Sharad Agarwal

14 Pages·2012·0.18 MB·English

The SAHARA Model for Service Composition Across
Multiple Providers
BhaskaranRaman,SharadAgarwal,YanChen,MatthewCaesar,WeidongCui,
PerJohansson?,KevinLai,TalLaviany,SridharMachiraju,Z.MorleyMao,
GeorgePorter,TimothyRoscoez,MukundSeshadri,JimmyShih,KeithSklower,
LakshminarayananSubramanian,TakashiSuzuki,ShelleyZhuang,
AnthonyD.Joseph,RandyH.Katz,IonStoica
UniversityofCaliforniaatBerkeley
Abstract. Services are capabilities that enable applications and are of crucial
importance to pervasive computing in next-generation networks. Service Com-
position is the construction of complex services from primitive ones; thus en-
ablingrapidandflexiblecreationofnewservices.Thepresenceofmultipleinde-
pendentserviceprovidersposesnewandsignificantchallenges.Managingtrust
across providers and verifying the performance of the components in compo-
sition become essential issues. Adapting the composed service to network and
userdynamicsbychoosingserviceprovidersandinstancesisyetanotherchal-
lenge. In SAHARA1, we are developing a comprehensive architecture for the
creation, placement, and management of services for composition across inde-
pendentproviders.Inthispaper,wepresentalayeredreferencemodelforcompo-
sitionbasedonaclassificationofdifferentkindsofcomposition.Wethendiscuss
thedifferentoverarchingmechanismsnecessaryforthesuccessfuldeploymentof
suchanarchitecturethroughavarietyofcase-studiesinvolvingcomposition.
1 Introduction
Pervasivecomputingdemandstheall-encompassingexploitationofservicesinsidethe
network. By services, we mean both the components of distributed applications and
the glue that interconnects them as they function across the network. Services range
fromprovidingbasicnetworkreachabilitytocreatingoverlaynetworkswithenhanced
qualitieslikepredictablelatenciesandsustainedbandwidths.Servicesalsoincludein-
stances of application building blocks, requiring processing and storage, judiciously
placedinthenetworktocontrolconnectionlatenciesandtoachievescalethroughload
sharing. Such services may be simple format translators, interworking functions, or
majorsubsystemsforcontentdistributionorInternetsearch,whichareoftenregarded
as applicationsin their own right. Compositionvia interconnectionof servicesallows
moresophisticatedservicesandapplicationstobeconstructedhierarchicallyfrommore
primitiveones.Sinceeconomicsmakesitunlikelythatanysingleserviceproviderwill
beabletoprovidealloftheconnectivity,applicationsbuildingblocks,processing,and
?EricssonBerkeleyWirelessCenter;yNortelNetworks;zIntelBerkeleyResearchLab
1ProjectsupportedbySprint,Ericsson,NTTDoCoMo,HRL,andCalif.MicroGrant#01-042.
Restaurant
Guide
Service
babelfish.altavista.com
Babelfish
Language JAL Travel User
Translator
Zagat’s
Restaurant Home Domain (Tokyo)
Guide
Foreign Domain (Salt Lake City)
User Utah, U.S.A.
Fig.1.ServiceComposition:AMulti-ProviderScenario
storage resources to effectively deploy a globe-spanning application, the composition
ofservicesacrossindependentprovidersisessential.Thispaperproposesacomprehen-
sivereferencemodelforcomposedservicesinsupportofpervasivecomputing.
