IEEE COMMUNICATIONS SURVEYS AND TUTORIALS, ACCEPTED …bdezfouli/publication/COMST-SDWLAN-2018.pdf · and improve the design and development of network con-trol mechanisms by decoupling

IEEE COMMUNICATIONS SURVEYS AND TUTORIALS ACCEPTED JULY 2018 1

A Review of Software-Defined WLANsArchitectures and Central Control Mechanisms

Behnam Dezfouli Member IEEE Vahid Esmaeelzadeh Student Member IEEEJaykumar Sheth Student Member IEEE and Marjan Radi Member IEEE

AbstractmdashThe significant growth in the number of WiFi-enabled devices as well as the increase in the traffic conveyedthrough wireless local area networks (WLANs) necessitate theadoption of new network control mechanisms Specifically densedeployment of access points client mobility and emerging QoSdemands bring about challenges that cannot be effectively ad-dressed by distributed mechanisms Recent studies show thatsoftware-defined WLANs (SDWLANs) simplify network controlimprove QoS provisioning and lower the deployment cost ofnew network control mechanisms In this paper we present anoverview of SDWLAN architectures and provide a qualitativecomparison in terms of features such as programmability andvirtualization In addition we classify and investigate the twoimportant classes of centralized network control mechanisms(i) association control (AsC) and (ii) channel assignment (ChA)We study the basic ideas employed by these mechanisms andin particular we focus on the metrics utilized and the problemformulation techniques proposed We present a comparison ofthese mechanisms and identify open research problems

Index TermsmdashSoftware-Defined Networking IEEE 80211 Ar-chitecture Mobility Interference Centralized Algorithms

I INTRODUCTION

W IRELESS local area networks (WLANs) have beendeployed broadly in various types of environments

such as enterprises universities airports shopping malls andhomes The IEEE 80211 standard commonly known as WiFiis becoming more popular and ubiquitous In particular thetraffic demand of WiFi networks is increasing due to theemergence of high-definition video streaming Internet ofThings (IoT) cellular data offloading online gaming andvirtual reality [1]ndash[6]

Statistics and forecasts show a continuous growth in thenumber of WiFi-enabled devices shipped as well as theamount of traffic conveyed through WLANs Figure 1(a)shows the number of WiFi-enabled devices including ac-cess points (APs) network interface cards (NICs) routersswitches laptops tablets smartphones and TVs that wereshipped from 2012 to 2017 The shipment of WiFi-enableddevices was 158 billion units in 2012 and is estimated to be491 billion units in 2017 [5]ndash[7] Furthermore the percentageof traffic conveyed by WiFi networks is growing continuouslyas illustrated in Figure 1(b) During 2014 41 of mobile

Behnam Dezfouli Vahid Esmaeelzadeh and Jaykumar Sheth are with theInternet of Things Research Lab Department of Computer EngineeringSanta Clara University Santa Clara CA 95053 USA E-mail bdezfoulivesmaeelzadeh jshethscuedu

Marjan Radi is with Western Digital Corporation Milpitas CA 95035USA E-mail marjanradiwdcedu

2012 2013 2014 2015 2016 2017Year

Billio

2014 2015 2016 2017 2018 2019Year

EthernetWiFiWiFi Offloaded from CellularCellular

(b)(a)

Fig 1 (a) Increase in the number of WiFi-enabled devices shipped during2012 to 2017 (b) Traffic growth of WiFi cellular and Ethernet networksduring 2014 to 2019 [8]

data traffic was exchanged through WiFi networks includingthe traffic offloaded from cellular networks into WLANs [4][8] [9] This traffic is expected to reach 53 by 2019 Inaddition the scientific community has paid great attention to80211 networks as demonstrated by the 165K citations thatthis topic has received until 2014 according to [10]

The increase in the number of WiFi-enabled devices andtheir growing traffic demand requires a dense installation ofAPs to improve spatial reuse and enhance network capacityHowever AP densification intensifies channel contention in-creases the interference level among APs and their associ-ated clients and exacerbates the challenges of client handoffamong APs Therefore effective association and interferencecontrol are both necessary to satisfy the QoS demands ofclients in terms of throughput delay reliability and energyefficiency [11]ndash[15] Satisfying these requirements is evenmore important and challenging when WLANs are used formission-critical application such as industrial process controlfactory automation and medical monitoring In particularthese applications require bounded packet delivery delay andhigh link reliability and in some cases energy efficiency is acritical performance metric as well [16] [17]

In order to tackle the challenges of controlling complexwired networks (such as ISPs and data centers) software-defined networking (SDN) has been proposed to simplifyand improve the design and development of network con-trol mechanisms by decoupling the control and data planeInstead of running network control mechanisms distributivelyon switching devices a controller configures the switches tooperate according to the decisions made centrally [18] [19]Consequently in addition to simplifying the deployment of

network control mechanisms as applications running on thecontroller these mechanisms can benefit from the controllerrsquosglobal network view Not only for wired networks SDNhas gained popularity for the design of WLANs and cellularnetworks as well [20]ndash[23]

A software-defined WLAN (SDWLAN) enables centralmonitoring and control of network operation A SDWLANcan be studied from two point of views as follows

ndash Architecture A SDWLAN architecture (i) definesthe various network components (eg controller APsswitches middleboxes) used to build the network aswell as the topology employed to connect these com-ponents (ii) implements application programming inter-faces (APIs) through which the controller communicateswith network devices to perform data collection anddistribute control commands (iii) specifies the separationlevel between the control plane and data plane The archi-tectural features highly affect the development flexibilityand performance of network control mechanisms [12][24]ndash[26] For example when medium access control(MAC) functionalities are transferred from the APs to thecontroller network control applications may implementAsC mechanisms with various capabilities and perfor-mance levels depending on the separation level and APIsprovided

ndash Control Mechanisms To benefit from the features ofSDWLAN architectures and in particular the global net-work view provided various control mechanisms suchas association control (AsC) channel assignment (ChA)transmission power control and CCA threshold adjust-ment have been proposed by the research communityIn this paper we particularly focus on centralized AsCand ChA mechanisms because they have been widelyproposed and adopted to benefit from the features of SD-WLANs AsC mechanisms address client-AP associationfor both static and dynamic clients to achieve variousobjectives such as seamless mobility fairness amongclients load balancing among APs and interference miti-gation ChA mechanisms address the problem of channelassignment to APs in order to minimize the interferencelevel experienced by APs and clients [27]ndash[29]

A Contributions

At a high level this paper studies SDWLANs from twoperspectives (i) SDWLAN architectures and (ii) the two maincentral network control mechanisms AsC and ChA that areemployed to control network operation More precisely thecontributions of this paper are as follows

ndash We first present an overview of software-defined net-working and the standard protocols used for networkmonitoring management and programming Next wereview SDWLAN architectures and reveal their main ob-jective components interconnection protocols and APIsprovided Based on their main contribution we classifyarchitectures into the following categories observabil-ity and configurability programmability virtualizationscalability and traffic shaping Then we compare these

architectures and highlight future research directionsconsidering the current and upcoming requirements ofSDWLANs

ndash We review the AsC mechanisms proposed for SD-WLANs We first classify these mechanisms into twogroups (i) seamless handoff mechanisms for reducingthe overhead and delay of client handoff and (ii) clientsteering mechanisms to optimize parameters such as theairtime allocated to clients Client steering approachesin turn are classified into two groups (i) centrally-generated hints where the controller relies on the globalnetwork view to generate hints for the association ofclients and (ii) centrally-made decisions where thecontroller makes the association decisions and enforcesthe clients to apply them In addition to discussing theimpact of architecture on supporting seamless handoffwe compare the reviewed mechanisms in terms of op-timization scope traffic awareness and the ability torecognize the QoS requirements of clients By discussingthe capabilities and weaknesses of AsC mechanismswe present research directions towards designing AsCsolutions for heterogeneous and dynamic SDWLANswith diverse QoS requirements We also identify howthe potential features of SDWLAN architectures can beemployed to design more efficient AsC mechanisms

ndash We review the ChA mechanisms proposed for SD-WLANs We classify these mechanisms into two groupstraffic-aware and traffic-agnostic based on whether theyincorporate APsrsquo and clientsrsquo traffic into the decisionmaking process We specifically focus on how ChAmechanisms define and measure their interference met-rics as well as the approaches employed for modelingand solving the formulated problems Furthermore wecompare these mechanisms in terms of factors like client-awareness dynamicity and the unique features of dif-ferent 80211 standards We propose research directionstowards designing mechanisms that address the demandsof emerging applications We also identify how thepotentials of SDWLANs can be employed to design moreefficient ChA mechanisms

When studying AsC and ChA mechanisms we reformulatethe problems using a consistent set of notations to ease theunderstanding and comparison of these mechanisms Table Ipresents the key acronyms and notations used in this paperFigure 2 shows the high-level organization of the paper

II MOTIVATION

In this section we overview the major application domainsand requirements of WLANs the challenges of wireless com-munication and the importance of centralization in addressingthe requirements of current and emerging applications

A Applications and Requirements

WLANs are being used in a variety of applications withvarious performance requirements WLANs installed in enter-prises campuses and public areas usually serve a small tomedium number of stationary and mobile users For voice

Section II Motivation

Section III Architectures

Section III-A Principles of Software-Defined Networking

Section III-B Standard Protocols

Section III-C SDWLAN Architectures

Section III-C Architectures Lessons Learned Comparison and Open Problems

Section IV Centralized Association Control (AsC)

Section IV-A Seamless Handoff

Section IV-B Client Steering

Section IV-C Association Control Lessons Learned Comparison and Open Problems

Section V Centralized Channel Assignment (ChA)

Section V-A Traffic-Agnostic Channel Assignment

Section V-A Traffic-Aware Channel Assignment

Section V-C Channel Assignment Lessons Learned Comparison and Open Problems

Section II-A Applications and Requirements

Section II-B Challenges

Section II-C Architectures and Control Mechanisms

Fig 2 High-level organization of the paper

TABLE IKEY ACRONYMS AND NOTATIONS

Acronym DefinitionAsC Association ControlAP Access PointAPI Application Programming InterfaceBSSID Basic Service Set ID (APrsquos MAC address)ChA Channel AssignmentCCA Clear Channel AssessmentDCF Distributed Coordination FunctionISP Internet Service ProviderLVAP Light Virtual Access PointMAC Medium Access ControlNFV Network Function VirtualizationNIC Network Interface CardPHY PhysicalRSSI Received Signal Strength IndexSDN Software-Defined NetworkSDWLAN Software-Defined WLANSINR Signal to Interference-and-Noise RatioVAP Virtual Access PointWLAN Wireless Local Area NetworkNotation DefinitionC = c1 ci Set of clientsAP = AP1 APi Set of access pointsCH = ch1 chi Set of channels

and video streams in addition to high throughput providingan uninterrupted service for mobile users is an importantperformance metric [30] For home WLANs achieving theseservice requirements is more challenging due to the adjacencyof multiple networks controlled individually [31] [32]

Industrial WLANs usually serve a larger number of nodesdeployed for monitoring and control purposes In additionwhile most of the nodes are static mobility is characterizedby the speed and movement scale of robot arms rotating partsand mobile robots [33] In these applications the networkinfrastructure must satisfy a set of unique requirements [16][34] [35] as follows First many of these applications requiredeterministic packet delivery delay This is particularly impor-tant when sensors and actuators communicate with controllers

because the delay of the control loop directly affects controlperformance Second a packet delivery reliability of 99 orhigher must be guaranteed to ensure the robust operation ofcontrol and monitoring systems In addition to sensors andactuators industrial wireless networks also include high-ratedevices such as cameras [36] For example it is importantto ensure that the longer transmission duration and higherpriority of video packets do not affect the timeliness ofshort control packets Similar to industrial applications othermission-critical systems such as medical monitoring mustsatisfy a similar set of requirements [17]

Besides user devices such as smartphones more energy-constraint devices like sensors and actuators are being usednowadays in various types of applications [1] [17] [37]ndash[39] For example wireless sensors and actuators used inindustrial environments specifically those installed in hightemperature and hard-to-access areas are usually battery-powered Furthermore a greater number of low-power IoTdevices and wearables are being used for applications suchas medical monitoring Therefore energy is an importantrequirement of current and emerging applications of WLANs

B Challenges

Wireless networks introduce a new set of network controlchallenges compared to wired networks The broadcast natureof the wireless medium causes interference resulting in linkquality and throughput variations For example an AP mayunder-utilize its spectrum and hardware capacity due to thehigh interference level caused by neighboring APs or clients[40]ndash[42] Interference level depends on dynamic factors suchas mobility traffic rate number of devices and environmentalproperties (eg path loss multipath effect obstacles etc)[11] [43]ndash[45] The high path loss of wireless signals andthe inability of wireless transceivers to detect collisions bringabout new challenges such as hidden and exposed terminalproblems [46]

From the network control point of view channel assignmentis a widely used mechanism to cope with interference enhancethroughput and reduce channel access delay Channel assign-ment refers to the process of assigning RF channels to APsbased on various factors such as the interference relationshipbetween APs and clients and the traffic demand of clients Thedynamics of wireless channel the variability of user trafficand variations in transmission power make channel assignmenta challenging process This process is further complicated bythe introduction of channel bonding supported by 80211nand 80211ac standards Channel bonding allows a device tocombine multiple channels in order to transmit at a higherrate over a wider channel However the performance benefitof this technique depends on several factors and requirescollaboration among APs [47] [48] For example using alarger bandwidth might increase contention with nearby APsand result in a longer channel access delay

In addition to affecting interference level and the load ofAPs mobility results in connection interruption because ofthe delay of re-association process [49] [50] Although mostof the large WLAN deployments employ a centralized user

authentication scheme the delay of association process is notcompletely eliminated In particular when a client associateswith a new AP the AP needs to establish the data structurespertaining to the new client However this process might resultin communication interruption

To increase the throughput of WLANs extend coverage andfacilitate reliable handoff network densification is employedin various types of environments such as enterprises andcampuses With this approach the coverage range of the APsis reduced and the density of APs deployed per square meter isincreased [14] Despite the potential benefits of this approachthe network dynamics caused by interference and mobilityare exacerbated by densification For example densificationincreases the number of re-associations as a client moves Inaddition due to the limited number of channels available inthe 24 and 5GHz bands channel assignment becomes a morechallenging process

From the energy point of view the 80211 standards providetwo mechanisms power saving mode (PSM) and automaticpower saving delivery (APSD) However the efficiency ofthese algorithms reduces as network density increases [51][52] For example the concurrent wake-up of multiple clientsresults in collision and higher energy consumption [53] Fur-thermore energy is wasted when a node wakes up during abad channel condition

C Architectures and Control Mechanisms

Central network control plays an important role to sat-isfy the requirements of existing and emerging applicationsFirst architectures are required to provide an infrastructurefor centralized network control Second centralized controlmechanisms are necessary to make control decisions basedon network-wide dynamics and requirements For exampleutilizing centralized network control facilitates addressing thefollowing challenges To offer seamless handoff we can elimi-nate the need for re-association by maintaining the associationinformation of clients at a central controller to enhancethroughput the controller can steer the association point ofclients to balance the load of APs the performance of channelassignment can be improved by utilizing centralization becausechannel assignment is in essence a graph coloring problem Inaddition to client steering seamless handoff and channel allo-cation which significantly affect throughput delay and energyconsumption of clients centralization enhances the adoptionand performance of other techniques such as scheduling tofurther improve the aforementioned metrics [17] [52]ndash[54]

III ARCHITECTURES

In this section we overview the existing SDWLAN ar-chitectures Before presenting these architectures we firstoverview the basic concepts of software-defined networkingand the standard protocols used for north and south-boundcommunication

A Principles of Software-Defined Networking

A software-defined network (SDN) decouples control mech-anisms from data forwarding paths In fact a SDN has two

planes a (i) data plane such as switches and APs and a (ii)control plane which is a controller that runs an operatingsystem (eg NOX [55] Floodlight [56]) and mechanismsthat control the network operation (eg AsC ChA) Thecontroller1 communicates with data plane equipment through asouth-bound protocol (eg OpenFlow [59] SNMP [60]) andexposes a set of north-bound APIs (eg REST) through whichnetwork control mechanisms are developed In fact the globalnetwork view of the controller and the reprogrammability ofdata plane significantly simplifies the design and deploymentof network control mechanisms Network control mechanismsare also referred to as network applications as they areactually applications running on the network operating system

SDN is an enabler of network virtualization aka networkslicing which refers to the abstraction slicing and sharingof network resources [61] For example an abstraction layer(eg FlowVisor [62]) is used to provide applications withan isolated view of resources Network virtualization providescontrol logic isolation as each application can see and controlonly the slice presented to it Therefore when a SDWLANis shared by multiple operators network slicing enables theenforcement of multiple access policies to the users associatedwith various network operators In addition virtual networkscorresponding to various services can be established to sup-port differentiated services and achieve higher QoS controlover resources Virtualization also expedites the design anddevelopment of novel wireless technologies For example oneor multiple experiments can be run on slices of a wirelessnetwork while the production network is in operation Threelevels of slicing are applicable to a SDWLAN (i) Spectrum(aka link virtualization) requires frequency time or spacemultiplexing (ii) Infrastructure the slicing of physical devicessuch as APs and switches (iii) Network the slicing of thenetwork infrastructure

Given the simplified configuration of data forwarding pathsSDN promotes the use of network function virtualization(NFV) Instead of using dedicated hardware NFV relieson the implementation of services using general computingplatforms NFV is the virtualization of network functions suchas domain name service (DNS) firewall intrusion detectionload balancing and network address translation (NAT) To thisend a network service is broken into a set of functions thatare executed on general purpose hardware These functionscan then be introduced to or moved between network deviceswhenever and wherever required As we will show latersoftware-defined radio (SDR) and virtual APs (VAPs) are thetwo most notable types of NFV in SDWLAN to offload theprocessing of APs into general computing platforms

B Standard Protocols

In SDN architectures the controller communicates with thedata plane through a south-bound protocol and provides north-bound APIs for application developers To avoid exposing thecomplexities of south-bound APIs the north-bound APIs areusually REST-based [63] [64] and network engineers use

1To achieve reliability a logical controller may represent multiple physicalcontrollers [57] [58]

language abstractions such as Frenetic [65] Pyretic [66] andNetCore [67] for application development [68] For exampleFloodLight [56] and OpenDaylight [69] both support RESTnorth-bound APIs In the rest of this section we review thewidely-adopted standard south-bound protocols

Simple Network Management Protocol (SNMP) [60][70]ndash[72] This is an application layer protocol used to monitorand configure network elements SNMP exposes managementdata in the form of variables on the managed systems Thenature of the communication is fetch-store either the managerfetches resources or stores the value of an object on the agentSNMP cannot directly implement actions For example torestart an agent the manager cannot send a command to rebootthe agent instead it sets the value of the reboot counter whichindicates the seconds until next reboot SNMP has been themost widely used management protocol in production net-works due to its simplistic nature and ease of usage AlthoughSNMPv3 introduces significant security enhancements SNMPhas been mostly used for monitoring rather than configuration[73] Furthermore due to the sequential execution of com-mands SNMP introduces high network overhead and mayresult in network failure in large-scale deployments [74] [75]

Netconf [76] Netconf uses remote procedure call (RPC) toexchange messages and apply configuration through CreateRetrieve Update and Delete (CRUD) operations In additionit enables the manager to define what action (eg continuestop roll-back) should be taken if a command fails Netconfhas been mostly used for configuration and the assumptionis that another protocol such as SNMP is used for monitoringpurposes [73] YANG is a tree-structured modelling languageused by Netconf to describe the management information ofnetwork elements Since YANG enables the definition of datamodels Netconf has been widely used by industry howevervendors may not support a consistent set of data models

OpenFlow [59] This protocol assumes that data planeelements are simple packet forwarding devices and theiroperation is determined by the forwarding rules received fromthe controller Whenever a new flow enters a switch theforwarding table of the switch is queried to find a matchingforwarding rule If no forwarding rule is found either thepacket is dropped or an inquiry is sent to the controller to re-trieve the required action OpenFlow 11 replaces actions withinstructions which enables packet modification and updatesactions The possibility of connecting to multiple controllershas been introduced in OpenFlow 12 OpenFlow 13 enablesper-flow rate control

Control And Provisioning of Wireless Access Points(CAPWAP) [77] This protocol enables a controller to es-tablish DTLS [78] connections with APs to exchange dataand control messages Data messages encapsulate wirelessframes and control messages are used for monitoring andcontrol purposes In addition to centralizing authentication andnetwork management CAPWAP defines two MAC operationmodes [79] local MAC (LM) and split-MAC (SM) Using theLM mode most of 80211 MAC operations are implementedin APs and the role of controller is almost negligible Usingthe SM mode the real-time operations of 80211 MAC are runon APs and the rest of operations such as the generation of

beacons and probe responses are handled by the controllerCPE WAN Management Protocol (CWMP) [80] (aka

TR-069) Published by Broadband Forum CWMP is a text-based protocol for communication between Auto Configura-tion Servers (ACS) and Customer Premise Equipment (CPE)such as modems APs and VoIP phones The primary featuresof CWMP are secure auto-configuration dynamic service pro-visioning softwarefirmware image management status andperformance monitoring and diagnostics One of the maincapabilities of CWMP is to enable the devices behind NAT tosecurely discover ACSs and establish connection In additionan ACS can request a session start from a CPE The commands(eg get set download upload etc) run over a SOAPHTTPSapplication layer protocol In fact a CPE acts as a client andthe ACS is the HTTP server Configuration and monitoringof CPEs is performed by setting and getting the deviceparameters which are defined as a hierarchical structureCWMP is considered to be an important management protocolin M2M architectures [81] [82] However this protocol doesnot address client mobility in WLANs

As OpenFlow has been primarily designed to configure theflow table of switches an additional protocol such as SNMPNetconf or a proprietary protocol should be used to enablethe control of wireless devices as we will see in Section III-CTherefore both SNMP and Netconf are widely supported bySDN controllers (eg OpenDaylight ONOS [83]) We providefurther discussion about protocols in Section III-D1

There are other protocols in addition to those mentionedin this section used for south-bound communication Forexample although REST APIs are mostly used for north-bound interactions REST is also employed for south-boundcommunication by controllers such as Ryu [84]

C SDWLAN Architectures

In this section we review the architectures proposed forSDWLANs We categorize these architectures based on themain contributed feature into five categories observabilityand configurability programmability virtualization scalabil-ity and traffic shaping However it should be noted that someof these architectures present multiple features as we willshow in Figure 7 and Table II Please note that the AsC andChA mechanisms proposed to benefit from these architectureswill be studied in Section IV and V respectively

1) Observability and Configurability The architectures ofthis section provide the means for central monitoring andconfiguration however they cannot be used to develop newnetwork control mechanisms

DenseAP This architecture [12] addresses the challengesof densely-deployed WLANs in terms of AsC and ChA Thetwo main components of this architecture are DenseAP APsand DenseAP Controller DenseAP APs are off-the-shelf PCsrunning Windows operating system Each AP runs a DenseAPdaemon which is a user-level service responsible for managingAP functionality This daemon monitors NIC and reports toDenseAP Controller periodically A report includes the listof associated clients sample RSSI values traffic pattern ofclients channel condition and a list of new clients requesting

to join the network The DenseAP Controller enables the userto tune parameters makes control decisions and informs thedaemon to apply the configuration We will review the AsCmechanism of DenseAP in Section IV-B2

Cisco Unified Wireless Network (CUWN) CUWN [85]proposes an architecture and a set of configurable servicesincluding seamless mobility control security and QoS provi-sioning CUWN uses either Lightweight Access Point Protocol(LWAPP) [86] or CAPWAP [77] as its south-bound protocolAs supported by CAPWAP CUWN provides two MAC modeslocal MAC (LM) and split-MAC (SM) The Radio ResourceManagement (RRM) [87] of CUWN provides several networkcontrol services For example RRM establishes RF groups anddetermines a leader controller for each group To this end eachAP periodically transmits Neighbor Discovery Protocol (NDP)messages on all channels NDP messages contain informationsuch as the number of APs and clients All APs forward thereceived NDP messages to the controller which establishesa network map groups APs into RF Groups and choosesa leader controller for each group We will review the AsCand ChA mechanisms of CUWN in Section IV-A and V-Brespectively

2) Programmability In addition to central monitoringthese architectures expose north-bound APIs that enable theimplementation of control mechanisms as applications runningon a controller We overview these architectures as follows

DIRAC This architecture [88] proposes a proprietary south-bound protocol using two main components a router core(RC) running on a PC and router agents (RA) running onAPs

RAs receive messages from RC and convey them to theNICrsquos device driver There are three types of interactionsbetween the RC and RAs events statistics and actions Sam-ple events are association re-association and authenticationwhich are reported by RAs to the RC RAs periodically reportstatistics such as packet loss rate and SNR The RC enforcesits policy by sending actions to RAs For example the RCmay use set_retransmissions() to limit the numberof retransmissions The control plane of the RC has twomain components (i) management mechanisms which are thenetwork applications and (ii) a control engine which includescomponents to simplify network application development

Trantor This architecture [89] is an improved versionof DenseAP [12] and its south-bound protocol extends theexchange of control messages with clients (in addition to APs)These messages include packet loss estimation the RSSI ofpackets received from APs channel utilization and neighbor-hood size Trantor also supports rdquoactive monitoringrdquo throughwhich clients and APs are directed to exchange packets tomeasure metrics that are used by network control mechanismsIn addition Trantor provides a set of APIs used to control theassociation channel transmission rate and transmission powerof devices For example using associate (APi) a clientis instructed to associate with APi Unfortunately no networkcontrol mechanism has been proposed to benefit from theseprimitives

Dyson This architecture [90] (which is an extension ofTrantor) enables both clients and APs to communicate with

the controller Clients are categorized into two groups Dyson-aware clients and legacy clients Only Dyson-aware clientscan communicate with the controller The measurement APIscan be used to collect information such as the sum of theRSSI values (total RSSI) of all received packets during ameasurement window the number of transmitted packets perPHY rate the number of transmission failures and the channelairtime utilization

Dyson establishes a network map based on the informationcollected from APs and Dyson-aware clients The networkmap provides the following components (i) node locationsreflects the location of APs (which are fixed) and clients (using[91]) (ii) connectivity information a directed graph that showsfrom which APsclients an APclient can receive packets (iii)airtime utilization reflects channel utilization in the vicinityof each node (iv) historical measurement a database to storemeasurements

Dysonrsquos APIs include commands to configure the channeltransmission power rate and CCA of APs and clients The APIset also includes commands to manage AP-client associationsA reduced set of control APIs is used to support legacy clientsThis set enables the controller to control the association pointand operating channel of these clients

AEligtherFlow This architecture [92] simplifies data planeprogrammability by extending OpenFlow based on its formalspecifications The control interfaces are categorized accordingto the OpenFlow specification as follows (i) Capabilities in-terfaces through which a controller inquires the capabilities ofan APrsquos radio interfaces These capabilities include the numberof channels transmission power levels etc (ii) Configurationthese interfaces enable the configuration of physical ports(eg channel transmission power) and logical wireless ports(eg SSID BSSID) (iii) Events interfaces for event report-ing (eg probe authentication association) to a controller(iv) Statistics these interfaces enable a controller to querywireless related statistics AEligtherFlow has been implementedin CPqD SoftSwitch [93] which is installed on OpenWRT[94] The performance and capability of this architecture hasbeen evaluated using a predictive AsC mechanism which wewill explain in Section IV

COAP Coordination framework for Open APs (COAP)[31] is a cloud-based architecture for residential deploymentsIn addition to using OpenFlow this architecture proposes anopen API to provide a cloud-based central control of homeAPs COAP improves client QoS by running central controlalgorithms as well as enabling cooperation among APs ACOAP AP implements three modules (i) APConfigManager(ii) DiagnosticStatsReporter and (iii) BasicStatsReporter AP-ConfigManager is used for AP configuration and the othertwo modules provide diagnostic and basic statistics for thecloud-based COAP controller The COAP controller has threemodules (i) StatsManager (ii) COAPManager and (iii) Con-figManager which are implemented within Floodlight Thecontroller collects the required information using the Stats-Manager module and performs control configurations throughConfigManager The COAPManager module allows networkadministrators to implement new strategies and policies

ResFi [32] enables information collection and sharing in

Virtual Machine 1

Policy Plane

Virtual Wired Interface

Wired NIC Wireless NIC

Virtual Wireless Interface

VAP Manager

hellip Virtual Machine n

Policy Plane

Fig 3 The architecture of an AP used by AP Aggregate [101]

residential networks ResFi Agent is a user-space programthat runs on APs and provides IP connectivity over the wiredbackhaul network This agent instructs the APs to exchangemessages through which APs discover each otherrsquos IP addressAfter the discovery phase each AP establishes a secure point-to-point connection to the agents of discovered APs usingthe wired backbone Each agent can then share informationwith other APs or a remote controller For example anAP can subscribe to the publish socket of its neighbors toreceive their updates Using ResFi Agents on an AP doesnot require any kernel or driver modification as it simplyrelies on the interfaces provided by hostapd [95] The APIsprovided enables both distributed and centralized managementof residential WLANs

