Utility Computing - University of...
8
Computer Networks Lecture 24: Cloud Computing and Data Center Networking Utility Computing August 2006: Amazon Elastic Compute Cloud, EC2+S3 • first successful IaaS offering IaaS == Infrastructure as a Service • swipe your credit card, and spin up your VM Provides utility computing: • computing resources as a metered service (“pay as you go”) • ability to dynamically provision virtual machines Why utility computing? • cost: CAPEX vs. OPEX • scalability: “infinite” capacity • elasticity: scale “out” (or in) on demand [Joshi&Lagar-Cavilla, Lin] I think there is a world market for about five computers. Evolution into PaaS Platform as a Service (PaaS) is higher level • simpleDB (relational tables) • simple queue service • elastic load balancing • flexible payment service PaaS diversity (and lock-in) • Amazon’s Elastic Beanstalk (upload your JAR) • Microsoft Azure: .NET, SQL • Google AppEngine: python, java, GQL, memcache • Heroku: ruby, python, node.js, php, java • Joyent: node.js and javascript [Joshi&Lagar-Cavilla] IaaS vs. PaaS Hardware-centric vs. API-centric Never care about drivers again • or sys-admins, or power bills You can scale if you have the money • you can deploy on two continents • and ten thousand servers • and 20TB of storage x86 JAR Byte Key Value IaaS PaaS [Joshi&Lagar-Cavilla]
Transcript of Utility Computing - University of...
Computer Networks
Lecture24:CloudComputingand
DataCenterNetworking
UtilityComputingAugust2006:AmazonElasticComputeCloud,EC2+S3• firstsuccessfulIaaSoffering
IaaS==InfrastructureasaService• swipeyourcreditcard,andspinupyourVM
Providesutilitycomputing:• computingresourcesasameteredservice(“payasyougo”)• abilitytodynamicallyprovisionvirtualmachines
Whyutilitycomputing?• cost:CAPEXvs.OPEX• scalability:“infinite”capacity• elasticity:scale“out”(orin)ondemand
[Joshi&Lagar-Cavilla,Lin]
Ithinkthereisaworldmarketforaboutfivecomputers.
EvolutionintoPaaSPlatformasaService(PaaS)ishigherlevel• simpleDB(relationaltables)• simplequeueservice• elasticloadbalancing• flexiblepaymentservice
PaaSdiversity(andlock-in)• Amazon’sElasticBeanstalk(uploadyourJAR)• MicrosoftAzure:.NET,SQL• GoogleAppEngine:python,java,GQL,memcache• Heroku:ruby,python,node.js,php,java• Joyent:node.jsandjavascript
[Joshi&Lagar-Cavilla]
IaaSvs.PaaSHardware-centricvs.API-centric
Nevercareaboutdriversagain• orsys-admins,orpowerbills
Youcanscaleifyouhavethemoney• youcandeployontwocontinents• andtenthousandservers• and20TBofstorage
x86 JAR
Byte KeyValue
IaaS PaaS
[Joshi&Lagar-Cavilla]
YourNewConcernsAppprovider:• howwillIhorizontallyscalemyapplication• howwillmyapplicationdealwithdistribution
• latency,partitioning,concurrency
• howwillIguaranteeavailability• failureswillhappen• dependenciesareunknown
Cloudprovider:• howwillImaximizemultiplexing?• canIscaleandprovideperformanceguarantees?• howcanIdiagnoseinfrastructureproblems?
[Joshi&Lagar-Cavilla]
FromCloud-User’sPOVCloudisliketheIPlayer• itprovidesabest-effortsubstrate• iscost-effective• ison-demand• providescomputeandstorageinfrastructure
Butyouhavetobuildyourownreliableservice• faulttolerance• availability,durability,QoS
[Joshi&Lagar-Cavilla]
EverythingasaServiceUtilitycomputing=InfrastructureasaService(IaaS)• whybuymachineswhenyoucanrentcycles?• examples:Amazon’sEC2,Rackspace
PlatformasaService(PaaS)• givemeaniceAPIandtakecareofthemaintenance,upgrades,…• example:GoogleAppEngine,Heroku
SoftwareasaService(SaaS)• Justrunitforme!• example:Gmail,GoogleDocs,Salesforce,Adobe’sCreativeCloud,Microsoft’sOffice365
[Lin]
CloudComputing:SummaryNIST’sdefinition:servicesaccessedoverastandardizednetworkwiththefollowingcharacteristics:
• on-demandself-service:acustomercanordercomputeresourceswithoutanyhumaninteractionwithprovider
• resourcepooling:provider’sphysicalandvirtualresourcespooledtoservemultiplecustomersdynamically
• rapidelasticity:resourcesappearunlimitedandcanbescaledupordownrapidly
• measuredservice:meteredusage(andbilling)
• broadnetworkaccess:availableovertheInternet,platformindependent:mobile,laptops,tablets
AnatomyofaDatacenter
Source:BarrosoandUrsHölzle(2009) Source: IEEE Spectrum and Google
HowMuchPowerNeeded?