Toillustrateourconceptofservicecompositionacrossserviceproviders,consider
thefollowingscenario(Figure1).Ms.TanakatravelsfromTokyotoSaltLakeCityto
attend the Winter Olympics. Her cellular provider, NTTDoCoMo, maintains roaming
agreementswithseveralforeignnetworkoperators,suchasSprint,soshecanmakeand
receivecallsintheU.S..Buthernewinformationappliance,builtbyEricsson,ismuch
more capable than just a phone – it is a gateway to extensive information, entertain-
ment, and messaging services. In Japan, she subscribes to NTTDoCoMo’s restaurant
recommendation service, and would like to use the same application in Utah. In the
U.S.,SprinthasanarrangementwithZagat’sGuidetopresentsuchinformationtoits
subscribers,butthetextisinEnglishandformattedforpresentationonadifferentkind
of display than Ms. Tanaka’s. A third party, JAL Travel, assemblesa special new ser-
viceforJapanesetouristsattheOlympicsfromcomponentservices:Zagat’sRestaurant
Guide,JapanesetranslationusingBabelfish,andreformattingforJapanese-styleinfor-
mationappliancedisplays.Ms.Tanakasubscribestothisservicefortwoweeks,andthe
usagechargesappearonherNTTDoCoMobillbackinJapan.
This scenario exemplifies several key points about services composition. In next
generation networks, users will demand enhanced services like restaurant recommen-
dations,buttheyexpectthemtolookandfeelthesamewhethertheyareathomeorina
foreignnetwork.Anewservice,composedfromlocalizedinformationsources,appro-
priatelanguagetranslators,andcontentreformatters,makestheunderlyingdifferences
transparenttotheuser.Suchaservicecanbecreatedquicklybysimplyconnectingthe
componentsacrossthenetwork.Butmanyentitiesparticipateintherealizationofthis
service,andtheymustbemanaged,authenticatedandcompensatedinsomeway.
In the context of object-based systems, programming by composition across the
network is hardly new. Yet there are critical new challenges. The first is composi-
tionsacrossindependentserviceproviders:Zagatownsthedatabaseandthewebsite,
NTTDoCoMomaintainstherelationshipwiththeclientandcollectshermonthlybill,
and Sprint provides the local wireless access as well as the gateway access to local
services. Creating the necessary mechanisms to support cooperative composition of
Virtual Operators
Dynamic Relationship
Static Relationship (exists today)
Mobile Network
Operators
Fig.2.Multi-providerScenarioinWirelessConnectivityService
servicesacrossindependentserviceproviders,eachwithitsownauthentication,autho-
rization,andaccountingmechanisms,isanessentialchallenge.
JAL Travel created a new service from pieces, some of which were provided by
other third parties, such as Babelfish, and all of which need to run on machines in-
terconnectedacrossthenetwork,spanningInternetDataCenterprovidersandInternet
ServiceProviders.Hereinliesthesecondchallenge:theabilityforthirdpartiestodis-
covercomponentsandtobrokernewservicesfromconstituentpieces,someofwhich
maynotevenbeawareofthecompositioninwhichtheyareparticipating.Asthequal-
itiesofacomposedservicearenobetterthanitsweakestcomponent,anessentialneed
isforbrokerstobeabletoverifytheperformanceandbehavioroftheassembledcom-
ponents, whether or not these underlying participants are aware of their role in com-
positions.Ifacomponentdoesnotmeetitsperformanceorbehavioralspecification,it
mustbe“composedout”,andanewinstancefromadifferentprovider“composedin”.
A third challenge for service providers is the need for an extensive set of new
service composition management tools. From a provisioning viewpoint, sufficient in-
stancesofthecomponentsneedtobeplacedatlocationswithinthenetworktoensure
scalableperformanceand high availability even in the face of site failuresor network
outages. Such placement also needs to ensure appropriate network and processing la-
tencies to achieve adequate responsiveness for the supported applications. Such tools
includeapolicymanagementmechanismforserviceproviderstoinformservicecom-
posers about how their instances in their network should be used for providing fault
tolerantandloadbalancedbehavior.Apervasivemonitoringandmeasurementinfras-
tructure is needed to detect changing access patterns and shifting workloads, to drive
redirection to unloaded service instances or to change the number and placement of
deployedinstances.Networktopology-awarenessisimportant,foravailabilityaswell
asperformance.Placementandconnectivityissuesarecomplicatedsincesomeservice
instances are anchored to fixed locations, while lighter weight services can be placed
closetotheusercommunity.
