudytechnologiesfordistributednetworkmanage-ment.Thegoalistobeflexible,scalable,andopentoextensions.

Oneofthepromisingapproachesisintelligentsoftwareagenttechnology,whichsup-portsnotonlydistributedapplicationsbutalsoapplicationmobility.However,despite

1

theresearcheffortsinthisarea,thereisnotmuchacceptanceofagent-basedtech-niquesinthepracticalnetworkmanagementworld1.Oneofthereasonsmightbethatavailableagentplatformsaregeneralintheirdesignandnotcustomizedtowardsthisdomain.

AspartoftheactivitiesattheNECC&CResearchLaboratoriesinthisdirection,anopenplatformfordistributedmanagementofmultiservicenetworks,INCA(Intelli-gentNetworkControlArchitecture),hasbeendeveloped.Themostimportantdif-ferencebetweenINCAandotheragentplatforms,amongseveralcustomizations,isreliability.

Thispapergivesadescriptionofthetechniquesappliedtothedesignofourarchitec-ture.INCAisbasedonsoftwareagentswhichcanmigrateoverthenetworktoperformdelegatedand/ormobilefunctionsinanintelligentmanner.Thereisalsosupportforopensignalling,allowingflexibleservicecreationaswellasextensibilityandmodularreplacementofelements,servicesandfunctions.

Theremainderofthispaperisorganizedasfollows:afterdiscussingrelatedworkinthefollowingsection,wemotivatetheuseofsoftwareagents,intelligence,andmobilityinmanagementofnetworksandservicesinSection3.TheINCAarchitectureisintroducedinSection4,andthefeatureswhichmakeitareliableplatform,well-suitedfornetworkmanagementandtelecommunicationsapplications,arediscussed.MeasurementsdemonstratingtheusefulnessofagentprioritizationarepresentedininSection5.

2RelatedWork

Thegeneraladvantagesofdecentralizedandagent-basedapproachestonetworkman-agementandtelecommunicationshaveoftenbeenaddressed,forexamplesee[6,11,10,3,1].

Currently,severalmobileagentenvironmentsareavailable,mostlyduetothewideacceptanceoftheJavaenvironmentwithitsinherentsupportforcodemobility;see[7]foranoverview.Fourofthemajorenvironmentsforagent-orientedJavaprogram-mingare:Aglets[8],Concordia[19],Odyssey2(aJava-basedsubsetofaplatformformerlyknownasTelescript[18]),andVoyager3.However,alltheseandmanyoth-ersaregeneralpurposeenvironmentsandnotcustomizedfornetworkmanagement,wherereliabilityandefficiencyareimportantissues.Wearguethatforhighreliabilitytransactioncapabilitiesforinteractionsofdistributedentitiesarerequired.Basedonsuchsafeinteractions,afault-tolerantcontroloftheitineraryofamobileagentispos-

sible.Forefficiency,agentprioritizationandsupportofdifferentagentcodetransferschemesisdesired.

ConcretenetworkmanagementapplicationsfocusingonagentintelligencehavebeendescribedbydaRochaandWestphall[4]andSomers[14],bothusingstationaryagents.

Sahaietal.havepresentedanagentenvironmentcustomizedfornetworkandsystemmanagement,calledAstrolog[13].Theyemploymobileagentstosupportthemobilityofthenetworkoperator.However,thecorenetworkmanagementsystemisbasedonahierarchyofstationaryagents.

Surprisingly,eveninthenetworkmanagementdomainthereisnoflexibleenvironmentformobileintelligentagentsofferingthemeansforreliabilityandefficiencymentionedabove.

3NetworkManagementwithAgents,IntelligenceandMobility

Toensureeffectiveandefficientfunctionoflargenetworksaswellasprovisionofservices,andfurthermoretoaddresstherequirementsofnetworkproviders,serviceproviders,retailersandusers,thereisastrongneedforacoherentframeworksupport-ingautomatedandmoreintelligentend-to-endmanagementsolutions.

