HAL Id: inria-00441366
https://hal.inria.fr/inria-00441366
Submitted on 15 Dec 2009
HAL is a multi-disciplinary open access
archive for the deposit and dissemination of
sci-entific research documents, whether they are
pub-lished or not. The documents may come from
teaching and research institutions in France or
abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est
destinée au dépôt et à la diffusion de documents
scientifiques de niveau recherche, publiés ou non,
émanant des établissements d’enseignement et de
recherche français ou étrangers, des laboratoires
publics ou privés.
Power Management in Grid Computing with Xen
Fabien Hermenier, Nicolas Loriant, Jean-Marc Menaud
To cite this version:
Fabien Hermenier, Nicolas Loriant, Jean-Marc Menaud. Power Management in Grid Computing with
Xen. 2006 ISPA Workshop on Xen in High-Performance Clusters and Grid Computing Environments
(XHPC’06), Dec 2006, Sorrento, Italy. pp.407-416, �10.1007/11942634_43�. �inria-00441366�
FabienHermenier,Ni olasLoriant,Jean-Mar Menaud
OBASCOproje t
É oledesMinesdeNantes-INRIA,LINA
4,rueAlfredKastler
44307NantesCedex3,Fran e
{fabien.hermenier,ni olas.lorian t,je an-ma r .me naud }emn .fr
Abstra t. While hip vendorsstill sti kto Moore'slaw, andthe
per-forman eperdollarkeepsgoingup,theperforman eperwatthasbeen
stagnantforlastfewyears.Moreoverenergypri es ontinuetorise
world-wide.Thisposesamajor hallengetoorganisationsrunninggrids,indeed
su har hite turesrequirelarge oolingsystems.Indeedtheone-year ost
ofa ooling systemand of thepower onsumption mayoutt thegrid
initialinvestment.
Weobserve,however,thatagriddoesnot onstantlyrunatpeak
perfor-man e.Inthispaper, we proposeaworkload on entration strategyto
redu egridpower onsumption.UsingtheXenvirtualma hinemigration
te hnology, our power management poli y an dispat h transparently
and dynami allyany appli ations of the grid.Our poli y on entrates
theworkloadtoshutdownnodesthatareunusedwithanegligibleimpa t
onperforman e.Weshowthroughevaluationsthatthispoli yde reases
theoverallpower onsumptionofthegridsigni antly.
1 Introdu tion
Thenumberofappli ationsrequiringgiganti storageand omputational
apa-bilitieshasin reasedtheinterestingrid omputinginthelastfewyears.Atthe
sametime, hipvendorshavebeenprodu ingmoreandmorepowerfulCPUsat
lowerpri es. Past work has mostlyfo used on designing e ient and s alable
grids. Nevertheless, energy onsumption has re entlybe ome amajor on ern
fororganisationsrunninggrids.Indeed,power onsumptionisanimportant
por-tionofthetotalownership ost,itdeterminesthesizeofthe oolingsystemand
oftheele tri alba kuppowergenerators.The ostofa oolingsystemwiththe
power onsumptionmaysometimesex eedthegridinitialinvestment.Moreover,
intensivepower onsumption in reases the han e of omponent failure. Thus,
power onsumptionmustbeakeyfeatureingriddesign.
Inthe meantime and despite numerous userssubmitting jobs, a grid must
rarelymaintainpeakperforman es onstantly.Forexample,theaverageloadof
thegridoftheÉ oledesMinesdeNantessubatomi resear hlabisabout70%.
Moreover,many appli ationshavespe i needse.g. middlewareoroperating
forageneri powermanagementsystemadaptedtogrid.
Inthis paper, we presentour prototypeon powermanagementfor grid
ar- hite tures. Ourprototypeis basedonXen virtualma hine hypervisor sothat
oursolutionistransparenttouser.Basedonworkloadprobes(CPUload,
mem-ory,NICsthroughputet .),thepoli ymigratesvirtualma hinesto on entrate
theworkloadonfewernodesofthegrid,thusallowingunusednodesto beshut
down.Wepresentperforman eresultsoverdierentappli ation workloadsand
showthee tivenessofouralgorithm.
The rest of the paper is organised as follows. Se tion 2 presents existing
powermanagementpoli iesfor luster omputing.Se tion 3reviewsthevirtual
ma hinete hnologyandthenSe tion4des ribesourimplementation.Se tion5
showsperforman eresultsandSe tion 6 on ludes.