Furtherchallengesarisewhenweconsideruserdynamics.Alargenumberoffor-
eign roamers like Ms. Tanaka converge in Salt Lake City, yielding flash crowds and
over-utilizedspectrum.Weneednewwaysofefficientlyallocatingresourcesinthecon-
text of new serviceprovider businessmodels.The 3rd GenerationPartnershipProject
(3GPP) defines the concept of a Mobile Virtual Network Operator (MVNO) [1], an
entitywithsubscribersbutnonetwork.TheMVNOprovideswirelessconnectivityser-
vice composedfrom the physical network resourcesof underlyingwirelessoperators.
In our example, NTTDoCoMo acts as a virtual operator using spectrum for its sub-
scribers from the already established network operators such as Sprint, Cingular, and
AT&T Wireless. We view such a multi-provider relationship as just another case of
servicecomposition.Today,therelationshipsbetweenthevirtualoperatorandtheMo-
bileNetworkOperator(MNO)arestaticandnegotiatedlonginadvance(e.g.,between
VirginMobileandOne2OneinUK[1],seeFig.2).Butthisisinefficientwhenuserdy-
namicsare considered;and we expect much more dynamicformationand dissolution
ofrelationshipsinthefuture.Thus,asafourthchallenge,weaddressissuesinefficient
resourceallocationacrossproviders,consideringthedynamicsofusercommunities.In
our MVNO scenario, this translates to dynamically selecting (“roaming”) among co-
locatedMNOs.Figure2illustratessuchdynamicallocationsintheMVNOcontext.We
envisionthatdynamicrelationshipswilllastforshorttime-scalesofminutestohours,
thusallowingforloadbalancing,andefficientresourceusage.
Our overall goal is to define a comprehensive reference model that is able to de-
scribetheassemblyfromcomponentsofend-to-endserviceswithdesirable,predictable,
enforceable properties, yet spanning potentially uncooperating service providers. We
are developing these concepts in the context of the SAHARA2 project, which is also
the name of our prototype architecture. The next section summarizes the discussion
above with the technical issues in composition. Sec. 3 presents a classification of the
differentkindsofcompositionandSec.4presentsalayeredreferencemodelforcompo-
sition.Wedescribethedifferentmechanismsweemploytoaddressthetechnicalissues
incompositioninSec.5.WediscussrelatedworkinSec.6andconcludeinSec.7.
2 TechnicalIssues
The scenario in our prior section can be understood along three dimensions in the
choices in service composition: (a) what set of services to use for composition, (b)
which service providers’ resources to use, and (c) which instances of each service to
use for a particular client session. In addition, we also have the issue of who makes
thesedecisions.Consideringthesethreedimensions,thetechnicalissuesthatmustbe
addressedinareferencemodelforservicecompositionare:
– Trustmanagementandbehaviorverification:Whenmultipleprovidersinteract,it
is importantto establishmutualtrust.This is not onlyfor the purposeof userau-
thenticationandbilling,butalsotoverifythebehaviorofthecomponentsincom-
position.Doesacomponentmeetitspromisesintermsoffunctionality,protocols,
performance,availability,orotherproperties?
– Adaptingtonetworkdynamics:Withnetworkdynamics,workloadscanshift,con-
gestion can arise, and reachability to services can be lost. This implies the need
forperformancemonitoring,modelingandprediction,andaperformance-sensitive
choiceofprovidersandinstances.Wetermthistheserviceselectionproblem.
2SAHARA: Service Architecture for Heterogeneous Access, Resources, and Applications;
http://sahara.cs.berkeley.edu
– Adapting to user dynamics: User dynamics can cause different providers to face
varying demand for physical resources from their users. Allocation of resources
(spectrum, network bandwidth, CPU, etc.) across providers, based on current de-
mandisimportanttoachievefair/utility-drivenresourceallocation.