Networkmanagementtodayistypicallybasedonaclient-servermodelwithSNMP[2]andCMIP[16]asthedefactostandardsformonitoringandformanaginganetworkofcomputersanddevices.Managedobjects,likehosts,routers,bridges,switches,printersetc.,provideread/writeaccesstoasetofvariablesthroughamanagementprocess.Amanagementstationcanthencommunicatewiththemanagementprocessesinaclient-serverrelationship.

Althoughsimple,thiscentralizedapproachtonetworkmanagementhassomeseveredrawbackswithregardtoscalability,flexibilityandperformance.Bydividingman-agementfunctionsintomobile,autonomousandintelligentcomputingentities(i.e.softwareagents),manyoftheproblemswithnetworkandservicemanagementcanbeaddressed:

Scalabilityisincreased,asmanagementisnotperformedsolelybythemanage-mentstationbutdelegatedtodistributedmanagementagents.

Repetitivetaskscanbeavoidedifsoftwareagentslearnfromexperience.Faulttolerancecanbeincreasedthroughtheautonomyandlearningcapabilityofmanagementagents.

3

Betterperformanceisachievedbymovingthemanagementfunctionalityclosertotheactualnetworkelement,thusreducingnetworktraffic.

Thelow-leveldetailsofdifferentdevicescanbehiddenbehindtheagentinter-face.

Legacysystemscanbeintegratedbyusinganagentforinter-operation.

Typicalscenarioslikeserviceprovisionthroughdifferentadministrativedomains,in-cludingaccountingandcharging,canbehandledbyamulti-agentbasedapproachtodistributednetworkandservicemanagementinamorescalableandcoherentwaythanwithacentralizedapproachtonetworkmanagement.

Thisisparticularlytrueforthecreationofnewservicesinthetelecommunicationsworld.Withtheadventofhigh-speednetworksbasedonATMswitching,thereisde-mandforbettersupportwhencreatingandmanagingnewservices.Proprietary,low-levelinterfacestonetworkdevicescurrentlypreventarapiddevelopmentoftelecom-municationapplications(see[17]).Controlandmanagementofscalablemultiservicenetworksisdifficultsincetheswitchingsoftwareisusuallytightlycoupledwiththeindividualswitchingdevices.

TINA[5],xbind[9]andDCAN[12,17]focusonprovidingmorecoherentandflexibleaccesstomultiservicenetworks.However,controlandmanagementofswitchingdevicesbasedonmobilesoftwareagentshasnotbeeninvestigatedsofar.

4TheArchitecture

INCAcanbedescribedasamobileagentplatformcustomizedfornetworkmanage-mentandcontrol.Thissectiongivesanoverviewofthearchitectureanddiscussesthemajordesigndecisions.Inparticular,wewillexplaintheconceptsofagentpriorities,transactioncapabilities,migrationcontrol,andmultipleagentcodetransferschemes,whichintheircombinationdistinguishINCAfromotheravailableagentenvironments.

4.1GeneralArchitecture

AninstanceoftheINCAplatformconsistsofasetofstationsandofservicesofferedlocallybyastationorgloballybytheplatform.4.1.1LocalServices

Astationislocaltoanetworkelementandprovidesthefollowingservices:

4

Concurrentexecutionofmultipleagents.Multipleagents–eachwithitsownthreadofcontrol–canbeexecutedconcurrently.TheagentsarescheduledaccordingtotheirpriorityattributebyanINCAspecificschedulingscheme,whichforexampletakesintoaccountthecurrentsystemload.Noagentoflowerprioritywillbeexecutedifthereisanagentwithhigherpriorityina’runnable’statepresentatastation.Loadingofagentcode.Agentcodecanbetransferredbetweenstations.Threeschemesforcodetransferaresupported:push,pull,andmigrate.Pushedcodeissentfromanagentcoderepositorytoadedicatedstation,pulledcodeisloadedbythestationfromarepository,andinmigratingcodeissentfromone(non-repository)stationtoanother.Allstationssupportthesamecodeformat,i.e.thesamecodecanbeexecutedatanystations.Inordertoimproveperformance,cachingofagentcodeisused.

