Intenational Jounal o Compute Science & Inomation Technology (IJCSIT) Vol 4, No, June 0 HIGH AVAILAILITY SOLUTION: ESOUCE USAGE MANAGEMENT IN VITUALIZED SOFTWAE AGING Aye Myat Myat aing and Ni La Thein Univesity o Compute Studies, Yangon amyatp@googlemail.com ASTACT As businesses inceasingly ely on IT o thei mission-citical opeations, continuous availability is a univesal concen. Nowadays, vitualized platom has become the popula option to deploy complex enough sevices. Deployed sevices ae expected to be always, available, but these long unning sevices ae especially sensitive to sue om sotwae aging phenomenon. Sotwae aging o vitual machine monitos (VMMs) is becoming citical because peomance degadation o cash ailue o a VMM aects all vitual machines (VMs) on it. To counteact this issue, we deploy sotwae ej uvenation methodology. To eliminate the sevice outage duing sotwae ejuvenation pocess, we combine the ejuvenation with live ation technology. Live VM ation enables a unning VM on a host see to move onto the othe host seve with vey small intetion o the execution. Live VM ation depends on VMs placement and eicient esouce available is equied. The idea behind ou pape is two-old. Fist, we pesent the optimization o the esouce usage as accepting as many sevices as in vitualized envionment which spot o VM live ation. Second, to demonstate how much it can impove system availability, the stochastic eti nets model a vitualized seve system in case o using time based sotwae ejuvenation o VMM is pesented. Finally, we peom the numeical analysis to evaluate the model. KEYWODS Availability, Clusteing Technology, Sotwae Aging, Sotwae ejuvenation, Stochastic eti Nets, Vitualization.. INTODUCTION Availability has long been a citical issue o online compute systems whose ailue can halt business pocess. This is paticulaly moe petaining to high demanded high available (HA) computing systems based on o-the-shel components. Exhaustion o system esouces, data cotion, and numeical eo accumulation ae the pimay symptoms o the degadation, which may eventually lead to peomance degadation o the sotwae, cash/hang ailue, o othe undesiable eects. That degadation o sotwae is known as sotwae aging. Sotwae aging has not only been obseved in sotwae used on a mass scale but also in specialized sotwae used in high-availability and saety citical applications []. System Vitualization techniques ae getting popula and gaining signiicant inteest in the entepise and pesonal computing spaces. The majoity o today s high availability (HA) clustes is based on eal physical hadwae and vitualization is coming moe and moe popula nowadays, one has to think about possible combinations o vitualization and high availability clusteing. Moden computes ae suiciently poweul to use VMs. Many ields, such as autonomic computing, seve consolidation, secuity and education publish esults that paise the beneits o DOI : 0.5/ijcsit.0.408 85
Intenational Jounal o Compute Science & Inomation Technology (IJCSIT) Vol 4, No, June 0 vitualization. Sotwae ejuvenation can be applied so as to mitigate advese eects o sotwae aging. Sotwae ejuvenation [] is a poactive ault management technique aimed at cleaning the system intenal state to pevent the occuence o moe sevee cash ailues in the utue. To eliminate the sevice outage duing sotwae ejuvenation pocess, we combine ejuvenation with vitualization technology. A vitualization laye also called vitual machine monito (VMM) is a sotwae laye that abstacts the physical esouces o use by the VMs. ecently, sotwae aging o VMMs is becoming citical because many VMs un on top o a VMM in one machine consolidating multiple seves and aging o the VMM diectly aects all the VMs. Without any ejuvenation, both the OS sotwae and the application sotwae unning on top o VM degade in peomance with time due to the exhaustion o system esouces such as ee physical memoy, and eventually cash, which is vey undesiable to high availability (HA) systems and atal to mission-citical applications. One o the pomises o vitualization is the ability to allow applications to dynamically move om one physical seve to anothe as the demands and esouce availabilities change, without sevice intetion. Moe ecently, Xen [] and VMWae [8] have implemented ``live'' ation o VMs that involve extemely shot downtimes. To ate, we must know which vitual machine needs to be ated and when this elocation has to be done and, moeove, which host must be