• 0.0003kWhtoansweratypicalGooglesearch• 0.05 kWtousealaptopforanhour• 0.1 kWtorunaceilingfanforanhour• 1.1 kWtouseacoffeemakerforanhour• Howmuchpoweris30 MW?• 6,000averagehomeswithcentralair(~5 kW/home)• 300fastfoodrestaurants• 45largeretailstores • 37grocerystores• 30largehomeimprovementstores• 1.5SearsTowers• 1computerdatacenter
DataCenterNetworksTenstohundredsofthousandsofhosts,oftencloselycoupled,incloseproximity:• e-commerce(e.g.,Amazon)• contentservers(e.g.,NetFlix,YouTube,Apple,Microsoft)• searchengines,datamining(e.g.,Google)Challenges:• multipleapplications,eachservingmassivenumbersofclients
• managing/balancingload,avoidingprocessing,networking,databottlenecks
Insidea40-ftMicrosoftcontainer,Chicagodatacenter
Server(blade)racks
Top-of-Rack(ToR)/edgeswitches
Tier-1/coreswitches
Tier-2/aggregationswitches
loadbalancer
loadbalancer
1 2 3 4 5 6 7 8
borderrouter
accessrouter
Internet
DataCenterNetworksLoadbalancer:layer-4“switch”• receivesexternalclientrequests• directsworkloadwithindatacenter• returnsresultstoexternalclient(hidingdatacenterinternalsfromclient)
PotentialNetworkBottleneckHost–ToR:1GbpsToR–Tier2andTier1–Tier2:each10Gbps10hostsonrack1eachtalktoadifferenthostonrack5Similarlybetweenracks2–6,3–7,and4–840flowssharethe10GbpsA–Blink,eachgetsonly10/40= 250Mbps,only¼ofthe1Gbpshost–ToRcapacity
B
1 2 3 4 5 6 7 8
A C10Gbps
10Gbps
1Gbps
Fat-treeTopologywithk = 4Richinterconnectionamongswitches,a.k.a.Closnetwork• increasedthroughputbetweenracksEqualCostMulti-Path(ECMP)routing• increasedreliabilityviaredundancy• originallyintendedfordatacenterwithoff-the-shelfparts
Server/hosts
Top-of-Rack(ToR)/edgeswitches
Tier-1/coreswitches
Tier-2/aggregationswitches
Fat-treeArchitecturek-aryfat-tree:three-layertopology• kpods,eachconsistsof(k/2)2hostsandtwolayersofswitches,eachlayerhas k/2k-portswitches
• eachToRswitchconnectstok/2hostsandk/2Tier-2switches• eachTier-2switchconnectstok/2ToRandk/2Tier-1switches• (k/2)2 Tier-1switches:eachconnectstoallkpods• supportsk3/4hosts,k < 256,fat-treedoesnotscaleindefinitely
Server/hosts
Top-of-Rack(ToR)/edgeswitches
Tier-1/coreswitches
Tier-2/aggregationswitches
[Beyer]
CostAnalysis
Maximumpossibleclustersizewithallhostscapableoffullyutilizinguplinkcapacity
Hierarchicaldesignuseshigher-speed,andmoreexpensive,switcheshigherupinthehierarchy(scaleup)
AddressinginFat-treeUse10.0.0.0/8privateaddressingblock
Podswitcheshaveaddress10.pod.switch.1 • podandswitchinrange[0, k-1], basedonpositionTier-1switcheshaveaddress10.k.i.j • iandjdenoteswitchpositionin(k/2)2Tier-1switches
Hostshaveaddress10.pod.switch.ID • IDinrange[2, (k/2) + 1],fork = 4,IDcanonlybe2or3
10.0.0.1
10.0.3.1
10.0.0.2 10.0.0.3 10.0.1.3 10.2.1.2 10.2.1.3
[Beyer]
ForwardinginFat-treeTier-1switchescontain(10.pod.0.0/16, port)entries• staticallyforwardsinter-podtrafficonspecifiedport
• 10.4.1.1’sroutingtable:
[Beyer]
Top-of-Rack(ToR)/edgeswitches
