Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
1
GLOBECOM, Dec. 2016
BySedef Savas
NetworksLab,April14,2017
ControlPlane MappingProblem
Forhighresiliencyà solvecontrollerplacement(bothnumber&placement.)
CPaffectsalloftheresiliencedisciplines
• survivability(asasupersetoffaulttolerance),
• dependability(asasupersetofreliability),
• security,
• performability,
• traffictolerance,and
• disruptiontolerance.
Focus isonthecontrollers’failureswhiletakingintosomefactorswhicharemissinginexistingworks
suchasflowsetuplatency,switches’demands,andthecapacityofthecontrollers.
RelatedWork
A.ControllerPlacementRound-trippropagationlatency,calculatedbyusingthe lengthoftheshortestpathbetweens-c
• FacilityLocationProblem(FLP)
• Capacitatedk-centerproblem:positioningthecontrollerstominimizethepropagation
latencywhileconsideringcapacityofthecontrollersanddemandoftheswitches.
• Pareto-basedOptimalCOntroller (POCO)placementframeworkprovidespareto-optimal
placementswithregardtodifferentmeasures,includingswitch-controllerlatency,
controller-controllerlatencyandcontrollerloadimbalance
RelatedWork(cont.)
B.ResilientControllerPlacementMainlyconcernedwithresilient controllerplacement toimprovetheresilience ofthesouth-
boundconnections(controller-switchconnections)inSDN,resilient controllerplacement to
enhance theresilience ofthecontrolplane(dealingwithcontrollernodefailures).
FaultTolerantControllerPlacement (FTCP)problemtoachievehighsouth- boundreliability. Inthe
proposedformulation,eachswitchisrequiredtosatisfyareliability constraintinawaythatthe
operationalroutetoanyofitsconnectedcontrollersremainswithatleastagivenprobability.
Thesimulationoutcomeofapplyingtheproposedheuristicalgorithmtoseveralnetwork
topologies,demonstratedthatbeingconnectedtotwocontrollerssufficesforeachswitch.[*]
[*]F.J.Ros andP.M.Ruiz,“FiveNinesofSouthboundReliabilityinSoftware-definedNetworks,”inProceedingsoftheThirdACMSIG- COMMWorkshoponHotTopicsinSoftwareDefinedNetworking,2014,pp.31–36.
ProblemDefinition
Tomeettheresilience constraintsaswellastoaddresstheperformance(mostly relatedtothepropagation
latency),thecost,andcapacitylimitation.
• Costlimitations areassociatedwiththenumberofrequiredcontrollersorhavingabudgetinterms
ofthethenumber ofcontrollerinstances, andtheinherentcostofdeployments (e.g.,CAPEX and
OPEX).
• Also, thecapacityconstraintsassistindealingwiththeloadonacontroller(CPU,memory,and
accessbandwidth).
Ifcontrollerbecomesoverloaded, processing latencywillgoup,affectthelatencybetweenaswitchandthe
controller(itbecomesanon-negligible partofthetotallatency).
Moreover,overloaded controllershaveahigherprobability offailure[6].
Assumptions
Allnodesaresuitableforcontrollerdeployment.
Switcheshavecertainload,controllershavecertaincapacity.
Allcontrollershaveauniformfailureprobabilityandtheyfailindependently.
Multi-levelbackupcontrollerlist
Objective:minimizetheswitch- controllerre-assignmentcostsafterthe
possiblecontroller(s)’failures.InRCP,rdenotestheresiliencelevelatwhicha
controllerservesagivenswitch.
Forinstance,r=0indicatesaprimaryassignmentofacontrollertoaswitch,
andr=1denotestheassignmentofthefirstbackupcontrollertoaswitch.
Todocapacitymanagementafterthereassignment– reassignseveryswitch.
PerformanceEvaluation
Networktopologies ofUScontinentaltier- 1serviceprovidersobtainedfromthe
InternetTopology Zoo.
PointofPresence(PoP)-leveltopologies areSprint,ATTNorthAmerica(twomaps,one
before2008andanotherin2008),PSINeT (nowpartofCogentCommunications), and
UUNET(nowpartofVerizonBusiness).