2 Energy management
Mostexistingworkonpowermanagementhasfo usedonCPUsandlaptops.
In-deed,batterydurationisa riti alproblemforusersandCPUisbyfarthemost
power onsuming omponent of a omputer
1
. Today, most pro essors feature
DVFS (Dynami Voltage and Frequen y S aling) even serverdesigned
pro es-sor.ADVFSpro essorhasasetoflevels(voltageandfrequen ypairs).Adefault
poli yintegratedintheBIOS anbeoverloadedbytheoperatingsystem.OSand
systemsele tthelevela ordingto performan eandtemperature onstraints.
Therest of this se tiondes ribesIVS, CVS and VOVO,three power
man-agementpoli ies forsingle systemor lusters and theproblem of omputation
migrationin grid omputing.
2.1 IVS IndependentVoltageS aling
Independent Voltage S alingis anode-lo al poli y for powermanagement. In
this approa h, every node must be a DVFS pro essor. Ea h node adjuts its
pro essorvoltageand frequen yto thelowest possiblevaluesu h that there is
not mu h impa t on performan e. IVS is very simple and it does not require
anyspe i environmentor information aboutrunning appli ations. Elnozahy
etal.[1℄reportpowersavingsofupto30%dependingonappli ations.Inorder
to minimize the impa t on performan e, Chen et al. [2℄ applied IVS only to
nodesnotinthe riti alpathofrunningappli ations.
2.2 CVS CoordinatedVoltage S aling
In ontrast to IVS, Coordinated Voltage S alingis a global poli y. Ea h node
periodi allyinformsamonitorofitsworkload.Themonitorevaluatestheoverall
1
low-nodesadapt theirvoltageandfrequen ytothisaverage.Thisstrategytendsto
de reasethevarian eof thefrequen ydistributiona rossthegrid.
CVS is ee tive when the workload is uniformly distributed among nodes.
Elnozahy et al. [1℄ evaluate the powersavings of CVS to be about 25%.
Nev-ertheless, the ee tiveness of CVS is questionable when workloaddistribution
is notuniform, indeed, nodes with an bigger workload are slowed down, thus
de reasingtheoverallperforman e.
2.3 VOVO Vary On,Vary o
Pinheiroet al.[3℄ andChaseet al.[4℄haveproposed VaryOn VaryO, whi h
dynami ally adapts thenumberof nodes available a ordingto the grid
work-load.
VOVOappliesaworkload on entrationstrategy.SimilarlytoCVS,a
mon-itor estimates theoverallworkload,from whi h VOVOdedu esthe numberof
nodesrequired.Whentoomanynodesareinuse,jobsaremigratedorbalan ed
to on entrate the workload. Then unused nodes are simply shut down. The
monitormaydeneaperforman etoleran etopossiblyoverloadsomenodesto
furtherin reasepowersavings.
VOVOhas been found to give a power savings of 31% to 86% [1℄, indeed
VOVOisfully ee tivewhenalarge lusteris almostunused. HoweverVOVO
requiresjob migrationte hnology,that anbeimplementedatseverallevel.At
the appli ation level, usershave to developtheir own load balan er. While at
theoperatingsystemlevel,aspe i OShasto beinstalledon everyma hines
and these must be identi al. Whilethis is rarelyan issuein lusters, grids are
moreheterogeneous.
Designing the migration me hanism at the appli ation[5℄, middleware, or
operating system[6℄ levels is hardly feasible, as energy saving is an issue for
organizationsrunninggrids,notforusersdevelopingappli ations.Inour ontext,
wesupposethatgridarebuiltbyaggregationsofprimary luster.Ea h lusteris
managedbydierentorganizationswiththeirspe i appli ations.Thus,various
frameworksandmiddlewareareusedtodevelopgridappli ations.Itisunfeasible
to impose one of them on the entire grid as these are spe i to the kind of
appli ationsdeveloped( omputationnal,data,servi es)[7℄.
3 Virtual Ma hines
ThegoalofavirtualVirtualma hinesistoprovideea hofmultipleuserswith
the illusion of having an entire omputer (a virtual domain), all on a single
physi al omputer(thehost).Thiste hnology,alsoknownasvirtualization,
dif-ferentiates from appli ation virtual ma hines su h as Sun Mi rosystem's Java
op-pervisororthevirtualma hinemonitormustenfor eisolationbetweenmultiple
virtualdomains andmust partitionresour es(CPUs, RAM,HDDs, NICs,et )
among them. The rest of this se tion des ribesvirtualization and
paravirtual-ization,thetwomajorwaystoimplementanhypervisor.
