SCALING VIRTUAL NETWORK FUNCTIONS TO MINIMIZE OPERATIONAL COST

Sabidur RahmanFriday Group Meeting, NetlabUC Davis

4/14/171

krahman@ucdavis.eduhttp://www.linkedin.com/in/kmsabidurrahman/

Happy Bengali New Year 1424!

4/14/172

AgendaDefineauto-scalingMotivationLiteraturereviewProblemstatementanddimensionsMethodsandprogress

4/14/173

Auto-scaling (1)“Autoscaling, also spelled auto scaling or auto-scaling, is a method usedin cloud computing, whereby the amount of computational resources in aserver farm, typically measured in terms of the number of active servers,scales automatically based on the load on the farm.”

4/14/174

1.https://en.wikipedia.org/wiki/Autoscaling

Amazon Web Services (AWS)NetflixMicrosoft's Windows AzureGoogle Cloud PlatformFacebook

Auto-scaling (2)“Auto Scaling helps you maintain application availability and allows you toscale your Amazon EC2 capacity up or down automatically according toconditions you define.….Auto Scaling can also automatically increase the number of Amazon EC2instances during demand spikes to maintain performance and decreasecapacity during lulls to reduce costs.Auto Scaling is well suited both to applications that have stable demandpatterns or that experience hourly, daily, or weekly variability in usage.”

4/14/175

2.https://aws.amazon.com/autoscaling/

Auto-scaling of network resources •BroadbandNetworkGateways(BNGs)•EvolvedPacketCore(EPC)•Firewalls•DeepPacketInspection(DPI)•Dataexfiltration systems•NATs•WebProxies•Loadbalancers•Contentcaching•Parentalcontrol

4/14/176

3.Palkar S,Lan C,HanS,JangK,PandaA,Ratnasamy S,RizzoL,Shenker S.E2:aframeworkforNFVapplications.InProceedings ofthe25thSymposiumonOperatingSystemsPrinciples2015Oct4(pp.121-136).ACM.

Network Function virtualization

4/14/177

4.http://www.alepo.com/white-papers/alepo-in-the-virtualized-core-network/5.GuptaA,Habib MF,Chowdhury P,Tornatore M,Mukherjee B.Jointvirtualnetworkfunctionplacementandroutingoftrafficinoperatornetworks.UCDavis,Davis,CA,USA,Tech.Rep.2015Apr20.

Service chaining

4/14/178

6.https://www.ntt-review.jp/archive/ntttechnical.php?contents=ntr201408fa2.html7.GuptaA,Habib MF,Chowdhury P,Tornatore M,Mukherjee B.OnservicechainingusingVirtualNetworkFunctionsinNetwork-enabledCloudsystems.In2015IEEEInternationalConferenceonAdvancedNetworksandTelecommuncationsSystems(ANTS)2015Dec15(pp.1-3).IEEE.

Network function outsourcing

4/14/179

8.Lan C,SherryJ,Popa RA,Ratnasamy S,LiuZ.Embark:securelyoutsourcingmiddleboxes tothecloud.In13thUSENIXSymposiumonNetworkedSystemsDesignandImplementation(NSDI16)2016Mar16(pp.255-273).9.Fayazbakhsh SK,ReiterMK,Sekar V.Verifiablenetworkfunctionoutsourcing:requirements,challenges,androadmap.InProceedings ofthe2013workshoponHottopicsinmiddleboxes andnetworkfunctionvirtualization2013Dec9(pp.25-30).ACM.

Motivation (1)•Autonomous•Bettermanagementandcontrol•Costsavings•Energyefficiency(Savingtheworld?)

4/14/1710

Motivation (2)•ContentDistributionNetworks(CDNs)[10]:Netflix,Akamai.•Telecomnetworks[11]:AT&T,Verizon.•DataCenterNetworks[13]:Google,Amazon,Facebook.•MobileVirtualNetworkOperators[12]:BoostMobile (Sprint), CricketWireless (AT&T), MetroPCS (T-MobileUS)•Software-definedDataCenter[14]•Networkfunctionoutsourcing

4/14/1711

10.Mandal U,Chowdhury P,LangeC,Gladisch A,Mukherjee B.Energy-efficientnetworkingforcontentdistributionovertelecomnetworkinfrastructure.OpticalSwitchingandNetworking.2013Nov30;10(4):393-405.11.ZhangY,Chowdhury P,Tornatore M,Mukherjee B.Energyefficiencyintelecomopticalnetworks.IEEECommunicationsSurveys&Tutorials.2010Oct;12(4):441-58.12.Zarinni F,Chakraborty A,Sekar V,DasSR,GillP.Afirstlookatperformanceinmobilevirtualnetworkoperators.InProceedings ofthe2014ConferenceonInternetMeasurementConference2014Nov5(pp.165-172).ACM.13.HellerB,Seetharaman S,Mahadevan P,Yiakoumis Y,SharmaP,Banerjee S,McKeown N.ElasticTree:SavingEnergyinDataCenterNetworks.InNSDI2010Apr28(Vol.10,pp.249-264).14.http://www.vmware.com/solutions/software-defined-datacenter.html?src=phd709