3) Virtualization In this section we review the architec-tures that offer virtualization to support seamless mobilityslicing and NFV In particular we discuss AP virtualizationwhich refers to the mapping of multiple virtual APs to aphysical AP (eg Odin [49] CloudMAC [50]) or mapping avirtual AP to multiple physical APs (eg SplitAP [96] BIGAP[97])

WVT Wireless Virtualization Testbed (WVT) [98] relieson AP virtualization to run isolated experiments on a testbedCorresponding to each experiment each physical AP runsa Node Agent where each Node Agent is controlled bya Node Handler running on the controller Node Handlersprovide the user with APIs for monitoring and controllingNode Agents In addition to Node Agents the APs run a NodeOverseer which is responsible for activating the Node Agentsin a round robin fashion using TDMA scheduling In factat each slot boundary the Node Overseer records the state(eg SSID channel) of the currently running Node Agentand restores the state of the next Node Agent Node Overseerscan be configured through the Master Overseer running on thecontroller This architecture uses an NTP time synchronizationserver which is required for TDMA scheduling

OpenRoads This architecture [99] [100] adds OpenFlowto APs In addition due to the limitations of OpenFlowOpenRoads uses SNMP for configuring the APsrsquo parame-ters Resource slicing is provided through using FlowVisorand SNMPVisor the former slices data path and the latterslices configuration commands through sending them to theappropriate data path element

AP Aggregate [101] presents a flexible architecture thatenables the aggregation of multiple virtual APs (VAPs) insidean AP as well as slicing resources on APs VAP aggregationis also used for interference reduction by turning off lightly-loaded APs Figure 3 shows the architecture of a physical APThe VAP Manager is responsible for creating and destroying

Controller

Odin Agent

Internet

Data Plane

Odin Control Plane

OpenFlow Control Plane

LVAPLVAP LVAPLVAP

Odin OpenFlow

OpenFlow Agent

Distribution Network(OpenFlow supported)

Applications

Fig 4 Odin [102] introduces the concept of lightweight virtual APs (LVAP)to reduce the overhead of handoffs

VAPs based on the commands received from the controllerThe Policy Plane is a set of settings that are applicable tothe VAP For example these policies may specify the QoSand firewall settings of the VAP A Virtual Wireless Interfacehas a SSID MAC and IP address Multiple Virtual WirelessInterfaces are multiplexed into the Wireless NIC using TDMABased on its collected information a controller moves VAPsbetween APs This architecture employs layer-2 tunnelingtechniques when VAPs are moved between subnets

Odin Figure 4 shows the architecture of Odin [49] [102]ndash[104] This architecture is composed of the following compo-nents (i) Odin Controller maintains a global network view in-cluding the status of APs clients and OpenFlow switches (ii)Odin Agents run on APs and enable communication througha proprietary south-bound protocol Time-critical operations(eg ACK transmission) are performed by APs and non-time-critical operations (eg association) are handled by the con-troller This operation results in a split-MAC protocol whichis inspired by CAPWAP (iii) Applications implemented onthe controller Applications may use the information providedby Odin Agents OpenFlow and SNMP The APIs support theimplementation of various control mechanisms such as AsCload balancing and hidden-terminal handling

Odin introduces the concept of light virtual AP (LVAP) toseparate clientsrsquo association states from physical APs therebyreducing the overhead of handoffs An LVAP is characterizedby the following four-tuple

ndash client IP addressndash a unique virtual basic service set ID (BSSID)ndash one or more service set ID (SSID)ndash a set of OpenFlow rules

For each client an LVAP resides in an AP (hosted byOdin Agent) to represent client-AP association When thecontroller decides to associate a client with a new AP it cansimply transfer the LVAP of that client to the new AP Froma programmerrsquos point of view multiple clients are connectedto different ports (the LVAPs) of a physical AP LVAPs alsosimplify client authentication through adding a session keyto the clientrsquos LVAP We will discuss about Odinrsquos AsCmechanism in Section IV-A

SplitAP [96] proposes an architecture to support networkvirtualization and manage clientsrsquo share of airtime especiallyfor uplink traffic Each AP runs a SplitAP controller which

ControllerController

OpenFlow Switch

Controller

Virtual Access Point (VAP)Virtual Access Point(VAP)

Virtual Machine Host

OpenFlow Controller

Policies

Data Plane

Control Plane

Virtual Access Point (VAP)

Virtual WLAN Card

InternetSwitching Table

Applications

Virtual WLAN Card

Applications

Virtual WLAN Card

Fig 5 CloudMAC [26] reduces the complexity of APs through runningVAPs on a cloud computing infrastructure

is responsible for VAP management and computing the uplinktraffic of each slice By relying on the VAP concept eachphysical AP broadcasts the beacons of independent virtualnetworks For example when three different ISPs utilize aninfrastructure each AP emulates three VAPs This architecturerequires client modification to enforce controller commandsSpecifically each client runs a client controller that adjuststraffic based on the commands received from the associatedAP When a SplitAP controller realizes that the usage of aparticular slice is higher than its threshold the AP broadcasts anew maximum uplink airtime utilization that can be consumedby clients in that slice All the APs are connected to a sharedbackhaul through which they receive channel allocation com-mands from a controller

VAN Virtualized Access Network (VAN) [105] proposesan architecture for central control of residential networks Thearchitecture is composed of the following components (i)residential APs (ii) a controller which is part of the ISPand (iii) content provider servers that are part of the contentprovider network (eg Netflix) The controller provides aset of APIs through which content providers control homenetworks These APIs are suitable for different traffic classessuch as video streaming and file transfer For example whena content provider receives the request of a user the con-tent provider communicates with the controller and reservesresources for the requested data flow

In this architecture IP address assignment and authentica-tion are performed centrally per user device therefore it ispossible to transfer the association of a user device betweenAPs More importantly this architecture enables the sharing ofAPsrsquo bandwidth between neighbors Given a flow bandwidthrequirement and the neighborhood of each client the controllerassociates clients to APs by employing a heuristic algorithm

CloudMAC Figure 5 illustrates the CloudMAC [26] [50]architecture CloudMAC decomposes AP operations into twomodules and employs OpenFlow switching tables to managethe communication between these modules The main compo-nents of CloudMAC are (i) APs (ii) VAPs (iii) an OpenFlowswitch and (iv) an OpenFlow controller APs only forwardMAC frames All other MAC functionalities are performed byVAPs that reside in a cloud computing infrastructure VAPs areoperating system instances on a hypervisor such as VSphereEach VAP can include multiple virtual WLAN cards where

each card is implemented as a driver in a VAP thereforestandard software tools can utilize these cards Virtual ma-chines are connected to physical APs by an OpenFlow-baseddistribution network The generation and processing of MACframes are performed by VAPs A physical AP which is aslim AP sends and receives raw MAC frames from clientsand handles time-sensitive operations (eg ACK generationand re-transmission)

APs and VAPs are connected by layer-2 tunnels [106] andthe OpenFlow switch The forwarding table of switches areused to decide how packets must be routed between physicalAPs and VAPs This mechanism simplifies the implementationof control mechanisms such as AsC For example whileOdin requires LVAP migration to handle mobility CloudMACsimply reconfigures the switching tables to change the dataforwarding path

When a control command is generated by a network appli-cation the command is processed by VAP and a configurationpacket is forwarded toward the OpenFlow switch The Open-Flow switch forwards the packet to the OpenFlow controller tocheck the legitimacy of the configuration command accordingto user-defined policies If the command is legitimate theOpenFlow controller forwards the packet to the AP

To study the performance of CloudMAC the round trip time(RTT) is measured between a client and a VAP Due to theprocessing overhead of the OpenFlow switch and the delayoverhead added by the tunnels the average RTT is increasedfrom 179 ms to 228 ms compared to a standard WLANHowever time-critical MAC frames (eg association responseframes) are delivered fast enough to satisfy the timelinessrequirements For large TCP packets throughput is decreasedby almost 85 compared to a standard WLAN This perfor-mance degradation is because of the tunnels implemented inuser space which requires context switching

EmPOWER EmPOWER [107] relies on the concept ofLVAP (proposed by Odin) in order to decrease the overhead ofclient mobility The controller runs Floodlight as the operatingsystem and FlowVisor to enable virtualization In this archi-tecture each AP is equipped with an Energino [108] add-onwhich is an open toolkit for energy monitoring This add-onprovides REST-based APIs that enable network administratorsto turn on and off the APs

EmPOWER2 An improved version of EmPOWER hasbeen proposed in [24] [109] We refer to this architec-ture as EmPOWER2 This architecture provides a full andopen set of Python-based APIs by introducing four keyabstractions (i) LVAPs a per-client VAP similar to Odin(ii) Resource pool each resource block is identified as((frequency bandwidth) time) For example an AP work-ing on channel 36 with bandwidth 40MHz is representedas ((36 40)infin) Similar notation is used to represent the80211 standard supported by LVAPs This representationprovides a mean for LVAP to AP mapping (iii) Channelquality and interference map provides network programmerswith a global network view in terms of the channel qualitybetween APs and LVAPs This view enables the programmersto allocate resources in an efficient manner (v) Port specifiesthe configuration of an AP-client link in terms of power

modulation and MIMO configuration We will study its AsCand ChA mechanisms in Section IV-B2 and V-A respectively

Sd-wlan The main idea of Sd-wlan [110] [111] is theimplementation of most MAC functionalities at a controllerwhich is essentially similar to Odin [102] To this end anextended version of OpenFlow is proposed as the south-bound protocol Using this protocol the controller can instructthe APs whether they should send ACK packets or notWhen a client needs to change its point of association thecontroller must install new rules on the new AP The controlleralso updates OpenFlow switches to direct the traffic beingexchanged between the controller and APs

BeHop The main components of BeHop [112] are (i) Be-Hop APs (ii) a BeHop collector and (iii) a BeHop controllerBeHop collector is responsible for monitoring and collectinginformation from the network The controller includes networkcontrol mechanisms in addition to processing and respondingto probe authentication and association requests The collectormay be implemented internally (inside the controller) orexternally as a separate component

A proprietary south-bound protocol has been implementedby running monitoring agents on APs This protocol is usedto collect and forward statistics (eg channel utilizationSNR RSSI and PHY rate) to the collector The controllercommunicates with the collector through a remote procedurecall (RPC) interface to update its status about BeHop APs

BeHop APs create a VAP per client and each AP maintainsa client table Each entry of the client table includes clientcontrol information such as client-VAP mapping and clientdata rate VAPs can be added to or removed from APsusing the south-bound protocol implemented Each BeHopAP operates as an OpenFlow switch and exposes APIs forcontrolling channels power and association APs forwardcontrol traffic (eg probe association) to the controller forfurther processing

RCWLAN This architecture [113] addresses the coexis-tence of multiple virtual networks (slices) For each virtualnetwork a VAP is created on the physical APs where eachVAP has its own virtual machine and MAC queue The set ofresources allocated to each slice is controlled by configuringthe MAC parameters of the VAPs The extension of RCWLANis vBS which we explain next

vBS [114] supports service-based wireless resource reser-vation which is similar to the concept of flow-based virtu-alization [62] Specifically by reserving resources for time-critical services (eg VoIP) vBS provides QoS guaranteeseven in the presence of background traffic In contrast withSplitAP which supports virtualization through broadcastingdifferent ESSIDs this architecture relies on the concept ofvirtual Base Station (vBS) A vBS is a virtual multichannelAP that uses the resources of multiple physical APs All ofthe APs share the same BSSID (MAC address) and ESSIDtherefore AP selection and handoff are completely handled bythe controller When a network administrator registers a newservice the controller creates a vBS for that service Whena client joins the network for the first time the controllerassociates the client with an AP of the common vBS Thecommon vBS provides a service for non-prioritized traffic

Controller

StatefullFirewall (FW)

Deep PacketInspection

(DPI)Agent

MiddleBoxDriver

Radio Driver

Radio Agent

Participatory Interface

Applications

Data Plane

RESTful API commands

WDTXWDTXWDTXLVAP

WDTXWDTXWDTX

LVAPvMB

WDTXWDTXWDTX

Internet

MiddleBox (MB)

Fig 6 OpenSDWN [25] enhances data plane programmability by introducingvirtual middle boxes (vMBs)

When the client initiates a service the controller performs aclient handoff to the vBS of that service The association statusof the client is transferred to the new AP and the OpenFlowswitch is updated

BIGAP The main goal of this architecture [97] is to supportseamless mobility by mapping physical APs to a single VAPBIGAP assumes that each AP has two wireless interfacesone for serving as AP and the second one for statisticscollection which works in promiscuous mode to periodicallyoverhear packets on all channels The collected informationis used by the controller to provide a fast AsC mechanismwhich will be explained in Section IV-A In order to reducethe complexities of the handoff process BIGAP requires allof the APs to share the same BSSID However to avoidcollisions BIGAP requires each AP to select a channel thatis not being used by its two-hop neighborhood Associationof a client with another AP is achieved through a channelswitch command and transferring clientrsquos status to the newAP BIGAP provides a rich south-bound protocol for statisticscollection and association control The AP-related APIs areimplemented by modifying hostapd [95] and the APIs areavailable as RPC by relying on ZeroRPC [115] and ZeroMQ[116]

OpenSDWN Figure 6 illustrates this architecture [25] Sim-ilar to Odin OpenSDWN provides per-client VAP howeverthe new concepts of OpenSDWN are virtual middle boxes(vMB) and wireless datapath transmission (WDTX) rules AvMB encapsulates a clientrsquos middle box state Due to the smallsize of vMBs their transfer across network does not introduceany significant overhead Each vMB includes the followinginformation (i) the list of tunable parameters of the client (ii)the states of clientrsquos active connections (iii) the packet-basedstatistics and (iv) an event list that defines the behavior of theclient

The controller implements a MB driver and Radio Driver tocommunicate with the agents of MBs and APs respectivelyThrough the Radio Driver the controller manages LVAPs andcollects the status of clients Furthermore the controller can

getset per-flow transmission rules through its radio interfaceTo this end OpenFlow rules are combined with WDTX ruleswithin APs WDTX rules provide per-flow service differen-tiation using a newly defined action that is compatible withOpenFlow protocol The controller manages vMBs gathersthe statistics of vMBs and receives the required events fromdifferent MBs The latter in particular is supported througha publishsubscribe interface to ensure controller update whenspecific events occur

A middlebox in OpenSDWN can have two types of in-terfaces (i) a stateful firewall (FW) agent and (ii) a deeppacket inspection (DPI) agent The FW agent keeps trackof all traffic passing through the MB in both directions andprovides statistics per client and per flow Also the FW agentruns the WDTX rules defined by the controller for differentflows and frames The DPI agent is responsible for detectingevents such as denial of service (DoS) attacks The agentsof a MB generate and send events that are of the interest tothe controller For instance a significant change in the loadof an AP may trigger sending a report to the controller Thecontroller can submit a report list to a MB in order to obtainonly requested events

The OpenSDWN controller exposes a participatory interfacethat enables network applications to define flow-based andclient-based priorities This is achieved by providing RESTfulAPIs through which LVAPs WDTXs and vMBs are monitoredand controlled

Through empirical evaluations the authors showed the smalldelay overhead of vMB migration among MBs to supportmobility and they demonstrated the effectiveness of usingWDTX rules to perform service differentiation

4) Scalability Scalability depends on topology as wellas factors such as the overhead of south-bound protocolthe virtualization mechanisms employed and the traffic andmobility pattern of clients The scalability of SDWLANs isparticularly important due to the dynamic nature of thesenetworks Specifically an architecture with a high controlplane delay cannot be used for quick reactions to networkdynamics

AeroFlux This architecture [117] [118] highlights thatper-flow or per-packet control mechanisms require short andbounded control plane delay For example implementing rateand power control mechanisms in a controller may resultin an overloaded and high-latency control plane To addressthis concern AeroFlux employs a 2-tier control plane GlobalControl (GC) plane and Near-Sighted Control (NSC) planeGC handles non real-time tasks and operations that require aglobal network knowledge For example authentication large-scale mobility and load balancing are handled by the GCNSCs are located close to the APs to perform time-criticaloperations such as rate control and power adjustment perpacket For example in the case of video streaming AeroFluxcan configure APs to use lower rates and higher transmissionpower values for key frames compared to regular frames

CUWN This architecture [85] employs split-MAC as wellas a multi-tier controller topology to address scalability Basedon the RF locations of APs CUWN establishes RF groupsand determines a leader controller for each RF group All

APs connected to a controller belong to an RF group In thisway the controllers are aware of RF location of APs and theirinterference relationship

Odin and CloudMAC These architectures [49] [50] relyon the the split-MAC technique proposed by CAPWAP Inother words to mitigate the negative effect of controller-APcommunication time-critical operations are handled by APsand non time-critical operations are offloaded to the controllerIn addition to utilizing split-MAC during an associationCloudMAC prevents the overhead of moving VAPs betweenAPs through configuring the switches of the distribution net-work to forward the clientrsquos data to another AP

5) Traffic Shaping The architectures of this section offermore than the regular programmability used for configuringAPs and switches In fact these architectures extend data planeprogrammability by enabling traffic shaping which is used forpurposes such as scheduling and scalability

CENTAUR [119] can be added to SDWLAN architecturesto improve the operation of the data plane in terms of channelaccess and contention resolution When data traffic passesthrough a controller or programmable middleboxes data planeprogrammability can be supported without AP modificationTo this end CENTAUR proposes central packet schedulingmechanisms to avoid hidden-terminal transmissions and ex-ploit the exposed-terminal condition For example when thetransmission of two APs to their associated clients may causea hidden-terminal collision the controller carefully adjusts theinterval between packet transmissions to avoid the problem Asalient feature of CENTAUR is that it requires minor changesto APs and no client modification is required CENTAURimproves UDP and TCP throughput by around 46 and615 compared to DCF [120]

CloudMAC This architecture [26] is capable of implement-ing downlink scheduling by configuring OpenFlow switchesto use simple rate shaping or time division algorithms Forinstance during a time slot the switch may forward only thepackets of a specific physical AP while queuing the packetsof other APs Switching rules may be changed per time slotto provide a time division protocol

Ethanol Ethanol [121] is similar to its predecessors (eg[26] [49]) in terms of implementing slim APs and shiftingmost of the MAC functionalities to a controller HoweverEthanol argues that the original OpenFlow protocol cannot beused for QoS provisioning in wireless networks Accordinglyto complement OpenFlow the Ethanol architecture proposesa customized protocol named Ethanol protocol to providecontrol interface for wireless components and QoS controlTherefore an Ethanol AP provides two interfaces (i) Ethanolthrough an Ethanol Agent and (ii) OpenFlow The EthanolAgent provides APIs for QoS control on the APs Ethanol usesthe APIs provided by the Ethanol Agent to exploit HierarchicalToken Bucket (HTB) [122] scheduling in order to performper-flow programmability Specifically different queues aredefined on Ethanol APs When a flow arrives it is assignedto the proper queue based on its class of service (eg voicevideo)

Ethanol [121] provides the details of API implementationfollowing an object-oriented approach Entities are defined as

physical or virtual objects that could be configured or queriedEach entity has its own properties For example an AP isa physical entity that includes properties such as beaconinginterval A flow is a virtual entity that includes propertiessuch as a packet counter The controller communicates withentities through their getset interfaces Entities may alsoinclude events which trigger the controller to take properactions Ethanol API has been implemented on OpenWRT[94] by exploiting Pantou [122] a software package enablingOpenFlow on OpenWRT

D Architectures Learned Lessons Comparison and OpenProblems

Figure 7 highlights the main features and Table II summa-rizes the properties of reviewed architectures In this sectionwe summarize the learned lessons study the proposed featuresand identify future research directions and potential solutions

1) Programmability A reconfigurable architecture onlyenables the adjustment of parameters pertaining to a set of pre-defined network control mechanisms For example CUWNrsquosAPs are like regular APs with added CAPWAP supportthereby it is not possible to implement a new AsC mech-anism as CUWN employs a set of proprietary mechanismsfor network control In fact we can argue that CUWN andDenseAP introduce a management plane instead of a controlplane On the other hand a programmable architecture suchas OpenSDWN and Odin provides north-bound APIs throughwhich control mechanisms running on the network operatingsystem are developed

After the standardization of OpenFlow most SDWLANarchitectures have included this standard in their south-boundcommunication as Table II shows However this protocolhas been mainly designed to configure the flow tables ofswitches therefore it cannot be used for the configurationof wireless data plane equipment Furthermore since Open-Flow is very low-level too many implementation details areexposed to network programmers [24] Therefore we canobserve the introduction of proprietary protocols (eg OdinEthanol) and the extensions of OpenFlow (eg AeroFluxAEligtherFlow) However this results in interoperability issuesAnother shortcoming of OpenFlow is the lack of supportingtransactions which means the operation of a network deviceduring receiving updates is unpredictable This has a particularimplication on wireless network control mechanisms Forexample when an AsC mechanism sends updates to multipleAPs network performance may significantly drop during theupdate process due to the inconsistency of APsrsquo software De-spite the significant number of studies on the performance andimprovement of OpenFlow for wired networks [57] [123]ndash[125] there is a very limited study on the use of OpenFlow andother standard protocols (eg CWMP SNMP Netconf) forSDWLAN design [24] [92] [126] [127] In addition it is notclear what programmability features are required to developarchitectures that include recent high throughput standardsFor example supporting the 80211ad [128] standard requiresadditional interfaces for central coordination and training ofantennas

Our review shows that data plane programmability is sup-ported at four levels controller (eg CENTAUR) switches(eg CloudMAC) APs (eg Ethanol) and clients (egDyson) While controller-based approaches provide higherflexibility they pose scalability and processing challengesSpecifically the variations of data plane delay may inad-vertently affect the accuracy of implemented mechanismsFor example deterministic controller-AP delay is an essentialrequirement when packet scheduling is implemented at acontroller in order to avoid collisions caused by concurrentdownlink transmissions On the other hand AP-based ap-proaches require low and bounded control plane delay toensure prompt enforcement of the decisions made centrallyMeanwhile the programmability of switching equipment canbe employed to improve scalability and cope with overheadissues For example when a client is associated with a new APCloudMAC configures the switches of the distribution networkto forward the clientrsquos data to another AP thereby preventingthe overhead of moving VAPs between APs We believe thatinvestigating the pros and cons of these design approachesis necessary and mechanisms are required to integrate thebenefits of these approaches

AP programmability at layer-1 and layer-2 requires the useof software-defined radio (SDR) platforms such as OpenRadio[129] Atomix [130] or Sora [131] When integrated into aSDWLAN architecture using SDRs enables us to offer aricher set of programmability options to network applicationdevelopers Additionally integrating SDRs with SDWLANssimplifies the upgrade of communication standards when theSDR hardware is capable of supporting the new technologyFor example when the digital signal processor (DSP) usedon a SDR is fast enough to support new modulation andcoding schemes the controller can update APs to match thecapabilities of clients Furthermore when signal processingoperations are offloaded to a cloud platform advanced signalprocessing techniques may be applied to the incoming signalsto cope with interference While SDWLAN-SDR integrationis an ongoing research trend it is also important to study theimplications of SDR platforms (eg FPGA DSP general-purpose processor) on architecture in terms of factors suchas control plane delay and overhead

Although most architectures only rely on the informa-tion collected from APs some architectures such as Dysonand SplitAP also require client modification to collect in-formation from the clientsrsquo point of view Extending dataplane programmability to clients enables the use of moresophisticated control mechanisms For example this wouldenable the control plane to adjust clientsrsquo MAC parametersbased on network dynamics and user demands which alsosimplifies resource allocation and network slicing Howeverin addition to client modification which may not be desirablefor public access networks extending control plane to clientsintroduces traffic overhead Although newly proposed mecha-nisms such as FlashBack [132] enable low-overhead exchangeof control information with clients the integration of thesemechanisms with SDWLANs is not clear Accordingly webelieve that client programmability has not been well studiedin the literature Additionally though the 80211r and 80211k

CUWN+ Observability and Controllability+ Split MAC+ Scalability (multi-tier architecture)

+ Programmability+ Centralized control of WLANs

Sd-wlan

+ Light Virtual AP (LVAP) to improve handoff+ Scalability (Split-MAC)

Trantor

+ Programmability+ Client control

SplitAP

+ Slicing

+ Virtual APs (Cloud-based)+ Traffic shaping (switch-level programmability)+ Scalability (Split-MAC)

CloudMAC

+ Virtualization (seamless handoff)+ Energy management

+ Virtualization of residential networks+ Seamless handoff

+ Scalability (Two-tier control plane)

AeroFluxDenseAP

+ Observability and Controllability+ Load balancing

+ Traffic shaping (AP-level programmability)+ Enhanced programmability

Ethanol

+ Programmability (OpenFlow extension)

AEtherFlow

+ Virtualization (AP and middle boxes)+ Programmability (data plane)

OpenSDWN

+ Slicing (virtual base stations)

+ Seamless mobility (a single big AP)

+ Programmability+ Cooperation among residential AP

EmPOWER

+ AP Virtualization+ Interference Mitigation

AP Aggregate

+ AP Virtualization for airtime control

RCWLAN

+ AP Virtualization+ Seamless handoff

+ AP Virtualization+ Slicing

WVTCoAP and ResFi

+ Slicing+ Adding OpenFlow to AP

OpenRoads

BeHop+ AP Virtualization+ Slicing

+ AP Virtualization + Enhancing programmability

EmPOWER2

+ Traffic shaping

CENTAUR

Fig 7 This figure summarizes the main features of SDWLAN architectures based on the categorization presented in Section III-C

amendments (which have been recently integrated into 80211standard [133]) realize richer interactions between clients andAPs the SDWLAN architectures did not benefit from thesefeatures

2) Mobility Mapping VAPs to physical APs is the maintechnique used by architectures to support seamless mobilityAs each VAP is represented by a simple data structure thatholds clientrsquos status exchanging VAPs between APs removesthe burden of re-association as offered by Odin To avoidthe overhead of VAP migration CloudMAC settles the APs incloud servers and programs the switches to route data betweenphysical APs and VAPs However performance evaluation ofCloudMAC shows that when traffic is transmitted through ashared network high-priority packets may be lost or delayedthereby reducing the chance of connection establishment to55 Although queue management has been proposed as aremedy the effectiveness of such approaches for differenttopologies mobility and traffic patterns is an open researcharea Meanwhile we believe that the match-action paradigmof OpenFlow must be exploited to improve scalability andhandoff performance through designing traffic shaping andprioritization mechanisms that ensure the real-time delivery oftime-critical packets However the literature does not presentany relevant contribution Another solution is to use networkslicing as we will explain in Section III-D3 On the otherhand most of the reviewed architectures (except CUWN andAeroFlux) use a single-tier controller topology and unfortu-nately they do not present an evaluation of handoff perfor-mance in a real-world scenario with tens of APs backgroundtraffic and variable mobility patterns For example OdinAEligtherFlow and BIGAP use testbeds with 10 2 and 2 APsrespectively Therefore a study of architectural implications onhandoff performance is missing We present further discussionabout the challenges of mobility management after the reviewof central control mechanisms in Sections IV-C and V-C

3) Network Virtualization Our review shows that the mostcommon approach of network slicing is through the use of

VAPs For example in Odin a slice is defined as a set ofAPs (Odin Agents) LVAPs corresponding to clients SSIDsand network applications Each SSID may belong to oneor multiple slices and each client joins the slice to whichthe SSID belongs to Since each application can only seethe LVAPs belonging to its slice slice isolation is logicallyprovided through the concept of LVAPs

Although VAP migration can be used for slicing airtimebandwidth and processing power it does not provide a finegranularity for resource allocation For example if the con-troller moves a clientrsquos VAP to another AP in order toincrease its available bandwidth the actual bandwidth offereddepends on the activity of other associated clients In order toachieve high granularity layer-1 and layer-2 resource slicingtechniques are required to support airtime frequency and spacemultiplexing For example a middleware should manage theaccess of VAPs to the APrsquos physical resources through TDMAscheduling However compared to wired networks the highlyvariable nature of wireless communications makes predictableslicing a challenging task Specifically it is hard to predictand manage the effect of one slice on another For exampleuser mobility and variations of traffic in one slice affectthe available bandwidth experienced by the users in anotherslice [113] Compared to LTE networks resource slicing ismore challenging in 80211 networks because 80211 relies onrandom access mechanisms (ie CSMA) and does not utilizea dedicated control plane to communicate with clients Theproblem is exacerbated when multiple networks managed bydifferent entities coexist To cope with these challenges theexisting network slicing techniques heavily rely on a local-ized network view [96] [134] For example while SplitAPsupports central network control each AP decides about itsuplink airtime allocation individually and through a distributedalgorithm Based on this discussion utilizing algorithms thatrely on global network view and the integration of centralizedAsC and virtualization are necessary towards improving thecapacity and QoS of virtualized networks

