2001年度日本オペレーションズ・リサーチ学会 秋季研究発表会
2−D−8
Optimal Software Rejuvenation Policies with Discounting
TadashiDohi(01307065),1もkashiDanjou,HiroyukiOkamura(01013754) DepartmentofInbrmationEngineerlng,HiroshimaUniversity thedistributionfunctionsofthetimetosoftwarerquve− nationanditssystemoverhead by F(t)and凡(t)(with densityふ(t)andmeanFLr(>0)),reSpeCtively・Aftercom− pletingtherepairortherejuvenation,thesoftwaresystem becomesasgoodasnew,andthesoftwareageisinitiated at the beginingofthe next high1y robust state・Conse− quently,Wedefinethetimeintervalfromthebeginingof the system operation to the next one as one cycle,and thesamecycleisrepeatedagalnandagalnOVeraninfinite
time horizon.
Note that the software rejuvenation cycle is measured fromthetimeinstantjustafterthesystementersStatel fromStateO.Ifweconsiderthetimetosoftwarerqiuve− nationasaconstantto,thenitb1lowsthat 〈 珊=U(f−lo)= 0 (1) Wecall亡0(≧0)舶亡んe叩βぴα代rり祝γeγlα如れβCんe血gein thispaperandU(・)istheunitstepfunction・Hence,the underlying stochastic process is a semi-hlarkov rprocess
Withfourregenerativestatest 3.Analysis 1.Introduction SoftwarerqjuvenationisapreventivemaintenanCeteCh− nlquethathasbeenextensivelystudiedintherecentliter− ature【1,2]・Inthispaper,WeCOnSiderageneralizedprob− 1emtoestimatetheoptimalsoftwarerqiuvenationsched− ule.Moreprecisely,thesoftwarerqjuvenationmodelsare formulatedviathesemiTMarkov processes,and theopti− malsoftware rquvenation schedule which minimizes the
expectedtotaldiscountedcostoveraninfinitetimehori− ZOnarederivedanalytically,
2.ModelDescription
Fb1lowingHuanget.al[1】andDohietal・(2】,WeCOn−
Siderthetwo−Stepfailuremodeltodescribetheaglngphe− nomenonin telecommunications bi11ingapplications・De−
finethefo1lowlngfourstates: StateO:high1yrobuststate(normaloperationstate) Statel:failureprobablestate State2:failurest,ate State3:SOftwarerqjuvenationstate
Supposethatallthestatesmentionedaboveareregen−
erationpoints.Morespecifically,1etZbetherandomtime intervalwhenthehigh1yrobuststatechangeStOthefail− ureprobablestate,havingthecommondistributionfunc− tionPr(Z≦t)=Fb(l)withdensityfo(t)and丘nitemean FLo(>0)・Justafterthestatebecomesthefailureprobable State,aSyStemfailuremayoccurwithapositiveprobabil− ity.Withoutanylossofgenerality,itisassumedthatthe randomvariableZisobservableduringthesystemopera− tion【1,2】・ DefinethefailuretimeX(fromStatel)andtherepair time Y,having thedistribution functions Pr(X≦t)= Fj(t)andPr(Y≦t)=FL(t),reSpeCtively,Where ff(t) andん(t)aretheassociatedprobabilitydensityfunctions and入f(>0)andFLa(>0)aretheirmeanvalues,reSpeC− tively.IfthesystemfailureoccursbeforetriggerlngaSOft− WarereJuVenation,thentherepalrisstartedimmediatelyat that time and is completed after the random time Y
elapses.Otherwise,thesoftwarerqjuvenationisstartedas apreventivemaintenanceofthesoftwaresystem.Denote DefinethefbllowlngCOStCOmpOnentS‥ Cs(>0):repaircostperunittime cp(>0):rqjuvenationcostperunittime β(>0):discounthctor・ Forconvenience,Wedefine
ノ:、㍉
exp(−α5)J(5)d5 (2) £(J(α))=for an arbitrary continuousfunction f(・)and acomplex numberα.Thatis,thefunCtionL:(f(α))istheLaplace transformofthefunctionf(・),
The discounted unit cost for one cycle,i・e・the net presentvalueofonedollarafteronecycle,isformulatedas ∂(fo)= ‘し ( 。 〃−ん ト・州、︰ ∞ 〝/ム 〝′−悪
∞ ′ム
♪−−JO + 上 e ̄β(坤0+5)d巧(ご)d凡(z)d凡(5). −234一 © 日本オペレーションズ・リサーチ学会. 無断複写・複製・転載を禁ず.(i)Ifq(0)<Oandq(∞)>0,thenthereexistsafinite andunlqueOptimalsoftwarereJuVenationschedulet岩 (0<t岩<∞)satisfyingq(t岩)=0,andtheminimum expectedtotaldiscountedcostis rC(t岩)= c5Z(Jα(β))rJ(f岩トcpZ(ん(β))(β+り(f岩)) (3) Theexpectedtotaldiscountedcostfbronecycleis c5e−β(トトエ+欄) V(£0) ×d亡dF′(諾)d凡(z)d凡(y) +上 上
.(.