Transferofagentstate.Thecurrentstateofanagentcanbetransferredfromonestationtoanother.Incombinationwithagentcodetransfer,thisserviceenablesagentmigration.

Transactioncapabilities.Forpeer-to-peercommunicationbetweenstationsareliableandfault-tolerantlayerisusedofferingtransactioncapabilities.

Accesstolocalresources.Accesstolocalresourcesatastationisprovidedbyman-agedobjects.Managedobjectsrepresentnetworkelements(orpartsofthese)thataretobemanaged.Theirinterfacespecifieswhichresourcecanbeaccessedandhow.Usually,astationcontainsatleastoneinstanceofamanagedobjectgivingaccesstotheresourcesofthelocalnetworkelement.Additionalmanagedobjectscanbecreateddynamicallyondemandofanagent.

Supportforlocationandstatusmonitoringofagents.Locationandstatusofanagentmigratingfromstationtostationmaybemonitoredbyacentralinstancecalledlocator.Formonitoredagentsthestationsendmessagestothelocatoronarrivalanddepartureorterminationoftheagent.4.1.2GlobalServices

Besidesservicesprovidedbyeachsinglestationthereareservicesofferedbytheentireplatform,orbyasetofdedicatedstations:

Control.ForeachINCAplatformthereisatleastonecontrolstation.Thisstationexe-cutesastationarycontrolagentofferingacontrolserviceforlaunchingandmonitoringagents.Usually,thelocationofthisstationistheconsoleofthenetworkmanager.Repository.Dedicatedstationsofferarepositoryservice.Thisserviceprovidesaccesstothecodeoffallagentsthatcanrunontheplatform.Therepositorycanbeaccessedviathecontrolservicetopushagentcodetoastationorbyanystationtopullagentcode.

5

Locator.Auniquestationrunsastationarylocatoragentofferingthelocationtrackingservice.Thelocatoragentreceivesmessagesaboutmonitoredagentsfromthestations.Basedonthisinformationthelocatoragentanswersrequestsfromthecontrolagentaboutlocationandstatusofotheragents.Furthermore,thelocatoragentsupervisespredefineditineraries(seebelow).

Itinerarycontrol.Thecontrolserviceallowstolinkamobileagentwithapredefineditinerary.Theitinerarybecomespartoftheagentstate,butisnotaccessedbytheagentitself.Itiscontrolledbyastationwheneveramobileagentistobemigrated,inordertodeterminethestationtosenditto.Whenanagentwithapredefineditineraryislaunchedbythecontrolagent,acopyoftheagentsitineraryissenttothelocatoragentwhichcomparestheitinerarywithlocationmessages.

Migrationcontrol.Bycombiningitinerarycontrolwithpersistenceofagentstatefaulttolerancecanbeincreasedforthemigrationofagents.

Inter-agentcommunication.Communicationbetweenagentsisofferedbyacommoninterfaceforinter-agentmessageinvocation.Thisfunctionalityincludesanamingservicefromwhichagentscangetreferencestootheragents.

4.2Priorities

Concurrentexecutionofmultipleagentsdoesnotonlyincreasethefunctionalityofanagentplatform,italsointroducestheproblemofschedulingagents.Weobservedthatafairandequalschedulingofallagentsinasystemdoesnotmatchtherequirementsofnetworkmanagementapplications.Thegroupsofagentslistedbelowillustratetheserequirements,forexample.

NetworkControlAgentssupportthenetworkmanagere.g.inconfigurationandfaultmanagementofthenetwork.Thetaskstheypursuearedemanddriven,perhapstriggeredbynetworkfaults,andoftenhavetobecarriedoutasfastaspossible.