TABLE IICOMPARISON OF SDWLAN ARCHITECTURES

Architecture ProgrammableArchitecture

Details ofProvided APIs VAP Network

SlicingSouth-bound

ProtocolSplit-MAC

SupportClient

ModificationImplemented Mechanisms

AsC ChADenseAP [12] times times times times Proprietary times times X timesCUWN [85] times times times times CAPWAP X times X XDIRAC [88] X X times times Proprietary times times X timesTrantor [89] X X times times Proprietary times X times timesDyson [90] X X times times Proprietary times X X times

AEligtherFlow [92] X X times times Extended OpenFlow times times X timesWVT [98] times times X X Proprietary times times times times

AP Aggregate [101] times times X X Proprietary times times X timesOpenRoads [99] [100] X X times X OpenFlow+SNMP times times X timesOdin [49] [102] [103] X X X X OpenFlow+Proprietary X times X times

SplitAP [96] times times X X Proprietary times X times timesCloudMAC [26] [50] X times X X OpenFlow X times X X

EmPOWER [107] X times X X OpenFlow+REST X times X timesEmPOWER2 [24] X X X X OpenFlow X times X X

Sd-wlan [110] X times X times Extended OpenFlow X times X timesBeHop [112] X times X times OpenFlow+Proprietary X times times times

RCWLAN [113] times times X X Proprietary times times times timesvBS [114] X times X X OpenFlow times times X times

BIGAP [97] X X X times Proprietary times times X XAeroFlux [117] X X X times Extended OpenFlow X times X timesEthanol [121] X X X times OpenFlow+Proprietary X times X times

OpenSDWN [25] X X X X OpenFlow+REST X times X timesVAN [105] X X times X OpenFlow times times times timesCOAP [31] X X times times OpenFlow+Proprietary times times times XResFi [32] X X X times Proprietary times times X X

Network slicing can also be used to tackle the challengesof control and data plane communication To this end anabstraction layer slices the resources of switching elementsbased on the communication demands of higher layer controlmechanisms For example based on the mobility patternand number of clients an AsC mechanism may request theabstraction layer for the allocation of switching resources Thisis an open research area

Network slicing would also enable the coexistence of M2Mcommunication with regular user generated traffic similar tothe 5G vision given in [135] For example an IoT devicemay utilize multiple slices of a network to transmit flows withdifferent QoS requirements Developing such architectures andtheir associated control mechanisms is an open research areaWe present further discussion about this after the review ofcontrol mechanisms in Section V-C4

4) Network Function Virtualization The most commonusage of NFV in SDWLANs is split-MAC which is mainlyimplemented through VAP Split-MAC was first introducedby the split-MAC protocol implemented in CAPWAP As ananother example CloudMAC separates MAC functionalitiesbetween physical AP and cloud platform Using the split-MAC strategy time critical operations run on APs whileother MAC operations are handled by the controller Thisstrategy enables exploiting the resources of remote computingplatforms and simplifies MAC and security updates withoutrequiring to replace APs as either no or minimal changes toAPs is required Despite the benefits of VAPs it is importantto establish a balance between flexibility and communicationoverhead because when more functionalities are shifted toa controller the overhead and delay of distribution systemincrease For example even for small networks the importanceof utilizing queuing strategies with OpenFlow switches hasbeen demonstrated [50] [121] [136] Consequently studiesare required to show the tradeoffs between centralization and

scalabilitySDR is an another example of NFV SDR platforms trans-

fer radio signal processing operations into a general-purposeprocessor on the same board or a remote server In addition toimproving programmability this enables the use of powerfulprocessors for centralized signal processing which can beused to improve signal decoding probability and coping withinterference challenges As mentioned in Section III-D1 theintegration of SDRs with SDWLANs is in its early stages

5) Home WLANs The importance of utilizing SDN-basedmechanisms in home networks is justified given the facts thatthe number of 80211 connected devices is being increased(eg home appliances lighting and HVAC medical monitor-ing) and some of these devices require timely and reliable dataexchange with APs [1] [17] [137] In addition the emergenceof 80211 mesh technologies that rely on the installationof multiple APs per home [138] makes central control ofhome networks even more important Compared to enterprisenetworks the topology of AP placement in home networksis random and variable Topology is even more randomizedwhen users or distributed algorithms modify the transmissionpower and channel of their APs This has serious implicationsWhile ISP-based control (eg VAN) simplifies the installationand operation of home networks the ISP may not be able tocollect enough information from the home network becausenot all the neighbors choose the same ISP On the other handcollaboration between homes (eg ResFi) introduces serioussecurity challenges Finally cloud-based control may not bevery effective because neighboring users may not subscribe tothe same service Therefore the opportunities and challengesof home SDWLANs requires further research

6) Security In addition to running new network controlmechanisms such as AsC and ChA programmability (seeSection III-D1) simplifies security provisioning For exam-ple when a security mechanism does not require hardware

replacement (eg 80211irsquos WPA) it can be implemented byreprogramming APs and clients However the decompositionof control and data plane exposes security threats due toenabling of remote programming of network equipment Forexample a DoS attack can be implemented by instructing alarge number of clients to associate with an AP Thereforeit is important to identify security threats and take them intoaccount during the architecture design process

As mentioned earlier virtualization through network slicingis an effective approach to control and isolate clientsrsquo access toresources However as users will be sharing an infrastructureachieving a strict isolation becomes more challenging Forexample a client may simply switch to the frequency bandof another slice to disrupt the capacity offered in that sliceUnfortunately secure virtualization has not been investigatedby the research community

SDWLANs enable network applications and administratorsto detect abnormal activities and network breaches In thearchitectures that extend their control plane to clients in orderto exchange monitoring and control packets (eg TrantorDyson SplitAP) the reports received from malicious clientsresult in control decisions that negatively affect networkperformance To cope with this challenge a malicious clientmay be detected using the flow statistics collected from APsand switches [139] In the next step client location may befound through analyzing RSSI measurements [140] Rogue(unauthorized) APs can be detected through collaborationamong APs For example an abnormal interference level orsniffing the packets being exchanged with a rogue AP can beused by the detection algorithm Depending on the architecturein use a network application can block the access of such APsthrough various mechanisms For example using Odin clientaccess can be blocked through updating OpenFlow switchesProposing security mechanisms that benefit from the featuresof SDWLAN architectures is an important future directionespecially for large-scale and public networks

IV CENTRALIZED ASSOCIATION CONTROL (ASC)

In this section we review centralized AsC mechanismsWe focus on the metrics employed as well as the problemformulation and solving approaches proposed This sectionalso highlights the impact of architecture on design whendetails are available We employ a consistent notation to easethe understanding and comparison of metrics and formula-tions Our review summarizes the performance improvementsachieved to reveal the benefits of these centralized mechanismscompared to distributed approaches2 Although we mostlyfocus on state-of-the-art centralized mechanisms we reviewthe seminal distributed mechanisms as well because of theiradoption as the baseline to evaluate the performance of cen-tralized mechanisms

At a high level we categorize AsC mechanisms from twoperspectives

2Note that in this section and the next section (Section V) we do not studyall the AsC and ChA mechanisms implemented by the architectures reviewedin Section III This is because some of these architectures only show thefeasibility of implementing control mechanisms and they do not propose anynew mechanism to benefit from the features of SDWLANs

ndash Seamless handoff refers to the mechanisms that theirobjective is to reduce the overhead and delay of clienthandoff

ndash Client steering refers to the mechanisms that adjustclient-AP associations to optimize parameters such as theload of APs and the airtime allocated to clients

Supporting seamless handoff is usually addressed by proposingarchitectures that reduce the overhead of re-association Incontrast client steering is performed through proposing AsCmechanisms that run on the control plane For example anAsC mechanism may propose an optimization problem tobalance the load of APs while handoff delays depend on thearchitectural properties From the client steering point of viewAsC is particularly important in dense topologies because thereare usually multiple candidate APs for a client in a givenlocation Therefore using a simple RSSI metric may resultin hot spots unbalanced load of APs and unfair resourceallocation to clients Hence one of the main goals of AsCis to achieve fairness among clients and APs

The distributed AsC mechanism employed by 80211 stan-dard is strongest signal first (SSF) [141] [142] UsingSSF each client decides about its association based on theRSSI of probe response and beacon messages Each clientassociates with the AP from which highest RSSI has beenreceived Least load first (LLF) [143] is an another widely-adopted distributed mechanism where APs broadcast theircurrent load through beacon messages to help the clientsinclude AP load when making association decisions The loadof an AP is represented through various metrics such as thenumber of associated clients3 In the following we review AsCmechanisms

A Seamless Handoff

Some of the architectures reviewed in Section III-C proposemechanisms to reduce the overhead of client handoff Handoffoverhead refers to (i) the packets exchanged between clientand AP to establish a connection and (ii) the delay incurred bythe client during this process [144] In this section we study thecontributions of SDWLAN architectures in terms of supportingseamless handoff

Cisco unified wireless network (CUWN) [85] enablesthe central configuration of RSSI threshold with hysteresis toperform seamless handoff of CCX [145] compatible clientsWhen the RSSI received from associated AP drops below thescan threshold the client increases its AP scanning rate toensure fast handoff to another AP when the difference betweenthe RSSI of the associated and new AP is equal or greaterthan the hysteresis value specified In addition to hysteresisspecifying the minimum RSSI value forces the clients to re-associate when their RSSI drops below a minimum RSSI

Three types of client roaming scenarios are handled byCUWN (i) intra-controller roaming (ii) inter-controller layer-2 roaming and (iii) inter-controller layer-3 roaming Thecontroller simply handles an intra-controller roaming by up-dating its client database with the new AP connected to

3The traffic indication map (TIM) of a beacon packet represents a bitmapthat indicates the clients for which the AP has buffered packets

the roaming client Inter-controller layer-2 roaming occurswhen a client associates with an AP that is controlled by adifferent controller belonging to the same subnet In this typeof roaming mobility messages are exchanged between the oldcontroller and the new one Then the database entry relatedto the roaming client is moved to the new controller Inter-controller layer-3 roaming occurs when a client is associatedwith an AP that is controlled by a controller belonging to adifferent subnet In layer-3 roaming the database entry of theroaming client is not moved to the new controller Rather theold controller marks the clientrsquos entry as anchor entry Theentry is copied to the new controller and marked as foreignentry Therefore the original IP address of the roaming clientis maintained by the old controller CUWN enables networkadministrators to establish mobility groups where each groupmay consist of up to 24 controllers A client can roam amongall the controllers in a mobility group without IP addresschange which makes seamless and fast roaming possible

Odinrsquos Mobility Manager (OMM) Odin [49] [102] en-ables seamless handoffs through LVAP migration betweenAPs The Odin controller maintains a persistent TCP con-nection per Odin Agent running on AP thereby switchingamong agents does not require connection reestablishmentIn addition the delay of a re-association equals the delay ofsending two messages from the Odin controller to the old andnew APs The first message removes an LVAP from the oldAP and the second message adds an LVAP to the new APAssuming that the messages are sent successfully the delayequals the longest round trip time (RTT) between the controllerand the two APs which depends on network size

To show the effectiveness of LVAPs Odin employs asimple RSSI-based AsC mechanism The mobility application(running on the controller) selects the Odin Agent with highestRSSI if client movement is detected by the controller Theeffect of handoff on TCP throughput has been evaluated usingtwo APs and a client While a period of throughput dropin regular 80211 layer-2 handoff is observable Odin doesnot show any throughput degradation Performance evaluationsalso show that executing 10 handoffs per second results in anegligible reduction in TCP throughput

AEligtherFlow [92] argues that handoff support through LVAPmigration (eg Odin [102] and OpenSDWN [25]) imposeshigh computational and communication overhead especiallyin large networks with many mobile clients They proposea predictive handoff strategy by relying on the extendedOpenFlow protocol proposed in this work (see Section III-C2)The handoff mechanism works in three phases

ndash Prediction The controller predicts an association whenthe RSSI of a client to its associated AP declines whileits RSSI to another AP increases

ndash Multicasting By updating OpenFlow tables the con-troller multicasts the packets of the client to both thecurrent AP and the predicted AP

ndash Redirection The controller will redirect the clientrsquos trafficto the new AP after the handoff completion If no handoffoccurs then the multicasting is stopped

Experiments using two APs and a client shows that thehandoff delay of AEligtherFlow is around 53 seconds compared

to the 71 seconds delay of 80211 standardBIGAP [97] uses the 25 non-overlapping channels of

5GHz band to form disjoint collision domains for handoffAs explained in Section III-C a separate NIC is used toperiodically overhear packets on all channels which enablesthe controller to compute the potential SNR values of client-AP links A handoff happens when a higher SNR would beachievable for a client However to avoid the ping-pong effectan 8dB hysteresis value is used

BIGAP performs handoff through client channel switchingSince all APs share the same BSSID in order to handoff aclient from AP1 to AP2 the controller instructs AP1 to senda channel switching command to the client The client thenswitches to the channel being used by AP2 Since both APsuse the same BSSID the client does not notice handoff

Performance evaluations (using two APs) show that theBIGAP handoff is about 32 times shorter than the regular80211 handoff In addition BIGAP results in lower energyconsumption because it moves the overhead of handoff toAPs and there is no need for the clients to scan the channelsIn addition while 80211 results in zero throughput for about 4seconds BIGAP shows only 5 throughput reduction duringthe handoff

B Client Steering

Based on the scope of the optimization problem employedwe categorize client steering mechanisms into two groupscentrally-generated hints and centrally-made decisions InAsC mechanisms using centrally-generated hints the con-troller relies on the global network view to generate hints forthe association of clients In other words the controller doesnot make the final decision about associations and insteadenables the clients to make more informed decisions throughthe hints conveyed On the other hand in AsC mechanismsusing centrally-made decisions the controller makes the as-sociation decisions and enforces the clients to apply themWe will discuss in Sections IV-B2 and IV-B3 the two sub-categories of mechanisms based on centrally-made decisions

1) Centrally-Generated Hints In this section we reviewAsC mechanisms that employ client steering through hintsgenerated centrally

BestAP [146] proposes an AsC mechanism based on theestimation of available bandwidth (ABW) for each client atevery AP in its vicinity The available bandwidth depends onchannel load which varies with packet loss and PHY rate Theestimated available bandwidth at PHY rate r is computed as

E[ABW (r)] =8Sdata(1minusB)sumn

k=0(1minus ps(r))kps(r)Ttx(r k) (1)

where Sdata is data size (bytes) ps is the probability ofsuccessful packet transmission Ttx(r k) is the transmissiontime of a packet during the kth transmission attempt withPHY rate r and B is the fraction of channel busy time Bis measured through using the CCA register of NIC ps(r) ismeasured using the statistics provided by the rate adaptationalgorithm All other parameters are configured statically

A scheduler is run on each client to allocate a measurementperiod (eg 50ms every 2s) during which the client sends data

to all nearby APs in order to update packet loss and channelbusy fraction A monitoring service is run on APs to collectthe statistics from clients calculate ABW and send a reportto the controller The controller sends the estimated ABWs ofthe best five APs to each client periodically BEST-AP onlyconsiders the ABW of downlinks

Testbed evaluations show that the delay overhead of ABWestimation is less than 50ms which is appropriate for adynamic AsC mechanism Compared to SSF [141] the pro-posed AsC mechanism shows 81 and 176 improvement inthroughput for static and mobile clients respectively

Ethanol This architecture [121] (see Section III-C5) hasbeen evaluated through running a load-aware AsC mechanismthat aims to balance the number of associated clients amongAPs When a client requests to join an AP with higher load thecontroller drops the request to force the client look for anotherAP A simple testbed with two APs and up to 120 clients showsthat the maximum difference between the number of clientsassociated to APs is two which is due to the concurrent arrivalof association requests

2) Individual Optimization through Centrally-Made Deci-sions In this section we review AsC mechanisms that employclient steering through decisions generated centrally Thesemechanisms however do not define a global optimizationproblem thereby they do not take into account the effect of anassociation on other clientsAPs Due to the individual natureof association control these mechanisms only improve theoverall network performance and cannot be used to enforcefairness

DenseAP The AsC mechanism proposed by DenseAP [12](see Section III-C1) works as follows The available capacitymetric is defined to rank all the APs a client could beassociated with A client associates to the AP with highestavailable capacity The available capacity of APi operating onchannel chi is defined as follows

APicj= F chi

APitimes rchi

APicj (2)

where F chi

APiis the free airtime of APi on channel chi

and rchi

APicjis the expected transmission rate of cj when

communicating with APi Free airtime is estimated by mea-suring the amount of time that the MAC layer contends forchannel access to send a high-priority packet The expectedtransmission rate is estimated using the RSSI of probe requestframes received at the AP A mapping table is used for thispurpose APs hear the probe requests of clients and sendreports to a controller APs also measure their free airtime andreport it to the controller The controller selects the AP withhighest available capacity for each client to associate withTo instruct a client associate with the selected AP only theselected AP responds to the clientrsquos probe message

DenseAP also proposes a dynamic load balancing algorithmthat periodically decides about associations In particular thecontroller checks the free airtime of APs every minute An APis overloaded if its free airtime is less than 20 and it hasat least one associated client If an overloaded AP exists thecontroller considers its clients as the potential candidates forassociation with APs experiencing lower load A candidate AP

ON OFF

Cclu(APi) lt Cminclu(APi)

and tinactive Tidle

Cclu(APi) Cminclu(APi)

and toff Toffline

Fig 8 Finite state machine (FSM) of energy manager for a slave AP inEmPOWER [107]

must satisfy these conditions (i) the expected transmission rateof clients when associated with the new AP must not be lowerthan the current transmission rate and (ii) the free airtimeof new AP must be at least 25 more than the current APDuring each decision period at most one client is allowed tobe associated with a new AP and two consecutive associationsfor a client is prohibited to avoid the ping-pong effect

Empirical evaluations show a 40 to 70 increase in per-client throughput compared to SSF Moreover the authorsconduct a small experiment using three clients and two APs toshow that the load balancing algorithm improves the through-put of clients by more than 200

Odinrsquos Load Balancing (OLB) This load-balancing mech-anism [102] periodically (every minute) inquires APs to collectthe RSSI relationship between clients and APs and thenLVAPs are evenly distributed between APs to balance theirloads The evaluations on a testbed with 10 APs and 32 clientsshow that around 50 and 15 of clients were able to receivea fair amount of throughput when the proposed mechanismwas enabled and disabled respectively

EmPOWER This mechanism [24] [107] relies on clientsteering to improve the energy efficiency of APs APs arepartitioned into clusters where each cluster has one masterand multiple slaves Master APs are always active and theyare manually selected during the network deployment phaseto provide a full coverage Slave APs are deployed to increasenetwork capacity These APs are turned onoff using the finitestate machine (FSM) depicted in Figure 8 In the ON modeall wireless interfaces of the AP are on In OFF mode onlythe Energino [108] module of the AP is on

Two metrics are defined for a slave APi belonging to clusterclu(APi) Cclu(APi) is the number of clients in the clusterand Cmin

clu(APi)is the minimum required number of clients in

cluster clu(APi) to keep APi active A slave APi transitionsfrom ON mode to OFF mode if (i) the number of its clusterrsquosclients is less than Cmin

clu(APi) and (ii) APi has been inactive

(ie no client associated) for at least Tidle seconds Also APi

transitions from OFF mode to ON mode if (i) the number ofits clusterrsquos clients is at least Cmin

clu(APi) and (ii) APi has been

OFF for at least Toffline secondsThe mobility manager associates a client to a new AP

with higher SNR However to establish a balance betweenperformance and energy consumption the mobility managermay associate a client to an AP with lower SNR but smallerCminclu(APi)

In this way the energy manager is able to turn offthe APs with higher Cmin

clu(APi)in order to decrease energy

consumption Re-association is performed if there is a betterAP in terms of SNR and Cmin

clu(APi) By relying on the Odin

[102] APIs the authors showed that this AsC mechanism was

implemented as a Java network application with only 120 linesof codes

Adaptive mobility control (AMC) [147] shows that usinga fixed RSSI threshold by clients results in an unbalanced loadof APs They propose a heuristic algorithm which uses RSSIand traffic load of APs to provide dynamic hysteresis marginson AP traffic load level The load of APi is defined as

LAPi =

B if |CAPi | = 008timesB + 02times |CAPi | if |CAPi | gt 0

where |CAPi| is the number of clients associated with APi and

B is channel busy time The algorithm defines three thresholdson AP load and RSSI low medium and high Using thesethresholds a client is associated with a new AP that satisfiesone of these conditions (i) higher signal strength and lowerload (ii) significantly higher signal strength and slightly higherload or (iii) significantly lower load and slightly lower signalstrength

The proposed adaptive mobility manager has been imple-mented using Odin Empirical results show more than 200improvement in TCP throughput compared to SSF

Associating to good enterprise APs (AGE) The mainobjective of AGE [148] is the reduction of clientsrsquo packetexchange delay over wireless links through client steeringAGE has two main phases learning and AP selection Duringthe learning phase the performance metrics and environmentalfactors are pulled every minute from all APs using SNMP [60]These metrics include AP-client RTT RSSI SNR numberof associated clients channel number frequency band APlocation day of week and time

Using the collected training set the authors use the randomforest [149] technique to generate a two-class learning modelfor classifying APs into high latency and low latency Figure 9shows an overview of AGE and its two main components (i)the AGE app that is installed on each clientrsquos mobile phoneand (ii) the AGE controller that uses the random forest modelto classify the APs in the vicinity of each client AGE operatesas follows

ndash The AGE application on a client device sends an AGErequest (including the list of achievable APs) to the AGEcontroller periodically

ndash The AGE controller pulls the SNMP data of the APsrequested by AGE app

ndash The latency class of each AP is predicted using thelearning technique mentioned earlier

ndash The AGE controller informs the AGE application aboutthe best AP nearby

ndash The client re-associates with a new AP using AGE appThe authors deployed AGE at Tsinghua University campus

where over 1000 devices used the network for 25 monthsMeasurements confirmed that the wireless data exchange delayof more than 72 of clients has been reduced by over 50

3) Global Optimization through Centrally-Made Deci-sions In this section we review AsC mechanisms that employclient steering by formulating a problem that aims to optimizeperformance parameters globally These mechanisms proposeheuristics to solve the NP-hard problems that usually aim toachieve network-wide fairness

AP1 AP2 APi

AGE app AGE controller Random forest model

Controller

Client

SNMPdata

AGE request

Specified AP

Get AGE factors of each AP AGE

factors

Predicted latency class for each AP

Re-associate to the specified AP

Fig 9 Associating to Good Enterprise APs (AGE) [148] Each client isinstructed to connect to the AP that provides minimum delay

Association Control for fairness and load balancing(ACFL) In [150] and [151] the authors address the unbal-anced load of APs as a result of using SSF or LLF To addressthis challenge an AsC problem is formulated to establish themax-min bandwidth fairness among APs Intuitively the loadof a client on its associated AP is inversely proportional to theeffective bit rate of AP-client link The load of APi (denotedas LAPi

) is modeled as the maximum of the aggregated loadsof its wireless and wired links generated by all clients c isin Cas follows

LAPi= max

sumforallcjisinC

ωcj timesXAPicj

rAPicj

sumforallcjisinC

ωcj timesXAPicj

where rAPicj is the transmission rate between APi and cj RAPi

is the transmission rate of wired interfaces of APiXAPicj isin 0 1 is the association state of cj to APi andωcj is the traffic volume size of client cj In other words theload of APi is defined as the period of time this AP requiresto handle the traffic of its associated clients APi providesbandwidth XAPicj times ωcjLAPi

to client cj Two approximation algorithms are introduced to solve the

formulated NP-hard max-min fair bandwidth allocation prob-lem The first algorithm solves the problem for unweightedgreedy clients The second algorithm proposes a solutionfor weighted and limited throughput demand of clients Thealgorithms are run periodically by a controller to update clientassociations Simulation results show over 20 improvementin terms of average per-client bandwidth compared to SSFand LLF

Association control for proportional fairness (ACPF)[152] argues that using throughput-based fairness (eg [153][154]) in multirate networks leads to low overall networkthroughput because clients with a low bit rate can occupythe channel longer than those with higher rate The authorsinvestigate the problem of achieving proportional fairness byintroducing the following objective function formulated basedon the effective bandwidth of clientssum

forallcjisinC

ρcj times log(βcj ) (5)

where ρcj and βcj are the priority and effective bandwidth ofclient cj respectively βcj is defined assum

forallAPiisinAP

XAPicj times tAPicj times rAPicj (6)

where XAPicj isin 0 1 is the association index tAPicj isin[0 1] is the effective normalized time and rAPicj is the

PHY rate between cj and APi Achieving fairness throughassociation is formulated as a linear-programming problemto maximize the objective function A centralized algorithmcalled non-linear approximation optimization for proportionalfairness (NLAO-PF) is proposed to solve the optimizationproblem It has been proven that the approximation ratio ofNLAO-PF is 50

The authors used the OMNet++ [155] simulator to evaluatethe effectiveness of NLAO-PF compared to cvapPF [156] interms of average throughput The average improvement isreported as 188 and 35 for uniformly distributed clientsas well as clients in a hotspot area respectively

Association control with heterogeneous clients (ACHC)Experimental analysis of the negative impacts of legacy clients(80211abg) on 80211n clients has been reported by [157]The authors propose a two-dimensional Markov model tocalculate the uplink and downlink throughput of heterogeneousclients MAC efficiency of a client cj is defined as follows

Υ(cj) =Thupcj + Thdown

min1 pupcj + pdowncj times rAPicj

where Thupcj and Thdowncj are the estimated uplink and down-

link throughput of client cj respectively pupcj and pdowncj are

respectively the uplink and downlink traffic probabilities ofcj rAPicj is the optimal PHY rate between cj and APi AsCis formulated through an optimization problem to provide abandwidth proportional to each clientrsquos achievable data ratewhich is obtained by maximizing

sumforallcjisinC log Υ(cj)

Two heuristic dynamic AsC algorithms are proposed tosolve the optimization problem FAir MAC Efficiency (FAME)and Categorized Instead of maximizing all MAC efficienciesthese algorithms maximize the minimum MAC efficiency Torun FAME the data rate and traffic load of all clients innearby BSSs must be collected from each clientrsquos point ofview With Categorized the type of all clients in nearbyBSSs must be collected as well However Categorized is lesssensitive to network dynamics because it tries to maximizethe number of similar-standard 80211 clients connected toeach AP Therefore Categorized takes advantage of densedeployment and alleviates the performance degradation of80211n clients that is caused by the presence of legacy clients

Simulation and testbed experiments confirmed the higherperformance of FAME and Categorized compared to SSFand ACFL in terms of TCP and UDP throughput MACefficiency and aggregated throughput of clients supportingdifferent standards Although the conducted experiments usedistributed execution of FAME and Categorized collectingthe information required by these algorithms imposes a highburden on clients Therefore these algorithms are mostlysuitable for SDWLANs

Association control and CCA adjustment (ACCA) Twocentral algorithms for AsC and CCA threshold adjustment areproposed in [158] for dense AP deployments The basic idea ofthe AsC mechanism is to utilize the SINR of a client perceivedby an AP in order to measure the interference (congestion)level of APs The network is modeled as a weighted bipartitegraph (WBG) and the AsC problem is formulated as a max-imum WBG matching combinatorial optimization problem

The weights of edges in WBG are the uplink SINR values Theoptimization problemrsquos objective is to maximize the sum ofthe clientsrsquo throughput where the throughput upper-bound iscomputed using the Shannon-Hartley formula [159] Per-BSSCCA level is adjusted by using a constraint of the optimizationproblem based on the cell-edge SINR of each BSS Theoptimization problem has been solved using Kuhn-Munkres[160] assignment algorithm

MATLAB simulations show a 15 to 58 improvement at the10th percentile of cumulative distribution of clients through-put compared to SSF Furthermore the CCA adjustmentalgorithm significantly improves the throughput of cell-edgeclients compared to the network with fixed CCA threshold

Demand-aware load balancing (DALB) [161] formulatesjoint AP association and bandwidth allocation as a mixed-integer nonlinear programming problem that includes thebandwidth demand of clients The objective is to maximizethe aggregated bandwidth of clients which establishes a trade-off between throughput gain and time-based user fairness Thekey idea of this work is the inclusion of clientsrsquo bandwidthdemands in the computation of APsrsquo load For a client cj bandwidth demand is defined as βcj = ScjT where Scj isthe size of data received for client cj by its associated APduring the allocable transmission time of the AP ie T Thetransmission time demand of cj associated to APi is definedas TAPicj = βcj times TrAPicj