– ResourceProvisioningandManagement:Foragivencommunityofusersandaset
ofperformance,availability,andadministrativeconstraints,howmanyinstancesof
a service are needed? How can this be optimized given a knowledge of the pro-
visioningdonebyotherserviceproviders,andnetworktopologyinformation?We
call this the service placement problem. Also, how can service providers enforce
theirlocalpoliciesoftheuseoftheirserviceinstancesinacompositionperformed
byathirdparty?Wetermthisthepolicymanagementissue.
– Interoperabilityacrossmultipleserviceproviders:Whencomposingservicesacross
providers,wehavetodealwithheterogeneityinprotocolsanddataformats,aswell
asauthenticationandauthorizationmechanisms.
3 ServiceCompositionModels
Wenowclassifyservicecompositionintotwomodelsbasedonthetypeofinteraction
between composed component service providers and analyze their pros and cons: (1)
CooperativeModel:Serviceprovidersinteractinadistributedfashion,withdistributed
responsibility, to provide an end-to-end composed service; (2) Brokered Model: A
single provider, the broker, uses the functionalities provided by underlying service
providersandencapsulatesthesetocomposetheend-to-endservice.
In either case, the end-user subscribes to only one provider. However, the differ-
ence lies in the way the responsibility for the composed service is apportioned. The
two possibilities represent different business models for composition. In the coopera-
tivemodel,thepropertiesofthecomposedservicesuchasfunctionality,performance,
andavailability,areguaranteedbythedesignofthedistributedinteraction,andthrough
service-levelagreementsbetweentheinteractingentities.Eachserviceproviderisonly
responsible for providing guarantees for the portion of the composed service within
itsdomain.Inthebrokeredmodel,thebrokerassumesresponsibilityfortheproperties
of the composed service. We can imagine a broker entering into contracts with ser-
viceproviders,andusingthesetoconstructend-to-endcomposedservices.Thebroker
verifiesthefunctionalityoftheindividualpiecesintheservicepath.Thisisbecausein-
dividualcomponentprovidersmaynottrusteachother–theymaylimittheinformation
aboutthestateoftheirservicetheyexposetotheotherprovidersinthecomposition,or
mayactivelyseektocheatonthequalityofservicetheyprovide.
Anexampleofcooperativecompositioniscellularroamingasaservice,composed
fromtheresourcesofmultiplemobilenetworkoperators,asinSec.1.Thedistributed
interactionbetweenoperators(NTTDoCoMoandSprint)enablesroaming.Anotherex-
ampleisend-to-endconnectivityserviceintheInternet.Aninter-domainroutingpro-
tocolallowscooperationbetweendomainsinadistributedfashion.Intheseexamples,
therearelong-term,static,negotiatedcontractsbetweentheparticipantproviders.