ServiceManagementAgentssupporttheserviceprovider,e.g.byaservicesub-scriptionagent.Theyusethenetworkenvironmenttogetherwiththeplatformservicestoinstallandestablishenduserservices,orreconfigurethem.NetworkMaintenanceAgentsperformrepeatedlyoccurringmanagementtasks,e.g.thegatheringofdatafromselectednetworkelementsorfaultanalysis.Theirtasksarecarriedoutrepeatedly,onlyrarelyrequiringadjustmentsbythenetworkmanagementinstance.

NetworkMonitoringAgentsperformroutinetaskssuchasthefilteringofrawdatacollectedfromnetworkelements.Often,suchagentsarestationary,sincetheirtaskistomonitoraconcreteelementofthenetwork,oralink.

6

Usually,networkcontrolagentsshouldbeexecutedfast,becausethenetworkmanageriswaitingfortheexecution.Incontrasttothis,monitoringagentstypicallyrunforalongtimeandshouldbeinterrupted,whenoneoftheotherkindsofagentsarrivestobeexecuted.

Weuseagentprioritiestosatisfytheserequirements.Anarrivingagentofhigherpriorityinterruptsanagentoflowerpriorityandnoagentoflowerpriorityisexecutedwhileanagentofhigherpriorityispresentandnotblocked.

Bythesemeanstheproblemabove,canbesolvedbygivingnetworkcontrolagentsthehighestpriority,servicemanagementagentsandnetworkmaintenanceagentsamediumpriority,andmonitoringagentsthelowestpriority.

AnapplicationdescribedinSection5demonstratestheusefulnessofthepriority-basedapproach.

4.3TransactionCapabilities

MostinteractionsinINCAarepeer-to-peercommunicationsbetweenstations.Intheplatformwehaveaclearseparationofdifferentlayersofcommunicationfacilities.INCAisdesignedtorunontopofanobject-orienteddistributedmiddlewareprovidingnamingserviceandtransparentaccessofobjectsregardlessoftheirphysicallocation.Betweenthemiddlewareandtheapplication,INCAcontainsanotherlayercalledtransactioncapabilitiesprovidingreliable,fault-tolerantinteractionsbetweentwosta-tions.ThedesignofthislayerisbasedontheSS7(CommonChannelSignallingSystemNo.7)transactioncapabilities[15].SS7transactioncapabilitiesusedistributedtransactionmonitorstoensurethereliableexchangeofmessagesbetweentwocom-municatingpeers.

IntheINCAplatformweusedthistechniquetosupportasynchronousexchangeofmessagesaswellassynchronousexchange.Ontheapplicationlevelthesetwoaresufficienttoexpressallinteractionsoftwostationsconveniently.Retransmissionat-temptsofmessages,time-outs,healthchecks,andfurthermeansprovidingreliabilityinteractions.

4.4MigrationControl

Whenstartinganagentwithapredefineditinerary,acopyisstoredatthelocator.Afailureatthestationcurrentlyhostingthemobileagentdoesnotnecessarilyleadtoabreakdownoftheagentstask,sinceanup-to-datecopyoftheagent’sitineraryiskept,andanewagentcanbecreatedinordertovisittheremainingstationsoftheitinerary.Ofcourse,thisfault-tolerantprocedureshouldonlybeappliedtoagentswhichhavebeendesignedaccordingtotheprocedure.Incontrasttothisconcept,mobileagents

7

mightnotmakeuseofsuchanitineraryatall,butinsteaddeterminethehoststobevisiteddynamically,atruntime.Itdependsontheparticularapplicationwhichofbothmethodsshouldbeemployed.Weobservethatapplicationsofmobileagentswhichrequirehighreliability,asitisthecaseformanynetworkmanagementtasks,canmakegooduseoftheitineraryconcept.Tounderstandwhythatisthecaseconsiderthefollowingexample.