The optimization problemrsquos objective is defined as followssumforallcjisinC

logsum

forallAPiisinAP

XAPicj times rAPicj times tAPicj

where the transmission time allocated to cj associated withAPi is expressed by tAPicj The authors proved that maxi-mizing Equation 8 subject to a constraint on the transmissiontime demand of clients ie 0 le tAPicj le TAPicj leads toproportional time fairness among clients

A two-step heuristic algorithm is introduced to solve theoptimization problem The first step associates a client toan AP with the lowest allocated transmission time at eachiteration until all clients are associated At each iteration thealgorithm selects a client with the largest bandwidth demandThe second step schedules the transmission time of APsto establish proportional transmission time fairness amongclients

Simulations show a 23 improvement in clientsrsquo throughputcompared to SSF ACPF and NLB [162] In addition thisimprovement has been achieved without sacrificing fairnessIn particular while bandwidth fairness is slightly better thanthat of SSF ACPF and NLB time fairness is significantlyimproved compared to these algorithms

Migration-cost-aware association control (MCAC)[163] investigates the problem of max-min fairness subject

to a migration cost constraint An integer linear program-ming problem is formulated with the aim of maximizingthe minimum throughput of clients in order to establishmax-min fairness The migration cost of clients is the mainconstraint of the formulation The load of APi is defined asLAPi =

sumforallcjisinC XAPicjrAPicj Assuming the achievable

throughput of clients is proportional to the inverse of theload of the associated AP (ie 1LAPi ) maximizing theminimum throughput of clients is equivalent to minimizingmaxforallAPiisinAP LAPi

The migration cost constraint is definedas sum

forallAPiisinAP

sumforallAPlisinAP

(YAPiAPltimesMcj ) leM (9)

where YAPiAPlisin 0 1 indicates whether a client cj cur-

rently associated with APi will be associated with APl Mcj

is the migration cost of client cj from APi to APl and M isthe total permissible migration cost

The authors showed that the proposed problem is NP-hardand they proposed an approximation algorithm named thecost-constrained association control algorithm (CACA) tosolve the optimization problem In this algorithm the problemis divided into two sub-problems client removal and clientre-association In the first sub-problem a subset of clients areremoved from their current AP with the aim of minimizing themaximum AP load considering the migration cost constraintIn the second sub-problem the algorithm performs client re-association to minimize the load of maximum-loaded AP

NS3 [164] simulations and experimental results show im-provements in terms of the number of re-associations through-put loss ratio and end-to-end delay compared to SSF ACPFand GameBased [165]

QoS-driven association control (QoSAC) [166] inves-tigates the problem of flow-level association to address theQoS demands of clients Specifically the backhaul capacity ofAPs is included in the association decision making This workproposes a mechanism for concurrent association of a clientto multiple APs and supports flow-level routing of traffic Thismechanism facilitates dedicated management of each flow Forexample a client may use AP1 for video streaming whileusing AP2 to upload a file

The main objective is to minimize the average inter-packetdelay of individual download flows The inter-packet delayfor a client associated to an AP is formulated based on thebi-dimensional unsaturated Markov model proposed in [157]An optimization problem is formulated to minimize the sumof the average inter-packet delay of all APs The backhaulcapacity of an AP should not be smaller than the sum of thearrival traffic rates of the AP ie the sum of all downloadflow rates This is included as a constraint of the optimizationproblem The proposed optimization problem is interpretedinto a supermodular set function optimization [167] whichis an NP-hard problem This problem has been solved usingtwo heuristics (i) greedy association and (ii) bounded localsearch association The greedy algorithm associates a clientwith the AP that minimizes the total inter-packet delay in eachiteration In this algorithm the association loop is continueduntil all clients are associated The bounded local search is apolynomial-time algorithm similar to the algorithm proposedin [168]

Simulation results confirm the reduced average inter-packetdelay achieved with this mechanism compared to SSF andFAME [157] The delay is considerably reduced when thebackhaul capacity is limited and clients form hotspots

C Association Control Learned Lessons Comparison andOpen Problems

Table III presents and compares the features of the AsCmechanisms Although most of the AsC mechanisms focus onclient steering we should note that client steering and seamlesshandoff are interdependent For example when associationdecisions are made centrally to balance the load of APs aquick reaction to network dynamics requires seamless handoffotherwise the effect of load balancing will be compromised bythe overhead and delay of re-associations In the following westudy the features of AsC mechanisms and identify researchdirections

1) Optimization Scope of Client Steering Based on ourreview we have classified the client steering approaches intothree categories (i) centrally-generated hints (ii) individualoptimization through centrally-made decisions and (iii) globaloptimization through centrally-made decisions Our reviewshows that most of the AsC mechanisms fall into the last twocategories where their optimization problems is classified asfollows

ndash Overall network performance To improve the overallperformance of clients and APs For example QoSACminimizes the sum of inter-packet delay for all APsand ACCA maximizes the sum of clientsrsquo achievablethroughput Note that improving overall performancedoes not necessarily require defining an optimizationproblem as Section IV-B2 shows

ndash Max-min fairness To maximize the minimum perfor-mance In other words max-min fairness indicates pro-viding a client with more resources would not be possiblewithout sacrificing the resources of other clients [169]For example the objective function of ACFL and MCACmaximizes the throughput of minimum-throughput client

ndash Proportional fairness Clients occupy the channel pro-portional to their transmission rate Our review showsthat proposing these approaches is motivated by thechallenges of achieving fairness in multi-rate networksMulti-rate is caused by both client heterogeneity (eg80211abn) and the rate adaptation mechanism of80211 standards Since the 80211 DCF protocol aimsto provide an equal channel access probability for allcontending clients low-rate clients occupy the mediumlonger than high-rate clients In this case the max-minthroughput fairness results in sacrificing the performanceof high-rate clients which leads to a problem calledperformance anomaly Performance anomaly in multi-rate networks may result in total throughput degradationeven in comparison to SSF [141] [157] [170] [171]A well-known solution is proportional fairness whichis usually achieved by defining a logarithmic objectivefunction and scheduling the transmission time of clientsto be proportional to their transmission rate ACPFACHC and DALB are examples of AsC mechanisms thatestablish proportional fairness

Using proportional fairness in multi-rate networks decreasesthe performance of low-rate legacy clients and may dissatisfytheir QoS requirements We suggest two approaches to ad-

TABLE IIICOMPARISON OF ASSOCIATION CONTROL (ASC) MECHANISMS

MechanismObjective Optimization Scope

Decision MetricTraffic Awareness Performance Evaluation

SeamlessHandoff

ClientSteering

Cent GenHints

Indv OptCent Made

Decisions

Glob OptCent Made

DecisionsDownlink Uplink Simulation Testbed

CUWN [85] X times times times RSSI times times times timesOMM [102] X times times times RSSI times times times X

AEligtherFlow [92] X times times times RSSI times times times XBIGAP [97] X times times times RSSI times times times X

BestAP [146] times BandwidthImprovement X times Available

bandwidth X times times X

Ethanol [121] times AP LoadBalancing X times Number of clients times times times X

DenseAP [12] times AP LoadBalancing times Overall

performance times Availablebandwidth X times times X

OLB [102] times AP LoadBalancing times Overall

performance times RSSINumber of clients times times times X

EmPOWER [107] times EnergyEfficiency times Overall

AMC [147] times AP LoadBalancing times Overall

performance times RSSI Free airtime X times times X

AGE [148] times Reducing PacketExchange Delay times Overall

performance times Latency times times times X

ACFL [151] times AP LoadBalancing times Max-min

fairnessFree airtime

Transmission rate X times X times

ACPF [152] times Fair BandwidthAllocation times Proportional

FairnessAvailablebandwidth X times X times

ACHC [157] times Fair BandwidthAllocation times Proportional

Fairness MAC efficiency X X X X

ACCA [158] times MaximizeThroughput times Overall

Performance Uplink SINR times X X times

DALB [161] times MaximizeThroughput times Proportional

Fairness Transmission time X times X times

MCAC [163] times Fair BandwidthAllocation times Max-min

FairnessThroughput

Re-association cost X times X X

QoSAC [166] times Reducing PacketExchange Delay times Overall

Performance Inter-packet delay X times X times

dress this problem (i) hybrid fairness and (ii) demand-awareproportional fairness In hybrid fairness a weighted fairnessmetric is defined by assigning weights for the max-min andproportional fairness metrics Therefore it is possible to adjustthe weights based on the demand of low-rate legacy clientsIn the second approach the demand of clients is merged withtheir rates when formulating the problem

2) Decision Metrics The most important component of anAsC mechanism is its decision metrics as they affect on bothaccuracy and overhead When using the AsC mechanism of80211 standard (ie SSF) a client associates with the AP thatis providing the highest RSSI The main shortcoming of SSFis that it does not recognize the load of APs and the demandof clients Therefore AsC mechanisms introduce additionalmetrics in their decision process

The rdquoDecision Metricrdquo column of Table III summarizesthe decision metrics employed Generally our review showsthat the most popular decision metrics are RSSI AP loadthroughput of clients and packet exchange delay experiencedby clients Balancing the load of APs is a widely-adoptedoptimization metric as it implicitly results in an improvedperformance of clients In the mechanisms relying on thismetric load is modeled through parameters such as thenumber of associated clients AP throughput and channelbusy time Although these mechanisms implicitly improve theperformance of clients explicit consideration of clientsrsquo trafficresults in higher performance

Based on this we can classify AsC mechanisms into twogroups

ndash Demand-agnostic Refers to the AsC mechanisms thatassociate each client with the best AP nearby with-out taking into account clientsrsquo demands For instanceDenseAP ACPF and BestAP associate each client withthe nearby AP that is providing the highest availablebandwidth independent of the current demand of clientsSimilarly AGE and MCAC associate each client with theAP that is providing the minimum delay without consid-ering the tolerable delay of clients in the decision makingprocess As a result demand-agnostic approaches mayperform unnecessary associations which incur handoffoverheads that result in performance degradation

ndash Demand-aware Refers to the AsC mechanisms that ex-plicitly include the demand of clients in their decisionmaking process Unfortunately the number of demand-aware AsC mechanisms (eg ACHC and DALB) is verylimited

Throughput demand in particular is composed of uplinkand downlink traffic Unfortunately most of the AsC mech-anisms assume downlink traffic is dominant when calculat-ing their association decision metrics as Table III showsHowever the emerging applications of SDWLANs justify theimportance of uplink traffic For example medical monitoringindustrial process control surveillance cameras and interac-tion with cloud storage require timely and reliable uplinkcommunication [17] [35] [54] [172]ndash[174]

These discussions reveal that designing AsC mechanismsto support both uplink and downlink requirements is an openproblem Additionally due to the resource-constraint nature

of IoT devices it is important to define compute and includemetrics such as channel contention intensity and packet lossrate in the decision making process

In addition to the mentioned challenges it is also importantto identify and address the effect of hardware on the metricsused For example the RSSI perceived by a tiny IoT devicewould be different from the value perceived by a multi-antenna smartphone under the same condition Therefore thecontroller cannot generate a realistic network map based onthe information collected from clients Similarly for environ-ments with highly asymmetric links (for example caused byshadowing) the information collected in the controller basedon the RSSI perceived at the APs does not reflect the realconnectivity and interference relationship Understating theeffect of metric evaluation irregularity on performance and thecalibration of metrics and inclusion in the AsC mechanismsare open research areas

3) Dynamicity and Overhead Dynamic and scalable AsCrequires minimizing overhead from the following point ofviews

ndash Measurement delay the delay of measuring decisionmetrics must be short to reflect network dynamics

ndash Bandwidth requirements the amount of bandwidth con-sumed for exchanging control data should be minimizedto enhance scalability Each association command re-quires communication between the controller and APs Inaddition in a VAP-based architecture such as Odin eachassociation requires a VAP transfer where the numberof transfers depends on the number of associationsTherefore both the frequency and the scale of reasso-ciation should be taken into account for various types ofarchitectures

ndash Deployment cost it is ideal to define decision metrics thatdo not require AP or client modification For examplealthough BestAP claims low measurement overhead interms of delay and bandwidth it requires the modificationof APs and clients

Unfortunately the existing AsC mechanisms do not takeinto account the effects of these three factors the scalabilityof the proposed mechanisms has not been studied thoroughlyand their effect on the energy consumption of smartphonesand resource constraint devices is unknown

Although some AsC mechanisms propose their own mea-surement techniques continuous and low-overhead networkmonitoring should be provided as a SDWLAN architecturalservice to facilitate the development and interoperability ofcontrol mechanisms For example as we will see in SectionV dynamic ChA mechanisms share some of the metricsused by AsC mechanisms We propose the followings toprovide network applications with low-overhead and con-tinuous measurement of metrics (i) designing metrics thatcan be used by multiple control mechanisms (ii) efficientencoding of control data conveyed by a south-bound protocol(iii) novel hardwaresoftware techniques deployed on networkinfrastructure devices for passive (ie without introducingextra traffic) measurement of network status and (iv) in-ference and prediction algorithms that extract new metricsand provide insight into the future status of the network

Compared to distributed implementations SDWLANs enablein-depth analysis of clientsrsquo traffic type and pattern by lookinginto packet headers [30] [175]ndash[178] This information couldbe employed to improve the accuracy of traffic predictionand interference modeling Specifically while the existingapproaches rely on the saturated demands of clients when esti-mating load and interference levels the throughput and airtimeof clients and APs strongly depend on the protocols used atthe transport and application layers [172] [179] [180] Forexample the periodic nature of an IoT devicersquos uplink trafficcould be exploited to improve AsC performance We believethat learning and prediction mechanisms must be employedby SDWLANs to provide network control applications (suchas AsC and ChA) with meaningful realistic and predictiveknowledge regarding network operation

Hybrid design is another approach towards achieving timelyreactions against dynamics The overhead of network mon-itoring and control can be improved through hybrid controlmechanisms as proposed by SoftRAN [181] Due to the highdynamics of wireless networks it is desirable to design controlmechanisms that can make control decisions locally whileimproving decisions based on the global network view as wellThis solution could also benefit from hierarchical controllertopologies to make decisions at multiple levels depending onthe dynamicity of variations For example AeroFlux (SectionIII-C) proposes an architecture to enable the design of multi-level control mechanisms Such mechanisms make controldecisions in a distributive manner at local controllers orby a central controller based on network dynamics timeconstraints and overhead of exchanging control messages

4) Security Considerations As centralized mechanismsrely on global network information and usually aim to achievea network-wide optimization receiving malicious reports hasa more severe effect on performance compared to distributedcontrol Verifying the legitimacy of network status reports andclientsrsquo demands have not yet been addressed by SDWLANarchitectures and AsC mechanisms

V CENTRALIZED CHANNEL ASSIGNMENT (CHA)In this section we review centralized ChA mechanisms

To be consistent with Section IV we focus on the metricsemployed as well as the problem formulation and solutionproposed by these mechanisms In addition we highlight theperformance improvements achieved compared to distributedmechanisms

Considering the broadcast nature of wireless communicationas well as the dense deployment of APs ChA is an importantnetwork control mechanism to reduce co-channel interferenceand improve capacity Basically a WLAN can be modeled asa conflict graph where the vertices represent APs and edgesreflect the interference level between APs The color of eachvertex is the channel assigned to that AP The objective ofa ChA mechanism is to color the graph using a minimumnumber of colors in order to improve channel reuse

A typical channel assignment algorithm relies on all ora subset of the following inputs (i) the set of APs andclients (ii) AP-client associations (iii) interference relation-ship among APs and clients and (iv) total number of available

Collect distinguishable beacons in each channel

Determine the number of clients associated with

each beacon

Is a non-overlapping channel available

Select the least congested channel

Extract traffic informationfor each client associated

with each beacon

Select the non-overlapping channel

Fig 10 Least Congested Channel Search (LCCS) algorithm [182]

channels Using the aforementioned information optimizationmetrics such as interference throughput spectral efficiencyand fairness are used to define an optimization problem Inthis section we study centralized ChA mechanisms

Figure 10 represents the flowchart of the least congestedchannel search (LCCS) [182] algorithm Although LCCS isnot a central ChA mechanism we briefly discuss its operationas it has been used by commercial off-the-shelf APs andadopted as the baseline for performance comparison of variousChA mechanisms In this algorithm APs scan all channelspassively after startup Each AP collects the beacon framesfrom its neighboring APs and extracts the number of connectedclients and their traffic information Then the AP selects anon-overlapping channel if it is available Otherwise the APselects the channel with lowest traffic sum To obtain client andtraffic information LCCS relies on the use of optional fields inbeacon packets Therefore the use of this mechanism requiresvendor support

A simpler approach which has been adopted as the compar-ison baseline by some ChA mechanisms is to simply samplethe RSSI of nearby APs on all channels and choose the channelwith the minimum RSSI level The literature refers to thisapproach as minRSSI

We overview ChA mechanisms categorized into twogroups (i) traffic-agnostic and (ii) traffic-aware as follows

A Traffic-Agnostic Channel Assignment

These mechanisms do not include the traffic load of APsand the traffic demand of clients in their channel assignmentprocess In fact these mechanisms mainly rely on the interfer-ence relationship to formulate their graph coloring problems

Weighted coloring channel assignment (WCCA) Thisseminal work [183] formulates channel assignment throughthe weighted graph coloring problem and improves channelre-use by assigning partially-overlapping channels to APsVertices are APs and edges represent the potential interferencebetween corresponding APs (see Figure 11) The colors ofthe vertices represent the assigned channel number to APsThe number of clients associated with two neighboring APsand the interference level between them are used to calculatethe weight of conflicting edges Site-Report is the proposedmethod to construct the overlap graph and capture networkdynamics This method enables APs to request their clientsperform channel scanning A list of active APs is generated

APi APjW (APi APj) I(chi chj)

chi chj

Fig 11 Weighted coloring channel assignment (WCCA) [183] formulatesa weighted graph coloring problem This figure represents the I-value of theconflicting edge between APi and APj

for each channel This list includes the APs in direct com-munication range and the APs whose associated clients are indirect communication with the client performing Site-Reportscan Based on the results of Site-Report W (APi APj) iscomputed as follows

W (APi APj) =NAPiAPj +NAPj APi

NAPi+NAPj

where NAPjis the number of site reports performed by the

clients of APj and NAPiAPjis the number of site reports

of APirsquos clients that reported interference with APj or itsassociated clients

The I-factor I(chi chj) is the normalized interferencefactor between two channels chi and chj The valueof I(chi chj) is computed as follows (i) for two non-overlapping channels (eg channel 1 and 6 in the 24GHzband) the value is 0 (ii) if chi = chj the value is 1 (iii)for two partially-overlapping channels chi and chj the valueis between 0 and 1 depending on the distance between theircentral frequency

The weight of the conflicting edge between APi and APj isdefined by Ivalue(APi APj) = W (APi APj)timesI(chi chj)Having K colors (channels) the problem is determining howto color the graph with minimum number of colors in orderto optimize the objective function Three different objectivefunctions Omax Osum and Onum were used Omax repre-sents the maximum I-value amongst all the links Osum is thesum of I-values of all conflicting edges The total number ofconflicting edges is represented by Onum In order to minimizethe objective functions two distributed heuristic algorithms areproposed (i) Hminmax aims to minimize Omax without APcoordination and (ii) Hsum aims to jointly minimize Omax

and Osum which requires coordination between APs in orderto minimize the interference level of all conflicting edgesAlthough the proposed algorithms were run distributively itwould be easier to provide them with their required informa-tion using a SDWLAN architecture

The proposed algorithms have been evaluated using NS2[184] simulator and testbed Two different scenarios are usedin simulations (i) three non-overlapping channels and (ii)partially overlapping channels (11 channels in the 24GHzband) For the first scenario the algorithms showed a 455and 56 improvement in interference reduction for sparse anddense networks respectively For the second scenario there isa 40 reduction in interference compared to LCCS

Client-driven channel assignment (CDCA) [185] formu-lates a conflict set coloring problem and proposes a heuristicalgorithm that increases the number of conflict-free clients

Client

RS(ci) = AP0 AP1 AP2IS(ci) = AP3 AP4 AP5 AP6

Fig 12 Client-driven channel assignment (CDCA) [185] assigns two setsper client range set (denoted by RS(ci)) and interference set (denoted byIS(ci)) APs are specified by numbered squares

at each iteration Here conflict refers to the condition inwhich two nodes (APs or clients) that belong to differentBSSs use the same channel To measure the interference levelexperienced by each client measurement points are selectedand the signal level received from each AP is measuredat those points Based on this information there are twosets assigned to each client (i) range set the APs that cancommunicate with the client directly and (ii) interference setthe APs that cannot communicate with the client but can causeinterference Figure 12 shows the range set and interferenceset for a given client

This work defines two objective functions (i) one thatmaximizes the number of conflict-free clients and (ii) one thatminimizes the total conflict in the network To achieve the firstobjective if a channel ch is assigned to an AP then no APin the range set and interference set of the clients associatedwith that AP should have channel ch assigned to them Aclient satisfying this condition is referred to as a rdquoconflict-freeclientrdquo Since it may not be possible to satisfy this requirementfor all the clients the algorithm minimizes the overall networkconflict

The main step of the channel assignment algorithm ran-domly selects an AP and assigns a channel that results in themaximum number of conflict-free clients This step is repeatedfor all APs and the order of APs is determined by a randompermutation This process is repeated as long as it increasesthe number of conflict-free clients

After the first step there might be clients that are left witha high level of interference To remedy this condition theconflict level of all non-conflict-free clients is balanced Thismechanism implicitly results in traffic load balancing

Frequency planning in large-scale networks (FPLN)[28] models the channel assignment problem as a weightedgraph coloring problem taking into account the external inter-ference generated by non-controllable APs The interferencelevel caused by APj on APi which is the weight of edgeAPj rarr APi is defined as follows

W (APi APj) = CactiveAPitimes I(chi chj)timesP (APi APj) (11)

where CactiveAPiis the number of active associated clients with

APi An associated client is called active in a time interval ifit is sendingreceiving data packets The interference factorbetween two channels assigned to APi and APj is givenby I(chi chj) The power received at APj from APi isrepresented by P (APi APj) The interference factor of APi is

defined assumforallAPjisinAPW (APi APj) The objective function

is defined as the sum of all APsrsquo interference factor includingcontrollable and non-controllable APs The ChA mechanismaims to minimize the objective function when the channels ofnon-controllable APs are fixed

The authors propose a two-phase heuristic algorithm tosolve the above optimization problem In the first phase thecontrollable APs are categorized into separate clusters basedon neighborhood relationships and the local optimizationproblem is solved for each cluster by finding an appropriatechannel per AP to minimize its interference level while thechannels of other APs (in the cluster) are fixed This processis started from the AP with largest interference factor In thesecond phase a pruning-based exhaustive search is run oneach cluster starting from the AP with highest interferencelevel The algorithm updates the channels of APs to reducethe total interference level of the cluster The pruning strategydeletes the failed channel assignments (ie the solutions thatincrease the interference level obtained in the first phase) fromthe search space FPLN uses CAPWAP (see Section III-C1) asits south-bound protocol Simulation and testbed experimentsshow about a 12x lower interference level compared toLCCS

CloudMAC The CloudMAC [26] [50] architecture allowsmultiple NICs of a physical AP to be mapped to a VAPwhere NICs may operate on different channels These NICsperiodically monitor and report channel utilization to the con-troller Unfortunately it is not clear how channel measurementis performed If a client is operating on a high-interferencechannel the controller sends a 80211h channel switch an-nouncement message to the client (mandatory for 80211anstandards) instructing it to switch to another channel withlower interference Since a client is associated with a VAPthe client does not need to re-associate and it can continue itscommunication through the same VAP using another NIC ofthe same physical AP Additionally this process does not re-quire any client modification Unfortunately the performanceevaluation of this mechanism has not been reported

Odin [102] proposes a simple channel assignment strategyFor each AP the channel assignment application (running onOdin controller) samples the RSSI value of all channels duringvarious operational hours The heuristic algorithm chooses thechannel with the smallest maximum and average RSSI for eachAP Unfortunately the performance of this mechanism has notbeen evaluated

Primary channel allocation in 80211ac networks (PCA)The hidden channel (HC) problem is addressed in [29] for80211ac networks The HC problem occurs when an APinterferes with the bandwidth of another AP This problemoccurs due to the heterogeneity of channel bandwidths differ-ent CCA thresholds for primary and secondary channels andbandwidth-ignorant fixed transmission power

Figure 13 illustrates an example of the HC problem Thebandwidth of AP1 is 80MHz including four 20-MHz channels1 2 3 and 4 where channel 1 is its primary channel Theprimary channel sensing range is greater than the secondarychannel sensing range due to the difference in CCA thresholdsfor primary and secondary channels in 80211ac AP2 uses

3AP1 AP2

InvaderAP

VictimAP

Collision1

Primary channel sensing rangeSecondary channels sensing range

Fig 13 An example of the hidden-channel (HC) problem in 80211ac [29]

a 20MHz bandwidth operating on channel 3 Suppose AP2

is communicating with its clients on channel 3 When AP1

performs carrier sensing on its channels it cannot detect thepresence of AP2 on channel 3 because AP2 is not in thesecondary sensing range of AP1 Therefore AP1 starts to senddata on channels 1 2 3 and 4 and invades AP2

In order to study the impact of the HC problem the authorsproposed a Markov chain-based analysis to show the impactof packet length and MAC contention parameters on theperformance of APs Furthermore the effect of various HCscenarios on packet error rate (PER) has been investigatedThe authors used the graph coloring problem where APs arevertices and channels are colors to model the primary channelallocation (PCA) problem The edges represent the invasionrelationship between APs It is assumed that a controllercollects information about the bandwidth of APs and theirinterference relationship The objective is to color the verticesin order to (i) minimize the interferenceinvasion relationshipbetween APs and (ii) maximize channel utilization Theproblem is formulated as an integer programming optimizationproblem which is NP-hard A heuristic primary channel allo-cation algorithm is introduced to solve the problem Simulationresults show the higher performance of PCA compared tominRSSI and random channel assignment Furthermore PCAis a close-to-optimal solution when compared to the optimalexhaustive search algorithm

EmPOWER2 As mentioned in Section III-C3 Em-POWER2 [24] establishes channel quality and interferencemap abstraction at the controller The proposed ChA mech-anism uses the APIs provided to traverse the map and detectuplinkdownlink conflicts between LVAP pairs There is anedge between two nodes in the interference map if they are inthe communication range of each other In addition the weightof each link corresponds to the channel quality between nodesThe graph coloring algorithm proposed in [186] is used toassign channels based on the conflict relationships

Wi-5 channel assignment (Wi5CA) [187] proposes a ChAmechanism as part of the Wi-5 project [188] A binary integerlinear programming problem is formulated with objectivefunction U = GtimesAT I where primetimesprime and primeprime are matrix multipli-cation and element-wise multiplication operators respectivelyG isin 0 1|AP|times|AP| represents the network topology If theaverage power strength of APi on APj exceeds a specificthreshold the value of GAPiAPj

will be 1 otherwise it will be

0 A isin 0 1|CH|times|AP| is the current channel assignment forAPs If channel chi is assigned to APj the value of AchiAPj

is 1 otherwise it is 0 I isin R|AP|times|CH| represents the predictedinterference matrix where IAPichi

is the interference levelpredicted for APi if channel chj is assigned to it Unfor-tunately the interference prediction method is unclear Thechannel assignment algorithm minimizessum

forallAPiisinAP

sumforallchjisinCH

UAPichj (12)

which is the network-wide interference level A controllerexecutes the ChA mechanism when the interference level ishigher than a threshold MATLAB simulations show 2dB and3dB reduction in the average interference level compared toLCCS and uncoordinated channel assignment respectively

B Traffic-Aware Channel Assignment

Traffic-aware channel assignment mechanisms include thetraffic of APs clients or both in their decision process Asmeasuring the effect of interference on performance is com-plicated these mechanisms explicitly include metrics such asthroughput delay and fairness in their problem formulations

AP placement and channel assignment (APCA) [189]addresses AP placement and channel assignment in order toimprove the total network throughput and the fairness estab-lished among clients The throughput of clients is estimated asa function of MAC-layer timing parameters network topologyand data rate of clients The data rate of each client (11 552 or 1Mbps) depends on the linkrsquos RSSI The fairness amongclients is defined as the throughput deviations of clients basedon the Jainrsquos fairness index [190] as follows

J =(sumforallciisinC Thci)

|C|sumforallciisinC (Thci)2 (13)

where Thci is the throughput of client ci Maximizing theobjective function defined as J times sumforallciisinC Thci leads tooptimizing both total throughput and fairness among clientsThe authors propose a heuristic local search method namedthe patching algorithm In each iteration of this algorithm anAP is placed on one of the predefined locations and a non-overlapping channel is assigned to the AP to maximize theobjective function This process is repeated until a predefinednumber of APs are placed Simulation and testbed results showthat the proposed patching algorithm provides close-to-optimalsolutions in terms of total throughput and fairness APCA isused only for the initial network configuration phase

