Lecture Notes in Control and Information Sciences 308 Editors:M.Thoma · M.Morari (cid:1) (cid:1) S. Tarbouriech C.T.Abdallah J. Chiasson (Eds.) Advances in Communication Control Networks With99Figures SeriesAdvisoryBoard A.Bensoussan·P.Fleming·M.J.Grimble·P.Kokotovic· A.B.Kurzhanski·H.Kwakernaak·J.N.Tsitsiklis Editors Dr.SophieTarbouriech Prof.ChaoukiT.Abdallah LAAS-CNRS TheUniversityofNewMexico 7,AvenueduColonelRoche DepartmentofElectricalandComputerEngineering 31077Toulousecedex4 EECEBldg.MSC011100 France Albuquerque,NM87131-0001 USA Prof.JohnChiasson UniversityofTennessee ECEDepartment 1508MiddleDrive Knoxville,TN37996-2100 USA ISSN0170-8643 ISBN3-540-22819-5 SpringerBerlinHeidelbergNewYork LibraryofCongressControlNumber:2004111522 Thisworkissubjecttocopyright.Allrightsarereserved,whetherthewholeorpartofthemate- rialisconcerned,specificallytherightsoftranslation,reprinting,reuseofillustrations,recitation, broadcasting,reproductiononmicrofilmorinotherways,andstorageindatabanks.Duplication ofthispublicationorpartsthereofispermittedonlyundertheprovisionsoftheGermanCopyright LawofSeptember9,1965,initscurrentversion,andpermissionforusemustalwaysbeobtained fromSpringer-Verlag.ViolationsareliabletoprosecutionunderGermanCopyrightLaw. SpringerisapartofSpringerScience+BusinessMedia springeronline.com ©Springer-VerlagBerlinHeidelberg2005 PrintedinGermany Theuseofgeneraldescriptivenames,registerednames,trademarks,etc.inthispublicationdoes notimply,evenintheabsenceofaspecificstatement,thatsuchnamesareexemptfromtherelevant protectivelawsandregulationsandthereforefreeforgeneraluse. Typesetting:Dataconversionbytheauthors. FinalprocessingbyPTP-BerlinProtago-TeX-ProductionGmbH,Germany Cover-Design:design&productionGmbH,Heidelberg Printedonacid-freepaper 62/3020Yu-543210 Preface The area of communicationand computernetworkshas become a veryactive area ofresearchbytheControlSystemscommunityinthelastfewyears.Therecentspe- cialissues in journals(e.g.,IEEETransactionsonAutomaticControl, Vol. 47,No. 6,June2002,Automatica,Vol.35,December,1999,andControlEngineeringPrac- tice, Vol. 11, 2003 ) and the many special sessions in control conferences certify to this strong interest (see, for example, the plenary sessions of T. Basar and of F. Kelly in the IEEE 2003 Conference on Decision and Control, and the 2003 Euro- peanControlConferences,respectively,aswellasthemorethan15sessionsinthe IEEE2003ConferenceonDecisionandControl).Asdescribedinaspecialissueof IEEEControlSystemsMagazine(vol.21,no.1,February2001),theareaofcommu- nicationnetworksis of greatinterestto controlresearchersbecauseits challenging problemsfallwithinthescopeandbackgroundofsystemsandcontrolengineering. Furthermore,the increasingneedto controldynamicalsystemsvia communication networksprovidesfertilegroundsforcontroltheoristsaswellaspractitioners. As a result, both analysis and synthesis approaches by the control community have appeared in the context of network control. Tools from convex optimization and controltheory are playingincreasing roles in efficientnetwork utilization, fair resourceallocation,andcommunicationdelayaccommodation.Ontheotherhand,as feedbackcontrolsystemsbecomeincreasinglydistributedanddependentonshared networksfortheirefficientoperation,thefieldofNetworkedControlSystems(NCS) is fast becoming a mainstay of control systems research and applications. In what follows,theterm“communicationcontrolnetworks”willrefertobothnetworksun- dercontrol(controlinnetworks)aswellasnetworkedcontrolsystems(controlover networks). Thecomplexityinthedesignandoperationofcommunication/controlnetworks, alongwiththeirreal-timerequirements,severelylimittheabilitytoobtainaccurate mathematicalmodels.Suchmodelsareusuallyhybrid(containingbothtime-driven andevent-drivendynamics),uncertainduetointentionalsimplification(fluidmodel approximations),or to parametric inaccuracies (uncertain delays), or still from the presenceofadditivedisturbances(disruptiveloadsandcongestion).Furthermore,in all practical situations, the