Supposethatamobilemaintenanceagentisusedtoperformaroutinetaskatallthenet-workelements,orafixedsubsetofit.Anexamplecouldbeanagentwhichexaminesthenetworkelementsandchecksthecurrentmemoryusageandthediskspaceused.Incaseofanalarmingsituationatthenetworkelement,itwillsendamessagetothenetworkmanager,otherwiseitproceedsalongitsitinerary.Insuchanexamplethereisactuallynostateinvolvedintheagentmigration,apartfromtheinformationcontainedintheitinerary.Nowconsiderabreakdownatoneoftheintermediatestations.With-outanyobservinginstancetheworkdonebythemobileagent,i.e.theconfirmationofacertainsystemstateatthestationsvisited,wouldbelost.Theplatformwouldhavetoinstantiateanotheragentandstartfromscratch,becauseitcannotbeawareofhowfartheagentcouldproceeditstask.Andincasecertainadjustmentsweremadebytheagentatthenetworkelements,thesewouldperhapshavetobeoverwritten.

Sincethesystememploysthelocator,however,itcanrecoverfromsuchsimpleerrorcasesandletanotherinstanceofthemaintenanceagentproceedthetaskfromthethestationfollowingtheonewhichexperiencedthebreakdown.Excessivesystemandnetworkloadcanbeavoided,becausethissecondinstancedoesnotneedtorepeatitspredecessor’swork.

Apparently,theideaofkeepingacopyofanagent’sitineraryisnotthefullstory,butonlythemostsimplifiedexplanationofamoregeneralrecoveryscheme.Whentheagentstateactuallymatters,asitisthecaseformorecomplicatedormoreintelligentagents,theknowledgeoftheitineraryaloneisinsufficient.Instead,thewholeagentstateshouldbepreservedatanysourcestationofanagentmigration.Then,afterthedestinationstationexperiencesafault,thelocatorinstancetimesoutinwaitingforthemessagethatconfirmstheagent’sarrival.Itcanthereforetakefurtheractionsandrestoretheagentfromthestatepreviouslycachedatthesourcestation.

4.5MultipleAgentCodeTransferSchemes

INCAusesmobilecodeforallagents.Forstationaryagentstheagentcodehastobetransferredtoasinglestationonly,whereasmobileagentsrequirethecodeonallvisitedstations.Whenmigrating,thestateofamobileagenthastobetransferredfromthestationtheagentisleavingtothestationtheagentisgoingtovisitnext.Butsincethecodeoftheagentisindependentofthecurrentagentinstanceanditsstate,itcanbetransferredindifferentways,anditcanbecachedatstations.INCAsupportsthreeschemesforcodedistribution:pull,push,andmigrate.

8

Figure1:Pushtypeagentcodedistribution

Pushtypeagentcodedistribution.ThisschemeisknownfromInternetserviceslikePointCast4orMarimba’sCastanet5.Whiletheseservicespushuserdatatoalistofsubscribers,thepushschemeinINCApushesagentcodetoalistofstations.Usually,thepushschemeisinconjunctionwithpredefineditineraries.Whentheagentislaunched,thecodeispushedtoallstationsoftheitinerary.So,whentheagentinstancearrivesatastation,thecorrespondingcodeisalreadythereanditcanstartexecutionimmediately.Thepushschemeisrealizedbysendingamessagestotherepositorystationwhichpushesthecodetothestations.

Pulltypeagentcodedistribution.Thepullschemeisprobablythemostwidespreadone,sinceitisusedbywebbrowserstodownloadJavaappletsandscripts.Whenthisschemeisused,astationdownloadsthecorrespondingagentcodefromarepositorystationafteranagentinstancehasarrived.Thisschemecanbeappliedtomobileagentswithflexibleitineraries.Theseagentsmaydecidedynamicallywhichstationtovisitnext.Thedrawbackofthisschemeisadecreaseofperformance.Whenanagentarrivesatastation,itsexecutionhastobedelayeduntiltheagentcodehasbeendownloaded.Agentcodecachingcanavoidthisdelayforsubsequentarrivalsofagentsofthesametype.Analternativewouldbepushingthecodeoftheagenttoallstations,

butthissolutionreducesscalability,producesunnecessaryloadonthenetwork,andmightincreasememoryrequirementsofthestations.