Traffic-aware channel assignment (TACA) [191] assignsweights to APs and clients proportional to their traffic de-mands The weighted channel separation (WCS) metric is

WCS =sum

ijisinAPcupCBSS(i)6=BSS(j)

W (i j)times Separation(i j) (14)

where BSS(k) is the BSS including AP or client kSeparation(i j) is the distance between operating channelsof node i and node j and W (i j) is a function of the trafficdemands of i and j Separation(i j) = min (|chi minus chj | 5)therefore the maximum separation value is 5 which is the

Interference graphmeasurement

Traffic demand collection amp prediction

Step 1Initial channel assignment

Step 2Improving channel assignment using

Simulated Annealing(SA)

New assignment

Old assignment6=

Update channel assignment

SNMP statistics

Interference estimation

Fig 14 Traffic-aware channel assignment (TACA) [191]

distance between orthogonal channels (eg channel 1 and 6)Since all nodes are included in the calculation of the WCSmetric all AP-AP AP-client and client-client interferences aretaken into account

A heuristic algorithm has been proposed to maximize WCSFigure 14 shows the flowchart of this algorithm The interfer-ence graph measurement is performed a few times per day Thetraffic demands of APs and clients are collected and predictedusing the SNMP statistics collected from APs In rdquoStep 1rdquo aninitial channel assignment is performed to maximize the WCSmetric considering APsrsquo traffic demands only In rdquoStep 2rdquo asimulated annealing (SA) approach is used to update channelassignment which maximizes WCS by taking into accountthe traffic demands of AP and clients This step randomlyselects an AP and its clients and assigns a new channel ateach iteration in order to maximize the WCS metric Empiricalresults (using a 25-node testbed) show 26 times increasein TCP throughput compared to a downgraded algorithmin which traffic demands are not included The authors alsoshowed that one channel switching per 5 minutes does notresult in throughput degradation

Virtual delayed time channel assignment (VDTCA) [27]introduces a new interference prediction model based on thesignal strength and traffic demands of APs and clients Theextra transmission delay caused by interference between twoAPs is represented as V DT = Tint minus Tn where V DT isvirtual delayed time The term rdquovirtualrdquo refers to the calcu-lation of this metric through interference prediction withoutactually changing channels Tint and Tn are the time requiredto transmit a given amount of traffic in the presence andabsence of interference respectively

The transmission time is a function of physical layer datarate which depends on the signal SINR computed throughmeasuring RSSI at receiver side On the other hand thetraffic demands of APs and clients have a great impact onthe transmission time of neighboring APs For instance whenthere is no traffic on an AP there will be no transmissiondelay overhead on the neighboring APs and clients ThereforeTint is calculated based on RSSI values and traffic rates ofall neighboring APs and their associated clients Using theproposed VDT model it is possible to measure the VDT ofeach AP pair Channel assignment is modeled as a graph vertexcoloring problem which uses V DT (BSSi BSSj) as the edgeweight between two interfering BSSs where BSSi and BSSj

are two APs and their associated clients Non-overlappingchannels are the colors of vertices The objective is to color the

graph with a minimum number of colors while minimizing thesum of edge weights to achieve interference minimization Thesemi-definite programming (SDP) [192] relaxation techniqueis used to solve this problem The authors show that SDP cansolve the channel assignment problem in the order of secondsfor a medium-size network For instance it takes 275 secondsto find the solution for a network with 50 APs and 9 non-overlapping channels Testbed results show 30 throughputimprovement compared to LCCS [182]

Cisco unified wireless network (CUWN) In the CUWNarchitecture [85] (see Section III-C1) based on the differentRF groups established and the cost metric calculated for eachAP each controller updates the assigned channels periodi-cally4 The controller calculates a cost metric for each AP torepresent the interference level of the AP and prepares a listcalled channel plan change initiator (CPCI) which includesthe sorted cost metric values of all APs The leader selectsthe AP with highest cost metric and assigns a channel tothe selected AP and its one-hop neighboring APs in order todecrease their cost metrics These APs are removed from theCPCI list and this process is repeated for all remaining APsin the list A sub-optimal channel assignment is calculatedthrough a heuristic algorithm5 that aims to maximize thefrequency distance of selected channels Note that a longer fre-quency distance results in a lower interference between APsCUWN also enables network administrator to set channelsmanually

Channel assignment with fairness approach (CAFA)[193] formulates ChA as a weighted graph coloring problemwhere weights are assigned to the set of edges and verticesThe distance of overlapping channels is determined by thenormalized interference factor ie I-factor introduced in[183] (see WCCA explained in Section V-A) The normalizedthroughput of each client is estimated using the approximationintroduced in [194] The weight of APi is defined as the totalnormalized throughput reduction of APi which is calculatedbased on clientsrsquo normalized throughput and I-factor Theweight of the edge between two interfering APs representsthe throughput reduction caused by their mutual interferenceFurthermore the authors formulate the fairness among nor-malized throughputs of APs using Jainrsquos fairness index [190]The objective function is defined as the joint minimizationof clientsrsquo throughput reduction and maximization of clientrsquosfairness index

An iterative heuristic algorithm is introduced to solve theoptimization problem The algorithm sorts the APs based ontheir weights (ie normalized throughput reduction) At eachiteration the AP with highest weight is selected for channelassignment The channel of the selected AP is assigned so thatthe throughput reduction of the neighboring APs is minimizedThis process is repeated until channel assignment is performedfor all APs Simulation results show 15 and 6 improvementin total throughput for 4-AP and 8-AP scenarios respectivelycompared to WCCA [183]

Throughput and fairness-aware channel assignment

4The period can be adjusted by network administrators5The details of ChA mechanism used by CUWN are not available

(TFACA) [195] uses spectrum monitoring information andthe traffic demand of APs This mechanism defines the utilityfunction of APi when operating on channel chj as

U(APi chj) =min F (APi chj) T (APi)

T (APi) (15)

where F (APi chj) is the free airtime of channel chj on APiand T (APi) is the total time required to send data of APi in atime unit Note that Equation 15 represents the percentage ofAPirsquos data that can be sent over channel chj in a given timeunit Two objective functions are defined

maxsum

forallAPiisinAPforallchjisinCH

U(APi chj) (16)

andmax( min

forallAPiisinAPforallchjisinCHU(APi chj)) (17)

A heuristic algorithm has been proposed to solve the aboveNP-hard problem in the following two steps (i) temporarychannel assignment which assigns a channel to each AP inde-pendent of other APs to maximize each APrsquos utility functionand (ii) channel reassignment of the APs with low utilityfunction (starting with the lowest one) without decreasing thesum of all utility function values The second step is repeatedfor a given number of iterations or until U(APi chj) = 1for all APs A fairness index similar to Equation 13 has beendefined to include the utility function of APs

The 80211g protocol with non-overlapping channels areused for performance evaluation through simulation andtestbed An 8 to 15 improvement in the fairness index andan 8 to 21 increase in the mean utility function are achievedcompared to minRSSI

Frequency and bandwidth assignment for 80211nac(FBWA) [196] addresses channel and bandwidth allocationto benefit from the channel bonding feature of 80211nacThe utility function of APi that works on primary channelchj and bandwidth β is defined as

U(APi chj β) =Th(APi chj β)

minThmax(APi chj β) Lmean

where Th(APi chj β) is the expected throughput of APi

operating on primary channel chj and channel bandwidth βThmax(APi chj β) is the maximum achievable throughput ofAPi and Lmean

APiis the mean traffic load on APi Through a

heuristic algorithm similar to TFACA [195] the appropriateprimary channels and bandwidths are determined for all APsto maximize the sum of the APsrsquo utility functions

Testbed results show a 65 to 89 reduction in the numberof overlapping BSSs and a mean throughput that is 3 timeshigher compared to random channel assignment and minRSSI

Normalized-airtime channel assignment (NATCA) [197]requires each AP to measure its busy time (referred to asnormalized airtime) and report to a controller periodicallyThe busy time of an AP is the percentage of time that achannel assigned to that AP is occupied by its neighborsAn AP measures this value when sending data towards itsclients A threshold value is defined to categorize APs into twogroups based on their busy time heavily congested and lightly

congested An optimization problem has been proposed tominimize the sum of the normalized airtime of all APs withoutmodifying the status of lightly congested APs A Tabu searchalgorithm is proposed to solve the optimization problem NS3simulations show a 15x throughput improvement comparedto LCCS [182] The measurement of normalized airtime forall APs is performed passively using the NS3 tracing systemwhich provides packet-level trace files for all APs

C Channel Assignment Learned Lessons Comparison andOpen Problems

Table IV presents and compares the features of ChA mech-anisms In the following we study these features and identifyresearch directions

1) Dynamicity and Traffic-Awareness As Table IV showsnot all the ChA mechanisms support dynamic channel re-allocation In addition most of the ChA mechanisms areeither traffic-agnostic or they do not recognize uplink trafficFurthermore many of the ChA mechanisms do not actuallydiscuss the data gathering policy employed and the rest focuson the mean traffic load of APs and the mean throughputdemand of clients These limitations are due to two mainreasons First the higher the rate of central interference mapgeneration the higher the overhead of control data exchangedby the controller Second although heuristic algorithms havebeen proposed to tackle the NP-hardness of the channel as-signment problems the execution duration of these algorithmsmight not be short enough to respond to network dynamicsThe two approaches earlier proposed in Section IV-C3 (ieprediction and hybrid design) can be employed to cope withthese challenges For example depending on the time windowand accuracy of predictions a ChA proactively responds todynamics which in turn reduces the burden on real-timenetwork mapping and fast algorithm execution The secondsolution ie hybrid design employs multi-level decisionmaking to support fast and low-overhead reaction to networkdynamics For the hybrid designs however the topology ofcontrollers may be tailored depending on the control mech-anism employed For example while a suitable controllertopology for a ChA mechanism depends on the interferencerelationship between APs a more suitable topology for an AsCmechanism is to connect all the APs of a hallway to a localcontroller to improve the QoS of clients walking in that areaTherefore it is important to design architectures that supportflexible communication between controllers

2) Joint ChA and AsC As channel switching may leadto re-association (depending on the architecture used) clientsmay experience communication interruption and violation oftheir QoS requirements [111] [198] [199] Therefore it isnecessary to measure and reduce the channel switching over-head of dynamic ChA mechanisms In this regard joint designof ChA and AsC is essential for uninterrupted performanceguarantee Such a joint mechanism for example monitors andpredicts clientsrsquo traffic pattern performs channel assignmentto maximize spatial reuse and moves VAPs between APs tosupport seamless handoff as the traffic changes In additionmobility prediction [200] [201] may be employed to minimize

TABLE IVCOMPARISON OF CHANNEL ASSIGNMENT (CHA) MECHANISMS

Mechanism Traffic-Aware Standards amp Channels Dynamic Performance EvaluationDownlink Uplink Standard Partially Overlapping Simulation Testbed

WCCA [183] times times 80211b X X X XCDCA [185] times times 80211b times X X XFPLN [28] times times 80211b X times X X

CloudMAC [26] times times 80211abgn X X times XOdin [102] times times 80211abgn X X times XPCA [29] times times 80211ac times X X times

EMPOWER2 [24] times times 80211abgnac times X times XWi5CA [187] times times 80211bg times X X timesAPCA [189] times X 80211b times times X timesTACA [191] X X 80211bg X X X XVDTCA [27] X X 80211ag times X times XCUWN [85] X X 80211abgnac X X times timesCAFA [193] X X 80211b X times X times

TFACA [195] X times 80211g times times X XFBWA [196] X times 80211anac times times X times

NATCA [197] X times 80211g times X X times

the delay of AsC and ChA by enabling the use of proactivecontrol mechanisms instead of using reactive mechanisms thatare triggered by performance drop Note that mobility predic-tion is specifically useful in SDWLANs as these architecturesenable the central collection and analysis of metrics such asRSSI to apply localization and mobility prediction algorithmsUnfortunately the existing joint ChA-AsC mechanisms (eg[202] [203]) operate distributively and do not provide anyperformance guarantee which is required for mission-criticalapplications such as medical monitoring or industrial control[17]

3) Partially Overlapping Channels and AP Density Man-agement Most ChA mechanisms designed for 24GHz net-works utilize only non-overlapping channels to decrease theinterference level (as shown in Table IV) However due tothe dense deployment of APs this approach limits the trade-off between interference reduction and channel reuse [204]To address this concern research studies have improved theperformance of channel assignment by dynamically assigningpartially-overlapping channels [193] These mechanisms forexample rely on the interference relationship between APsto increase the distance between assigned channels as thepairwise interference increases However these mechanismsrequire the underlying SDWLAN to provide APIs for efficientcollection of network statistics

In addition to utilizing partially-overlapping channels ChAmechanisms can further improve network capacity throughAP topology management which is achieved by three mainstrategies (i) AP placement the site survey is performed to de-termine the location of APs during the installation phase [205](ii) AP power control a SDWLAN controller dynamicallyadjusts the transmission power level of APs [206] (iii) APmode control a SDWLAN controller dynamically turns onoffAPs [107] Note that the second and third strategy requirearchitectural support It is also important to coordinate theaforementioned strategies with AsC to avoid clientsrsquo serviceinterruption For example by relying on the global networkview established clients association may be changed beforechannel assignment or transmission power control cause clientdisconnection Although these problems have been addressedin isolation or distributively [40] [207]ndash[209] centralized and

integrated solutions are missing4) Integration with Virtualization Our review of ChA and

AsC mechanisms shows that these mechanisms are obliviousto virtualization Specifically they do not take into accounthow the pool of resources is assigned to various network slicesHowever virtualization has serious implications on networkcontrol For example mobility management becomes morechallenging when network virtualization is employed When aclient requires a new point of association due to its mobility inaddition to parameters such as fair bandwidth allocation theavailable resources of APs should also be taken into accountMore specifically for a client belonging to slice n the AsCmechanism should ensure that after the association of thisclient with a new AP the QoS provided by slice n and otherslices is not violated However as this may require clientsteering the re-association cost of other clients should beminimized As an another example a ChA mechanism mayassign non-overlapping and overlapping channels to a mission-critical slice and a regular slice respectively The problembecomes even more complicated when multiple controllerscollaborate to manage network resources The integration ofnetwork slicing and control mechanisms would pave the wayto use SDWLANs in emerging applications such as mission-critical data transfer from mobile and IoT devices

In addition to network-based slicing of resources it is desir-able to support personalized mobility services as well To thisend AsC mechanisms require architectural support to enableclient association based on a variety of factors As an examplea clientrsquos VAP stores the clientrsquos handset features and exposesthat information to the AsC mechanism to associate the clientwith APs that result in minimum energy consumption of theclient Such interactions between architectural components andcontrol mechanisms are unexplored

Another important challenge of slicing at low level isthe complexity of coordination For example assume thatdifferent transmission powers are assigned to flows In thiscase simply changing the transmission power of APs workingon the same channel would cause significant interference anddisrupt the services being offered Therefore offering low-level virtualization requires coordination between the networkcontrol mechanisms

5) Coexistence of High-Throughput and Legacy ClientsAlthough new WiFi standards (eg 80211nac) supportinghigh throughput are broadly used in various environmentswe should not neglect the existence of legacy devices (eg80211b) especially for applications that require low through-put [1] [37] [38] In heterogeneous networks the throughputof high-rate 80211ac devices is jeopardized by legacy clientsdue to the wider channel widths used by 80211nac throughchannel bonding [210] [211] Specifically the channel bond-ing feature of 80211nac brings about new challenges such ashidden-channel problem partial channel blocking and middle-channel starvation [29] [211] [212] The bandwidth of a80211nac device (80160MHz) is blocked or interrupted bya legacy client (20MHz) due to the partial channel blockingand hidden-channel problem In the middle-channel starvationproblem a wideband client is starved because its bandwidthoverlaps with the channels of two legacy clients Hencethe wideband client can use its entire bandwidth only ifboth legacy clients are idle Despite the heterogeneity ofWLAN environments our review shows that addressing thechallenges of multi-rate networks to satisfy the diverse setof QoS requirements is still premature In fact only PCA[29] proposes a central solution and its performance has beencompared with distributed mechanisms

6) Security Wireless communication is susceptible to jam-ming attacks through which malicious signals generated ona frequency band avoid the reception of user traffic on thatchannel The vulnerability of 80211 networks to such de-nial of service (DoS) attacks have been investigated throughvarious empirical measurements [213]ndash[215] Although theresearch community proposes various mechanisms (such asfrequency hopping [216]) to cope with this problem theexisting approaches do not benefit from a central network viewIn particular by relying on the capabilities of SDWLANsclients and APs can periodically report their operating channelinformation (such as channel access time noise level andpacket success rate) to the controller and mechanisms arerequired to exploit this information to perform anomaly de-tection

VI RELATED SURVEYS

In this section we mention the existing surveys relevant toSDN and WLAN

A comprehensive study of software-defined wireless cellu-lar sensor mesh and home networks is presented in [217]This work classifies the existing contributions based on theirobjective and architecture A survey of data plane and controlplane programmability approaches is given in [218] where theauthors review the components of programmable wired andwireless networks as well as control plane architectures andvirtualization tools [219] presents a survey of virtualizationand its challenges in wireless networks A survey of OpenFlowconcepts abstractions and its potential applications in wiredand wireless networks is given in [220] [3] and [221] reviewthe 80211 standards for high-throughput applications TVwhite spaces and machine-to-machine communications [39]surveys energy-efficient MAC protocols for WLANs A surveyof cellular to WLAN offloading mechanisms is given in [4]

Channel assignment strategies are reviewed in [222] and thedynamics of channel assignment in dense WLANS are inves-tigated in [223] QoS support in dense WLANs through layer-1 and layer-2 protocols are surveyed in [224]ndash[226] Client-based association mechanisms and AP placement strategiesare surveyed in [227] Distributed load-balancing mechanismsare reviewed in [228] [229] provides a survey on spectrumoccupancy measurement techniques in order to generate in-terference maps in WLANs In [230] the future demands ofWLANs and the potential new features of ongoing 80211axstandard are investigated and reviewed A review of wirelessvirtualization mechanisms is given in [231] This work alsopresents the business models enabling technologies and per-formance metrics of wireless virtual networks

Our review of existing surveys shows that there is nostudy of WLANs from a SDN point of view Thereforethe main motivation for conducting this study was the lackof a comprehensive survey on SDWLAN architectures andessential centralized network control mechanisms

VII CONCLUSION

SDWLANs enable the implementation of network controlmechanisms as applications running on a network operatingsystem Specifically SDWLAN architectures provide a set ofAPIs for network monitoring and dissemination of controlcommands Additionally network applications can benefitfrom the global network view established in the controllerto run network control mechanisms in a more efficient waycompared to distributed mechanisms In this paper we cat-egorized and reviewed the existing SDWLAN architecturesbased on their main contribution including observabilityand configurability programmability virtualization scalabil-ity traffic shaping and home networks Through comparingthe existing architectures as well as identifying the growingand potential applications of SDWLANs we proposed severalresearch directions such as establishing trade-off betweencentralization and scalability investigating and reducing theoverhead of south-bound protocols exploiting layer-1 andlayer-2 programmability for virtualization and opportunitiesand challenge of SDWLANs with respect to security provi-sioning

After reviewing architectures we focused on associationcontrol (AsC) and channel assignment (ChA) which are thetwo widely-employed network control mechanisms proposedto benefit from the features of SDWLAN architectures We re-viewed centralized AsC and ChA mechanisms in terms of theirmetrics used problem formulation and solving techniques andthe results achieved compared to distributed mechanisms

At a high level we categorized AsC mechanisms into twogroups seamless handoff used to reduce the overhead of clienthandoff and client steering to improve the performance ofclients Furthermore we specified several research directionstowards improving AsC mechanisms such as the design andmeasurement of metrics to include both uplink and downlinktraffic into the decision making process establishing propor-tional demand-aware fairness among clients and including theapplication-level demand and behavior of clients in formulat-ing AsC problems

We reviewed central ChA mechanisms and we focused onthe approaches employed by these mechanisms to model theinterference relationship among APs and clients Based onthe inclusion of traffic in the decision making process wecategorized ChA mechanisms into traffic-aware and traffic-agnostic Our comparisons revealed that given the hetero-geneity of clients the new ChA mechanisms should addressthe coexistence of low- and high-throughput devices Wealso discussed the challenges and potential solutions for theintegration of virtualization with AsC and ChA as well as theimportance and potential benefits of AsC and ChA co-design

VIII ACKNOWLEDGMENT

This research is partially supported by a grant from theCypress Semiconductor Corporation (Grant No CYP001)

REFERENCES

[1] S Tozlu M Senel W Mao and A Keshavarzian ldquoWi-Fi enabledsensors for internet of things A practical approachrdquo IEEE Communi-cations Magazine vol 50 no 6 pp 134ndash143 2012

[2] M Baird B Ng and W Seah ldquoWiFi Network Access Control forIoT Connectivity with Software Defined Networkingrdquo in 8th ACM onMultimedia Systems Conference (MMSys) 2017 pp 343ndash348

[3] H A Omar K Abboud N Cheng K R Malekshan A T Gamageand W Zhuang ldquoA survey on high efficiency wireless local areanetworks Next generation wifirdquo IEEE Communications Surveys ampTutorials vol 18 no 4 pp 2315ndash2344 2016

[4] Y He M Chen B Ge and M Guizani ldquoOn wifi offloading inheterogeneous networks Various incentives and trade-off strategiesrdquoIEEE Communications Surveys amp Tutorials vol 18 no 4 pp 2345ndash2385 2016

[5] ldquoWi-fi alliance publishes 7 for rsquo17 wi-fi predictionsrdquo[Online] Available httpwwwwi-fiorgnews-eventsnewsroomwi-fi-alliance-publishes-7-for-17-wi-fi-predictions

[6] ldquoEricsson mobility reportrdquo 2016 [Online] Available httpswwwericssoncomresdocs2016ericsson-mobility-report-2016pdf

[7] ldquoGlobal wi-fi enabled equipment shipments from 2012 to 2017 (inbillion units)rdquo [Online] Available httpswwwstatistacomstatistics282512global-wi-fi-enabled-equipment-shipments

[8] ldquoCisco visual networking index Global mobile data trafficforecast update 2015-2020 white paperrdquo [Online] Availablehttpwwwciscocomcenussolutionscollateralservice-providervisual-networking-index-vnimobile-white-paper-c11-520862html

[9] K Lee J Lee Y Yi I Rhee and S Chong ldquoMobile data offloadingHow much can wifi deliverrdquo IEEEACM Transactions on Networkingvol 21 no 2 pp 536ndash550 2013

[10] P Serrano P Salvador V Mancuso and Y Grunenberger ldquoEx-perimenting with commodity 80211 hardware Overview and futuredirectionsrdquo IEEE Communications Surveys amp Tutorials vol 17 no 2pp 671ndash699 2015

[11] S Biswas J Bicket E Wong R Musaloiu-E A Bhartia andD Aguayo ldquoLarge-scale measurements of wireless network behaviorrdquoin SIGCOMM 2015 pp 153ndash165

[12] R Murty J Padhye R Chandra A Wolman and B Zill ldquoDesigninghigh performance enterprise wi-fi networksrdquo in 5th USENIX Sympo-sium on Networked Systems Design and Implementation April 2008pp 73ndash88

[13] I Broustis K Papagiannaki S V Krishnamurthy M Faloutsos andV P Mhatre ldquoMeasurement-driven guidelines for 80211 wlan designrdquoIEEEACM Transactions on Networking vol 18 no 3 pp 722ndash7352010

[14] K Sui S Sun Y Azzabi X Zhang Y Zhao J Wang Z Li andD Pei ldquoUnderstanding the impact of ap density on wifi performancethrough real-world deploymentrdquo in IEEE International Symposium onLocal and Metropolitan Area Networks (LANMAN) 2016

[15] J Choi and K G Shin ldquoQos provisioning for large-scale multi-apwlansrdquo Ad Hoc Networks vol 10 no 2 pp 174ndash185 2012

[16] F Tramarin S Vitturi M Luvisotto and A Zanella ldquoOn the use ofieee 80211 n for industrial communicationsrdquo IEEE Transactions onIndustrial Informatics vol 12 no 5 pp 1877ndash1886 2016

[17] B Dezfouli M Radi and O Chipara ldquoREWIMO A Real-Time andLow-Power Mobile Wireless Networkrdquo ACM Transactions on SensorNetworks (TOSN) vol 13 no 3 2017

[18] Y Jarraya T Madi and M Debbabi ldquoA survey and a layered taxonomyof software-defined networkingrdquo IEEE Communications Surveys ampTutorials vol 16 no 4 pp 1955ndash1980 2014

[19] D B Rawat and S R Reddy ldquoSoftware defined networking architec-ture security and energy efficiency A surveyrdquo IEEE CommunicationsSurveys amp Tutorials vol 19 no 1 pp 325ndash346 2017

[20] L E Li Z M Mao and J Rexford ldquoToward software-defined cellularnetworksrdquo in EWSDN 2012 pp 7ndash12

[21] H Ali-Ahmad C Cicconetti A De la Oliva M Draxler R GuptaV Mancuso L Roullet and V Sciancalepore ldquoCrowd an sdn approachfor densenetsrdquo in European Workshop on Software Defined Networks(EWSDN) 2013 pp 25ndash31

[22] T Chen M Matinmikko X Chen X Zhou and P AhokangasldquoSoftware defined mobile networks concept survey and researchdirectionsrdquo IEEE Communications Magazine vol 53 no 11 pp 126ndash133 2015

[23] I T Haque and N Abu-ghazaleh ldquoWireless software defined net-working A survey and taxonomyrdquo IEEE Communications Surveys ampTutorials vol 18 no 4 pp 2713ndash2737 2016

[24] R Riggio M K Marina J Schulz-Zander S Kuklinski andT Rasheed ldquoProgramming abstractions for software-defined wirelessnetworksrdquo IEEE Transactions on Network and Service Managementvol 12 no 2 pp 146ndash162 2015

[25] J Schulz-Zander C Mayer B Ciobotaru S Schmid and A Feld-mann ldquoOpensdwn Programmatic control over home and enterprisewifirdquo in SIGCOMM Symposium on Software Defined NetworkingResearch ACM 2015 p 16

[26] J Vestin P Dely A Kassler N Bayer H Einsiedler and C PeyloldquoCloudmac Towards software defined wlansrdquo ACM SIGMOBILE Mo-bile Computing and Communications Review vol 16 no 4 pp 42ndash45February 2013

[27] Y Liu W Wu B Wang T He S Yi and Y Xia ldquoMeasurement-basedchannel management in wlansrdquo in IEEE Wireless Communication andNetworking Conference (WCNC) April 2010 pp 1ndash6

[28] M Bernaschi F Cacace A Davoli D Guerri M Latini andL Vollero ldquoA capwap-based solution for frequency planning in largescale networks of wifi hot-spotsrdquo Computer Communications vol 34no 11 pp 1283ndash1293 2011

[29] S Jang and S Bahk ldquoA channel allocation algorithm for reducingthe channel sensingreserving asymmetry in 80211ac networksrdquo IEEETransactions on Mobile Computing vol 14 no 3 pp 458ndash472 March2015

[30] L Oliveira K Obraczka and A Rodrguez ldquoCharacterizing UserActivity in WiFi Networks University Campus and Urban Area CaseStudiesrdquo in MSWiM 2016 pp 190ndash194

[31] A Patro and S Banerjee ldquoCoap A software-defined approach forhome wlan management through an open apirdquo in 9th ACM workshopon Mobility in the evolving internet architecture (MobiArch rsquo14)September 2014 pp 31ndash36

[32] S Zehl A Zubow M Doring and A Wolisz ldquoResfi A secureframework for self organized radio resource management in residentialwifi networksrdquo in IEEE 17th International Symposium on A World ofWireless (WoWMoM) 2016 pp 1ndash11

[33] J-Q Li F R Yu G Deng C Luo Z Ming and Q Yan ldquoIndustrialinternet A survey on the enabling technologies applications andchallengesrdquo IEEE Communications Surveys amp Tutorials vol 19 no 3pp 1504ndash1526 2017

[34] M Luvisotto Z Pang and D Dzung ldquoUltra high performance wirelesscontrol for critical applications Challenges and directionsrdquo IEEETransactions on Industrial Informatics vol 13 no 3 pp 1448ndash14592017

[35] Y h Wei Q Leng S Han A K Mok W Zhang and M TomizukaldquoRt-wifi Real-time high-speed communication protocol for wirelesscyber-physical control applicationsrdquo in IEEE 34th Real-Time SystemsSymposium (RTSS) 2013 pp 140ndash149

[36] J Silvestre-Blanes L Almeida R Marau and P Pedreiras ldquoOnlineqos management for multimedia real-time transmission in industrialnetworksrdquo IEEE Transactions on Industrial Electronics vol 58 no 3pp 1061ndash1071 2011

