StabilityandInstabilityofFluidModelsforRe-EntrantLinesJ.G.Dai1andG.Weiss2GeorgiaInstituteofTechnologyAbstractRe-entrantlinescanbeusedtomodelcomplexmanufacturingsystemssuchaswaferfabricationfacilities.Asthe rststeptotheoptimalornear-optimalschedulingofsuchlines,weinvestigatetheirstability.InlightofarecenttheoremofDai(1995)whichstatesthataschedulingpolicyisstableifthecorresponding uidmodelisstable,westudythestabilityandinstabilityof uidmodels.TodothisweutilizepiecewiselinearLyapunovfunctions.WeestablishstabilityofFirst-Bu er-First-Served(FBFS)andLast-Bu er-First-Served(LBFS)disciplinesinallre-entrantlines,andofallwork-conservingdisciplinesinanythreebu erre-entrantlines.Forthefourbu ernetworkofLuandKumarwecharacterizethestabilityregionoftheLuandKumarpolicy,andshowthatitisalsotheglobalstabilityregionforthisnetwork.WealsostudystabilityandinstabilityofKelly-typenetworks.Inparticular,weshowthatnotallwork-conservingpoliciesarestableforsuchnetworks;however,allwork-conservingpoliciesarestableinaringnetwork.RunningTitle:Stabilityfor uidmodelsAMS1991subjectclassi cation:Primary60K25,90B22;Secondary60K20,90B35.Keywordsandphrases:Stability,unstablenetworks, uidmodels,piecewiselinearLya-punovfunctions,re-entrantlines,multiclassqueueingnetworks,schedulingpolicies,Harrisrecurrence.February6,1994RevisedNovember,1994FinalversionJanuary19,1995MathematicsofOperationsResearch,Vol.21,115{135(1996)1ResearchsupportedbyNSFgrantsDMS-9203524andDDM-9215233,andtwograntsfromtheTexasIn-strumentsCorporation2ResearchsupportedbyNSFgrantsDDM-8914863andDDM-9215233,andthefundforthepromotionofresearchattheTechnion1IntroductionConsideramulticlassqueueingnetworkwithasinglerouteforallcustomers.ThereareIstations(nodes)inthenetwork.Allthecustomersfollowa xeddeterministicKstageroutethroughthenetwork.Weshallnumberthestagessothateachcustomerwillenterthesysteminstage1,andoncompletionofstagekwillmovetostagek+1,k=1;:::;K 1,andleavethesystemoncompletionofstageK.Wedesignatethosecustomersonthekthstageoftheroute(thekthvisitalongtheroute)asclasskcustomers.Weenvisionclasskcustomerswaitinginbu erk,whichisassumedtohavein nitecapacity(whencustomersareservedinFIFOrule,itisenoughtohaveonebu erforeachstation).Foreachklet (k)bethestationnumberthatclasskcustomersvisit(thestationservingstagek).Thismodeliscalledare-entrantline,seeKumar(1993).Adistinctivefeatureofare-entrantlineisthatcustomersmayvisitaparticularstationmorethanonce,sothateachnodeservesseveralclasses.Letf (n);n 1gbeasequenceofpositiverandomvariables.Thenthrandomvariable (n)isinterpretedastheinterarrivaltimebetweenthe(n 1)thcustomerarrivalandthenthcustomerarrivalfromoutside;the rstcustomerarrivesattime (1).Theservicetimesatdi erentstages(visits)forthenthcustomerare 1(n),:::, K(n).Wemakethefollowingthreeassumptions.Firstweassumethatf( (n); 1(n);:::; K(n));n 1gisaniidsequence.(1.1)Thenweassumethatalltherandomvariableshave nite rstmoments.Thatis,E[ (1)]1andE[ k(1)]1fork=1;:::;K:(1.2)Finally,weassumethattheinterarrivaltimesareunboundedandtheirdistributionisspreadout.Thatis,foranyx0,Probaf (1) xg0;(1.3)andforsomeintegern0andsomefunctionp(x) 0onR+withR10p(x)dx0,Proba(a nXi=1 (i) b) Zbap(x)dx;forany0 ab:(1.4)Notethatin(1.1),weallowtheservicetimesatdi erentstagesofvisitstohavearbitraryde-pendency.Thisfeatureisusefulforcertainapplications,notablyincomputercommunicationsandmanufacturingsystems.Therethelengthofacomputermessageorthesizeofamanu-facturinglotmayberandom.However,theservicetimesingeneralareproportionaltothemessagelengthorlotsize,andthereforearepositivelycorrelated.Wealsoallowdependenceoftheservicetimesonthepreviousinterarrivaltime.LetCi=fk: (k)=ig.ThesetCiiscalledtheconstituencyofstationi.LetCbetheI Kincidencematrix,Cik=(1if (k)=i0otherwise.(1.5)1Withoutlossofgenerality,weassumethatE[ (1)]=1:(1.6)Letmk=E[ k(1)]bethemeanservicetimeforclasskcustomers.Weassumethatthereisasinglereliableserverateachstation.Foreachi=1;:::;I,let i=Xk2Cimk:Wecall ithenominalworkloadforserveriperunitoftime(recallthatthearrivalrateisnormalizedtobeonein(1.6)).Throughoutthispaper,weassume i1fori=1;:::;I.(1.7)Nothinghasbeensaidyetaboutaqueueingdiscipline,whichdictatestheorderinwhichcustomersareservedateachstation(wewillusequeueingdisciplineandschedulingpolicyinterchangeably).Speci cqueueingdisciplineswillbediscussedinlatersections.Weassumethatalldisciplinesarework-conserving.Thatis,serveriworksatfullspeedwheneverthereisworktodoatstationi.InDai(1995),theauthorintroducedastochasticprocessfX(t);t 0gthatdescribesthedynamicsofthequeueingnetworkunderaspeci cqueueingdiscipline.Foreacht 0,X(t)=(X1(t);:::;XI(t)),whereXi(t)isthestateatstationiattimet.Theexactde nitionofstatedependsontheparticularqueueingdiscipline.Forexample,if rst-in- rst-out(FIFO)disciplineisusedateachstation,thenoneneedstotakeXi(t)= (xi1(t);:::;xini(t)(t));u(t);vxi1(t)(t) ;whereni(t)isthequeuelength(includingthepossibleonebeingserved)atstationi,xij(t)istheclassnumberofthejthcustomeratthestation,u(t)istheresidualtimeforthenextexternalcustomertoarrive,vxi1(t)(t)istheresidu