network devices and systems are limited in their range VI Preface of operation (limited queue lengths and rates) and therefore are subject to ampli- tude and rate saturation. Hence, large amplitude disturbances or uncertainties may drive the states and/or the controls into saturation where the system may operate in a modefromwhich itmay be difficultto preserveperformanceor evenstability requirements.Finally,theinherentpresenceofpacketsloss,delays,andthelimited- capacity communication medium are importantfeatures that need to be taken into accounttoguaranteetheperformanceandstabilityofsuchnetworks. Ourobjectiveineditingthisbookwithsolicitedcontributionsfromexpertsinthe variousareasofcommunication/controlnetworks,istopresentinterestingandcom- plementarytechniquesthattreatproblemsarisinginthecontrolofnetworksaswell asincontrolacrossnetworks.Wealsoseethisbookasamodestattempttoreversethe trendoffragmentationandspecializationoftherelatedfieldsofcontrol,communi- cation,andcomputing.Thebookhaditsgenesisinamultidisciplinarycollaborative effortbetweenFranceandtheUSA,undertheauspicesofCNRSandNSF,concern- ingproblemsinstability,stabilization,andoptimizationfortime-delaysystemsand related applications including the control of communication networks. Hence, the developments proposed in the sixteen chapters are interdisciplinary as they cover variousresearchfieldsincludingControls, Communications,AppliedMathematics, andComputerScience. Thebookisorganizedasfollows. • Part1isdevotedtoFluid/FlowModelsandconsistsofchapters1through5. • Part2isdevotedtoCongestionandNonlinearSystemsandconsistsofchapters 6through10. • Part3 isdevotedto LoadBalancingandTeleoperationandconsistsofchapters 11through13. • Part4isdevotedto“EmergingControlTheoryandInformationComplexity”and consistsofchapters14to16. Note that this partition is somewhatarbitraryas most of the chapters are inter- connected,andmainlyreflectstheeditors’biasesandinterests. Wehopethatthisvolumewillhelpinclaimingmanyofthecommunicationsand networkingproblemsfor controlsresearchers,and to alert graduatestudentsto the many interesting ideas at the boundary between communications, computing, and controls. Acknowledgements The idea of this edited bookwas formedthrougha series of e-mailexchangesand face-to-facediscussionswhenthethreeofusmetatconferencesaspartoftheproject “Time-delaysystems:analysis,computeraideddesign,andapplications,”startedin 2002andjointlyfundedbyCNRS(France)andNSF(USA).Indeed,manycontrib- utors ar e amongst the participantsof that project, and most of the French partici- pantsal so belongto CNRS GDR Automatique:Delay systems, a researchnetwork Preface VII inAutomaticControlinexistencesince1995.However,wealsoinvitedresearchers outsidetheseresearchteamsinordertoprovideadeeperandmorebalancedrepre- sentationoftheareas. Firstandforemost,wewouldliketothankallthecontributorsofthebook.Wit houttheirencouragement,enthusiasm,andpatience,thisbookwouldhavenotbeen possible. A list of contributors is provided at the end of the book. Professors S.I. Niculescu and K. Gu in particular have contributed more than their share to the birthing of this volume. Next, we would like to thank CNRS and NSF, and more specificallyourprogrammanagersClaireGiraud,Jean-LucCle´ment(CNRS),Rose Gombay and Kishan Baheti (NSF) for funding the joint research which made this book possible. We would also like to thank Luis Farinas Del Cerro (DRI-CNRS). ThanksalsogotoGDRAutomatique(France),MENRT(France)andthelaboratory LAAS-CNRS.Finally,thankalso IsabelleQueinnec(LAAS-CNRS) forherhelpre- gardinglatexproblems. We also wish to thank Springer for agreeing to publish this book. We wish to express our gratitude to the Editor-in-Chief Dr. Manfred Thoma, to Dr. Thomas Ditzinger (Engineering Editor), and Ms. Heather King (International Engineering Editorial)for their carefulconsiderationand helpfulsuggestionsregardingthe for- matandorganizationofthebook. Toulouse,France,May2004, SophieTarbouriech Albuquerque,USA,May2004, ChaoukiT.Abdallah Knoxville,USA,May2004, JohnN.Chiasson Contents ControlofCommunicationNetworksUsingInfinitesimalPerturbation AnalysisofStochasticFluidModels ChristosPanayiotou,ChristosG.Cassandras,GangSun,YoraiWardi........ 