[37] ldquoCYW43907 WICED IEEE 80211 abgn SoC with an EmbeddedApplications Processorrdquo [Online] Available httpwwwcypresscomfile298236download

[38] ldquoBCM4343 Single-Chip 80211 bgn MACBasebandRadiordquo[Online] Available httpwwwcypresscomfile298081download

[39] S-L Tsao and C-H Huang ldquoA survey of energy efficient macprotocols for ieee 80211 wlanrdquo Computer Communications vol 34no 1 pp 54ndash67 2011

[40] V P Mhatre K Papagiannaki and F Baccelli ldquoInterference mitigationthrough power control in high density 80211 wlansrdquo in INFOCOM2007 pp 535ndash543

[41] A Kashyap U Paul and S R Das ldquoDeconstructing interferencerelations in wifi networksrdquo in SECON June 2010 pp 1ndash9

[42] S Gollakota F Adib D Katabi and S Seshan ldquoClearing the RFSmog Making 80211 Robust to Cross-Technology Interferencerdquo inACM SIGCOMM 2011 pp 170ndash181

[43] K Pelechrinis I Broustis S V Krishnamurthy and C GkantsidisldquoA measurement-driven anti-jamming system for 80211 networksrdquoIEEEACM Transactions on Networking vol 19 no 4 pp 1208ndash12222011

[44] A Zanella ldquoBest practice in rss measurements and rangingrdquo IEEECommunications Surveys amp Tutorials vol 18 no 4 pp 2662ndash26862016

[45] B Dezfouli M Radi S A Razak H-P Tan and K A BakarldquoModeling low-power wireless communicationsrdquo Journal of Networkand Computer Applications vol 51 pp 102ndash126 may 2015

[46] D Giustiniano D Malone D J Leith and K Papagiannaki ldquoMea-suring transmission opportunities in 80211 linksrdquo IEEEACM TransNetw vol 18 no 5 pp 1516ndash1529 2010

[47] L Deek E Garcia-Villegas E Belding S-J Lee and K AlmerothldquoJoint rate and channel width adaptation for 80211 mimo wirelessnetworksrdquo in Sensor Mesh and Ad Hoc Communications and Networks(SECON) 2013 10th Annual IEEE Communications Society Confer-ence on IEEE 2013 pp 167ndash175

[48] A Faridi B Bellalta and A Checco ldquoAnalysis of dynamic channelbonding in dense networks of wlansrdquo IEEE Transactions on MobileComputing vol 16 no 8 pp 2118ndash2131 2017

[49] L Suresh J Schulz-Zander R Merz A Feldmann and T VazaoldquoTowards programmable enterprise wlans with odinrdquo in HotSDNAugust 2012 pp 115ndash120

[50] J Vestin and A Kassler ldquoQos enabled wifi mac layer processing asan example of a nfv servicerdquo in NetSoft April 2015 pp 1ndash9

[51] K-C Ting F-C Kuo B-J Hwang H Wang and C-C Tseng ldquoApower-saving and robust point coordination function for the transmis-sion of voip over 80211rdquo in International Symposium on Parallel andDistributed Processing with Applications (ISPA) IEEE 2010 pp283ndash289

[52] Z Zeng Y Gao and P Kumar ldquoSofa A sleep-optimal fair-attentionscheduler for the power-saving mode of wlansrdquo in ICDCS IEEE2011 pp 87ndash98

[53] L Liu X Cao Y Cheng and Z Niu ldquoEnergy-efficient sleep schedul-ing for delay-constrained applications over wlansrdquo IEEE Transactionson Vehicular Technology vol 63 no 5 pp 2048ndash2058 2014

[54] B Dezfouli M Radi and O Chipara ldquoMobility-aware real-timescheduling for low-power wireless networksrdquo in 35th Annual IEEEInternational Conference on Computer Communications (INFOCOM)2016 pp 1ndash9

[55] N Gude T Koponen J Pettit B Pfaff M Casado N McKeown andS Shenker ldquoNox towards an operating system for networksrdquo ACMSIGCOMM Computer Communication Review vol 38 no 3 pp 105ndash110 2008

[56] ldquoFloodlightrdquo [Online] Available httpwwwprojectfloodlightorgfloodlight

[57] D Levin A Wundsam B Heller N Handigol and A FeldmannldquoLogically centralized State distribution trade-offs in software definednetworksrdquo HotSDN pp 1ndash6 2012

[58] S H Yeganeh A Tootoonchian and Y Ganjali ldquoOn scalabilityof software-defined networkingrdquo IEEE Communications Magazinevol 51 no 2 pp 136ndash141 2013

[59] N McKeown T Anderson H Balakrishnan G Parulkar L PetersonJ Rexford S Shenker and J Turner ldquoOpenflow Enabling innovationin campus networksrdquo ACM SIGCOMM Computer CommunicationReview vol 38 no 2 pp 69ndash74 2008

[60] S Waldbusser M Rose J Case and K McCloghrie ldquoProtocoloperations for version 2 of the simple network management protocol(SNMPv2)rdquo [Online] Available httpstoolsietforghtmlrfc1448

[61] A Wang M Iyer R Dutta G N Rouskas and I Baldine ldquoNetworkvirtualization Technologies perspectives and frontiersrdquo Journal ofLightwave Technology vol 31 no 4 pp 523ndash537 2013

[62] R Sherwood G Gibb K-K Yap G Appenzeller M CasadoN McKeown and G Parulkar ldquoFlowvisor A network virtualizationlayerrdquo OpenFlow Switch Consortium Tech Rep vol 1 p 132 2009

[63] W Zhou L Li M Luo and W Chou ldquoREST API design patterns forSDN northbound APIrdquo in WAINA IEEE 2014 pp 358ndash365

[64] L Richardson and S Ruby RESTful web services OrsquoReilly MediaInc 2008

[65] N Foster R Harrison M J Freedman C Monsanto J RexfordA Story and D Walker ldquoFrenetic A network programming languagerdquoin ACM Sigplan Notices vol 46 no 9 ACM 2011 pp 279ndash291

[66] J Reich C Monsanto N Foster J Rexford and D Walker ldquoModularsdn programming with pyreticrdquo Technical Report of USENIX 2013

[67] C Monsanto N Foster R Harrison and D Walker ldquoA compilerand run-time system for network programming languagesrdquo in ACMSIGPLAN Notices vol 47 no 1 ACM 2012 pp 217ndash230

[68] C Trois M D Del Fabro L C E de Bona and M Martinello ldquoASurvey on SDN Programming Languages Toward a Taxonomyrdquo IEEECommunications Surveys amp Tutorials vol 18 no 4 pp 2687ndash27122016

[69] J Medved R Varga A Tkacik and K Gray ldquoOpendaylight Towardsa model-driven sdn controller architecturerdquo in World of WirelessMobile and Multimedia Networks (WoWMoM) 2014 IEEE 15th In-ternational Symposium on a IEEE 2014 pp 1ndash6

[70] J D Case M Fedor M L Schoffstall and J Davin ldquoSimple networkmanagement protocol (snmp)rdquo Tech Rep 1990

[71] D Harrington R Presuhn and B Wijnen ldquoRFC 3411 Anarchitecture for describing simple network management protocol(SNMP) management frameworksrdquo 2002 [Online] Available httpstoolsietforghtmlrfc3411

[72] W Stallings SNMP SNMPv2 SNMPv3 and RMON 1 and 2 Addison-Wesley Longman Publishing Co Inc 1998

[73] A Clemm Network management fundamentals Cisco Press 2006[74] M K Kona and C-Z Xu ldquoA framework for network management

using mobile agentsrdquo in ipdps 2002[75] F A Silva L B Ruiz T R M Braga J M S Nogueira and A A F

Loureiro ldquoDefining a wireless sensor network management protocolrdquoin LANOMS 2005 pp 39ndash50

[76] R Enns M Bjorklund and J Schoenwaelder ldquoNetwork configurationprotocol (netconf)rdquo 2011 [Online] Available httpstoolsietforghtmlrfc6241html

[77] D Stanley P Calhoun and M Montemurro ldquoControl and provisioningof wireless access points (CAPWAP) protocol specificationrdquo [Online]Available httpstoolsietforghtmlrfc5415

[78] N Modadugu and E Rescorla ldquoDatagram transport layer security(DTLS)rdquo [Online] Available httpstoolsietforghtmlrfc4347

[79] C Shao H Deng R Pazhyannur F Bari R Zhang and S Mat-sushima ldquoIEEE 80211 Medium Access Control (MAC) Profile forControl and Provisioning of Wireless Access Points (CAPWAP)rdquo TechRep 2015

[80] Broadband Forum ldquoCPE WAN Management Protocol (CWMP)rdquo[Online] Available httpswwwbroadband-forumorgtechnicaldownloadTR-069pdf

[81] J Swetina G Lu P Jacobs F Ennesser and J Song ldquoToward astandardized common m2m service layer platform Introduction toonem2mrdquo IEEE Wireless Communications vol 21 no 3 pp 20ndash262014

[82] A Al-Fuqaha A Khreishah M Guizani A Rayes and M Moham-madi ldquoToward better horizontal integration among iot servicesrdquo IEEECommunications Magazine vol 53 no 9 pp 72ndash79 2015

[83] P Berde M Gerola J Hart Y Higuchi M Kobayashi T KoideB Lantz B OrsquoConnor P Radoslavov W Snow et al ldquoONOS towardsan open distributed SDN OSrdquo in Proceedings of the third workshopon Hot topics in software defined networking ACM 2014 pp 1ndash6

[84] ldquoRyu SDN Frameworkrdquo [Online] Available httposrggithubioryu[85] J Smith J Woodhams and R Marg Controller-Based Wireless

LAN Fundamentals An end-to-end reference guide to design deploymanage and secure 80211 wireless networks Cisco Press 2011

[86] R Suri N C Winget M Williams S Hares B OrsquoHara andS Kelly ldquoLightweight Access Point Protocol (LWAPP)rdquo [Online]Available httpstoolsietforghtmlrfc5412

[87] ldquoRadio resource management (RRM) under unified wirelessnetworksrdquo [Online] Available httpwwwciscocomcenussupportdocswireless-mobilitywireless-lan-wlan71113-rrm-newhtml

[88] P Zerfos G Zhong J Cheng H Luo S Lu and J J-R Li ldquoDirac Asoftware-based wireless router systemrdquo in MobiCom September 2003pp 230ndash244

[89] R Murty J Padhye A Wolman and M Welsh ldquoAn architecture forextensible wireless lansrdquo in ACM Hotnets 2008 pp 79ndash84

[90] R Murty J Padhye A Wolman and M Welsh ldquoDyson An architec-ture for extensible wireless lansrdquo in USENIX June 2010

[91] R Chandra J Padhye A Wolman and B Zill ldquoA location-basedmanagement system for enterprise wireless lansrdquo in NSDI April 2007

[92] M Yan J Casey P Shome A Sprintson and A Sutton ldquoaeligtherflowPrincipled wireless support in sdnrdquo in IEEE 23rd International Con-ference on Network Protocols (ICNP) 2015 pp 432ndash437

[93] ldquoCPqD SoftSwitchrdquo [Online] Available httpsgithubcomCPqDofsoftswitch13

[94] ldquoOpenWRTrdquo [Online] Available httpsopenwrtorg[95] ldquoHostapdrdquo [Online] Available httpsw1fihostapd[96] G Bhanage D Vete I Seskar and D Raychaudhuri ldquoSplitap Lever-

aging wireless network virtualization for flexible sharing of wlansrdquo inGLOBECOM December 2010

[97] A Zubow S Zehl and A Wolisz ldquoBIGAP - Seamless handover inhigh performance enterprise IEEE 80211 networksrdquo in IEEEIFIPNetwork Operations and Management Symposium (NOMS) 2016 pp445ndash453

[98] G Smith A Chaturvedi A Mishra and S Banerjee ldquoWirelessvirtualization on commodity 80211 hardwarerdquo in Proceedings of thesecond ACM international workshop on Wireless network testbedsexperimental evaluation and characterization ACM 2007 pp 75ndash82

[99] K-K Yap R Sherwood M Kobayashi T-Y Huang M ChanN Handigol N McKeown and G Parulkar ldquoBlueprint for intro-ducing innovation into wireless mobile networksrdquo in Second ACMSIGCOMM workshop on Virtualized infrastructure systems and archi-tectures (VISA10) Septemebr 2010 pp 25ndash32

[100] K-K Yap M Kobayashi R Sherwood T-Y Huang M ChanN Handigol and N McKeown ldquoOpenroads Empowering research inmobile networksrdquo ACM SIGCOMM Computer Communication Reviewvol 40 no 1 pp 125ndash126 2010

[101] T Nagai and H Shigeno ldquoA framework of ap aggregation usingvirtualization for high density wlansrdquo in INCoS IEEE 2011 pp350ndash355

[102] J Schulz-Zander P L Suresh N Sarrar A Feldmann T Huhn andR Merz ldquoProgrammatic orchestration of wifi networksrdquo in USENIX2014 pp 347ndash358

[103] R Riggio C Sengul L Suresh J Schulzzander and A FeldmannldquoThor Energy programmable wifi networksrdquo in IEEE Conference onComputer Communications Workshops April 2013 pp 21ndash22

[104] ldquoOdinrdquo [Online] Available httpsdninettu-berlinde[105] V Sivaraman T Moors H H Gharakheili D Ong J Matthews

and C Russell ldquoVirtualizing the access network via open apisrdquo inCoNEXT December 2013 pp 31ndash42

[106] ldquoCapsulatorrdquo [Online] Available httpsgithubcompeymankCapsulator

[107] R Riggio T Rasheed and F Granelli ldquoEmpower A testbed fornetwork function virtualization research and experimentationrdquo in IEEESDN for Future Networks and Services (SDN4FNS) November 2013pp 1ndash5

[108] ldquoEnergino projectrdquo [Online] Available httpwwwenergino-projectorg

[109] ldquoEmPOWERrdquo [Online] Available httpempowercreate-netorg[110] D Zhao M Zhu and M Xu ldquoSdwlan A flexible architecture of

enterprise wlan for client-unaware fast ap handoffrdquo in InternationalConference on Computing Communication and Networking Technolo-gies (ICCCNT) July 2014 pp 1ndash6

[111] D Zhao M Zhu and M Xu ldquoSupporting one big ap illusion in en-terprise wlan An sdn-based solutionrdquo in 6th International Conferenceon Wireless Communications and Signal Processing (WCSP) October2014 pp 1ndash6

[112] Y Yiakoumis M Bansal A Covington J van Reijendam S Kattiand N McKeown ldquoBehop A testbed for dense wifi networksrdquo ACMSIGMOBILE Mobile Computing and Communications Review vol 18no 3 pp 71ndash80 2015

[113] K Nakauchi Y Shoji and N Nishinaga ldquoAirtime-based resourcecontrol in wireless lans for wireless network virtualizationrdquo in ICUFNIEEE 2012 pp 166ndash169

[114] K Nakauchi and Y Shoji ldquoWifi network virtualization to control theconnectivity of a target servicerdquo IEEE Transactions on Network andService Management vol 12 no 2 pp 308ndash319 2015

[115] J Petazzoni ldquoBuild reliable traceable distributed systems with zeromq(zerorpc)rdquo [Online] Available httppycon-2012-notesreadthedocsorgenlatest-dotcloudzerorpchtml

[116] ldquoZeromqrdquo [Online] Available httpzeromqorg[117] J Schulz-Zander N Sarrar and S Schmid ldquoAeroflux A near-sighted

controller architecture for software-defined wireless networksrdquo in OpenNetworking Summit (ONS) 2014

[118] J Schulz-Zander N Sarrar and S Schmid ldquoTowards a scalable andnear-sighted control plane architecture for wifi sdnsrdquo in third workshopon Hot topics in software defined networking (HotSDN) 2014 pp 217ndash218

[119] V Shrivastava N Ahmed S Rayanchu S Banerjee S KeshavK Papagiannaki and A Mishra ldquoCentaur Realizing the full potentialof centralized wlans through a hybrid data pathrdquo in MobiCom 2009pp 297ndash308

[120] G Bianchi and I Tinnirello ldquoRemarks on ieee 80211 dcf performanceanalysisrdquo IEEE communications letters vol 9 no 8 pp 765ndash7672005

[121] H Moura G V C Bessa M A M Vieira and D F MacedoldquoEthanol Software defined networking for 80211 wireless networksrdquoin IFIPIEEE International Symposium on Integrated Network Man-agement (IM) May 2015 pp 388ndash396

[122] ldquoPantourdquo [Online] Available httpswwwsdxcentralcomprojectspantou-openwrt

[123] M Jarschel S Oechsner D Schlosser R Pries S Goll and P Tran-Gia ldquoModeling and performance evaluation of an openflow architec-turerdquo in Proceedings of the 23rd international teletraffic congress2011 pp 1ndash7

[124] L Vanbever J Reich T Benson N Foster and J Rexford ldquoHotswapCorrect and efficient controller upgrades for software-defined net-worksrdquo in Proceedings of the second ACM SIGCOMM workshop onHot topics in software defined networking ACM 2013 pp 133ndash138

[125] A Lara A Kolasani and B Ramamurthy ldquoNetwork innovationusing openflow A surveyrdquo IEEE communications surveys amp tutorialsvol 16 no 1 pp 493ndash512 2014

[126] A Rao ldquoTowards a software defined network based controller formanagement of large scale wlanrdquo Masterrsquos thesis Indian Institute ofTechnology Bombay 2015

[127] M Słabicki and K Grochla ldquoPerformance evaluation of snmp netconfand cwmp management protocols in wireless networkrdquo in CECNET2014 p 377

[128] T Nitsche C Cordeiro A B Flores E W Knightly E Perahia andJ C Widmer ldquoIEEE 80211ad directional 60 GHz communicationfor multi-Gigabit-per-second Wi-Firdquo IEEE Communications Magazinevol 52 no 12 pp 132ndash141 2014

[129] M Bansal J Mehlman S Katti and P Levis ldquoOpenRadio AProgrammable Wireless Dataplanerdquo in First Workshop on Hot Topics inSoftware Defined Networks (HotSDN rsquo12) August 2012 pp 109ndash114

[130] M Bansal A Schulman and S Katti ldquoAtomix A framework fordeploying signal processing applications on wireless infrastructurerdquo in12th USENIX Symposium on Networked Systems Design and Imple-mentation (NSDI rsquo15) May 2015 pp 173ndash188

[131] K Tan H Liu J Zhang Y Zhang J Fang and G M VoelkerldquoSora High-performance software radio using general-purpose multi-core processorsrdquo Communications of the ACM vol 54 no 1 pp 99ndash107 January 2011

[132] A Cidon K Nagaraj S Katti and P Viswanath ldquoFlashback Decou-pled lightweight wireless controlrdquo in ACM SIGCOMMrsquo12 conferenceon Applications technologies architectures and protocols for com-puter communication August 2012 pp 223ndash234

[133] ldquoIEEE Std 80211rdquo 2016 [Online] Available httpsstandardsieeeorgfindstdsstandard80211-2016html

[134] K Guo S Sanadhya and T Woo ldquoVifi Virtualizing wlan usingcommodity hardwarerdquo in 9th ACM workshop on Mobility in theevolving internet architecture September 2014 pp 25ndash30

[135] Alliance NGMN ldquo5G white paperrdquo 2015 [Online]Available httpwwwngmndefileadminngmncontentdownloadsTechnical2015NGMN 5G White Paper V1 0pdf

[136] J Vestin and A Kassler ldquoQoS Management for WiFi MAC LayerProcessing in the Cloud Demo Descriptionrdquo in Proceedings of the11th ACM Symposium on QoS and Security for Wireless and MobileNetworks 2015 pp 173ndash174

[137] Z Qin G Denker C Giannelli P Bellavista and N Venkatasubrama-nian ldquoA software defined networking architecture for the internet-of-thingsrdquo IEEEIFIP Network Operations and Management SymposiumManagement in a Software Defined World pp 1ndash9 2014

[138] X Jing A P Patil A Liu and D N Nguyen ldquoMulti-tier wirelesshome mesh network with a secure network discovery protocolrdquo 2016US Patent 9444639

[139] M Yu L Jose and R Miao ldquoSoftware defined traffic measurementwith opensketchrdquo in 10th USENIX conference on Networked SystemsDesign and Implementation April 2013 pp 29ndash42

[140] M Kotaru K Joshi D Bharadia and S Katti ldquoSpotfi Decimeter levellocalization using wifirdquo ACM SIGCOMM Computer CommunicationReview vol 45 no 5 pp 269ndash282 2015

[141] I Papanikos and M Logothetis ldquoA study on dynamic load balance forieee 80211b wireless lanrdquo in COMCON 2001

[142] ldquoIEEE Std 80211rdquo 2007 [Online] Available httpieeexploreieeeorgstampstampjsparnumber=4248378

[143] K Sinha S C Ghosh and B P Sinha Wireless Networks and MobileComputing CRC Press 2015

[144] S Pack J Choi T Kwon and Y Choi ldquoFast-handoff support in ieee80211 wireless networksrdquo IEEE Communications Surveys amp Tutorialsvol 9 no 1 pp 2ndash12 2007

[145] ldquoCisco compatible extensionrdquo [Online] Available httpswwwciscocomwebpartnersdownloads765ccxComp Ext Cust Presopdf

[146] P Dely A Kassler L Chow N Bambos N Bayer H Einsiedler andC Peylo ldquoBest-ap Non-intrusive estimation of available bandwidthand its application for dynamic access point selectionrdquo ComputerCommunications vol 39 pp 78ndash91 February 2014

[147] Y Han and K Yang ldquoAn adaptive mobility manager for software-defined enterprise wlansrdquo in Eighth International Conference on Ubiq-uitous and Future Networks (ICUFN) 2016 pp 888ndash893

[148] K Sui M Zhou D Liu M Ma D Pei Y Zhao Z Li andT Moscibroda ldquoCharacterizing and improving wifi latency in large-scale operational networksrdquo in Annual International Conference onMobile Systems Applications and Services (MobiSys) 2016 pp 347ndash360

[149] T K Ho ldquoRandom decision forestsrdquo in Third International Conferenceon Document Analysis and Recognition (ICDAR rsquo95) 1995 pp 278ndash282

[150] Y Bejerano S j Han and L E Li ldquoFairness and load balancing inwireless lans using association controlrdquo in 10th Annual InternationalConference on Mobile Computing and Networking 2004 pp 315ndash329

[151] Y Bejerano S J Han and L Li ldquoFairness and load balancing inwireless lans using association controlrdquo IEEEACM Transactions onNetworking vol 15 no 3 pp 560ndash573 2007

[152] W Li S Wang Y Cui X Cheng R Xin M A Al-Rodhaan andA Al-Dhelaan ldquoAp association for proportional fairness in multiratewlansrdquo IEEEACM Transactions on Networking vol 22 no 1 pp191ndash202 2014

[153] M Heusse F Rousseau G Berger-Sabbatel and A Duda ldquoPerfor-mance anomaly of 80211brdquo in 21th Annual Joint Conference of theIEEE Computer and Communications Societies (INFOCOM) 2003 pp836ndash843

[154] G Tan and J Guttag ldquoTime-based fairness improves performance inmulti-rate wlansrdquo in Annual Conference on USENIX Annual TechnicalConference 2004 pp 23ndash36

[155] ldquoOMNeT++rdquo [Online] Available httpwwwomnetpporg[156] L E Li M Pal and Y R Yang ldquoProportional fairness in multi-rate

wireless lansrdquo in The 27th Conference on Computer Communications(INFOCOM rsquo08) April 2008 pp 1004ndash1012

[157] D Gong and Y Yang ldquoOn-line ap association algorithms for 80211nwlans with heterogeneous clientsrdquo IEEE Transactions on Computersvol 63 no 11 pp 2772ndash2786 2014

[158] P B Oni and S D Blostein ldquoAp association optimization and ccathreshold adjustment in dense wlansrdquo in IEEE Globecom WorkshopsDecember 2015 pp 1ndash6

[159] T S Rappaport Wireless communications principles and practiceNew Jersey prentice hall PTR 1996 vol 2

[160] G A Mills-Tettey A Stentz and M B Dias ldquoThe dynamic hungarianalgorithm for the assignment problem with changing costsrdquo RoboticsInstitute School of Computer Science Carnegie Mellon UniversityPA USA Tech Rep CMU-RI-TR-07-27 2007

[161] L Yang Y Cui H Tang and S Xiao ldquoDemand-aware load balancingin wireless lans using association controlrdquo in IEEE Global Communi-cations Conference (GLOBECOM rsquo15) December 2015 pp 1ndash6

[162] F Xu X Zhu C C Tan Q Li G Yan and J Wu ldquoSmartassocDecentralized access point selection algorithm to improve throughputrdquoIEEE Transactions on Parallel and Distributed Systems vol 24 no 12pp 2482ndash2491 2013

[163] W Wong A Thakur and S H G Chan ldquoAn approximation algorithmfor ap association under user migration cost constraintrdquo in 35thAnnual IEEE International Conference on Computer Communications(INFOCOM rsquo16) April 2016 pp 1ndash9

[164] ldquoNS-3rdquo [Online] Available httpswwwnsnamorg[165] L-H Yen J-J Li and C-M Lin ldquoStability and fairness of native

ap selection games in ieee 80211 access networksrdquo in 7th Interna-

tional Conference on Wireless and Optical Communications Networks(WOCN) 2010 pp 1ndash5

[166] J Chen B Liu H Zhou Q Yu G Lin and X Shen ldquoQos-drivenefficient client association in high-density software defined wlanrdquo IEEETransactions on Vehicular Technology vol 66 no 8 pp 7372 ndash 73832017

[167] G Calinescu C Chekuri M Pal and J Vondrak ldquoMaximizing a sub-modular set function subject to a matroid constraintrdquo in InternationalConference on Integer Programming and Combinatorial Optimization2007 pp 182ndash196 lecture Notes in Computer Science

[168] X Zhe F Zhang W Wang and S He ldquoA polynomial time algorithmfor minimizing a nondecreasing supermodular set function and itsperformance guaranteerdquo International Journal of Pure and AppliedMathematics vol 52 no 5 pp 637ndash643 2009

[169] T Bonald L Massoulie A Proutiere and J Virtamo ldquoA queueinganalysis of max-min fairness proportional fairness and balanced fair-nessrdquo Queueing Systems vol 53 no 1 pp 65ndash84 2006

[170] D Gong and Y Yang ldquoAp association in 80211n wlans with heteroge-neous clientsrdquo in IEEE INFOCOM rsquo12 March 2012 pp 1440ndash1448

[171] B Bellalta A Checco A Zocca and J Barcelo ldquoOn the interactionsbetween multiple overlapping wlans using channel bondingrdquo IEEETransactions on Vehicular Technology vol 65 no 2 pp 796ndash8122016

[172] I-G Lee and M Kim ldquoInterference-aware self-optimizing wi-fi forhigh efficiency internet of things in dense networksrdquo Computer Com-munications vol 89-90 pp 60ndash74 2016

[173] K Sood S Yu and Y Xiang ldquoSoftware-defined wireless networkingopportunities and challenges for internet-of-things A reviewrdquo IEEEInternet of Things Journal vol 3 no 4 pp 453ndash463 2016

[174] D Singh G Tripathi and A J Jara ldquoA survey of Internet-of-ThingsFuture vision architecture challenges and servicesrdquo in IEEE WorldForum on Internet of Things (WF-IoT) 2014 pp 287ndash292

[175] G Gibb G Varghese M Horowitz and N McKeown ldquoDesignprinciples for packet parsersrdquo in Proceedings of the ninth ACMIEEEsymposium on Architectures for networking and communications sys-tems 2013 pp 13ndash24

[176] I F Akyildiz A Lee P Wang M Luo and W Chou ldquoA roadmapfor traffic engineering in sdn-openflow networksrdquo Computer Networksvol 71 pp 1ndash30 2014

[177] B Ng M Hayes and W K Seah ldquoDeveloping a traffic classificationplatform for enterprise networks with sdn Experiences amp lessonslearnedrdquo in IFIP IEEE 2015 pp 1ndash9

[178] S Yoon T Ha S Kim and H Lim ldquoScalable Traffic Sampling usingCentrality Measure on Software-Defined Networksrdquo IEEE Communi-cations Magazine - special issue on Traffic Measurements for CyberSecurity no July pp 43ndash49 2017

[179] M Sinky A Dhamodaran B Lee and J Zhao ldquoAnalysis of H264Bitstream Prioritization for Dual TCPUDP Streaming of HD videoover WLANsrdquo in CCNC 2015 pp 576ndash581

[180] T Hobfeld R Schatz M Varela and C Timmerer ldquoChallenges of QoEmanagement for cloud applicationsrdquo IEEE Communications Magazinevol 50 no 4 pp 28ndash36 2012