J5
cpe榊0再) β〈β姉(β)}+[瑚(β))−qん(β)}]γ∫㈲)■ (8) (ii)Ifq(0)≧0,thentheoptimalsoftwarer毎uvenation SCheduleis t岩 =0,i.e.itis optimalto start the SOftware rQjuvenationjust after entering the failureprobablestate,andtheminimumexpectedtotaldis− COuntedcostisgivenbyTC(0). (iii)Ifq(∞)≦0,thentheoptimalsoftwarerQjuvenation SCheduleist岩→∞,!.e.itisoptimalnottocarryout theso軋warerquvenation,andtheminimumexpected totaldiscountedcostisgivenbyTC(∞)・ (2)SupposethatthefailuretimedistributionisDFR(de− CreaSingfailurerate)undertheassumption(A−1)・Then, the expected totaldiscounted costfunction TC(to)is a
COnCaVefunctionoflo,andtheoptimalsoftwarerQjuvena− tionscheduleisl岩=00rt岩→00.
From the theorem above,itis seen that the optimal SOftware rquvenation policy t岩can be calculatedif the underlyingfailuretimedistributionFj(t)andothermodel
parametersaregiven.Ontheotherhand,iftheknowledge
OfFj(t)isnot available,anOnparametricmethod based On the modified totaltime on test statistics[3】can be appliedtoestimatet岩fromthecompletesampleoffailure
timedata・These detailed resultswi11be reportedin the
conference.
×dtd巧(ェ)d穐(z)dn(5) (4)
Theexpected totaldiscounted cost over aninfinite time horizonisglVenby Cく:) rc(to)=∑v(瑚(fo)k=咋0)/和0)・ た=0 (5) Thentheproblemistoseektheoptimalsoftwarerqiuve− nationschedulet岩whichminimizesTC(to). DefinethenumeratorofthederivativeofTC(壬0)with respecttoto,dividedbythefactor戸,(to)e▼βto,aSq(to), 慮.e. £(ふ(β)) ト恥(β)輌0)−Cp恥(β)} 9(to) o ))]恥)+瑚(β)) ×(β+り(f ×[(拙(β)}一姉(β)})γ㈲ −β£(ん(β)) V(to), (6) whererf(t)=fI(t)/戸f(t)isthehilurerateandingeneral 石(・)=1−4)(・).Itisassumedthatrf(t)iscontinuousand di鮎rentiablewithrespecttot. Wemakethefo1lowlngaSSumption: (A−1)C諾嶽鵠ヱ嵩謂)>rC(fo)・ In[2】,itisprovedthroughthereductionargument that (cβ〃α一年〃r)/(〝α−〝r)>1imβ→0β・rC(fo)bralltoif twoparametricassumptionsFLa>〃,andcs>cphold・In theassumption(A−1),takingβ→Oyields Refもrences 【1】Huang,Y・,Kintala,C・,Kolettis,N・andFunton,N・D・ (1995),SoftwarerQjuvenation‥analysis,mOduleand applications,PrDC・25thlEEEInt,lSymp・On
nlerant Computin9,381−390,IEEE CS Press,Los Alamitos,CA.
【2]Dohi,T・,Go畠eva−Popstojanova,K・andTヒivedi,K・S, (2000),Estimatingsoftwarerqjuvenationschedulein highassuranCe SyStemS,77Le ContputerJoumal(in press).
【3]Bergman,B・and Kle勾6,B・(1983),TTT−tranSform and age replacements with discounted costs,Naval
月e5.⊥0タ由.Q祝αr豪.,30,63ト639. c占Z(揖β))−CpZ(〃β))_C町 (7) 1illl β二らβ【£(ム(β))−£(ん(β))】 〃α∼〃r 丘omthel’Hospital’stheorem.Inotherwords,theassump− tion(A−1)may holdin the most cases with su瓜ciently
smal1discount,factor.
Thefbllowlng reSultgives theoptimalsoftware rejuve− nationpolicywithdiscounting. Theorem:(1)Supposethatthefailuretimedistribution isstrictlyIFR(increasingfailurerate)undertheassump− tion(A−1)・ ー235− © 日本オペレーションズ・リサーチ学会. 無断複写・複製・転載を禁ず.