Transcription
Elastic Virtual NetworkFunction PlacementCloudNet 2015M . G HAZN AVI , A. KHAN , N . S HAHR I AR , KH. ALS U B HI , R . AHM E D, R . B O U TAB ADAVI D R . C HE R I TON SC HOOL OF C OM PU T E R SC I E N C EU N I VE R SI T Y OF WAT E R LOO
OutlineIntroductionState of the ArtProblem: Elastic Virtual Network Function PlacementSolution: Simple Lazy Facility LocationEvaluationConclusion2 / 30
IntroductionM I DDLE -B OXE SN E T W OR K FU N C T I ON VI R T U ALI ZAT I ONVN F SE R VI C E S I N C LOU D3 / 30
Middle-Boxesโany intermediary device performing functions other than the normal, standard functions of anIP router on the datagram path between a source host and destination hostโ [1]Expensive hardwarefor in-network capabilities [9]. The lack of extensibility in middleboxes today inevitablyleads to further apMiddle-boxpliance sprawl, with associated increases in capital andHard to modifyoperating expenses.๐ก๐๐ the๐Despite these ๐ ๐๐concerns, administratorsreiteratedvalue they find in such appliances, particularly in supHard to scaleSource hostDestination hostporting new applications (e.g., IDSteleconferencing), inFirewallcreasing security (e.g., IDS), and improvingperformanceIPSProvision for peak-load(e.g., WAN optimizers). Hard to deploy3CoMb: Overview and OpportunitiesNormalized utilizationWAN optimizerProxy1Load BalancerFirewall0.80.60.40.20:00:00:00:00:009,06 7-09,17 7-10,04 7-10,15 7-11,020000Time (mm-dd,hr)07-0Figure 1: Middleboxutilizationpeakat differenttimesThe previous discussion shows that even though middleMiddle-boxutilizationpeak atdifferenttimes[2]boxes are a critical part of the network infrastructure,VPN WanOpt IDS Proxythey remain expensive, closed platforms that are diffiFirewallcult to extend, and difficult to manage. This motivates us[1] CARPENTER, B., AND BRIM, S. Middleboxes: Taxonomy and Issues. RFC 3234, https://tools.ietf.org/rfc/ rfc3234.txt, 2002.to rethink how middleboxes are designed and managed.[2] V. Sekar, N. Egi, S. Ratnasamy, M. K. Reiter, and G. Shi. Design and implementation of a consolidated middlebox architecture. In Proceedings of NSDI 12, 2012.Protocol'Parsers'We envision an alternative architecture, called CoMb,wherein software-centric implementations of middleSession'Management'box applications are consolidated to run on a shared4 / 30
Network Function VirtualizationVirtualization (Softwarization) of middle-boxesSoftware middle-boxes are called Virtual Network Function (VNF)NFV โinvolves the implementation of network functions in software that can run on a range ofindustry standard server hardware, and that can be moved to, or instantiated in, variouslocations in the network as required, without the need for installation of new equipment.โ[1]VNF๐ ๐๐Source host๐ฃ๐ก๐๐Target host[1] "Network Functions Virtualization". ISG web portal: https://portal.etsi.org/nfv/nfv white paper.pdf5 / 30
Network Function VirtualizationMIDDLE-BOXESVIRTUAL NETWORK FUNCTIONSExpensive hardwareLow-cost softwareHard to deployEasy to deployHard to modifyEasy to modifyHard to scaleEasy to scaleProvision for peak-loadScale resources on demand6 / 30
VNF Services in CloudOffered by cloud providers IBM Bluemix Microsoft Azure Amazon EC2Services Riverbed STEELHEAD WAN optimizer [1] McAfee Next Generation firewall [2] Virtual LoadMaster load balancer [3]ClientCloud-ProviderVNF ServiceRequest[1] http://media-cms.riverbed.com/documents/Spec Sheet - Steelhead Family - 05.06.2015.pdf[2] nt/live/PRODUCT DOCUMENTATION/25000/PD25151/en US/NGFW 57 HW Requirements.pdf[3] ntation/Datasheets/VLM-AWS.pdf7 / 30
VNF Services in CloudWHAT CLOUD PROVIDER SHOULD SUPPORTCHALLENGES OF CLOUD PROVIDERPay per useMinimizing Costs: Clients pay only for real used resourcesElasticity Scale resources on demand Upon arrival or departure of service request Variation of workload of admitted service request Trade-off between Host & Bandwidth ResourcesElasticity Which mechanisms to apply Elasticity benefit vs. its overhead8 / 30