Fig.1.Ar hite tureofaXenvirtualisedSystem
In ontrasttoemulation,virtualizationreprodu esthehostar hite ture
iden-ti ally forthe virtualma hines. Hypervisorssu hasVMWare [8℄run most
in-stru tionsofthevirtualdomainsdire tly onthehostpro essor.Thehypervisor
musttrapsensitiveinstru tionsthat annotbeexe utedsafelysu hasTLB
op-erations.Whilevirtualizationrestri tsthevirtualar hite turetothehostone,it
stillsallowsrunningunmodiedOSimageswithmu hbetterperforman ethan
emulation.
Paravirtualization[9,10℄hasbeenintrodu edtoover omestheperforman e
loses ofvirtualizationdue to sensitiveoperations.It expli itlyrequires porting
the OS to a parti ular hypervisor. The ar hite ture presented to the virtual
ma hines slightlydiersfromthehostar hite tureinorderto eliminate
oop-erativelysensitiveoperations.
Virtualizationthusmakesitpossiblehostingalotofvirtualdomainsonthe
same ma hine without onstraints. Ea h domain ould use its own operating
system,middlewareorappli ationswithoutriskof oni tswithotherssystem.
Thisadvantagesmakethatvirtualizationappre iableingrid omputing:
Oursolutionis adynami pla ementsystemforvirtualdomains onagrid
vir-tualizedwithXen(anhypervisorforea hnode).Thepla ementalgorithmaims
tosaveenergyby on entratingvirtualdomainsonthesmallestsubsetofnodes
possible.Nodesnothostingvirtualdomains anbestopped,in orderto redu e
theoverallpower onsumption.Atpresent,thedistribution riterionforvirtual
domainspla ementisbasedonCPUusagebut ouldbeextendedto takeother
riteria into a ount, e.g. network tra or memory usage. Distribution of
virtualdomainismadewithaminimalservi einterruptionwithxenmigration
me hanism[12℄.Theproblem ofdata migrationis solvedby usingales server
that exportsoperatingsystemsanddatasforea hvirtualdomain.
Oursolutionistransparentforusersofvirtualdomainsanddoesnotrequire
any adaptationsor spe i operating system
2
. Thear hite ture is basedon a
lient/servermodel.Astandardpre-madedomainisrunningonea hnode.That
domain, alledDomain0,runsatool,theharvestingagent,thatmonitorsCPU,
memoryandnetwork.Thissupervisormonitorstheresour eusageofthenode,
andhowthisresour esaredividedamongthenode'svirtualdomains.
Ea hharvestingagentperiodi allysendsareporttothede isionagentwhi h
analysesthemto getanoverviewofthegridandofitsvirtualdomain resour e
usage. Thenanalgorithm determinesa tionsto betaken bythevarious nodes
(virtualdomainmigration,boot,shutdown).
4.1 Resour e olle tion
Theharvestingagentusesfun tionalitiesoeredbythexenstatlibraryprovides
withXen.Thislibraryallowsthedomain-0toobtainstatisti sgeneratedbythe
hypervisor,forea hvirtualdomain,in ludinginformation on erningallo ated
memory,bytessentand/orre eivedbythevirtualnetworkinterfa eorthe
quan-tityoftimeusedbythevirtualCPU.Byretrievingthisinformationperiodi ally,
itispossibletoknowtheper entageofCPU onsumedbyea hvirtualdomains
and, by summing them, the global CPU usage of the node (in this ase, the
resultobtaineddoesnot onsidertheloadofthehypervisor.)Areportismade
andsenttothede isionagent.Theharvestingagentisresponsiblefornotifying
thede isionagentof hangesinanode'sstate:Whenanodeisstoppedorready
to a eptvirtual domains,its harvesting agentsend adeparture orarrival
no-ti ation,respe tively.Thus,thede isionagentknowsthepla ementofvirtual
domainsamongthegrid,andwhatnodesareavailableoroine.