Figure2:Pulltypeagentcodedistribution

Migratetypeagentcodedistribution.Thisschemeisthemostobviousonetouseformobileagents.Codeismovedtogetherwiththeagentinstancefromstationtosta-tion.Usually,theperformanceislowerthanthepushschemeperformance,buthigherthanthepullschemeperformance.Comparedtothepushschemethemigrationtimeislonger,becauseagentinstanceandcodehavetobetransferred.Comparedtothepullscheme,communicationwithanrepositorystationisonlyrequiredwhenlaunchinganagent.Afterlaunchtheagentismoreflexible,becauseofbeingindependentoftheavailabilityofarepository.

Sincethereareadvantagesaswellasshortcomingsinallofthesethreedistributionschemes,itdependsontheapplication,onthetypeoftaskthemobileagenthastoworkon,whichschemeispreferable.Theavailabilityofmorethanoneschemealsogivesmoreflexibilityforcachemanagement.Allstationsprovidecachingofagentcode,inordertoimproveagentperformanceandtoreducecommunicationloadonthenetwork.Butsincethecachesizeislimited,cachedcodehastobedroppedfromtimetotime.Now,ifpushtypeagentcodehasbeenloaded,thiscodemaybedroppedbythecachemanager-ifrequired-evenbeforetheagentwhichisgoingtouseithasarrived.Inthiscase,theagentcodeisloadedcorrespondingtothepullschemewhen

10

Figure3:Migratetypeagentcodedistribution

theagentarrives.Therearegoodreasonstoconsiderdroppingpushedagentcodefromthecachebeforetheagentarrived.Theagentmightbedelayedbecauseofcongestioninthenetworkorbecauseofitslowpriority,oritmighthaveterminatedirregularly.Inthelattercaseitwillneverarriveatsomestationsanditwouldbewrongtokeepthecodeinthecacheinfinitely.

4.6Implementation

ThecurrentimplementationofINCAisbasedontheJavalanguage.Javaalreadysupportscodemobilityanddynamicclassloading.However,wereplacedtheJavaclassloaderinordertobeabletoloadcodefromtherepositorystationandtosupportmultiplecodemobilityschemes.AscommunicationinfrastructureweuseJavaRMI,butCORBAisalsosupported.

AgentsareimplementedasJavathreadswithagentprioritiestranslatedintoJavathreadpriorities.ThismappingfollowstheINCAsystempolicyimposedonvariousagenttypes(seeSection4.2),andtakesintoaccounttheloadlevelofthecurrentstation(e.g.thememoryusage).AssuchitismorepowerfulthanJava’spriority-basedscheduling.AgentinstancetransferisbasedonJavaobjectserialization.Thestateofanagentisrepresentedbyanobjectwhichcanbesentfromstationtostationinitsserializedrepresentation.

11

Asetofmanagedobjectsisprovidedasalibrary.NetworkelementslikeworkstationsorserverswhichcanhostanINCAstationaresupportedbymanagedobjectswithdirectaccesstothenetworkelement.NetworkelementswithoutcapabilitiestorunaJavavirtualmachine,likeroutersandATMswitchesaresupportedbyproxymanagedobjectsaccessingtheseelementsviathelocalareanetwork.However,ourexperienceswithINCAshowedthatformanyapplicationsco-developmentofagentsandmanagedobjectsisdesirable.

5ImpactofPrioritization

Inordertodemonstratetheusefulnessofagentprioritizationfornetworkmanagementapplications,wechoseasimplescenario.ItconsistsofasetofconnectedLANs.AteachLANthereisonenetworkelementhostinganINCAstation.EachINCAstationprovidesasetofmanagedobjectsallowingagentstoaccessthenetworkelementsintheLAN.Furthermore,weconsideramanagementstationlaunchingagentsformon-itoringandcontrol.MonitoringagentsarestationaryandmonitordedicatednetworkelementsofaLAN.Thus,theretypicallyaremorethanonemonitoringagentsatastation.Differenttothesecontinuouslyrunningmonitoringagents,therearecontrolagentswhicharelaunchedinteractivelybyanetworkoperatortoperformausuallyshorttask.Fortheseagentsinstantaneoustaskcompletionisdesired.