Tier-1/coreswitches
Tier-2/aggregationswitches
Prefix Output port10.0.0.0/16 010.1.0.0/16 110.2.0.0/16 210.3.0.0/16 3
01 2
3
10.2.1.2 10.2.1.3
10.2.1.1
10.2.0.2 10.2.0.3
Tier-2’sTwo-LevelLookupTablePrefixtablecontains(10.pod.switch.0/24, port)entries• switchvalueistheToRswitchnumber• usedforforwardingintra-podtraffic
Suffixtableusedforforwardinginter-podtraffic
0 12
3
[Beyer]
Recall:fork = 4,hostIDcanonlybe2or3
01 2
3
Tier-2’sForwardingAlgorithm
Prefixtablepreventsintra-podtrafficfromleavingpod
Suffixtableforinter-podtrafficbasedoffhostIDs:• ensuresspreadoftrafficacrossTier-1switches• preventspacketreorderingbyassigningasinglestaticpathforeachhost-to-hostcommunication• betterthanhavingasinglepathbetweensubnets
[Beyer]
ToRSwitch’sForwarding
Inter-racktrafficreliesonswitch’soriginalbackwardlearningalgorithmAssumesforwardingtablesgeneratedbyacentralcontrollerwithfullknowledgeoftopology• centralcontrolleralsoresponsiblefordetectingswitchfailuresandre-routingtraffic
• andforansweringARPandDHCPrequests
[Beyer]
Fat-treeRoutingExample
Server/hosts
Top-of-Rack(ToR)/edgeswitches
Tier-1/coreswitches
Tier-2/aggregationswitches
Packetsfromsource10.0.1.2todestination10.2.0.3takethedashedpath
Two-LevelLookupImplementationImplementedinhardwareusingaTCAM• TCAM:Ternary(0,1,don’tcare)Content-AddressableMemory• canperformparallellookupsacrosstable• storesdon’tcarebits,suitableforvariablelengthprefixes
Prefixespreferredoversuffixes
lookupincomingaddress
[Beyer]
of switch 10.2.2.1 in the example network
TopologyPower/HeatDissipation
PackagingProblemFat-treehassignificantcablingoverhead• 1GigEswitchesusedtoreducecost• lackof10GigEportsleadstomorecabling
Apackagingsolutionfork = 48• generalizestoothervaluesofk
Cablingingeneralcanbeaproblemindatacenternetworks....
[Beyer]
VL2:• alsobasedon
Closnetwork• buthasamoreflexibleaddressingscheme
• runslink-staterouting• doesnetworkloadbalancing
Othertopologieshavethehoststhemselvesalsoserveasrouters
OtherDCNetworkTopologies
NetworkSecurityEvolvedVirtualprivateclouds• internalVLANswithincloud• virtualnetworkfunctions(VNFs):virtualgateways,virtualfirewalls:middleboxesimplementedinsoftware• removeexternaladdressability• MPLSVPNconnectiontocloudgateway• butdoesn’tprotectexternalfacingassets
• providers:Amazon,Google,Microsoft,etc.
[AmazonAWS]
InformationLeakageIsyourtargetinacloud?• traceroute• networktriangulation
EveryVMgetsitsprivate/publicIPAreyouonthesamemachineastarget?• IPaddresses• latencychecks• sidechannels(cacheinterference)
Canyougetonthesamemachine?• pigeon-holeprinciple(nitems,mcontainers,n > m⇒somecontainersmustbeshared)
• placementlocality[Joshi&Lagar-Cavilla]
Source:VoasandZhang,“CloudComputing:NewWineorJustaNewBottle?”ITProfessional,11(2):15–17,March2009
[Joshi&Lagar-Cavilla]
IBM PC�1981
Ethernet 802.3�1983
Commercial Internet�1995
Amazon EC2�2006
Datacenternetworkasaswitch
SETI@home�1999
Thecircleisnowcomplete…