destined. The main contibutions o this pape ae two-old. Fistly, we pesent the optimization o the esouce usage as accepting as many sevices as possible on the vitualized envionment. Secondly, we pesent VM ation based ejuvenation policy to oe the high availability by peventing ailues due to sotwae aging. We use a stochastic eti nets based appoach to builds models and evaluate though both analytical and SHAE (Symbolic Hieachical Automated eliability and eomance Evaluato) tool simulation [6]. The oganization o this pape is as ollows. In Section we discuss the elated wok. The poposed esouce management algoithm o enhancing system availability and stochastic eti nets model o system availability ae ollow in Section and 4. Finally we conclude ou pape in Section 5.. ELATED WOK In this section we eview selected publications elated to ou wok. Thee ae many beneits to vitualization technologies. The authos [] pesented a mixed sotwae ejuvenation policy o an opeational sotwae system with multiple degadation states, which consideed both the histoy inomation and the cuent unning state. y this policy, the system is ejuvenated when it achieves to a degadation theshold o it comes to pe-detemined ejuvenation inteval. Continuous-time Makov chains wee used to descibe the multiple degadation states model. Machida et. al [7] pesented compehensive availability models o thee ejuvenation techniques o a seve vitualized system with time-based ejuvenations o VM and VMM. The esult o sensitivity analysis showed that Migate-VM ejuvenation achieves the best steady-state availability as long as VM live ation is ast enough and othe seves have capacity to eceive the ated VM. Howeve they do not conside the VMs placement on othe physical seves when VM ation is needed. ezaei et al. [9] poposed a new ejuvenation technique o high available vitualized systems that is applied at both VM and VMM levels and yet it does not equie any modiications to applications. They have consideed a typical vitualized system that acts as a hot-standby machine containing two VMs that un on a single VMM. 86
Intenational Jounal o Compute Science & Inomation Technology (IJCSIT) Vol 4, No, June 0 High available systems using vitualization technology and sotwae ejuvenation methodology ae poposed by Thein et al. [, ]. In [], they povide stochastic pocess based models to evaluate availability o the system in case o without vitualization technology and in case when vitualization and sotwae ejuvenation ae used. Vitual machine based sotwae ejuvenation amewok to oe high availability o application seves is pesented in []. Kim et. al [4] have pesented availability models o a nonvitualized and vitualized system using two level hieachical stochastic models with ault tee and CTMC. In vitualized system model, HA sevice and VM live ation ae incopoated. In ode to minimize the downtime o a seve vitualized system, a ast sotwae ejuvenation technique called Wam-VM eboot o VMM [5] was pesented by intoducing the on-memoy suspend technique and quick eload mechanism. In ou wok, to analyze the availability o vitualized seve system we pesent stochastic eti nets model o time-based ejuvenation policy o VMM. To spot live VM ation, esouce usage management algoithm is poposed.. OOSED ESOUCE MANAGEMENT ALGOITHM FO ENHANCING SYSTEM AVAILAILITY To counteact the sotwae aging phenomenon o VMM in vitualized cluste system, we combine sotwae ejuvenation and live VM ation. Duing the live VM ation, we need to conside the pocess o dynamically allocating VMs to othe Ms in esouce pool and to meet thei esouce equiements. Theeoe, esouce management algoithm is poposed o guaanteeing the availability o potential sevices and is accepted as many sevices as possible accoding to the available esouce capacity. This section pesents system achitectue and esouce management in vitualized envionment... System Achitectue The system consists o vitualized cluste seves and management seve. Example system achitectue is shown in Figue. Clusteing spots two o moe seves unning multiple VMs. To enable live VM ation, all Ms ae connected on the same netwok and the shaed disk image. A heatbeat keep-alive system is used to monito the health o the M between them. The management seve contains thee sevice povides: esouce manage, aging detecto and ejuvenation manage. The aging detecto o management seve is esponsible o the detection o sotwae aging. When sotwae aging happens in active VMM, the ejuvenation manage will tigge the ejuvenation action. eoe ejuvenation, the esouce manage is esponsible o getting esouce inomation, and how to allocate the VMs to accept the maximum numbe o sevices in vitualized envionment accoding to the esouces available. 87