Anexampleofbrokeredcompositionistherestaurantguideserviceassembledby
JAL Travel in our earlier scenario. It assumes responsibility for the functionality and
Inter−provider, or Provider−customer interaction
Possible data flow
Broker
C
E(cid:0)(cid:1)(cid:0)(cid:1)(cid:0)(cid:1)(cid:0)(cid:1)(cid:0)(cid:1)(cid:2)(cid:1)(cid:2)(cid:1)(cid:2)(cid:1)(cid:2)(cid:1)(cid:2)(cid:1)nd(cid:0)(cid:1)(cid:0)(cid:1)(cid:0)(cid:1)(cid:0)(cid:1)(cid:0)(cid:1)(cid:2)(cid:1)(cid:2)(cid:1)(cid:2)(cid:1)(cid:2)(cid:1)(cid:2)(cid:1)−(cid:0)(cid:1)(cid:0)(cid:1)(cid:0)(cid:1)(cid:0)(cid:1)(cid:0)(cid:1)(cid:2)(cid:1)(cid:2)(cid:1)(cid:2)(cid:1)(cid:2)(cid:1)(cid:2)(cid:1)H(cid:0)(cid:0)(cid:0)(cid:0)(cid:0)(cid:2)(cid:2)(cid:2)(cid:2)(cid:2)EPonastdth−EndA B SInetrevriaccDet iPornoviderEnd(cid:3)(cid:1)(cid:3)(cid:1)(cid:3)(cid:1)(cid:3)(cid:1)(cid:3)(cid:1)(cid:4)(cid:1)(cid:4)(cid:1)(cid:4)(cid:1)(cid:4)(cid:1)(cid:4)(cid:1)−(cid:3)(cid:1)(cid:3)(cid:1)(cid:3)(cid:1)(cid:3)(cid:1)(cid:3)(cid:1)(cid:4)(cid:1)(cid:4)(cid:1)(cid:4)(cid:1)(cid:4)(cid:1)(cid:4)(cid:1)h(cid:3)(cid:3)(cid:3)(cid:3)(cid:3)(cid:4)(cid:4)(cid:4)(cid:4)(cid:4)ost A B C
Service Provider
Provider−customer Provider−customer Component Services
boundary boundary
(a) Cooperative Model of Composition (b) Brokered Model of Composition
Fig.3.ServiceCompositionModels
performanceofthecomposedservice.AnotherexampleistheYahooportalservicethat
composesthirdpartyservicessuchastheGooglesearchengine,stocktickerandnews.
Fig. 3 illustrates these alternate composition models. The models of composition
say nothingaboutthe data flow, onlythe natureof interaction,or the businessmodel,
between providers. As an example, in the brokered model, we could have data flow
through the broker, who assembles it (shown by the dotted lines on the top). Instead,
the broker could set up the data exchange and not be in the data path (shown by the
dottedlineatthebottomofthefigure).
The same composed service could be implemented in either model. Consider the
provision of a connectivity service with QoS guarantees between two points on the
Internet.In a cooperative model,ISPsenterintoservicelevel agreementsthatspecify
mutualQoSguarantees.Thesemaybestitchedtogetherinadistributedfashiontoof-
ferend-to-endguarantees(www.merit.edu/working.groups/i2-qbone-bb).Inabrokered
model, a provider like InterNAP (www.internap.com) purchases pipes with specified
guaranteesfromindividualISPsandusesthemtoprovideQoStoitscustomers.
Eachofthesemodelsissuitedtoaparticularenvironment.Inthecooperativemodel,
providers work together and can share performance information to ensure end-to-end
propertiesofthecomposedservice.However,theymustrelyoneachother,whichleads
toissuesoftrust.Sincetheresponsibilityforthecomposedserviceisdistributed,each
provider must continuously verify that the others with whom it has agreements meet
theirservicespecifications.Thesespecificationsareintermsoffunctionality,protocol,
performance,andavailability.Suchcomprehensiveverificationisabsentinthecooper-
ativecompositionofInternetconnectivityserviceacrossnetworkdomainstoday.
Inthebrokeredmodel,thebrokercomposesanend-to-endservicebyselectingindi-
vidualservicesresidingindifferentdomains.Thissimplifiesservicedeploymentsince
themembersofthecompositionneednotagreeamongthemselves,onlywiththebro-
ker.Thisalsoenablesthecompositionofservicesacrosscompetingserviceproviders.
The broker assumesresponsibilityfor constructingthe entireend-to-endservicepath.