CostofaController
Thehigherthedegreeofapotentialnode(location)forthecontroller,the
lessthecostis.
• nodeswithbetterconnectivityarereachablefrommorenodes,and
placingthecontrollersonsuchnodesmostprobablydecreasesthe
costintermsofthenumberofcontrollers.
In-bandcontrol,i.e.,nodedicatedlinksbetweencontrollersandswitchesfor
controltraffic.Also,shortest-pathlengthbetweenaswitchandacontrolleris
thesumofthepropagationlatenciesofallthelinksalongthepath.
3scenarios
1) homogeneous switches(i.e.,switcheswithhomogeneous demandsequalto500kiloreq/s)
andcontrollers(i.e.,controllerswithhomogeneouscapacitiesequalto5,000kiloreq/s).
Theprobabilityofthenodefailure(i.e.,pf)isassumedtoberandomlychosenfrom[0.010.25]
2)Thecorrespondingdemandsoftheswitchesareconfiguredwiththeminimumandmaximum
valuesof150kiloreq/sand1,000kiloreq/s(withstepsizeof50),respectively.Thecapacitiesof
thecontrollersandpfarethesameasthefirstscenario.
3)demandsoftheswitchesareuniformlydistributedin[2001,000]kiloreq/sandcontrollers’
capacitieshavevaluesin[1,8008,000]kiloreq/swithfixedpfas0.05.
Results
1)numberofassignedcontrollers,
2)propagationlatencybetweens-c,
3)controllerlocationsatdifferentresiliencelevels,and
4)distributionoftheloadsamongthecontrollers.
It can be seen that having a higher resilience
level is more cost-effective (in terms of the
number of required controllers) for the UUNET
which has larger network size as well as more
redundant paths and higher node degree
Regardlessoftheassignedprobability offailuretothenodes, thenumber of
requiredcontrollersarequitesimilarinthefirstandsecond scenarios.
Results(1)- #ofassignedcontrollers
Results(2)
Propagationlatencyasthemaincontributortotheflow-setuplatency
• thedistancebetweeneachtwonodesanapproximationoftheair-line
distance.Notactualfiberlengthsusedinashortestpath.
• Thepropagationlatencythreshold(ontheshortestpath)fromaswitchto
itsassignedcontrolleratanyresiliencelevelisassumedtobe250ms
Considering thehighestresiliencelevel(i.e.,m=2),themaximumpropagation
latenciesforallofthetopologies arebelow50ms,whichisfarlessthanthe
latencythreshold andleavestheroomforothercontributorsoftheflow-setup
latency,including transmission delay,processing delayandprobably thedelay
incurredbycongestioninthenetwork.
Result(2)cont.
In most cases, when the resilience level isincreased (e.g., from m = 0 to m = 1), themaximum latency between a switch and itsassigned controller in a topology goes up.
But this is not always the case; for scenario 3and ATT NA (2), the maximum latencydecreases when the resilience level increasesfrom 1 to 2 in the second experiment.
Results(3)- Controllerlocation
Due to their higher connectivity (higher node degree) and subsequently, better reachability from other nodes. This is reflected by our assumed location-dependent deployment cost (fc) which increases in inverse proportion to the node degree.
Kansas City (common in all scenarios) has a strategic location in most of topologies since it connects east and west of the US in the maps.
Results(4)– controllerloaddistribution
4-tuple shows the minimum, maximum, mean and standard deviation for the number of switches managed by the controllers in a given topology.
While increasing the resilience level results in more load imbalance for some topologies in different scenarios, it leads to less load imbalance for the others. This again confirms the reliance of the solution on the network topology.
Conclusion
Resilientcontrollerplacementproblem, takesintoaccountthecapacityofthe
controllersaswellasthedemandsoftheswitches.
Minimizing thetotalcost(including thecostofdeployment,thepropagationlatency,andthenumberofrequiredcontrollers)achievedwhileconsideringdifferentresiliencelevelstoenhancetheresilienceofthecontrollerplane.
Futureresearchdirections involvedesigning heuristic/approximationalgorithms to
dealwithlargenetworksizesaswellastestingmorerandom/synthetictopologies to
gainusefulinsightsontheresults.