1 StabilizedVegas HyojeongChoe,StevenH.Low ...................................... 27 RobustControllerDesignforAQMandH∞-PerformanceAnalysis PengYan,HitayO¨zbay ............................................. 49 Models and Methods for Analyzing Internet Congestion Control Algorithms RayadurgamSrikant ............................................... 65 DelayEffects ontheAsympt oticStabilityofVari ous FluidModels in High-PerformanceNetworks Silviu -IulianNiculescu,Wim Michiels,DanielMelchor- Aguillar, TatyanaLuzyanina,Fre´de´ricMazenc,KeqinGu,FabienChatte´............. 87 ASlidingModeApproach toTrafficEngineeringinComputer Networks BernardoA.Movsichoff,ConstantinoM.Lagoa,HaoChe .................111 ModelingandDesigningtheInternetCongestionControl SaverioMascolo ..................................................137 SaturatedController DesignofanABRExplicit RateAlgorithmfor ATMSwitches SophieTarbouriech,MarcoAriola,ChaoukiT.Abdallah ..................159 State-SpaceModels for ControlandIdentification Henri-Franc¸ois Raynaud,CarolineKulcsa´r,RimHammi ..................177 X Contents GlobalStability ofNonlinear CongestionControl withTime-Delay ZhikuiWang,FernandoPaganini .....................................199 On theOptimizationofLoadBalancinginDistributedNetworks in the PresenceofDelay Sagar Dhakal,MajeedM.Hayat,JeanGhanem,ChaoukiT.Abdallah,Henry Jerez,JohnChiasson,J.Douglas Birdwell..............................223 ClosedLoopControlofaLoadBalancingNetwork withTimeDelays andProcessor ResourceConstraints ZhongTang,J. Douglas Birdwell, JohnChiasson,ChaoukiT. Abdallah, MajeedM.Hayat .................................................245 PositionandForceTrackinginBilateralTeleoperation NikhilChopra,MarkW.Spong,RomeoOrtega,NikitaE.Barabanov .........269 SuboptimalControlTechniques for NetworkedHybridSystems SorinC.Bengea,Peter F.Hokayem,RayA.DeCarlo,ChaoukiT.Abdallah....281 CommunicationRequirements for NetworkedControl Sekhar Tatikonda,NicolaElia .......................................303 AnIntroduction toNonlinear Fault Diagnosis withanApplication toa CongestedInternet Router MichelFliess,Ce´dricJoin,Hugues Mounier ............................327 List ofContributors ..............................................345 Control of Communication Networks Using Infinitesimal Perturbation Analysis of Stochastic Fluid Models ChristosPanayiotou1,ChristosG.Cassandras2,GangSun3,andYoraiWardi4 1 Dept.ofElectricalandComputerEngineering,UniversityofCyprus,Nicosia,Cyprus, [email protected] 2 Dept.ofManufacturingEngineeringandCenterforInformationandSystemsEngineering, BostonUniversity,Brookline,MA02446,[email protected] 3 Dept.ofManufacturingEngineeringandCenterforInformationandSystemsEngineering, BostonUniversity,Brookline,MA02446,[email protected] 4 SchoolofElectricalandComputerEngineering,GeorgiaInstituteofTechnology,Atlanta, GA30332,[email protected] 1 Introduction Managingand operatinglargescale communicationnetworksisa challengingtask and it is onlyexpectedto getworse as networksgrowlarger.The difficultiesasso- ciated with networkmanagementstem from the factthat modelingand analysisof large scale communicationnetworks is an excessively difficult task. On one hand, theenormoustrafficvolumeintoday’sInternetmakespacket-by-packetanalysisin- feasible.Ontheotherhand,queueingsystems(thenaturalmodelingframeworkfor packet-basedcommunicationnetworks)arelargelybasedonPoissonprocessesand does not capture the bursty nature of realistic traffic. Moreover, the discovery of self-similar patterns in the Internet traffic distribution [1] and the resulting inade- quacies of Poisson traffic models [2] further complicate queueing analysis. At the sametimeweneedtoaccountforthefactthatthestochasticprocessesinvolvedare time-varying,i.e.,nostationarityassumptionshold.Inaddition,weneedtoexplicitly