Literature review (1)

•Focus:Contentdistributionovertelecomnetwork•Energyconsumptionmodel,analysisandcontent-placementtechniquestoreduceenergycost•Storagepowerconsumptionandtransmissionpowerconsumption•Time-varyingtrafficirregularities•Morecontentreplicasduringpeakloadandlessreplicasduringoff-peakload

4/14/1712

10.Mandal U,Chowdhury P,LangeC,Gladisch A,Mukherjee B.Energy-efficientnetworkingforcontentdistributionovertelecomnetworkinfrastructure.OpticalSwitchingandNetworking.2013Nov30;10(4):393-405.

Literature review (2)

•Focus:Datacenternetworks•Scaleupanddowntosaveenergy•Dynamicallyadjustlinkandswitchestosatisfychangingtrafficload•Optimizermonitorstraffictochoosesetofelementsneededtomeetperformanceandfaulttolerancegoals.•Formalmodel,Greedybin-packer,topology-awareheuristicanddemandprediction-basedmethod

4/14/1713

13.HellerB,Seetharaman S,Mahadevan P,Yiakoumis Y,SharmaP,Banerjee S,McKeown N.ElasticTree:SavingEnergyinDataCenterNetworks.InNSDI2010Apr28(Vol.10,pp.249-264).

Literature (3)

• Aproactivesystemscaleup/scaledown technique•Machinelearningmodelsforpredictingfailurescausedbyaccumulationofanomalies(Software/Hardware)•WhenaVMjoins(orleaves)aregion,theregionworkloadisautomaticallyspreadacrosslocalVMs

15.Avresky DR,DiSanzo P,Pellegrini A,Ciciani B,ForteL.ProactiveScalabilityandManagementofResourcesinHybridCloudsviaMachineLearning.InNetworkComputingandApplications(NCA),2015IEEE14thInternationalSymposiumon2015Sep28(pp.114-119).IEEE.

§ VNFs can be dynamically scale-in/out to meet the performance desire§ Auto-scaling algorithm for desired characteristics with low operation cost and low latency§ Tradeoff between performance and operation cost§ NFV enabled Evolved Packet Core (EPC) is modeled as queueing model§ Legacy network equipment are considered as reserved a block of servers§ VNF instances are powered on and off according to the number of job requests present.

4/14/1715

Phung-Duc T,Ren Y,ChenJC,YuZW.DesignandAnalysisofDeadlineandBudgetConstrainedAutoscaling(DBCA)Algorithmfor5GMobileNetworks.arXiv preprintarXiv:1609.09368.2016Sep29.

Literature review (4)

§ Provision and orchestration of physical and virtual resource is crucial for both Quality of Service (QoS) guarantee and cost management in cloud computing environment.§ SLA-aware and Resource-efficient Self-learning Approach (SRSA) for auto-scaling policy decision§ Busy-and-idle scenario and burst-traffic scenario

4/14/1716

TangP,LiF,ZhouW,Hu W,YangL.EfficientAuto-ScalingApproachintheTelcoCloudUsingSelf-LearningAlgorithm.In2015IEEEGlobalCommunicationsConference(GLOBECOM)2015Dec6(pp.1-6).IEEE.

Literature review (5)

Problem statementGiven:Networktopology,networktrafficdata,SLAObjective:PredictthefuturetrafficandscaleVNFstominimizenetworkoperationcost(ornetworkleasingcost).

4/14/1717

High level design

4/14/1718

Scalingalgorithm Actuator(s)

SLA

MetricsandStats

Trafficprediction

Usecases1) Network (or VNF) leaser (CDN, MVNO): Lower usage = lower rents2) Network (or VNF) owner (AT&T, Time Warner): Lower usage = lower

OPEx.

4/14/1719

Topologies1) Data Center hosting VNFs (Fat Tree)2) Access Network with COs (Compute node hosting VNFs) and

WAN

Scaling techniques1) Heuristic: threshold based; allocate enough VNFs to serve traffic2) Reinforcement learning algorithm

4/14/1720

Traffic prediction1) Machine learning2) Deep learning

Summary§ Machine learning can help with the prediction§ Reinforcement learning can help with the scaling decision§ Cost analysis from both network owner and leaser perspective would be interesting.

4/14/1721

4/14/1722

Questions?krahman@ucdavis.eduhttp://www.linkedin.com/in/kmsabidurrahman/