[181] A Gudipati D Perry L E Li and S Katti ldquoSoftran Software definedradio access networkrdquo in Proceedings of the second ACM SIGCOMMworkshop on Hot topics in software defined networking ACM 2013pp 25ndash30

[182] M Achanta ldquoMethod and apparatus for least congested channel scanfor wireless access pointsrdquo 2006 US Patent No 20060072602

[183] A Mishra S Banerjee and W Arbaugh ldquoWeighted coloring basedchannel assignment for wlansrdquo ACM SIGMOBILE Mobile Computingand Communications Review vol 9 no 3 pp 19ndash31 2005

[184] ldquoNS-2rdquo [Online] Available httpswwwisiedunsnamns[185] A Mishra V Brik S Banerjee A Srinivasan and W Arbaugh ldquoA

client-driven approach for channel management in wireless lansrdquo in25th IEEE International Conference on Computer Communications(INFOCOM) April 2006 pp 1ndash12

[186] P San Segundo ldquoA new DSATUR-based algorithm for exact vertexcoloringrdquo Computers amp Operations Research vol 39 no 7 pp 1724ndash1733 2012

[187] M Seyedebrahimi F Bouhafs A Raschell M Mackay and Q ShildquoSdn-based channel assignment algorithm for interference managementin dense wi-fi networksrdquo in European Conference on Networks andCommunications (EuCNC) June 2016 pp 128ndash132

[188] ldquoWi-5rdquo [Online] Available httpwwwwi5eu[189] X Ling and K L Yeung ldquoJoint access point placement and channel

assignment for 80211 wireless lansrdquo IEEE Transactions on WirelessCommunications vol 5 no 10 pp 2705ndash2711 2006

[190] R Jain D M Chiu and W Hawe ldquoA quantitative measure of fairnessand discrimination for resource allocation in shared computer systemsrdquoDEC-Research Report TR-301 Tech Rep 1984

[191] E Rozner Y Mehta A Akella and L Qiu ldquoTraffic-aware channel as-signment in enterprise wireless lansrdquo in IEEE International Conferenceon Network Protocols October 2007 pp 133ndash143

[192] L Vandenberghe and S Boyd ldquoSemidefinite programmingrdquo SIAMReview vol 38 no 1 pp 49ndash95 1996

[193] H Zhang H Ji and W Ge ldquoChannel assignment with fairness formulti-ap wlan based on distributed coordination functionrdquo in IEEEWireless Communications and Networking Conference (WCNC) March2011 pp 392ndash397

[194] K Duffy D Malone and D J Leith ldquoModeling the 80211 distributedcoordination function in non-saturated conditionsrdquo IEEE communica-tions letters vol 9 no 8 pp 715ndash717 2005

[195] B A H S Abeysekera K Ishihara Y Inoue and M MizoguchildquoNetwork-controlled channel allocation scheme for ieee 80211 wire-less lans Experimental and simulation studyrdquo in IEEE 79th VehicularTechnology Conference (VTC Spring) 2014 pp 1ndash5

[196] B A H S Abeysekera M Matsui Y Asai and M MizoguchildquoNetwork controlled frequency channel and bandwidth allocationscheme for ieee 80211anac wireless lans Ratoprdquo in IEEE 25thAnnual International Symposium on Personal Indoor and Mobile RadioCommunication (PIMRC) 2014 pp 1041ndash1045

[197] Y Zhang C Jiang Y Wang J Yuan and J Cao ldquoUnited channelassignments in residential environmentsrdquo in IEEE Global Communi-cations Conference (GLOBECOM) December 2015 pp 1ndash6

[198] M F Tuysuz and H A Mantar ldquoSmart channel scanning withminimized communication interruptions over IEEE 80211 WLANsrdquoin IEEE 24th Annual International Symposium on Personal Indoorand Mobile Radio Communications (PIMRC) 2013 pp 2218ndash2222

[199] S Jin M Choi L Wang and S Choi ldquoFast scanning schemesfor IEEE 80211 WLANs in virtual AP environmentsrdquo ComputerNetworks vol 55 no 10 pp 2520ndash2533 2011

[200] J Manweiler N Santhapuri R R Choudhury and S NelakuditildquoPredicting length of stay at wifi hotspotsrdquo in INFOCOM IEEE2013 pp 3102ndash3110

[201] Q Dong and W Dargie ldquoEvaluation of the reliability of rssi for indoorlocalizationrdquo in ICWCUCA IEEE 2012 pp 1ndash6

[202] W Xu C Hua and A Huang ldquoChannel Assignment and UserAssociation Game in Dense 80211 Wireless Networksrdquo in ICC 2011pp 1ndash5

[203] R Zheng and C Hua Channel Selection and User Association in WiFiNetworks Springer International Publishing 2016 pp 103ndash115

[204] W Zhao H Nishiyama Z Fadlullah N Kato and K HamaguchildquoDAPA Capacity Optimization in Wireless Networks Through a Com-bined Design of Density of Access Points and Partially OverlappedChannel Allocationrdquo IEEE Transactions on Vehicular Technologyvol 65 pp 3715ndash3722 May 2016

[205] S Zvanovec P Pechac and M Klepal ldquoWireless lan networks designsite survey or propagation modelingrdquo Radioengineering vol 12 no 4pp 42ndash49 2003

[206] W Li Y Cui X Cheng M A Al-Rodhaan and A Al-DhelaanldquoAchieving proportional fairness via ap power control in multi-ratewlansrdquo IEEE Transactions on Wireless Communications vol 10no 11 pp 3784ndash3792 2011

[207] Y Bejerano and S-J Han ldquoCell breathing techniques for load balanc-ing in wireless lansrdquo IEEE Transactions on Mobile Computing vol 8no 6 pp 735ndash749 2009

[208] S Wang J Huang X Cheng and B Chen ldquoCoverage adjustmentfor load balancing with an ap service availability guarantee in wlansrdquoWireless networks vol 20 no 3 pp 475ndash491 2014

[209] S Huang X Liu and Z Ding ldquoDistributed power control for cognitiveuser access based on primary link control feedbackrdquo in INFOCOM2010 pp 1ndash9

[210] Y Zeng P H Pathak and P Mohapatra ldquoA first look at 80211acin action Energy efficiency and interference characterizationrdquo IFIPNetworking Conference pp 1ndash9 2014

[211] S Han X Zhang and K G Shin ldquoFair and Efficient Coexistence ofHeterogeneous Channel Widths in Next-Generation Wireless LANsrdquoIEEE Transactions on Mobile Computing vol 15 no 11 pp 2749ndash2761 November 2016

[212] X Zhang and K G Shin ldquoAdaptive subcarrier nulling Enabling partialspectrum sharing in wireless lansrdquo in ICNP 2011 pp 311ndash320

[213] K Pelechrinis M Iliofotou and S V Krishnamurthy ldquoDenial ofService Attacks in Wireless Networks The Case of Jammersrdquo IEEECommunications Surveys Tutorials vol 13 no 2 pp 245ndash257 2011

[214] E Bayraktaroglu C King X Liu G Noubir R Rajaraman andB Thapa ldquoOn the performance of ieee 80211 under jammingrdquo inINFOCOM 2008 pp 1265ndash1273

[215] K Pelechrinis I Broustis S V Krishnamurthy and C GkantsidisldquoAres an anti-jamming reinforcement system for 80211 networksrdquo inCoNEXT 2009 pp 181ndash192

[216] K Pelechrinis C Koufogiannakis and S V Krishnamurthy ldquoOnthe efficacy of frequency hopping in coping with jamming attacksin 80211 networksrdquo IEEE transactions on wireless communicationsvol 9 no 10 pp 3258ndash3271 2010

[217] I T Haque and N Abu-Ghazaleh ldquoWireless software defined net-working A survey and taxonomyrdquo IEEE Communications Surveys ampTutorials vol 18 no 4 pp 2713ndash2737 2016

[218] D F Macedo D Guedes L F Vieira M A Vieira and M NogueiraldquoProgrammable networksfrom software-defined radio to software-defined networkingrdquo IEEE communications surveys amp tutorialsvol 17 no 2 pp 1102ndash1125 2015

[219] M Richart J Baliosian J Serrat and J-L Gorricho ldquoResource slicingin virtual wireless networks A surveyrdquo IEEE Transactions on Networkand Service Management vol 13 no 3 pp 462ndash476 2016

[220] F Hu Q Hao and K Bao ldquoA survey on software-defined network andopenflow From concept to implementationrdquo IEEE CommunicationsSurveys amp Tutorials vol 16 no 4 pp 2181ndash2206 2014

[221] B Bellalta L Bononi R Bruno and A Kassler ldquoNext generation ieee80211 wireless local area networks Current status future directionsand open challengesrdquo Computer Communications vol 75 pp 1ndash252016

[222] S Chieochan E Hossain and J Diamond ldquoChannel assignmentschemes for infrastructure-based 80211 wlans A surveyrdquo IEEE Com-munications Surveys amp Tutorials vol 12 no 1 pp 124ndash136 2010

[223] A Baid and D Raychaudhuri ldquoUnderstanding channel selection dy-namics in dense wi-fi networksrdquo IEEE Communications Magazinevol 53 no 1 pp 110ndash117 2015

[224] E Charfi L Chaari and L Kamoun ldquoPhymac enhancements andqos mechanisms for very high throughput wlans A surveyrdquo IEEECommunications Surveys amp Tutorials vol 15 no 4 pp 1714ndash17352013

[225] Q Ni L Romdhani and T Turletti ldquoA survey of qos enhancementsfor ieee 80211 wireless lanrdquo Wireless Communications and MobileComputing vol 4 no 5 pp 547ndash566 2004

[227] A Ozyagci K W Sung and J Zander ldquoAssociation and deploymentconsiderations in dense wireless lansrdquo in IEEE 79th Vehicular Tech-nology Conference (VTC Spring) May 2014 pp 1ndash5

[228] L H Yen T T Yeh and K H Chi ldquoLoad balancing in ieee 80211networksrdquo IEEE Internet Computing vol 13 no 1 pp 56ndash64 2009

[229] M Hoyhtya A Mammela M Eskola M Matinmikko J KalliovaaraJ Ojaniemi J Suutala R Ekman R Bacchus and D RobersonldquoSpectrum Occupancy Measurements A Survey and Use of Interfer-ence Mapsrdquo IEEE Communications Surveys amp Tutorials vol 18 no 4pp 2386ndash2414 2016

[230] B Bellalta ldquoIEEE 80211ax High-efficiency WLANsrdquo IEEE WirelessCommunications vol 23 no 1 pp 38ndash46 2016

[231] C Liang and F R Yu ldquoWireless network virtualization A surveysome research issues and challengesrdquo IEEE Communications Surveysamp Tutorials vol 17 no 1 pp 358ndash380 2015

I Introduction

I-A Contributions

II Motivation

II-A Applications and Requirements

II-B Challenges

II-C Architectures and Control Mechanisms

III Architectures

III-A Principles of Software-Defined Networking

III-B Standard Protocols

III-C SDWLAN Architectures

III-C1 Observability and Configurability

III-C2 Programmability

III-C3 Virtualization

III-C4 Scalability

III-C5 Traffic Shaping

III-D Architectures Learned Lessons Comparison and Open Problems

III-D1 Programmability

III-D2 Mobility

III-D3 Network Virtualization

III-D4 Network Function Virtualization

III-D5 Home WLANs

III-D6 Security

IV Centralized Association Control (AsC)

IV-A Seamless Handoff

IV-B Client Steering

IV-B1 Centrally-Generated Hints

IV-B2 Individual Optimization through Centrally-Made Decisions

IV-B3 Global Optimization through Centrally-Made Decisions

IV-C Association Control Learned Lessons Comparison and Open Problems

IV-C1 Optimization Scope of Client Steering

IV-C2 Decision Metrics

IV-C3 Dynamicity and Overhead

IV-C4 Security Considerations

V Centralized Channel Assignment (ChA)

V-A Traffic-Agnostic Channel Assignment

V-B Traffic-Aware Channel Assignment

V-C Channel Assignment Learned Lessons Comparison and Open Problems

V-C1 Dynamicity and Traffic-Awareness

V-C2 Joint ChA and AsC

V-C3 Partially Overlapping Channels and AP Density Management

V-C4 Integration with Virtualization

V-C5 Coexistence of High-Throughput and Legacy Clients

V-C6 Security

VI Related Surveys

VII Conclusion

VIII Acknowledgment

References

network control mechanisms as applications running on thecontroller these mechanisms can benefit from the controllerrsquosglobal network view Not only for wired networks SDNhas gained popularity for the design of WLANs and cellularnetworks as well [20]ndash[23]

A software-defined WLAN (SDWLAN) enables centralmonitoring and control of network operation A SDWLANcan be studied from two point of views as follows

ndash Architecture A SDWLAN architecture (i) definesthe various network components (eg controller APsswitches middleboxes) used to build the network aswell as the topology employed to connect these com-ponents (ii) implements application programming inter-faces (APIs) through which the controller communicateswith network devices to perform data collection anddistribute control commands (iii) specifies the separationlevel between the control plane and data plane The archi-tectural features highly affect the development flexibilityand performance of network control mechanisms [12][24]ndash[26] For example when medium access control(MAC) functionalities are transferred from the APs to thecontroller network control applications may implementAsC mechanisms with various capabilities and perfor-mance levels depending on the separation level and APIsprovided

ndash Control Mechanisms To benefit from the features ofSDWLAN architectures and in particular the global net-work view provided various control mechanisms suchas association control (AsC) channel assignment (ChA)transmission power control and CCA threshold adjust-ment have been proposed by the research communityIn this paper we particularly focus on centralized AsCand ChA mechanisms because they have been widelyproposed and adopted to benefit from the features of SD-WLANs AsC mechanisms address client-AP associationfor both static and dynamic clients to achieve variousobjectives such as seamless mobility fairness amongclients load balancing among APs and interference miti-gation ChA mechanisms address the problem of channelassignment to APs in order to minimize the interferencelevel experienced by APs and clients [27]ndash[29]

A Contributions

At a high level this paper studies SDWLANs from twoperspectives (i) SDWLAN architectures and (ii) the two maincentral network control mechanisms AsC and ChA that areemployed to control network operation More precisely thecontributions of this paper are as follows

ndash We first present an overview of software-defined net-working and the standard protocols used for networkmonitoring management and programming Next wereview SDWLAN architectures and reveal their main ob-jective components interconnection protocols and APIsprovided Based on their main contribution we classifyarchitectures into the following categories observabil-ity and configurability programmability virtualizationscalability and traffic shaping Then we compare these

architectures and highlight future research directionsconsidering the current and upcoming requirements ofSDWLANs

ndash We review the AsC mechanisms proposed for SD-WLANs We first classify these mechanisms into twogroups (i) seamless handoff mechanisms for reducingthe overhead and delay of client handoff and (ii) clientsteering mechanisms to optimize parameters such as theairtime allocated to clients Client steering approachesin turn are classified into two groups (i) centrally-generated hints where the controller relies on the globalnetwork view to generate hints for the association ofclients and (ii) centrally-made decisions where thecontroller makes the association decisions and enforcesthe clients to apply them In addition to discussing theimpact of architecture on supporting seamless handoffwe compare the reviewed mechanisms in terms of op-timization scope traffic awareness and the ability torecognize the QoS requirements of clients By discussingthe capabilities and weaknesses of AsC mechanismswe present research directions towards designing AsCsolutions for heterogeneous and dynamic SDWLANswith diverse QoS requirements We also identify howthe potential features of SDWLAN architectures can beemployed to design more efficient AsC mechanisms

ndash We review the ChA mechanisms proposed for SD-WLANs We classify these mechanisms into two groupstraffic-aware and traffic-agnostic based on whether theyincorporate APsrsquo and clientsrsquo traffic into the decisionmaking process We specifically focus on how ChAmechanisms define and measure their interference met-rics as well as the approaches employed for modelingand solving the formulated problems Furthermore wecompare these mechanisms in terms of factors like client-awareness dynamicity and the unique features of dif-ferent 80211 standards We propose research directionstowards designing mechanisms that address the demandsof emerging applications We also identify how thepotentials of SDWLANs can be employed to design moreefficient ChA mechanisms

When studying AsC and ChA mechanisms we reformulatethe problems using a consistent set of notations to ease theunderstanding and comparison of these mechanisms Table Ipresents the key acronyms and notations used in this paperFigure 2 shows the high-level organization of the paper

II MOTIVATION

In this section we overview the major application domainsand requirements of WLANs the challenges of wireless com-munication and the importance of centralization in addressingthe requirements of current and emerging applications

A Applications and Requirements

WLANs are being used in a variety of applications withvarious performance requirements WLANs installed in enter-prises campuses and public areas usually serve a small tomedium number of stationary and mobile users For voice

B Challenges

III ARCHITECTURES

Virtual Machine 1

Policy Plane

VAP Manager

Policy Plane

Controller

Odin Agent

Internet

Data Plane

Odin Control Plane

LVAPLVAP LVAPLVAP

Odin OpenFlow

OpenFlow Agent

Applications

OpenFlow Switch

Controller

OpenFlow Controller

Policies

Data Plane

Control Plane

Virtual WLAN Card

Applications

Virtual WLAN Card

Applications

Virtual WLAN Card

Controller

(DPI)Agent

MiddleBoxDriver

Radio Driver

Radio Agent

Applications

Data Plane

WDTXWDTXWDTXLVAP

WDTXWDTXWDTX

LVAPvMB

WDTXWDTXWDTX

Internet

MiddleBox (MB)

Sd-wlan

Trantor

SplitAP

+ Slicing

CloudMAC

AeroFluxDenseAP

Ethanol

AEtherFlow

OpenSDWN

EmPOWER

AP Aggregate

RCWLAN

WVTCoAP and ResFi

OpenRoads

EmPOWER2

+ Traffic shaping

CENTAUR

SlicingSouth-bound

ProtocolSplit-MAC

SupportClient

A Seamless Handoff

B Client Steering

APicj= F chi

APitimes rchi

APicj (2)

where F chi

and rchi

ON OFF

and tinactive Tidle

and toff Toffline

LAPi =

AP1 AP2 APi

Controller

Client

SNMPdata

AGE request

Specified AP

factors

LAPi= max

sumforallcjisinC

ωcj timesXAPicj

rAPicj

sumforallcjisinC

ωcj timesXAPicj

forallcjisinC

forallAPiisinAP

logsum

forallAPiisinAP

sumforallAPlisinAP

SeamlessHandoff

ClientSteering

Cent GenHints

Indv OptCent Made

Decisions

Glob OptCent Made

FairnessThroughput

each beacon

with each beacon

chi chj

NAPi+NAPj

Client

3AP1 AP2

InvaderAP

VictimAP

Collision1

forallAPiisinAP

sumforallchjisinCH

UAPichj (12)

WCS =sum

New assignment

Old assignment6=

SNMP statistics

T (APi) (15)

maxsum

U(APi chj) (16)

andmax( min

VI RELATED SURVEYS

VII CONCLUSION

VIII ACKNOWLEDGMENT

REFERENCES

I Introduction

I-A Contributions

II Motivation


II-B Challenges


III Architectures







III-C4 Scalability




III-D2 Mobility



III-D5 Home WLANs

III-D6 Security





















V-C6 Security

VI Related Surveys

VII Conclusion

VIII Acknowledgment

References

B Challenges

III ARCHITECTURES

Virtual Machine 1

Policy Plane

VAP Manager

Policy Plane

Controller

Odin Agent

Internet

Data Plane

Odin Control Plane

LVAPLVAP LVAPLVAP

Odin OpenFlow

OpenFlow Agent

Applications

OpenFlow Switch

Controller

OpenFlow Controller

Policies

Data Plane

Control Plane

Virtual WLAN Card

Applications

Virtual WLAN Card

Applications

Virtual WLAN Card

Controller

(DPI)Agent

MiddleBoxDriver

Radio Driver

Radio Agent

Applications

Data Plane

WDTXWDTXWDTXLVAP

WDTXWDTXWDTX

LVAPvMB

WDTXWDTXWDTX

Internet

MiddleBox (MB)

Sd-wlan

Trantor

SplitAP

+ Slicing

CloudMAC

AeroFluxDenseAP

Ethanol

AEtherFlow

OpenSDWN

EmPOWER

AP Aggregate

RCWLAN

WVTCoAP and ResFi

OpenRoads

EmPOWER2

+ Traffic shaping

CENTAUR

SlicingSouth-bound

ProtocolSplit-MAC

SupportClient

A Seamless Handoff

B Client Steering

APicj= F chi

APitimes rchi

APicj (2)

where F chi

and rchi

ON OFF

and tinactive Tidle

and toff Toffline

LAPi =

AP1 AP2 APi

Controller

Client

SNMPdata

AGE request

Specified AP

factors

LAPi= max

sumforallcjisinC

ωcj timesXAPicj

rAPicj

sumforallcjisinC

ωcj timesXAPicj

forallcjisinC

forallAPiisinAP

logsum

forallAPiisinAP

sumforallAPlisinAP

SeamlessHandoff

ClientSteering

Cent GenHints

Indv OptCent Made

Decisions

Glob OptCent Made

FairnessThroughput

each beacon

with each beacon

chi chj

NAPi+NAPj

Client

3AP1 AP2

InvaderAP

VictimAP

Collision1

forallAPiisinAP

sumforallchjisinCH

UAPichj (12)

WCS =sum

New assignment

Old assignment6=

SNMP statistics

T (APi) (15)

maxsum

U(APi chj) (16)

andmax( min

VI RELATED SURVEYS

VII CONCLUSION

VIII ACKNOWLEDGMENT

REFERENCES

I Introduction

I-A Contributions

II Motivation


II-B Challenges


III Architectures







III-C4 Scalability




III-D2 Mobility



III-D5 Home WLANs

III-D6 Security





















V-C6 Security

VI Related Surveys

VII Conclusion

VIII Acknowledgment

References

III ARCHITECTURES

Virtual Machine 1

Policy Plane

VAP Manager

Policy Plane

Controller

Odin Agent

Internet

Data Plane

Odin Control Plane

LVAPLVAP LVAPLVAP

Odin OpenFlow

OpenFlow Agent

Applications

OpenFlow Switch

Controller

OpenFlow Controller

Policies

Data Plane

Control Plane

Virtual WLAN Card

Applications

Virtual WLAN Card

Applications

Virtual WLAN Card

Controller

(DPI)Agent

MiddleBoxDriver

Radio Driver

Radio Agent

Applications

Data Plane

WDTXWDTXWDTXLVAP

WDTXWDTXWDTX

LVAPvMB

WDTXWDTXWDTX

Internet

MiddleBox (MB)

Sd-wlan

Trantor

SplitAP

+ Slicing

CloudMAC

AeroFluxDenseAP

Ethanol

AEtherFlow

OpenSDWN

EmPOWER

AP Aggregate

RCWLAN

WVTCoAP and ResFi

OpenRoads

EmPOWER2

+ Traffic shaping

CENTAUR

SlicingSouth-bound

ProtocolSplit-MAC

SupportClient

A Seamless Handoff

B Client Steering

APicj= F chi

APitimes rchi

APicj (2)

where F chi

and rchi

ON OFF

and tinactive Tidle

and toff Toffline

LAPi =

AP1 AP2 APi

Controller

Client

SNMPdata

AGE request

Specified AP

factors

LAPi= max

sumforallcjisinC

ωcj timesXAPicj

rAPicj

sumforallcjisinC

ωcj timesXAPicj

forallcjisinC

forallAPiisinAP

logsum

forallAPiisinAP

sumforallAPlisinAP

SeamlessHandoff

ClientSteering

Cent GenHints

Indv OptCent Made

Decisions

Glob OptCent Made

FairnessThroughput

each beacon

with each beacon

chi chj

NAPi+NAPj

Client

3AP1 AP2

InvaderAP

VictimAP

Collision1

forallAPiisinAP

sumforallchjisinCH

UAPichj (12)

WCS =sum

New assignment

Old assignment6=

SNMP statistics

T (APi) (15)

maxsum

U(APi chj) (16)

andmax( min

VI RELATED SURVEYS

VII CONCLUSION

VIII ACKNOWLEDGMENT

REFERENCES

I Introduction

I-A Contributions

II Motivation


II-B Challenges


III Architectures







III-C4 Scalability




III-D2 Mobility



III-D5 Home WLANs

III-D6 Security





















V-C6 Security

VI Related Surveys

VII Conclusion

VIII Acknowledgment

References

Virtual Machine 1

Policy Plane

VAP Manager

Policy Plane

Controller

Odin Agent

Internet

Data Plane

Odin Control Plane

LVAPLVAP LVAPLVAP

Odin OpenFlow

OpenFlow Agent

Applications

OpenFlow Switch

Controller

OpenFlow Controller

Policies

Data Plane

Control Plane

Virtual WLAN Card

Applications

Virtual WLAN Card

Applications

Virtual WLAN Card

Controller

(DPI)Agent

MiddleBoxDriver

Radio Driver

Radio Agent

Applications

Data Plane

WDTXWDTXWDTXLVAP

WDTXWDTXWDTX

LVAPvMB

WDTXWDTXWDTX

Internet

MiddleBox (MB)

Sd-wlan

Trantor

SplitAP

+ Slicing

CloudMAC

AeroFluxDenseAP

Ethanol

AEtherFlow

OpenSDWN

EmPOWER

AP Aggregate

RCWLAN

WVTCoAP and ResFi

OpenRoads

EmPOWER2

+ Traffic shaping

CENTAUR

SlicingSouth-bound

ProtocolSplit-MAC

SupportClient

A Seamless Handoff

B Client Steering

APicj= F chi

APitimes rchi

APicj (2)

where F chi

and rchi

ON OFF

and tinactive Tidle

and toff Toffline

LAPi =

AP1 AP2 APi

Controller

Client

SNMPdata

AGE request

Specified AP

factors

LAPi= max

sumforallcjisinC

ωcj timesXAPicj

rAPicj

sumforallcjisinC

ωcj timesXAPicj

forallcjisinC

forallAPiisinAP

logsum

forallAPiisinAP

sumforallAPlisinAP

SeamlessHandoff

ClientSteering

Cent GenHints

Indv OptCent Made

Decisions

Glob OptCent Made

FairnessThroughput

each beacon

with each beacon

chi chj

NAPi+NAPj

Client

3AP1 AP2

InvaderAP

VictimAP

Collision1

forallAPiisinAP

sumforallchjisinCH

UAPichj (12)

WCS =sum

New assignment

Old assignment6=

SNMP statistics

T (APi) (15)

maxsum

U(APi chj) (16)

andmax( min

VI RELATED SURVEYS

VII CONCLUSION

VIII ACKNOWLEDGMENT

REFERENCES

I Introduction

I-A Contributions

II Motivation


II-B Challenges


III Architectures







III-C4 Scalability




III-D2 Mobility



III-D5 Home WLANs

III-D6 Security





















V-C6 Security

VI Related Surveys

VII Conclusion

VIII Acknowledgment

References

Virtual Machine 1

Policy Plane

VAP Manager

Policy Plane

Controller

Odin Agent

Internet

Data Plane

Odin Control Plane

LVAPLVAP LVAPLVAP

Odin OpenFlow

OpenFlow Agent

Applications

OpenFlow Switch

Controller

OpenFlow Controller

Policies

Data Plane

Control Plane

Virtual WLAN Card

Applications

Virtual WLAN Card

Applications

Virtual WLAN Card

Controller

(DPI)Agent

MiddleBoxDriver

Radio Driver

Radio Agent

Applications

Data Plane

WDTXWDTXWDTXLVAP

WDTXWDTXWDTX

LVAPvMB

WDTXWDTXWDTX

Internet

MiddleBox (MB)

Sd-wlan

Trantor

SplitAP

+ Slicing

CloudMAC

AeroFluxDenseAP

Ethanol

AEtherFlow

OpenSDWN

EmPOWER

AP Aggregate

RCWLAN

WVTCoAP and ResFi

OpenRoads

EmPOWER2

+ Traffic shaping

CENTAUR

SlicingSouth-bound

ProtocolSplit-MAC

SupportClient

A Seamless Handoff

B Client Steering

APicj= F chi

APitimes rchi

APicj (2)

where F chi

and rchi

ON OFF

and tinactive Tidle

and toff Toffline

LAPi =

AP1 AP2 APi

Controller

Client

SNMPdata

AGE request

Specified AP

factors

LAPi= max

sumforallcjisinC

ωcj timesXAPicj

rAPicj

sumforallcjisinC

ωcj timesXAPicj

forallcjisinC

forallAPiisinAP

logsum

forallAPiisinAP

sumforallAPlisinAP

SeamlessHandoff

ClientSteering

Cent GenHints

Indv OptCent Made

Decisions

Glob OptCent Made

FairnessThroughput

each beacon

with each beacon

chi chj

NAPi+NAPj

Client

3AP1 AP2

InvaderAP

VictimAP

Collision1

forallAPiisinAP

sumforallchjisinCH

UAPichj (12)