4.2 The de ision agent
This agent runs on its own ma hine. A ording to reports and noti ations
pa kets sent bynodes, the agent onstru tstworepresentations ofthe grid. A
2
address,andstate(onlineoroine)ofea hnodes.Avirtualdomainbasedview
givesinformation abouttheirlo ation,their urrentresour e onsumption and
thelevelofsaturationoftheirhost.Thede isionagentthenusesthisinformation
to on entrate virtual domains in the smallest possible subset of nodes while
avoidingsaturation.Thealgorithm isbasedon twovariables: asaturationand
an underload threshold. De isionsare taken by omparing the urrent loadof
ea hnodewithitsthresholds.
Inorder to on entratevirtual domains withoutoverloading nodes, the
al-gorithmsele ts virtualdomains that onsumesthehighest amount ofCPU on
the lowest loaded node then move it on a more loaded node. As we suppose
anhomogeneousenvironment,CPU onsumptionofthatdomainshouldremain
almost identi alonthe newnode.So, thedestinationnode is hosensu h that
thesumofthe urrentdestinationnodeloadandthe urrentdomainloadis
un-derthe overload threshold.Therst a eptablenode is pi ked.This operation
is repeated in the nextiteration if thenode is still underloaded and ifamore
appropriatepla eisavailableforvirtualdomains.
Virtualdomains on entration ouldbedangerous in a ertainway, as
vir-tualdomainsdoesnot ontinuallydothesamejob andthusitsCPU
onsump-tion may vary. Then, it is possible that if some domains in rease their CPU
onsumption,thenodesthat hostthemwillbeoverloaded.Tolimitthe
perfor-man edegradation,weredu etheloadinthefollowingway:thedomainhaving
thelowestloadwillbemovedtoanodethat anhostit.Thedestination node
shouldhavethemaximumCPUusagewithoutbeingoverloaded(beforeand
af-ter themigration).If nonodeisavailable,anewnodeintheoinenodesset
willbeinitialisedanddomainwillbemigratedon.
Whenade isionismadebytheagent,itwillsendthe orrespondinga tion
to theappropriate node. A tions are shell ommands, that will beexe utedin
Domain 0. As we wantto migrate virtual domains to savepowerby aVOVO
system,threekindsofa tions anbesenttoanode.Themostuseda tionisto
migrateaspe i virtualma hinetoanewnode.Othersa tions,spe i tothe
VOVOaspe tofoursolution auseanodeto behaltedorbooted.
5 Evaluation
Theevaluationofoursolutionaimstoshowthat dynami pla ementofvirtual
domainsallowsagoodenergysavingwithatolerableperforman edegradation.
We ompareenergye onomyandperforman edegradationofoursolutionand
atraditionalIVSsolution.
Inordertoevaluateoursolutioninapredi tableenvironment,wedeveloped
aben hmarkthat sendsloadrequest to lients.This toolwastunedto make
ea h lient onsumesa ertainamountofCPUtime.Forea hexperimentweran
theteststhreetimes.First,inanativeenvironment,ea happli ationrunsona
migrated betweena variable amountof a tives nodes, depending on the CPU
resour eneeds).
Our grid is omposed by four Sun V20Z stations. Ea h ma hine ontains
4GBofRAM,2Opteron250andrun aGNU/Linuxdistribution withXen
3.0-testing(basedona2.6.16kernel).Allappli ationsare ompiledin32bitsmode.
A leserver ontainsall appli ationsand OSsused bynodes in orderto make
administration easier. All ma hines (4 nodes and 1 le server) are onne ted
throughtaGigabitnetwork.TheIVSsoftwareusedforthese ondtestisbasedon
powernowdandthenodes' overloadand underloadthreshold hosento migrate
virtualma hines wererespe tively80%and70%.Inallsituations, wemeasure
theee tiveCPU onsumptionofea hnode,distributionofCPUtimebetween
domainsandWatts onsumedbythegrid(not in ludingtheleserver).
Ourben hmarkevaluatesthedynami behaviour(migration, turningnodes
on and o) of our strategy and its benets in omparison to IVS. We adapt
it to reate four dierent s enarios(one per ben hmark lient). Ea h s enario
reateshigh-loadpeaksorlow-usageperiodsat dierenttimes.Thisevaluation
simulatesagridsupportingavarietyofdierentservi es.Figures2(a)and2(b),
and Table 1 show the nodes' CPU utilization and the domain lo ation over
time. First, we observe that the IVS strategy has CPU average load that is
almostequalforallfournodes(approximately60%),whereasoursolutiontends
to on entrateall thework onthree nodes.Se ond, theCPU average loadfor
ea hnodeislessvariablein timeforoursolutionduetothepoli y'sde isions.