VNF Services in CloudWhere to place VNF instances?๐ฃWhich request must be assigned to which VNF instance?VNF instance๐ ๐๐Source of Service Traffic๐ก๐๐Target of Service TrafficNetwork ๏ฟฝ๏ฟฝ๏ฟฝ๐๐(๐ ๐๐ ๐ก๐๐ ๐ ๐๐*๐ก๐๐*9 / 30
VNF Services in CloudA solution can be๐ฃ ๐ฃ) serves the first and second service traffics ๐ฃ( serves the third and forth service trafficsVNF instance๐ ๐๐Source of Service Traffic๐ก๐๐Target of Service TrafficNetwork ๏ฟฝ๏ฟฝ๏ฟฝ(๐ฃ(๐ก๐๐(๐ ๐๐ ๐ก๐๐ ๐ ๐๐*๐ก๐๐*10 / 30
State of the ArtCOMPA RISON OF STAT E OF T HE A RT11 / 30
Comparison of State of the ArtPaperHost Res. Cost Bandwidth Res. CostElasticityElastic Virtual Network Function Placement (EVNFP) Elasticity in Cloud [1, 2, 3] Dynamic VM Placement [2, 4] Network Aware VM Placement [5, 6, 7] Virtual DPI Placement [8] [1] Z. Gong, X. Gu, and J. Wilkes. Press: Predictive elastic resource scaling for cloud systems. In IEEE CNSM, 2010[2] U. Sharma, P. Shenoy, S. Sahu, and A. Shaikh. A cost-aware elasticity provisioning system for the cloud. In IEEE ICDCS 2011.[3] Z. Shen, S. Subbiah, X. Gu, and J. Wilkes. Cloudscale: Elastic resource scaling for multi-tenant cloud systems. In ACM SoCC, 2011.[4] A. Verma, P. Ahuja, and A. Neogi. pmapper: Power and migration cost aware application placement in virtualized systems. In ACM/IFIP/USENIX Middleware, 2008.[5] O.Biranetal.A stable network-aware vm placement for cloud systems. In CCGRID, pages 498โ506, 2012.[6] V. Mann, A. Kumar, P. Dutta, and S. Kalyanaraman. Vmflow: Leveraging vm mobility to reduce network power costs in data centers. In IFIP NETWORKING, 2011.[7] X. Meng, V. Pappas, and L. Zhang. Improving the scalability of data center networks with traffic-aware virtual machine placement. In IEEE INFOCOM, 2010.[8] M. Bouet, J. Leguay, and V. Conan. Cost-based placement of vdpi functions in nfv infrastructures. In NetSoft, 2015.12 / 30
Problem:Elastic Virtual Network FunctionPlacement (EVNFP)SC OPE AN D ASSU M PT I ON SC ON FLI C T I N G OB JE C T I VE SE LAST I C I T Y M E C HAN I SM S AN D OVE R HE AD13 / 30
Scope and AssumptionsSCOPEASSUMPTIONSSingle cloud providerOne VNF instance-typeSingle data-centerMulti-tenancyCentralized optimizationElasticity Mechanisms Horizontal Scaling Migration of VNF instances Reassignment of workload14 / 30
Conflicting ObjectivesMinimizing the bandwidth cost, and๐ฃMinimizing the number of installed VNFsVNF instance๐ ๐๐Source of Service Traffic๐ก๐๐Target of Service TrafficNetwork ๏ฟฝ๏ฟฝ๏ฟฝ๐๐(๐ ๐๐ ๐ก๐๐ ๐ ๐๐*๐ก๐๐*15 / 30
Conflicting ObjectivesMinimizing the bandwidth cost:๐ฃ 12 Unit of Bandwidth over 12 Links 4 VNF instancesVNF instance๐ ๐๐Source of Service Traffic๐ก๐๐Target of Service TrafficNetwork ๏ฟฝ๏ฟฝ๏ฟฝ(๐ฃ)๐ก๐๐(๐ ๐๐ ๐ฃ ๐ฃ*๐ก๐๐ ๐ ๐๐*๐ก๐๐*16 / 30
Conflicting ObjectivesMinimizing the number of installed VNFs๐ฃ 1 VNF instance 34 Unit of Bandwidth over 20 LinksVNF instance๐ ๐๐Source of Service Traffic๐ก๐๐Target of Service TrafficNetwork ๏ฟฝ๏ฟฝ๏ฟฝ(๐ก๐๐(๐ ๐๐ ๐ก๐๐ ๐ ๐๐*๐ก๐๐*17 / 30
Elasticity Mechanisms and OverheadMECHANISMSOVERHEADHorizontal Scaling of VNF instanceMigration overhead Installing a new VNF instance Removing an existing VNF instanceReassignment overheadMigration of a VNF instanceReassignment of workload to another VNFinstance18 / 30
Elasticity Mechanisms and Overhead1Initial Placement๐ฃ Installation and Reassignment2VNF instanceSource of service trafficTarget of service trafficService traffic increaseService traffic decrease34Migration of ๐ฃRemoving ๐ฃ19 / 30
Solution:Simple Lazy FacilityLocation(SLFL)ID EASLFL: SIMPLE LA ZY FA CILIT Y LOCAT ION20 / 30
IdeaArrival and departure of a request, or workload variation alter the locality๐ฃSLFL locally optimizes the placement of VNF instances in a greedy manner๐ ๐๐Source of Service Traffic๐ก๐๐Target of Service TrafficVNF instanceNetwork ๏ฟฝ๏ฟฝ๐๐ ๐ก๐๐ ๐ ๐๐*๐ก๐๐*21 / 30