Brokering is a powerful tool to construct services from providers who are not neces-
sarilyawarethattheyareparticipatinginalargerend-to-endservice(e.g.,inFigure1,
Zagat’s restaurant guide service does not know it is being composed with Babelfish’s
language translation service). However, because the broker has limited visibility into
the underlying provider resources, sub-optimal utilization of provider resources may
End−User Applications
e Composed service @
n
a layer i
Application Services n Pl
o Measmt.−based Dynamic res. Trust mgt./
ati adaptation allocation Verification
on Middleware Services plic Interoperability Policy
ositi End−to−End Network Ap Management
Service Comp with DEEnneshhiaarannbcceeledd P LPraiontphkessrties ectivity Plane ServicOei−tsh 1aetlr al sayeyerrevriices @ Unadfnoedrrl cymoinemgcp htoaesncihitsinmoinqsues
n
n
o Component services
IP Network C
(a) (b)
Fig.4.AlayeredreferencemodelforServiceComposition
result. The nature of the composition and the relationships of the underlying service
providerswilldeterminewhichmodelismostappropriate.
4 ServiceComposition:ALayeredReferenceModel
The previous section classified service composition by how responsibility is shared
across providers. Here we present a layered reference model for composition across
differentlayersofconcern.
ComposedservicesbuildontopofconnectivityprovidedbytheIPlayer,thebottom
sliceinFig.4(a).WeassumethatIPprovidesreachability,andbuildseveral“desirable”
propertiesintheconnectivityplane.Theseincludefeaturessuchasperformanceguaran-
tees(e.g.,latency,bandwidth,loss-ratebounds),availabilityguarantees(e.g.,available
99.99%ofthetime),aswellasfunctionalityorprotocolguarantees(e.g.,guaranteeing
thataparticularadvertisedInternetrouteisvalid).Weachievethemthroughcomposi-
tionattheconnectivityplane,acrossmultipleproviders.Thisresultsinan“end-to-end
networkwithdesirableproperties”,asshowninthemiddlesliceofthefigure.
Theconnectivityplaneisfurtherdividedintotwolayers:Enhancedlinks,andEn-
hancedpaths.Theenhancedlinksabstractionisbuiltbetweentwointeractingentities:
betweentwoserviceproviders,orbetweenaproviderandtheend-user.Thisabstraction
achieves desirable performance oriented properties, and verification of routing proto-
cols between peering entities. Verification checks if the routes advertised by a peer
are valid. An example of this composition is of the MVNO choosing between multi-
ple MNOs. Here the enhancement is the improved performance through reduced call
blockingrate,duetodynamicloadsharingacrossmultipleMNOs.3
3AlthoughitisnotobviousherehowthisexamplebuildsonIPtechnology,itwillbeclearafter
adiscussionofthemechanismsforresourceallocation,inSec.5.
Enhancedpathsbuildonenhancedlinks,andprovidedesirablepropertiesinanend-
to-end path between points on the Internet. The path spans multiple service domains,
and can be chosen adaptively via resource allocation across providers to meet perfor-
mance or availability constraints. Our reference model is independent of whether the
performanceguaranteesarestrict,orsimply“enhancements”tothebest-effortInternet.
Alternativeenhancementsmightbeappropriatefordifferentendapplications.
The top half of the figure represents the application plane. These layers support
end-userapplicationsandareinturnbuiltontopoftheend-to-endnetwork.Themiddle-
wareservicesareenablers,suchastheBabelfishlanguagetranslationservice,content-
distribution networks or video/audio transcoders. The application services layer con-
sists of services useful to end-users, such as the Zagat’s restaurant guide, search en-
ginesoravoice-mailservice.Compositionattheapplicationplaneresultsinenhanced
functionality.Inourexample,theenhancedfunctionalityisthatoftherestaurantguide
appearingintheuser’snativelanguage(Japanese)andpresentationstyle(NTTDoCoMo
userinterface)inaforeignnetwork.
Servicecompositioncanbeappliedwithinandacrosstheselayers,asillustratedin
Fig.4(a).Fig.4(b)showsthismoreexplicitly.Acomposedserviceatahigherlayeris
composedofmultipleservicesinthesamelayer,orofservicesatthelayersbelow.In
Fig. 1, composition takes place across the application (restaurant guide), middleware
(languagetranslation),andconnectivityplane(roamingserviceenabledbySprint)lay-
ers.InFig.2,compositionisattheenhancedlinklayer,usingthecomponentconnec-
tivityservicesofferedbyco-locatedMNOs.