Intenational Jounal o Compute Science & Inomation Technology (IJCSIT) Vol 4, No, June 0 M anagem ent Seve V itualized Seve C luste e s o u c e M a n a g e M a n a g e - V M S e v ic e - V M M a n a g e - V M S e v ic e - V M S e v ic e - V M S tan d b y - V M A g in g D e te c to ejuvenation M a n a g e V M M V M M V M M S h a e d S to a g e Figue. System Achitectue The sequence diagam o poposed system is shown in Figue. When some potential anomaly happens, in ode not to lose any in-light equest o session data at the time o ejuvenation o VMM, all the VMs on that aging aected M ae ated to one o the selected M om vitualized cluste system using esouce management algoithm. Ate the VM ation, a ejuvenation opeation will be tiggeed o VMM. : Monito : Aging Detecto : esouce Manage : ejuvenation Manage Monitoing ( ) Detection ( ) Detection pocess ( ) etun no M Selection ( ) esouce management ( ) Migation ( ) ejuvenation ( ) Figue. Sequence Diagam o oposed System 88
Intenational Jounal o Compute Science & Inomation Technology (IJCSIT) Vol 4, No, June 0.. esouce Management in Vitualized Envionment This sub-section descibes how to solve the sotwae aging ailue in VMM using sotwae ejuvenation and live VM ation. Live VM ation technology allows us to ate a unning VM without any o minimal outage, om one physical machine to anothe. As time degades, the VMM unning on the M becomes sotwae aging. When the sotwae aging o some potential anomaly happens in one o the M in the esouce pool, the ejuvenation manage o management seve will tigge a ejuvenation opeation. I the aging aected M is about to be ejuvenated, all the new equests and sessions ae ated om the VMs om the aging aected M to VMs om othe Ms in the esouce pool. When the ongoing equests ae inished in aging inected M, these VMM will be ejuvenated. With the advent o the vitualization, live ation can be peomed without sevice intetion. In ode to cay out the appopiate VM live ations, the poposed esouce Management Algoithm caies out how to allocate the VMs to accept the maximum numbe o sevices on the vitualized inastuctue accoding to the esouces available. The esouce management algoithm is pesented in Figue. The assumptions used in the esouce management algoithm ae as ollow. It does not allowed same sevice in one M because it can be decided to which M o each VM must be allocated such that numbe o VMs is maximized the sevices and VMs equiements ae satisied without exceeding M s capacity and whole inastuctue capacity limits. One o the Ms in the esouce pool can sue sotwae aging at a time. egin. Calculate the maximum numbe o VMs accoding to esouce capacity o Ms in the esouce pool. Detect the VMM aging that needed to ejuvenate. Fo othe Ms in the esouce pool do 4. Select M which has maximum size o available esouce capacity 5. Fo all VMs on aging aected M do 6. I ((size o ceated VMs size o ate VM) < size o available M esouce capacity) AND (Selected M s VM s sevice! Migate VM s sevice ) then 7. VM Migation to selected M 8. End i 9. End o 0. End o End Figue. esouce Management Algoithm The example scenaio is descibed o calculating the maximum numbe o sevices as the input o the esouce management algoithm taking into account the VM and the M eatues as shown in Table. 89
Intenational Jounal o Compute Science & Inomation Technology (IJCSIT) Vol 4, No, June 0 Table. Example esouce Capacity Desciption o VM and M ole M, M, M Vitualization Laye (VMWae ESX) Sevice VM Manage-VM Standby VM Memoy Size 684 M 4096 M 048 M 5 M 048 M In this basic scenaio, the VM size is only computed with the memoy equied and based on vitualized clusteing achitectue. The VM s size is pedeined and ixed on the available physical esouces. The example inastuctue is composed o thee physical machines, and Manage-VM is deployed sepaately with sevice VMs: then the maximum numbes o VMs is computed as ollowings: Available Memoy Memoy size Memoy size usage o o each M o each M - VMM () Total Available Memoy memoysize o M m m i TotalAvailableMemoysize Maximum numbe o VMs Memoysizeo each VM (m numbe o Ms) () () Accoding to the Table I, we can calculate available memoy o each M: 88 M and the example inastuctue is composed o Ms then we get 6864 M total memoy available. I we only deploy sevice VM (basic scenaio): 6864/048 8 VMs ae available to be un simultaneously as shown in Table. Fo the scenaio o Manage VM associated with evey Sevice VM, and a load balanced scenaio o Manage-VM and two sevice VMs, the maximum numbe o sevices can be deployed as shown in Table. Table. Maximum Numbe o VMs accepted by dieent scenaio Maximum Allowed VMs asic Scenaio 8 Scenaio with Manage -VM 4 Load alanced Scenaio with L(with Manage- VM and Standby VM) 8 90