Table1.CPUaverageutilizationwithdynami evaluation
IVS
node
1 = 51.16%(δ = 37.73)
Virtualizationnode
1 = 82.69%(δ = 20.35)
node
2 = 65.42%(δ = 36.23)
node
2 = 84.08%(δ = 6.91)
node
3 = 68.11%(δ = 22.13)
node
3 = 09.52%(δ = 9.58)
node
4 = 59.59%(δ = 33.39)
node
4 = 65.68%(δ = 28.3)
Thefourdierents enariosarebuiltto overdierentsituations that
illus-trate advantages and drawba ks of domain distributions. A positive ee t of
on entrationappearsaround1100s(Figure5),wheretwoofthefournodesare
disa tivatedbymigratingvirtualDomains1and4toNode2.Thesemigrations
arepossiblebe ausethetwodomains onsumesalmostnoresour esatthat
mo-ment. An overview of a virtual domain repartition appears around 1600s. On
Node2fromwhi h virtualDomain3wasmigrated toNode1,in ordertoleave
CPUtimeforvirtualDomain2.Thus,domainsthatneedanimportantamount
ofresour eshaveahigherprioritythansmalldomains.Thisex lusionofsmall
domainsavoidperforman eloss.
A side ee ts of migration anbe seen on Node 4, at about 400s. To free
some resour eon Node 1, Node 4 was booted, but during the boot time, the
0
25
50
75
100
0
500
1000
1500
2000
2500
% CPU Node-1
Domain-1
0
25
50
75
100
0
500
1000
1500
2000
2500
% CPU Node-2
Domain-2
0
25
50
75
100
0
500
1000
1500
2000
2500
% CPU Node-3
Domain-3
0
25
50
75
100
0
500
1000
1500
2000
2500
% CPU Node-4
Time (in sec.)
Domain-4
(a)NativeExe utionwithIVS
0
25
50
75
100
0
500
1000
1500
2000
2500
% CPU Node-1
0
25
50
75
100
0
500
1000
1500
2000
2500
% CPU Node-2
0
25
50
75
100
0
500
1000
1500
2000
2500
% CPU Node-3
0
25
50
75
100
0
500
1000
1500
2000
2500
% CPU Node-4
Time (in sec.)
boot time
Domain-4
Domain-3
Domain-2
Domain-1
(b)VirtualEnvironmentwithmigrations
Fig.2.GlobalCPU onsumption
minuteslater.Asitisimpossibletopre iselypredi tthefutureneedsofdomains,
it is hard to avoid su h useless boots. Nevertheless, it would be possible to
deneaminimumuptimeto avoidnoderebooting onstantly.
Table 2.Exe utiontimeandaverageenergy onsumption
Environment
Exe utiontime Energy
Cumulated
∆
withNativeConsumptionGain3
Native 118:39 - 651.12W
-IVS 119:01 +0.3% 636.57W 1.93%
Virtual 139:37 +17.67% 498.13W 9.97%
We anseeinTable2andFigure5,thattheenergysavingofoursolutionis
appre iable but there isasmall performan edegradation(the umulated time
ofexe utionin reasesby20minutes).Thisisduetothetimeneededbyanode
to boot (approximately 3 minutes). To evaluate the time lost due to reboot,
we re-ran the experiment with virtualenvironment without shutting down or
300
400
500
600
700
800
0
5
10
15
20
25
30
35
40
Energy (Watts)
Time (min)
Native execution
IVS execution
prototype
Fig.3.Globalenergy onsumption
bootingnodes.Migrationisstillee tivebutboottimeisredu edtozero.Inthis
situation,theperforman ede reasesonlyby8.53%(witha umulatedexe ution
timein reasedby10minutes).
This evaluation shows that our solution allows important energy savings,
espe ially in astable environment, but in a very dynami situation, the boot
time of nodes slightly redu es the benet of our solution. About 50% of the
overheadofoursolutionisduetoboottime.Softwaresolutionsu hashibernate
or suspend-to-rammaygreatlyminimisethat issue.