Wenotethatthislayerizationisonlyareferencemodel,andcompositionsneednot
strictly adhere to it. Some application services can be composed directly of enhanced
links, without using the enhanced paths abstraction or middleware services. This will
occur typically in performance sensitive applications, where the composer needs full
visibility via a flat rather than a hierarchical composition. We next discuss the tech-
niquesandmechanismsusedinenablingcompositionasshowninFig.4(b).
5 MechanismsforServiceComposition
TheissueswelistedinSec.2appearindifferentflavorsinthealternatemodelsofcom-
position.SAHARAis our architecturalprototypeto explore the mechanismsrequired
toaddresstheseissues.Tounderstandthevariousmechanisms,weareworkingonsev-
eral case studies of composition. These cover the dimensions of classifications in last
twosections.Wenowdescribethevariousmechanismsthatwearedesigningtoaddress
the challenges presented earlier and how they are used in the individual case studies.
Wehaveperformedin-depthevaluationsofseveralofthesemechanisms.Duetospace
limitations,wesummarizetheevaluationsofonlyasubsetofthesemechanisms.
5.1 Measurement-BasedAdaptation
For service composition, it is desirable to dynamically choose service providers, and
serviceinstancesbasedoncurrentnetworkandserverloads.Measurementscanbecar-
riedoutbyathirdpartymeasurementservice–acommonelementoftheserviceinfras-
tructure, or by the composer itself. This applies to both the cooperative and brokered
modelsofcomposition.Wenowdescribetwoservicesthatwearedevelopinginvolving
measurement-basedadaptation.
We are developing a general end-to-end Internet host distance estimation service.
Suchaserviceisespeciallyusefulinabrokeredcompositionsincethebrokerdoesnot
haveinsightintothenetworkcharacteristicsofindividualproviders.Giventhepotential
largenumbersofserviceprovidersandinstances,toscalethemeasurementservice,we
clustertheendhoststobemonitoredbasedonthesimilarityoftheirperceiveddistance
to the measurement points. The cluster center is then used as a single measurement
targetforfuturemonitoring.SimulationswithrealInternetmeasurementdatashowthat
ourschemehasgoodpredictionaccuracyandstabilitywithasmallcommunicationand
computationcost[2].
AnapplicationservicethatwehavedevelopedistheUniversalInbox[3],ametaphor
forany-to-anycommunicationacrossheterogeneousdevicesandnetworks.Datatrans-
formationservicessuchasaudio/videotranscodersandtext$speechenginesareex-
tensively used to adapt content between communicating devices. For example, in our
earlier scenario, Ms. Tanaka’s email service could be composed with a text-to-speech
conversionservicesothatshecanlistentoemailsoverhercellular-phone.Weusethe
brokered model of composition here. To adaptively choose service providers and in-
stances, we have designed a middleware measurement layer that exchanges network
andserverloadusingalink-statealgorithm[4].Thisexchangetakesplaceacrossser-
vice execution platforms, enabling a dynamic choice of service instances, possibly in
themiddleoftheusersessions,tohidenetworkandserverfailuresfromend-users.
Acriticalchallengethatweaddressinthecontextofsuchanapplicationisavailabil-
ity.Whencomposedservicesspanmultipleproviders,datacouldtraversethewide-area
Internet. We detect and recover from Internet path failures quickly by using the mid-
dlewaremeasurementlayertochoosealternateserviceinstancesfortheclientsession.
Ourmeasurementsin[4]showthatInternetpathfailuresthathappenduetocongestion
or other factors can be detected reliably within about 2 seconds. Further, subsequent
recovery by using alternate service instances can be completed in a few hundred mil-
liseconds.Thus,networkpathfailureslastingseveraltensofsecondstominutes[5]can
becompletelymaskedfromtheendclient.