Intenational Jounal o Compute Science & Inomation Technology (IJCSIT) Vol 4, No, June 0 4. STOCHASTIC ETI NETS MODEL FO SYSTEM AVAILAILITY In this section, we pesent stochastic eti nets (SN) model with time-based ejuvenation policy o VMM in vitualized cluste system as shown in Figue 4. It has n tokens which epesented n VMs and m tokens which epesented m Ms. In this policy, the ejuvenation inteval is detemined by using clock. I the active VMM is about to be ejuvenated cause o sotwae aging, all the VMs on aging aected M ae ated to othe Ms accoding to VMs placement though esouce management algoithm and then will be stated o the new equests and sessions. It can etun back to the oiginal M ate the completion o the VMM ejuvenation though live VM ation. In initial condition all Ms ae healthy woking states, indicated by a token in place. As time pogesses, each VMM eventually tansits to ailue pobably states in place F though the tansition T p but the ailue pobably ates ae dieent o each VMM. The VMM ae still opeation in this state but VMs need to ate when token is placed in F and beoe VMM ejuvenation. When the VMs ation tansition T is enabled, the VMs om VM ae ated to anothe M and a token is moved to a place Mig. Ate the ation, VMs will be estated on anothe M. u p m c lo c k T m ig - b k n n T p T c lo c k V M T e p M ig n F T e j T e j T m ig F T Figue. Stochastic eti Nets Model o VMM ejuvenation laces Desciption : Healthy state o i th M F : Failue obably state o i th M i,, (numbe o M) F : Failue state : ejuvenation state o i th M Mig : VMs Migation state om aging M to selected M o ation clock : ejuvenation Inteval state o i th M VM : Available VMs state on i th M In this model, ejuvenation inteval is detemined by using a clock and thee ae tokens in the place clock. Thee ae VMMs to be ejuvenated (a token is placed in F and ), the tansitions T clock is enabled. This tansition is competitively enabled with T p and ies when the clock expies 9
Intenational Jounal o Compute Science & Inomation Technology (IJCSIT) Vol 4, No, June 0 i T p has not ied by that time. Once its ies, token moves in the place and the activity elated with sotwae ejuvenation. Duing the ejuvenation, evey othe activity in the VMM is suspended. This is modeled by inhibito acs om to tansitions T p and T. Ate VMMs have been ejuvenated, it goes back to healthy state with tansition T ej and a token is also eset to clock. When thee is a system is cash (i.e., thee is a token is placed in place F by using tansition T. Fom a ull system outage, the system can be epaied though the tansition T ep and all VMMs ae in healthy state in place. 4.. eachability Gaph We constuct the eachability gaph o SN model with VMs and Ms (n, m). Let 7 tle (, F, clock,, Mig, F and VM ) denote the making with x, i a token is pesented in place x, and zeo othewise. A making is eachable om anothe making i thee exists a sequence o tansition iings stating om the oiginal making that esult in the new making. The making pocess is mapped into a continuous time Makov chain (CTMC) with state space isomophic to the eachability gaph o the SN as shown in Figue 5. μ 0000 λ λ μ 0000 μ F μ μ 000 λ F 000 λ λ 000 000 μ F λ λ 0000 λ 0000 μ Mig μ VM 0000 0000 λ Mig μ VM 00400 0000 λ Mig 000500 λ λ μ λ VM 0000 λ F 000 λ F 000 λ F 0000 μ μ μ Figue. eachability Gaph o SN Model Figue 5 illustates the eachability gaph with squaes epesenting the makings and acs epesenting possible tansition between the makings. Let λ, λ, λ, λ, µ, μ, and μ be the tansition ates associated with T p, T clock,t, T, T, T ej and T ep espectively. y mapping though actions to this eachability gaph with stochastic pocess, we get mathematical steady-state solution o the chain. We may compute the steady-state pobability by ist witing down the steady-state balance equations o igue as ollows. 9