modelbufferoverflowphenomenawhichtypicallydefytractableanalyticalderiva- tions.Consequently,performanceanalysistechniquesthatdonotdependondetailed trafficdistributionalinformationarehighlydesirable. An alternative modeling paradigm based on fluid models has become increas- inglyattractive.Theargumentleadingtothe popularityoffluidmodelsisthatran- dom phenomenamay play differentroles at differenttime scales. When the varia- tionson the faster time scale have less impactthan those onthe slower time scale, theuseoffluidmodelsisjustified.Theefficiencyofafluidmodelrestsonitsability toaggregatemultipleevents.Byignoringthemicro-dynamicsofeachdiscreteentity andfocusingonthechangeoftheaggregatedflowrateinstead,afluidmodelallows S.Tarbouriechet al.(Eds.):Advances inCommunicationControlNetworks,LNCIS308,pp.1–26,2004. ©Springer-VerlagBerlinHeidelberg 2005 2 ChristosPanayiotouetal. theaggregationofeventsassociatedwiththemovementofmultiplepacketswithina timeperiodofaconstantflowrateintoasingleratechangeevent.Introducedin[3] and later proposedin [4] for the analysis of multiplexeddata streams and network performance[5],fluidmodelshavebeenshowntobeespeciallyusefulforsimulating variouskindsofhighspeednetworks[6, 7, 8, 9].AStochasticFlow Model(SFM) has the extra feature that the flow rates are treated as generalstochastic processes, which distinguishesitself fromthe approachadoptedin [10, 11, 12] that dealwith deterministicorMarkovmodulatedfluidrates. Ontheotherhand,thefluidmodelingparadigmforgoestheidentityanddynam- icsofindividualpacketsandfocusesinsteadontheaggregateflowrate.Asaresult, this paradigm is more suitable for network-relatedmeasures, such as buffer levels andpacketlossvolumes,ratherthanpacket-relatedmeasuressuchassojourntimes (althoughit is still possible to define fluid-basedsojourntimes [13]). A Quality of Service (QoS) metric that depends on the identity of certain packets, for example, cannot be obviously captured by a fluid model. Furthermore, for the purpose of performance analysis of networks with QoS requirements, the accuracy of SFMs depends on traffic conditions, the structure of the underlying system, and the na- ture of the performance metrics of interest. Moreover, some metrics may depend onhigher-orderstatisticsofthedistributionsoftheunderlyingrandomvariablesin- volved,whichafluidmodelmaynotbeabletoaccuratelycapture. Inthischapter,ourgoalistoexploretheuseofSFMsforthepurposeofcontrol andoptimizationratherthan performanceanalysis. In thiscase, it is notunreason- abletoexpectthatonecanidentifythesolutionofanoptimizationproblembasedon amodelwhichcapturesonlythosefeaturesoftheunderlying“real”systemthatare neededtoleadtotherightsolution,withouttheneedtoestimatethecorresponding optimal performance with accuracy. Even if the exact solution cannot be obtained bysuch“lower-resolution”models,onecanstillobtainnear-optimalpointsthatex- hibitrobustnesswithrespecttocertainaspectsofthemodeltheyarebasedon.Such observationshavebeenmadeinseveralcontexts(e.g.,[14]),includingrecentresults relatedtoSFMsreportedin[15]whereaconnectionbetweentheSFMandqueueing- system-basedsolutionisestablishedforvariousoptimizationproblemsinqueueing systems. UsingtheSFMmodelingframework,anewapproachfornetworkmanagement isbeingdevelopedwhichisbasedonInfinitesimalPerturbationAnalysis(IPA)[16, 17, 18, 19] (IPA is coveredin detailin [20, 21]). Inthis approach,we estimate the gradientoftheperformancemeasureofinterest(e.g.,packetlossrate)withrespect tothecontrolparametersofinterest(e.g.,bufferthresholds)andusetheminstandard stochasticapproximationalgorithmstodeterminetheoptimalparametersetting.This approachhassomeveryimportantadvantages. • The gradientestimation is done on-line thus the approachcan be implemented ontherealsystem:astheoperatingconditionschange,itwillaimatcontinuously seekingtooptimizeagenerallytime-varyingperformancemetric. • Thegradientestimationprocessdoesnotrequireanyknowledgeofthesystem’s underlyingstochasticprocesses;inotherwords,itismodelfree.