In addition to these two services, we have also designed and implemented a mea-
surementmethodology[6]toimproveDNS-basedserverselection,whichisacommon
techniqueusedbyContentDistributionNetworks(CDNs)today.Ourtechniqueenables
thecollectionofclienttolocalDNSservermappingstoallowmoreaccurateserverse-
lection based on a client’s local DNS server. Understanding the distribution of HTTP
requestscorrespondingtolocalDNSserversalsoenablesbetterloadpredictiongivena
DNSrequestandthusimprovedserverselectionmechanisms.
5.2 Utility-BasedResourceAllocationMechanisms
Inamulti-providerenvironment,differentprovidersmayexperiencedifferentdemands
for resources due to user dynamics. Demand or utility-based resource allocation can
beappliedwithinaserviceprovidertomanageitsinstances.Inthebrokeredmodel,it
canbeusedtoallocateresourcesacrossproviders.InSAHARA,weareexploringtwo
resourceallocationmechanisms.
0.18
Random Redirection
0.16 Congestion & QoS Sensitive Redirection
0.14
bility 0.12
a
ob 0.1
Pr
g 0.08
n
ocki 0.06
Bl
0.04
0.02
0
0 0.2 0.4 0.6 0.8 1
Relative Weight of Congestion Sensitivity
Fig.5.Congestion-sensitivityinQoS-basedredirection
Auctionsareonewayofconstructingamarketplacewherearesource,suchasband-
widthorphysicalspectrum,canbedynamicallyallocated.Auctionsallocateresources
toconsumersbasedontheirbids,whichrepresentthevalueofthegoodtothem.Fur-
thermore,theresourcecanbesubdividedintounits,andmultiplebidderscanbeallo-
catedtheresourceuntiltheresourcepoolisexhausted.Auctionscanoccurinrounds,
where the allocation determined by each round can be for some future time period.
Auction-basedallocationsinamulti-providerenvironmentcanprovidethemechanism
fordemand-basedresourceallocation.Forinstance,intheMVNOscenarioinSection1,
spectrumresourcescouldbeauctionedoffeveryfewminutestocompetingvirtualop-
eratorsbasedontheircurrentuser-loadintheareaofcoverage.
Congestionpricingisanallocationmechanismthatassignsscarceresourcestocon-
sumers using the abstraction of price as a means to moderate demand. During high
demand,suchamarketensuresthatthepriceincreases.Onlythoseconsumerswiththe
greatest need and having sufficient currency will obtain the needed resource. During
low demand, the price drops, and access to the resource with be cheap and plentiful.
Thisapproachshouldyieldanassignmentofresources(supply)totheneed(demand)
thatadaptstoinstantaneousdemands[7].
As an application that uses the congestion pricing mechanism, we have looked at
the selection of Voice-over-IP (VoIP) gateways across multiple providers [8]. These
gateways are deployed by independent entities, and exchange dynamic pricing infor-
mationaswellasthepeeringrelationshipsbetweentheproviderentitiesbasedonthe
IETF TRIP (Telephony Routing over IP) protocol. The price is decided based on the
load,orcongestion,ateachgateway;andtheusergetstochoosebetweenseveralgate-
waysbasedonthepriceandtherequiredqualityofservice.Thisachievespricing-based
load-sharing among the gateways. In [9], we looked at the trade-off between using
QoS-basedredirection(nearestVoIPgateway)andcongestion-basedredirection.Thisis
showninFigure5whereweseethatincorporatingcongestion-sensitivityinQoS-based
redirectioncanimprovecallblockingratebyasmuchasafactorofthreeincomparison
torandomredirection.Wealsoobservethatincreasingcongestion-sensitivitydoesnot
significantlydegradetheQoSoftheVoIPcalls.