SLFL: Simple Lazy Facility LocationUPON REQUEST ARRIVAL ORWORKLOAD INCREASEUPON REQUEST DEPARTURE ORWORKLOAD DECREASEInstallation potentialRemoving potential Installing a VNF instance Set of reassignments The difference of operational cost before andafter installing the VNF instance andreassignmentsMigration potential Migration of a VNF instance The difference of operational cost before andafter migration of the VNF instance Removing a VNF instance Set of reassignments The difference of operational cost before andafter removing the VNF instance andreassignmentsEmigration potential Migration of a VNF instance The difference of operational cost before andafter migration of the VNF instance22 / 30
SLFL: Simple Lazy Facility LocationUPON REQUEST ARRIVAL ORWORKLOAD INCREASEUPON REQUEST DEPARTURE ORWORKLOAD DECREASEApply the best action among:Apply the best action among: Installing a VNF instance Considering the installation potential Migrating a VNF instance Considering the migration potential of the VNF instance Removing a VNF instance Considering the installation potential Migrating a VNF instance Considering the emigration potential of the VNF instance Assign to one of existing VNFs Considering bandwidth cost23 / 30
EvaluationE XPE R I M E N TAL SE T U P AN D OB JE C T I VE SAC C E PTAN C E R AT I O AN D OPE R AT I ON AL C OSTR E SOU R C E U T I LI ZAT I ON24 / 30
Experimental Setup and ObjectivesEXPERIMENTAL SETUPOBJECTIVESNetworkEvaluating Fat-tree of 99 nodes 54 hosts with 8 Core CPU 1 GB full bisection bandwidthVNF Bro IDS [2]: 80 Mbps, 1 vCPU, 1GB of memoryRequests 20,000 requests Arrival: Poisson distribution Duration: Exponential distribution The acceptance ratio Operational cost Balancing bandwidth and host resource costs Resource Utilization Balancing bandwidth and host resource utilization ?Comparison to Random Placement First-Fit Placement25 / 30
Acceptance Ratio andOperational CostACCEPTANCE RATIOTOTAL OPERATIONAL 000005010152025303540Time (s)SLFL accepts 2 workload vs basic algorithmsSLFL97% acceptance ratioRandom 48% acceptance ratioFirstFit45% acceptance ratioCost ( per 200 000000 500 000 500 000 500 0005010122334Time (s)SLFL accepts 2 workload with less cost9% operational cost less than Random4% operational cost less than FirstFit26 / 30
Resource UtilizationBANDWIDTH RESOURCE UTILIZATION100HOST RESOURCE me (s)Time (s)82% Utilization of bandwidth resources91% Utilization of host resources27 / 30
ConclusionSUMMA RY28 / 30
SummaryElastic Virtual Network Function Problem Bandwidth and host resources cost trade-off Elasticity OverheadSimple Lazy Facility Location Balancing the bandwidth and host resource cost trade-offCarefully selecting the correct elasticity mechanismsOptimizing the elasticity overheadAccepting 2 workload vs basic algorithms29 / 30
30
Acceptance Ratio andResource 0000000000005010152025303540Bandwidth Resource Util.Acceptance 0000000000005010152025303540Time (s)Time (s)100RandomSLFLFirstFit806040200VNFs Util.Host Resource 0000000005010152025303540Time (s)0000000000000000000000005010152025303540Time (s)
RandomSLFLFirstFit000000000000 500 000 500 000 500 0005010122334Cost ( per 200 sec.)140120100806040200Bandwidth Resource CostTotal Operational CostCost ( per 200 sec.)Operational 00 500 000 500 000 500 0005010122334Time (s)RandomSLFLFirstFit000000000000 500 000 500 000 500 0005010122334Time (s)101ReassignmentMigration100 per 200 sec.140120100806040200Elasticity Overhead CostHost Resource CostCost ( per 200 sec.)Time (s)10 110 210 310 410 50000000000000000050 100 150 200 250 300 350 400
Assumptions-Horizontal ScalingWhy horizontal scaling and ignoring vertical scaling On the fly vertical resource scaling is not supported in most cases Might require system reboot SLA violation
Assumptions-One VNF One small flavorMultiple flavorsHost Res. - Worse BetterBandwidth Res. Better- WorseInstallationIn a same machine Better- WorseRemovalIn a same machine Better- WorseMigration overhead Better- WorseReassign. overhead Equal Equal
WAN optimizer and IDS at two time instants t 1,t 2 are 10,50 packets and 50,10 packets respectively. Today each hardware device must be provisioned to handle its peak load resulting in a total provisioning cost corre-sponding to 2 max{10,50} 100 packets. A