Figure5a)showsmeasurementswiththreeINCAstations,eachofthemhostingthesamevariablenumberofmonitoringagents.Theseagentsallhavethesamepriority.Thenacontrolagentislaunchedtovisiteachofthesestationsoncetoperformhistask.Ifthisagenthasthesamepriorityasthemonitoringagents(dashedline)thetimerequiredforitscompletionincreaseslineartothenumberofthreadsperstation.Ifotherwisehehasahigherprioritythanthemonitoringagents(solidline)thetimerequiredforitsexecutionisalmostindependentfromthenumberofmonitoringagentsandthereactiontotheoperatorisalwaysasfastaswithoutanymonitoringagents.Figure5b)showsasimilarmeasurement.Herethenumberofthreadsperstationisaconstantoftwoforallstations,butthenumberofstationsisvaried.Again,thead-vantageofhighprioritiesforcontrolagentscanbeobserved.ThemeasurementswereconductedonWindowsNT4.0workstationswithPentium166processors,connectedby10BaseTethernet.

6Conclusion

INCA,theIntelligentNetworkControlandManagementArchitecture,isaninfras-tructuretosupportapplicationsintheintegratednetworkandservicemanagementandtelecommunicationareas.INCAisespeciallytargetedatdistributedmanagementap-12

Figure4:MeasurementSetup

plicationsbasedonintelligentmobileagents.Thistechnologyovercomesseveraloftheinherentdifficultiesofcentralized,client-serverbasedmanagementsystems.Thearchitectureprovidesprioritizedagents,transactioncapabilities,migrationcon-trol,andmultipleagentcodetransferschemestosupportthecreationofhighlyreliableandefficientapplications.Theimpactofagentprioritizationhasbeendemonstratedforanexamplenetworkmanagementapplication.

7FutureWork

Sofar,INCAagentsareimplementedinastraightforwardmanner,withthealgorithmscodeddirectlyintothem.Althoughthisapproachissufficientforsmallerapplicationsandquickprototyping,wearegoingtoextendtheINCAplatformbylibrariesforagentintelligence.Hence,ourmainworkiscurrentlyfocusedonchoosingapropergoalrepresentationandproblemsolvingformalism.

Alsoweareawareoftheimportanceofsecurityinagentsystems,inparticularinnetworkmanagement.Varioussecuritymechanismshavealreadybeenproposed,butitisstillsubjecttoresearchwhichofthemissuitedbestformobileagentsinnetwork

13

execution time / seconds60604040202000246a) threads per station80135b) number of stations7Figure5:Executiontimeofacontrolagentmanagement.

8Acknowledgments

Thisworkwascarriedoutaspartofoureffortstowardsmaintainablehighspeednet-worksformultimediacommunications,attheNECC&CResearchLaboratoriesinBerlin.WearegratefultoS.IwasakiandN.Elshiewyforvaluablefeedbackandcom-ments,andwewouldliketothankS.RobidouforimplementingINCAapplicationsandperformingthemeasurementsusedinthispaper.

References

[1]BALDI,M.,GAI,S.,ANDPICCO,G.P.Exploitingcodemobilityindecentral-izedandflexiblenetworkmanagement.InProc.FirstInt.WorkshoponMobileAgents(Berlin,1997),K.RothermelandR.Popescu-Zeletin,Eds.,vol.1219ofLectureNotesinComputerScience,Springer-Verlag,Berlin,pp.13–26.[2]CASE,J.,FEDOR,M.,SCHOFFSTALL,M.,ANDDAVIN,J.Asimplenetwork

managementprotocol(SNMP);RFC-1157.InternetRequestforComments,1157(May1990).[3]CHARLTON,P.,CHEN,Y.,MAMDANI,E.,OLSSON,O.,PITT,J.,SOMERS,F.,

ANDWEARN,A.Anopenagentarchitectureforintegratingmultimediaservices.InProceedingsofthe1stInternationalConferenceonAutonomousAgents(NewYork,February5–81997),W.L.JohnsonandB.Hayes-Roth,Eds.,ACMPress,pp.522–523.