WCS =sum

New assignment

Old assignment6=

SNMP statistics

T (APi) (15)

maxsum

U(APi chj) (16)

andmax( min

VI RELATED SURVEYS

VII CONCLUSION

VIII ACKNOWLEDGMENT

REFERENCES

I Introduction

I-A Contributions

II Motivation


II-B Challenges


III Architectures







III-C4 Scalability




III-D2 Mobility



III-D5 Home WLANs

III-D6 Security





















V-C6 Security

VI Related Surveys

VII Conclusion

VIII Acknowledgment

References

Virtual Machine 1

Policy Plane

VAP Manager

Policy Plane

Controller

Odin Agent

Internet

Data Plane

Odin Control Plane

LVAPLVAP LVAPLVAP

Odin OpenFlow

OpenFlow Agent

Applications

OpenFlow Switch

Controller

OpenFlow Controller

Policies

Data Plane

Control Plane

Virtual WLAN Card

Applications

Virtual WLAN Card

Applications

Virtual WLAN Card

Controller

(DPI)Agent

MiddleBoxDriver

Radio Driver

Radio Agent

Applications

Data Plane

WDTXWDTXWDTXLVAP

WDTXWDTXWDTX

LVAPvMB

WDTXWDTXWDTX

Internet

MiddleBox (MB)

Sd-wlan

Trantor

SplitAP

+ Slicing

CloudMAC

AeroFluxDenseAP

Ethanol

AEtherFlow

OpenSDWN

EmPOWER

AP Aggregate

RCWLAN

WVTCoAP and ResFi

OpenRoads

EmPOWER2

+ Traffic shaping

CENTAUR

SlicingSouth-bound

ProtocolSplit-MAC

SupportClient

A Seamless Handoff

B Client Steering

APicj= F chi

APitimes rchi

APicj (2)

where F chi

and rchi

ON OFF

and tinactive Tidle

and toff Toffline

LAPi =

AP1 AP2 APi

Controller

Client

SNMPdata

AGE request

Specified AP

factors

LAPi= max

sumforallcjisinC

ωcj timesXAPicj

rAPicj

sumforallcjisinC

ωcj timesXAPicj

forallcjisinC

forallAPiisinAP

logsum

forallAPiisinAP

sumforallAPlisinAP

SeamlessHandoff

ClientSteering

Cent GenHints

Indv OptCent Made

Decisions

Glob OptCent Made

FairnessThroughput

each beacon

with each beacon

chi chj

NAPi+NAPj

Client

3AP1 AP2

InvaderAP

VictimAP

Collision1

forallAPiisinAP

sumforallchjisinCH

UAPichj (12)

WCS =sum

New assignment

Old assignment6=

SNMP statistics

T (APi) (15)

maxsum

U(APi chj) (16)

andmax( min

VI RELATED SURVEYS

VII CONCLUSION

VIII ACKNOWLEDGMENT

REFERENCES

I Introduction

I-A Contributions

II Motivation


II-B Challenges


III Architectures







III-C4 Scalability




III-D2 Mobility



III-D5 Home WLANs

III-D6 Security





















V-C6 Security

VI Related Surveys

VII Conclusion

VIII Acknowledgment

References

OpenFlow Switch

Controller

OpenFlow Controller

Policies

Data Plane

Control Plane

Virtual WLAN Card

Applications

Virtual WLAN Card

Applications

Virtual WLAN Card

Controller

(DPI)Agent

MiddleBoxDriver

Radio Driver

Radio Agent

Applications

Data Plane

WDTXWDTXWDTXLVAP

WDTXWDTXWDTX

LVAPvMB

WDTXWDTXWDTX

Internet

MiddleBox (MB)

Sd-wlan

Trantor

SplitAP

+ Slicing

CloudMAC

AeroFluxDenseAP

Ethanol

AEtherFlow

OpenSDWN

EmPOWER

AP Aggregate

RCWLAN

WVTCoAP and ResFi

OpenRoads

EmPOWER2

+ Traffic shaping

CENTAUR

SlicingSouth-bound

ProtocolSplit-MAC

SupportClient

A Seamless Handoff

B Client Steering

APicj= F chi

APitimes rchi

APicj (2)

where F chi

and rchi

ON OFF

and tinactive Tidle

and toff Toffline

LAPi =

AP1 AP2 APi

Controller

Client

SNMPdata

AGE request

Specified AP

factors

LAPi= max

sumforallcjisinC

ωcj timesXAPicj

rAPicj

sumforallcjisinC

ωcj timesXAPicj

forallcjisinC

forallAPiisinAP

logsum

forallAPiisinAP

sumforallAPlisinAP

SeamlessHandoff

ClientSteering

Cent GenHints

Indv OptCent Made

Decisions

Glob OptCent Made

FairnessThroughput

each beacon

with each beacon

chi chj

NAPi+NAPj

Client

3AP1 AP2

InvaderAP

VictimAP

Collision1

forallAPiisinAP

sumforallchjisinCH

UAPichj (12)

WCS =sum

New assignment

Old assignment6=

SNMP statistics

T (APi) (15)

maxsum

U(APi chj) (16)

andmax( min

VI RELATED SURVEYS

VII CONCLUSION

VIII ACKNOWLEDGMENT

REFERENCES

I Introduction

I-A Contributions

II Motivation


II-B Challenges


III Architectures







III-C4 Scalability




III-D2 Mobility



III-D5 Home WLANs

III-D6 Security





















V-C6 Security

VI Related Surveys

VII Conclusion

VIII Acknowledgment

References

Controller

(DPI)Agent

MiddleBoxDriver

Radio Driver

Radio Agent

Applications

Data Plane

WDTXWDTXWDTXLVAP

WDTXWDTXWDTX

LVAPvMB

WDTXWDTXWDTX

Internet

MiddleBox (MB)

Sd-wlan

Trantor

SplitAP

+ Slicing

CloudMAC

AeroFluxDenseAP

Ethanol

AEtherFlow

OpenSDWN

EmPOWER

AP Aggregate

RCWLAN

WVTCoAP and ResFi

OpenRoads

EmPOWER2

+ Traffic shaping

CENTAUR

SlicingSouth-bound

ProtocolSplit-MAC

SupportClient

A Seamless Handoff

B Client Steering

APicj= F chi

APitimes rchi

APicj (2)

where F chi

and rchi

ON OFF

and tinactive Tidle

and toff Toffline

LAPi =

AP1 AP2 APi

Controller

Client

SNMPdata

AGE request

Specified AP

factors

LAPi= max

sumforallcjisinC

ωcj timesXAPicj

rAPicj

sumforallcjisinC

ωcj timesXAPicj

forallcjisinC

forallAPiisinAP

logsum

forallAPiisinAP

sumforallAPlisinAP

SeamlessHandoff

ClientSteering

Cent GenHints

Indv OptCent Made

Decisions

Glob OptCent Made

FairnessThroughput

each beacon

with each beacon

chi chj

NAPi+NAPj

Client

3AP1 AP2

InvaderAP

VictimAP

Collision1

forallAPiisinAP

sumforallchjisinCH

UAPichj (12)

WCS =sum

New assignment

Old assignment6=

SNMP statistics

T (APi) (15)

maxsum

U(APi chj) (16)

andmax( min

VI RELATED SURVEYS

VII CONCLUSION

VIII ACKNOWLEDGMENT

REFERENCES

I Introduction

I-A Contributions

II Motivation


II-B Challenges


III Architectures







III-C4 Scalability




III-D2 Mobility



III-D5 Home WLANs

III-D6 Security





















V-C6 Security

VI Related Surveys

VII Conclusion

VIII Acknowledgment

References

Sd-wlan

Trantor

SplitAP

+ Slicing

CloudMAC

AeroFluxDenseAP

Ethanol

AEtherFlow

OpenSDWN

EmPOWER

AP Aggregate

RCWLAN

WVTCoAP and ResFi

OpenRoads

EmPOWER2

+ Traffic shaping

CENTAUR

SlicingSouth-bound

ProtocolSplit-MAC

SupportClient

A Seamless Handoff

B Client Steering

APicj= F chi

APitimes rchi

APicj (2)

where F chi

and rchi

ON OFF

and tinactive Tidle

and toff Toffline

LAPi =

AP1 AP2 APi

Controller

Client

SNMPdata

AGE request

Specified AP

factors

LAPi= max

sumforallcjisinC

ωcj timesXAPicj

rAPicj

sumforallcjisinC

ωcj timesXAPicj

forallcjisinC

forallAPiisinAP

logsum

forallAPiisinAP

sumforallAPlisinAP

SeamlessHandoff

ClientSteering

Cent GenHints

Indv OptCent Made

Decisions

Glob OptCent Made

FairnessThroughput

each beacon

with each beacon

chi chj

NAPi+NAPj

Client

3AP1 AP2

InvaderAP

VictimAP

Collision1

forallAPiisinAP

sumforallchjisinCH

UAPichj (12)

WCS =sum

New assignment

Old assignment6=

SNMP statistics

T (APi) (15)

maxsum

U(APi chj) (16)

andmax( min

VI RELATED SURVEYS

VII CONCLUSION

VIII ACKNOWLEDGMENT

REFERENCES

I Introduction

I-A Contributions

II Motivation


II-B Challenges


III Architectures







III-C4 Scalability




III-D2 Mobility



III-D5 Home WLANs

III-D6 Security





















V-C6 Security

VI Related Surveys

VII Conclusion

VIII Acknowledgment

References

Sd-wlan

Trantor

SplitAP

+ Slicing

CloudMAC

AeroFluxDenseAP

Ethanol

AEtherFlow

OpenSDWN

EmPOWER

AP Aggregate

RCWLAN

WVTCoAP and ResFi

OpenRoads

EmPOWER2

+ Traffic shaping

CENTAUR

SlicingSouth-bound

ProtocolSplit-MAC

SupportClient

A Seamless Handoff

B Client Steering

APicj= F chi

APitimes rchi

APicj (2)

where F chi

and rchi

ON OFF

and tinactive Tidle

and toff Toffline

LAPi =

AP1 AP2 APi

Controller

Client

SNMPdata

AGE request

Specified AP

factors

LAPi= max

sumforallcjisinC

ωcj timesXAPicj

rAPicj

sumforallcjisinC

ωcj timesXAPicj

forallcjisinC

forallAPiisinAP

logsum

forallAPiisinAP

sumforallAPlisinAP

SeamlessHandoff

ClientSteering

Cent GenHints

Indv OptCent Made

Decisions

Glob OptCent Made

FairnessThroughput

each beacon

with each beacon

chi chj

NAPi+NAPj

Client

3AP1 AP2

InvaderAP

VictimAP

Collision1

forallAPiisinAP

sumforallchjisinCH

UAPichj (12)

WCS =sum

New assignment

Old assignment6=

SNMP statistics

T (APi) (15)

maxsum

U(APi chj) (16)

andmax( min

VI RELATED SURVEYS

VII CONCLUSION

VIII ACKNOWLEDGMENT

REFERENCES

I Introduction

I-A Contributions

II Motivation


II-B Challenges


III Architectures







III-C4 Scalability




III-D2 Mobility



III-D5 Home WLANs

III-D6 Security





















V-C6 Security

VI Related Surveys

VII Conclusion

VIII Acknowledgment

References

Sd-wlan

Trantor

SplitAP

+ Slicing

CloudMAC

AeroFluxDenseAP

Ethanol

AEtherFlow

OpenSDWN

EmPOWER

AP Aggregate

RCWLAN

WVTCoAP and ResFi

OpenRoads

EmPOWER2

+ Traffic shaping

CENTAUR

SlicingSouth-bound

ProtocolSplit-MAC

SupportClient

A Seamless Handoff

B Client Steering

APicj= F chi

APitimes rchi

APicj (2)

where F chi

and rchi

ON OFF

and tinactive Tidle

and toff Toffline

LAPi =

AP1 AP2 APi

Controller

Client

SNMPdata

AGE request

Specified AP

factors

LAPi= max

sumforallcjisinC

ωcj timesXAPicj

rAPicj

sumforallcjisinC

ωcj timesXAPicj

forallcjisinC

forallAPiisinAP

logsum

forallAPiisinAP

sumforallAPlisinAP

SeamlessHandoff

ClientSteering

Cent GenHints

Indv OptCent Made

Decisions

Glob OptCent Made

FairnessThroughput

each beacon

with each beacon

chi chj

NAPi+NAPj

Client

3AP1 AP2

InvaderAP

VictimAP

Collision1

forallAPiisinAP

sumforallchjisinCH

UAPichj (12)

WCS =sum

New assignment

Old assignment6=

SNMP statistics

T (APi) (15)

maxsum

U(APi chj) (16)

andmax( min

VI RELATED SURVEYS

VII CONCLUSION

VIII ACKNOWLEDGMENT

REFERENCES

I Introduction

I-A Contributions

II Motivation


II-B Challenges


III Architectures







III-C4 Scalability




III-D2 Mobility



III-D5 Home WLANs

III-D6 Security





















V-C6 Security

VI Related Surveys

VII Conclusion

VIII Acknowledgment

References

SlicingSouth-bound

ProtocolSplit-MAC

SupportClient

A Seamless Handoff

B Client Steering

APicj= F chi

APitimes rchi

APicj (2)

where F chi

and rchi

ON OFF

and tinactive Tidle

and toff Toffline

LAPi =

AP1 AP2 APi

Controller

Client

SNMPdata

AGE request

Specified AP

factors

LAPi= max

sumforallcjisinC

ωcj timesXAPicj

rAPicj

sumforallcjisinC

ωcj timesXAPicj

forallcjisinC

forallAPiisinAP

logsum

forallAPiisinAP

sumforallAPlisinAP

SeamlessHandoff

ClientSteering

Cent GenHints

Indv OptCent Made

Decisions

Glob OptCent Made

FairnessThroughput

each beacon

with each beacon

chi chj

NAPi+NAPj

Client

3AP1 AP2

InvaderAP

VictimAP

Collision1

forallAPiisinAP

sumforallchjisinCH

UAPichj (12)

WCS =sum

New assignment

Old assignment6=

SNMP statistics

T (APi) (15)

maxsum

U(APi chj) (16)

andmax( min

VI RELATED SURVEYS

VII CONCLUSION

VIII ACKNOWLEDGMENT

REFERENCES

I Introduction

I-A Contributions

II Motivation


II-B Challenges


III Architectures







III-C4 Scalability




III-D2 Mobility



III-D5 Home WLANs

III-D6 Security





















V-C6 Security

VI Related Surveys

VII Conclusion

VIII Acknowledgment

References

A Seamless Handoff

B Client Steering

APicj= F chi

APitimes rchi

APicj (2)

where F chi

and rchi

ON OFF

and tinactive Tidle

and toff Toffline

LAPi =

AP1 AP2 APi

Controller

Client

SNMPdata

AGE request

Specified AP

factors

LAPi= max

sumforallcjisinC

ωcj timesXAPicj

rAPicj

sumforallcjisinC

ωcj timesXAPicj

forallcjisinC

forallAPiisinAP

logsum

forallAPiisinAP

sumforallAPlisinAP

SeamlessHandoff

ClientSteering

Cent GenHints

Indv OptCent Made

Decisions

Glob OptCent Made

FairnessThroughput

each beacon

with each beacon

chi chj

NAPi+NAPj

Client

3AP1 AP2

InvaderAP

VictimAP

Collision1

forallAPiisinAP

sumforallchjisinCH

UAPichj (12)

WCS =sum

New assignment

Old assignment6=

SNMP statistics

T (APi) (15)

maxsum

U(APi chj) (16)

andmax( min

VI RELATED SURVEYS

VII CONCLUSION

VIII ACKNOWLEDGMENT

REFERENCES

I Introduction

I-A Contributions

II Motivation


II-B Challenges


III Architectures







III-C4 Scalability




III-D2 Mobility



III-D5 Home WLANs

III-D6 Security





















V-C6 Security

VI Related Surveys

VII Conclusion

VIII Acknowledgment

References

B Client Steering

APicj= F chi

APitimes rchi

APicj (2)

where F chi

and rchi

ON OFF

and tinactive Tidle

and toff Toffline

LAPi =

AP1 AP2 APi

Controller

Client

SNMPdata

AGE request

Specified AP

factors

LAPi= max

sumforallcjisinC

ωcj timesXAPicj

rAPicj

sumforallcjisinC

ωcj timesXAPicj

forallcjisinC

forallAPiisinAP

logsum

forallAPiisinAP

sumforallAPlisinAP

SeamlessHandoff

ClientSteering

Cent GenHints

Indv OptCent Made

Decisions

Glob OptCent Made

FairnessThroughput

each beacon

with each beacon

chi chj

NAPi+NAPj

Client

3AP1 AP2

InvaderAP

VictimAP

Collision1

forallAPiisinAP

sumforallchjisinCH

UAPichj (12)

WCS =sum

New assignment

Old assignment6=

SNMP statistics

T (APi) (15)

maxsum

U(APi chj) (16)

andmax( min

VI RELATED SURVEYS

VII CONCLUSION

VIII ACKNOWLEDGMENT

REFERENCES

I Introduction

I-A Contributions

II Motivation


II-B Challenges


III Architectures







III-C4 Scalability




III-D2 Mobility



III-D5 Home WLANs

III-D6 Security





















V-C6 Security

VI Related Surveys

VII Conclusion

VIII Acknowledgment

References

APicj= F chi

APitimes rchi

APicj (2)

where F chi

and rchi

ON OFF

and tinactive Tidle

and toff Toffline

LAPi =

AP1 AP2 APi

Controller

Client

SNMPdata

AGE request

Specified AP

factors

LAPi= max

sumforallcjisinC

ωcj timesXAPicj

rAPicj

sumforallcjisinC

ωcj timesXAPicj

forallcjisinC

forallAPiisinAP

logsum

forallAPiisinAP

sumforallAPlisinAP

SeamlessHandoff

ClientSteering

Cent GenHints

Indv OptCent Made

Decisions

Glob OptCent Made

FairnessThroughput

each beacon

with each beacon

chi chj

NAPi+NAPj

Client

3AP1 AP2

InvaderAP

VictimAP

Collision1

forallAPiisinAP

sumforallchjisinCH

UAPichj (12)

WCS =sum

New assignment

Old assignment6=

SNMP statistics

T (APi) (15)

maxsum

U(APi chj) (16)

andmax( min

VI RELATED SURVEYS

VII CONCLUSION

VIII ACKNOWLEDGMENT

REFERENCES

I Introduction

I-A Contributions

II Motivation


II-B Challenges


III Architectures







III-C4 Scalability




III-D2 Mobility



III-D5 Home WLANs

III-D6 Security





















V-C6 Security

VI Related Surveys

VII Conclusion

VIII Acknowledgment

References

LAPi =

AP1 AP2 APi

Controller

Client

SNMPdata

AGE request

Specified AP

factors

LAPi= max

sumforallcjisinC

ωcj timesXAPicj

rAPicj

sumforallcjisinC

ωcj timesXAPicj

forallcjisinC

forallAPiisinAP

logsum

forallAPiisinAP

sumforallAPlisinAP

SeamlessHandoff

ClientSteering

Cent GenHints

Indv OptCent Made

Decisions

Glob OptCent Made

FairnessThroughput

each beacon

with each beacon

chi chj

NAPi+NAPj

Client

3AP1 AP2

InvaderAP

VictimAP

Collision1

forallAPiisinAP

sumforallchjisinCH

UAPichj (12)

WCS =sum

New assignment

Old assignment6=

SNMP statistics

T (APi) (15)

maxsum

U(APi chj) (16)

andmax( min

VI RELATED SURVEYS

VII CONCLUSION

VIII ACKNOWLEDGMENT

REFERENCES

I Introduction

I-A Contributions

II Motivation


II-B Challenges


III Architectures







III-C4 Scalability




III-D2 Mobility



III-D5 Home WLANs

III-D6 Security





















V-C6 Security

VI Related Surveys

VII Conclusion

VIII Acknowledgment

References

logsum

forallAPiisinAP

sumforallAPlisinAP

SeamlessHandoff

ClientSteering

Cent GenHints

Indv OptCent Made

Decisions

Glob OptCent Made

FairnessThroughput

each beacon

with each beacon

chi chj

NAPi+NAPj

Client

3AP1 AP2

InvaderAP

VictimAP

Collision1

forallAPiisinAP

sumforallchjisinCH

UAPichj (12)

WCS =sum

New assignment

Old assignment6=

SNMP statistics

T (APi) (15)

maxsum

U(APi chj) (16)

andmax( min

VI RELATED SURVEYS

VII CONCLUSION

VIII ACKNOWLEDGMENT

REFERENCES

I Introduction

I-A Contributions

II Motivation


II-B Challenges


III Architectures







III-C4 Scalability




III-D2 Mobility



III-D5 Home WLANs

III-D6 Security





















V-C6 Security

VI Related Surveys

VII Conclusion

VIII Acknowledgment

References

forallAPiisinAP

sumforallAPlisinAP

SeamlessHandoff

ClientSteering

Cent GenHints

Indv OptCent Made

Decisions

Glob OptCent Made

FairnessThroughput

each beacon

with each beacon

chi chj

NAPi+NAPj

Client

3AP1 AP2

InvaderAP

VictimAP

Collision1

forallAPiisinAP

sumforallchjisinCH

UAPichj (12)

WCS =sum

New assignment

Old assignment6=

SNMP statistics

T (APi) (15)

maxsum

U(APi chj) (16)

andmax( min

VI RELATED SURVEYS

VII CONCLUSION

VIII ACKNOWLEDGMENT

REFERENCES

I Introduction

I-A Contributions

II Motivation


II-B Challenges


III Architectures







III-C4 Scalability




III-D2 Mobility



III-D5 Home WLANs

III-D6 Security





















V-C6 Security

VI Related Surveys

VII Conclusion

VIII Acknowledgment

References

SeamlessHandoff

ClientSteering

Cent GenHints

Indv OptCent Made

Decisions

Glob OptCent Made

FairnessThroughput

each beacon

with each beacon

chi chj

NAPi+NAPj

Client

3AP1 AP2

InvaderAP

VictimAP

Collision1

forallAPiisinAP

sumforallchjisinCH

UAPichj (12)

WCS =sum

New assignment

Old assignment6=

SNMP statistics

T (APi) (15)

maxsum

U(APi chj) (16)

andmax( min

VI RELATED SURVEYS

VII CONCLUSION

VIII ACKNOWLEDGMENT

REFERENCES

I Introduction

I-A Contributions

II Motivation


II-B Challenges


III Architectures







III-C4 Scalability




III-D2 Mobility



III-D5 Home WLANs

III-D6 Security





















V-C6 Security

VI Related Surveys

VII Conclusion

VIII Acknowledgment

References

each beacon

with each beacon

chi chj

NAPi+NAPj

Client

3AP1 AP2

InvaderAP

VictimAP

Collision1

forallAPiisinAP

sumforallchjisinCH

UAPichj (12)

WCS =sum

New assignment

Old assignment6=

SNMP statistics

T (APi) (15)

maxsum

U(APi chj) (16)

andmax( min

VI RELATED SURVEYS

VII CONCLUSION

VIII ACKNOWLEDGMENT

REFERENCES

I Introduction

I-A Contributions

II Motivation


II-B Challenges


III Architectures







III-C4 Scalability




III-D2 Mobility



III-D5 Home WLANs

III-D6 Security





















V-C6 Security

VI Related Surveys

VII Conclusion

VIII Acknowledgment

References

each beacon

with each beacon

chi chj

NAPi+NAPj

Client

3AP1 AP2

InvaderAP

VictimAP

Collision1

forallAPiisinAP

sumforallchjisinCH

UAPichj (12)

WCS =sum

New assignment

Old assignment6=

SNMP statistics

T (APi) (15)

maxsum

U(APi chj) (16)

andmax( min

VI RELATED SURVEYS

VII CONCLUSION

VIII ACKNOWLEDGMENT

REFERENCES

I Introduction

I-A Contributions

II Motivation


II-B Challenges


III Architectures







III-C4 Scalability




III-D2 Mobility



III-D5 Home WLANs

III-D6 Security





















V-C6 Security

VI Related Surveys

VII Conclusion

VIII Acknowledgment

References

Client

3AP1 AP2

InvaderAP

VictimAP

Collision1

forallAPiisinAP

sumforallchjisinCH

UAPichj (12)

WCS =sum

New assignment

Old assignment6=

SNMP statistics

T (APi) (15)

maxsum

U(APi chj) (16)

andmax( min

VI RELATED SURVEYS

VII CONCLUSION

VIII ACKNOWLEDGMENT

REFERENCES

I Introduction

I-A Contributions

II Motivation


II-B Challenges


III Architectures







III-C4 Scalability




III-D2 Mobility



III-D5 Home WLANs

III-D6 Security





















V-C6 Security

VI Related Surveys

VII Conclusion

VIII Acknowledgment

References

3AP1 AP2

InvaderAP

VictimAP

Collision1

forallAPiisinAP

sumforallchjisinCH

UAPichj (12)

WCS =sum

New assignment

Old assignment6=

SNMP statistics

T (APi) (15)

maxsum

U(APi chj) (16)

andmax( min

VI RELATED SURVEYS

VII CONCLUSION

VIII ACKNOWLEDGMENT

REFERENCES

I Introduction

I-A Contributions

II Motivation


II-B Challenges


III Architectures







III-C4 Scalability




III-D2 Mobility



III-D5 Home WLANs

III-D6 Security





















V-C6 Security

VI Related Surveys

VII Conclusion

VIII Acknowledgment

References

New assignment

Old assignment6=

SNMP statistics

T (APi) (15)

maxsum

U(APi chj) (16)

andmax( min

VI RELATED SURVEYS

VII CONCLUSION

VIII ACKNOWLEDGMENT

REFERENCES

I Introduction

I-A Contributions

II Motivation


II-B Challenges


III Architectures







III-C4 Scalability




III-D2 Mobility



III-D5 Home WLANs

III-D6 Security





















V-C6 Security

VI Related Surveys

VII Conclusion

VIII Acknowledgment

References

T (APi) (15)

maxsum

U(APi chj) (16)

andmax( min

VI RELATED SURVEYS

VII CONCLUSION

VIII ACKNOWLEDGMENT

REFERENCES

I Introduction

I-A Contributions

II Motivation


II-B Challenges


III Architectures







III-C4 Scalability




III-D2 Mobility



III-D5 Home WLANs

III-D6 Security





















V-C6 Security

VI Related Surveys

VII Conclusion

VIII Acknowledgment

References

VI RELATED SURVEYS

VII CONCLUSION

VIII ACKNOWLEDGMENT

REFERENCES

I Introduction

I-A Contributions

II Motivation


II-B Challenges


III Architectures







III-C4 Scalability




III-D2 Mobility



III-D5 Home WLANs

III-D6 Security





















V-C6 Security

VI Related Surveys

VII Conclusion

VIII Acknowledgment

References

VI RELATED SURVEYS

VII CONCLUSION

VIII ACKNOWLEDGMENT

REFERENCES

I Introduction

I-A Contributions

II Motivation


II-B Challenges


III Architectures







III-C4 Scalability




III-D2 Mobility



III-D5 Home WLANs

III-D6 Security





















V-C6 Security

VI Related Surveys

VII Conclusion

VIII Acknowledgment

References

VIII ACKNOWLEDGMENT

REFERENCES

I Introduction

I-A Contributions

II Motivation


II-B Challenges


III Architectures







III-C4 Scalability




III-D2 Mobility



III-D5 Home WLANs

III-D6 Security





















V-C6 Security

VI Related Surveys

VII Conclusion

VIII Acknowledgment

References

I Introduction

I-A Contributions

II Motivation


II-B Challenges


III Architectures







III-C4 Scalability




III-D2 Mobility



III-D5 Home WLANs

III-D6 Security





















V-C6 Security

VI Related Surveys

VII Conclusion

VIII Acknowledgment

References

I Introduction

I-A Contributions

II Motivation


II-B Challenges


III Architectures







III-C4 Scalability




III-D2 Mobility



III-D5 Home WLANs

III-D6 Security





















V-C6 Security

VI Related Surveys

VII Conclusion

VIII Acknowledgment

References

I Introduction

I-A Contributions

II Motivation


II-B Challenges


III Architectures







III-C4 Scalability




III-D2 Mobility



III-D5 Home WLANs

III-D6 Security





















V-C6 Security

VI Related Surveys

VII Conclusion

VIII Acknowledgment

References

I Introduction

I-A Contributions

II Motivation


II-B Challenges


III Architectures







III-C4 Scalability




III-D2 Mobility



III-D5 Home WLANs

III-D6 Security





















V-C6 Security

VI Related Surveys

VII Conclusion

VIII Acknowledgment

References

Documents

IEEE COMMUNICATIONS SURVEYS AND TUTORIALS, ACCEPTED …bdezfouli/publication/COMST-SDWLAN-2018.pdf · and improve the design and development of network con-trol mechanisms by decoupling