6 Con lusion and future work
In this paper we have show that the Xen live migration allows to adapt the
VOVOpowersavingstrategytogrid omputing onsideringits omplex
infras-tru ture.Contraryto traditionnalVOVOsystems,theabstra tionlayeroered
byXen provides ageneri omputationmigration pro ess basedonvirtual
do-mains pla ement. The omputation on entration is done by dynami ally
mi-gratingvirtualdomains, onsideringtheirresour eneeds.This on entrationof
virtual domains on the minimal subset of nodes allows to shut down unused
nodes,thustosaveenergy.Byseparatingthemigrationandthepla ement
on- ernat thehypervisorlevel,oursolutionis ompletlytransparenttodevelopers
andusers.Evaluationsshowourpowermanagmentpoli y basedsolelyonCPU
loadmaysigni alyde rease power onsumption in gridwith variously loaded
ma hine.
Asfuturework,weintendtoextendourpowermanagementpoli ytoa
multi- riterion system, espe ially to in lude network tra information. Indeed, the
networkhardwareofagridalso onsumesanonnegle tibleamountofele tri al
power. Ourpoli y ouldtighten orregroupona singlenode, highly
ommuni- atingvirtualdomainsinordertoredu enetworktra .Utilizationofsoftware
1. Elnozahy,E.N.,Kistler,M.,Rajamony,R.:Energy-e ientserver lusters.In:
Pro- eedingofthese ondWorkshoponPowerAwareComputingSystems,Cambridge,
MA,USA(2002)179196
2. Chen, G., Malkowski, K., Kandemir, M., Raghavan, P.: Redu ing power with
performan e onstraintsforparallelsparseappli ations.In:IPDPS'05:Pro eedings
of the 19thIEEE International Parallel and Distributed Pro essing Symposium
(IPDPS'05)-Workshop11,Washington,DC,USA,IEEEComputerSo iety(2005)
231.1
3. Pinheiro,E.,Bian hini,R.,Carrera,E.,Heath,T.:Dynami lusterre onguration
for powerand performan e. InBenini,L., Kandemir, M., Ramanujam, J.,eds.:
Compilers and OperatingSystems for Low Power. KluwerA ademi Publishers
(2002)
4. Chase,J.S.,Anderson,D.C.,Thakar,P.N.,Vahdat,A.M.,Doyle,R.P.:Managing
energyand serverresour esinhosting enters. In:SOSP'01:Pro eedingsof the
eighteenth ACM Symposium on Operating Systems Prin iples, Ban, Alberta,
Canada,ACMPress(2001)103116
5. le Mouël, F., André, F., Segarra, M.T.: AeDEn : An adaptive framework for
dynami distributionovermobileenvironments.Annalsoftele ommuni ations57
(2002)11241148
6. Gallard, P., Morin, C.: Dynami streams for e ient ommuni ations between
migrating pro essesina luster. In:Euro-Par 2003: ParallelPro essing. Volume
2790ofLe t.NotesinComp.S ien e.,Klagenfurt,Austria,Springer-Verlag(2003)
930937
7. Krauter,K.,Buyya,R.,Maheswaran,M.: Ataxonomyandsurveyofgridresour e
managementsystemsfordistributed omputing.SoftwarePra ti eandExperien e
32(2002)135164
8. Ber ,L.,Wadia,N.,Edelman,S.: VMWareESXserver2:Performan eand
s ala-bilityevaluation. Te hni alreport,IBM(2004)
9. Whitaker, A.,Shaw,M., Gribble,S.D.: S aleandperforman eintheDenali
iso-lationkernel. In:Pro eedingsofthe5thsymposiumonOperatingsystemsdesign
andimplementation(OSDI),Boston,MA,USA(2002)195209
10. Barham, P.,Dragovi ,B., Fraser, K., Hand,S., Harris, T., Ho,A., Neugebauer,
R.,Pratt, I.,Wareld, A.: Xen andthe artofvirtualization. In:Pro eedings of
the19thACMSymposiumonOperatingSystemsPrin iples,BoltonLanding,NY,
USA,ACMPress(2003)164177
11. Foster,I.,Kesselman,C.,Tue ke,S.: Theanatomyofthegrid:Enablings alable
virtualorganizations. Le tureNotesinComputerS ien e2150(2001)1??
12. Clark, C., Fraser, K., Hand, S., Hansen, J.G., Jul, E., Limpa h, C., Pratt,
I., Wareld, A.: Live migration of virtual ma hines. In: Pro eedings of the
2nd USENIX Symposium on Networked Systems Design and Implementation