14

[4]

ROCHA,M.A.,ANDWESTPHAL,C.B.Proactivemanagementofcom-puternetworksusingartificialintelligenceagentsandtechniques.InTheFifthIFIP/IEEEInternationalSymposiumonIntegratedNetworkManagement(1997),Chapman&Hall,pp.610–621.

DA

[5]DELAFUENTE,L.A.,PAVON,J.,ANDSINGER,N.ApplicationofTINA-C

architecturetomanagementservices.LectureNotesinComputerScience851.[6]GOLDSZMIDT,G.,ANDYEMINI,Y.Distributedmanagementbydelegation.In

15thInternationalConferenceonDistributedComputingSystems(1995),IEEEComputerSociety.[7]KINIRY,J.,ANDZIMMERMAN,D.Ahands-onlookatjavamobileagents.IEEE

InternetComputing1,4(1997),21–30.[8]LANGE,D.B.,OSHIMA,M.,KARJOTH,G.,ANDKOSAKA,K.Aglets:Pro-grammingmobileagentsinjava.In1stInt’lConf.onWorldwideComputingandItsApplications’97(WWCA97)(Mar.1997),T.Masuda,Y.Masunaga,andM.Tsukamoto,Eds.,LNCS,SV.[9]LAZAR,A.A.Programmingtelecommunicationnetworks.IEEENetwork

(1997),8–18.[10]MAGEDANZ,T.,ROTHERMEL,K.,ANDKRAUSE,S.Intelligentagents:An

emergingtechnologyfornextgenerationtelecommunications?InIEEEINFO-COM1996(1996).[11]MEYER,K.,ERLINGER,M.,BETSER,J.,SUNSHINE,C.,GOLDSZMIDT,G.,

ANDYEMINI,Y.Decentralizingcontrolandintelligenceinnetworkmanage-ment.InIntegratedNetworkManagementIV(1995),A.S.Sethi,Y.Raynaud,andF.Faure-Vincent,Eds.,Chapman&Hall,pp.4–16.[12]ROONEY,S.Thehollowmananinnovativeatmcontrolarchitecture.InIM

(1997).[13]SAHAI,A.,BILLIART,S.,ANDMORIN,C.Astrolog:Adistributedanddy-namicenvironmentfornetworkandsystemmanagement.InProceedingsofthe1stEuropeanInformationInfrastructureUserConference(February1997).[14]SOMERS,F.Hybrid:Intelligentagentsfordistributedatmnetworkmanagement.

IATAworkshopatECAI’96,Budapest,August1996.[15]SignallingSystemNo.7–FunctionalDescriptionofTransactionCapabilities,

March1993.ITU-TRecommendationQ.771.[16]STALLINGS,W.SNMP,SNMPv2andCMIP:ThePracticalGuidetoNetwork-ManagementStandards.Addison-Wesley,Reading,1993.

15

[17]

MERVE,K.,ANDLESLIE,I.M.Switchletsanddynamicvirtualatm

networks.InIM(1997).

VANDER

[18]WHITE,J.E.Telescripttechnology:Thefoundationfortheelectronicmarket-place.Whitepaper,GeneralMagic,Inc.,2465LathamStreet,MountainView,CA94040,1994.[19]WONG,D.,ETAL.Concordia:AnInfrastructureforCollaboratingMobile

Agents.InProc.FirstInt.WorkshoponMobileAgents(Berlin,1997),K.Rother-melandR.Popescu-Zeletin,Eds.,vol.1219ofLectureNotesinComputerSci-ence,Springer-Verlag,Berlin,pp.86–97.

16