Intenational Jounal o Compute Science & Inomation Technology (IJCSIT) Vol 4, No, June 0 F Mig (4) (5) Mig F F F F F Mig F F F F (6) (7) (8) (9) F F F F F F VM Mig Mig Mig Mig VM Mig VM Mig Mig Mig Mig VM Mig VM Mig Mig Mig VM F F F F F F The consevation equation o Figue 5 is obtained by summing the pobabilities o all states in the system and the sum o equation is. (0) () () () (4) (5) (6) (7) (8) (9) (0) Fi Migi i Fi i i i i i VM i () Combining the above-mentioned balance equations with the consevation equation, and solving these simultaneous equations, we acquie the closed-om solution o the system. 9
Intenational Jounal o Compute Science & Inomation Technology (IJCSIT) Vol 4, No, June 0 F F A A () () (4) (5) (6) F Mig VM Mig VM (7) (8) (9) (0) () () Mig VM F F F () (4) (5) (6) (7) 94
Intenational Jounal o Compute Science & Inomation Technology (IJCSIT) Vol 4, No, June 0 95 A A (8) Whee A ) ( A A The meaning o the pobabilities as ollows: : The pobability o the VMM is in healthy state : The pobability o the VMM is in ejuvenation state om VMM healthy state i : The pobability o the VMM is in ejuvenation state om VMM ailue pobably state (i,, whee i numbe o M) F i : The pobability o the VMM is in ailue pobably state Mig i : The pobability o the VMs ation state F i : The pobability o VMM is in ailue state i VM : The pobability available VMs on M state 4.. Availability and Downtime Analysis Availability is a pobability o a system which povides the sevices in a given instant time. ased on this, system availability and unavailability can be computed. In the poposed model, sevices ae not available when both VMMs ae in ailue state F i. We also deine the availability o the SN model as: Availability -Unavailability Availability i Fi (9)
Intenational Jounal o Compute Science & Inomation Technology (IJCSIT) Vol 4, No, June 0 Downtime is the expected total downtime o the application with ejuvenation in an inteval o T time units is: Downtime T F i i (40) 4.. Numeical Analysis In this subsection, we illustate the applicability o the SN model and solution methodology though numeical analysis. The exact model tansition iing ates o the model ae not known, a good estimate value o a ange o model tansition iing ates is assumed. Fo this pupose, we peom numeical analysis using the ollowing ailue poile mentioned in Table [,0]. Table. Tansition Fiing ates Tansitions Fiing ates (h - ) T p T T clock time/a day times/ months time/ month /T mins /T ej 0 mins T epai time/ hous The inluence o VM ation tansition iing ates along with dieent ejuvenation ates on availability is shown in Figue 6. The mean time to VM ation tansition is assumed mins and mins. The highe VM ation ates, the highe availability o ou system model can be achieved. Theeoe, the availability is dependent on the mean time to VM ation tansition. Availability 0.9999990 0.9999988 0.9999986 0.9999984 0.999998 0.9999980 0.9999978 0.9999976 0.9999974 0.0008 0.0078 0.0047 0.008 ejuvenation ates (pe hou) λmins (Deivation) λmins (Deivation) λmins (SHAE) λmins (SHAE) Figue. Availability vs. dieent ejuvenation ates and dieent mean time to VM ation The Figue 7 shows the dieences in downtime with dieent mean time to VM ation tansition and dieent ejuvenation tansition iing ates. Fom the esult, it is appaent that by combining vitualization technology and sotwae ejuvenation mechanism can enhance the availability o vitualized IT inastuctue. 96
Intenational Jounal o Compute Science & Inomation Technology (IJCSIT) Vol 4, No, June 0.000 Downtime (min pe yea).0000 0.8000 0.6000 0.4000 0.000 0.0000 0.0008 0.0078 0.0047 0.008 ejuvenation ates (pe hou) λmins (Deivation) λmins (Deivation) λmins (SHAE) λmins (SHAE) Figue 4. Downtime vs. dieent ejuvenation ates and dieent mean time to VM ation The change in availability o a system with dieent ejuvenation ates and dieent VMM ailue ates is plotted in Figue 8. The inluence o ailue ates along with dieent ejuvenation ates on availability can be seen om this igue. We assume the ailue ates ae one time pe month and two times pe thee months. High availability systems equie ewe ailues and aste epai. We obseve om Figue 8 that the ailue ate o the system acts as an impotant acto in the availability o the vitualized system. Availability 0.9999985 0.9999984 0.999998 0.999998 0.999998 0.9999980 0.9999979 0.9999978 0.9999977 0.9999976 0.9999975 0.0008 0.0078 0.0047 0.008 ejuvenation ates (pe hou) λtimes/months (Deivation) λtime/a month (Deivation) λtimes/months (SHAE) λtime/a month (SHAE) Figue 5. Availability vs. dieent ejuvenation ates and dieent VMM ailue ates 97
Intenational Jounal o Compute Science & Inomation Technology (IJCSIT) Vol 4, No, June 0 Dow ntim e (m in pe yea).000.0000 0.8000 0.6000 0.4000 0.000 0.0000 0.0008 0.0078 0.0047 0.008 e juve nation ate s (pe hou) λtimes/ months (Deivation) λtime/ a month (Deivation) λtimes/ months (SHA E) λtime/ a month (SHA E) Figue 6. Downtime vs. dieent ejuvenation ates and dieent VMM ailue ates Figue 9 plotted the downtime as a unction o the dieent ejuvenation ates. Accoding to the Figue 9, we can see that the ewe ailue, the lowe downtime can be achieved. SHAE [6] is well known package in the ield o eliability and peomability. It is possible to use dieent kinds o models hieachically o dieent physical o abstact levels o the system and to use dieent kinds o models to validate each othe s esults. So o ou models validity we use this tool. Accoding to Figue 6, 7, 8 and 9, it is ound that the deivation esults and SHAE tool simulation esults ae the same. 5. CONCLUSION Sotwae can cause sotwae aging as time degades. Sotwae aging deceases the availability o the system. It is widely undestood that the technique o ejuvenation povides bette esults, esulting in highe availability and lowe cost. In this pape we pesented possible combinations o vitualized clusteing and sotwae ejuvenation in ode to counteact the sotwae aging and also pesented a esouce management algoithm to guaantee the availability o the sevices deployed and optimize the esouces available in the inastuctue. Stochastic eti nets model o analyzing VMM time based sotwae ejuvenation in continuously unning applications ae pesented and expess availability and downtime and in tems o the tansitions iing ates in the model. The numeical esults ae validated with the evaluation esults though SHE tool. It is ound that the deivation esults and the SHAE esult ae the same. The poposed combine method can be applied to any cluste seve coniguations without any additional cost. The obtained esults show that the use o vitualization, clusteing technology and sotwae ejuvenation mechanism can povide a vey ast ecovey to cut down the mean time to ecovey to the minimum. It can achieve minimize downtime even in case o sevice estat. EFEENCES [] C. Clak, K. Fase, S. Hand, J. G. Hansen, E. Jul, C. Limpach, I. att, anda. Waield. Live ation o vitual machines. In oc. NSDI '05, May 005. [] X. Du, Y. Qi, D. Hou, Y. Chen, X.Zhong, A Mixed Sotwae ejuvenation olicy o Multiple Degadations Sotwae System, oceedings o th IEEE Intenational Coneence on High eomance Computing and Communications 009, pp.76-8,009 98
Intenational Jounal o Compute Science & Inomation Technology (IJCSIT) Vol 4, No, June 0 [] Y. Huang, C. Kintala, N. Kolettis, N. D. Fulton, Sotwae ejuvenation: Analysis, module and application, oc. the 5th Int. Symp., Fault-Toleant Computing, asadena, CA, June. 995, pp.8-90. doi: 0.09/FTCS.995.46696 [4] D. Kim, F. Machida, and K. S. Tivedi, Availability modeling and analysis o a vitualized system, In oc. o IEEE Int l Symp. aciic im Dependable Computing (DC 009), 009. [5] K. Kouai and S. Chiba, Fast sotwae ejuvenation o vitual machine monito, In IEEE Tans. on depnedable and secue computing, 00. [6] K. S. Tivedi, SHAE 00: Symbolic Hieachical Automated eliability and eomance Evaluato. In oc. Int. Coneence on Dependable Systems and Netwoks, 00, pp. 544. [7] F. Machida, D. Kim, and K. S. Tivedi, Modeling and Analysis o Sotwae ejuvenation in a Seve Vitualized System.In oc. o the d Intenational Wokshop on Sotwae Aging and ejuvenation,(wosa00),00. [8] F. Machida, D. Kim, J. ak and K. S. Tivedi, Towad Optimal Vitual Machine lacement and ejuvenation Scheduling in a Vitualized Data Cente, In oc. o the st Int l Wokshop on Sotwae Aging and ejuvenation (WoSA008), 008. [9] A. ezaei and M. Shaii, ejuvenating High Available Vitualized Systems, The 00 Intenational Coneence on Availability, eliability and Secuity, IEEE, 00 [0] Sotwae ejuvenation. Depatment o Electical and Compute Engineeing, Duke Univesity, [Online] Available om: http://www.sotwae-ejuvenation.com [] T. Thein and J. S. ak, Availability Analysis o Application Seves Using Sotwae ejuvenation and Vitualization, J. Compute Science and Technology, 4(), 009,pp. 9-46. [] T. Thein, S. Chi, J. ak, Impoving Fault Toleance by Vitualization and Sotwae ejuvenation, In oceedings o the Second Asia Intenational Coneence on Modeling & Simulation, Kuala Lumpu, Malaysia, 008, pp. 855 860. 99