A Distributed Simulator Coordination Platform and itsApplication for Integrating an IEEE 802.11 Effective

Capacity-based Admission Control Algorithm

Emmanouil KafetzakisInstitute of Informatics &

TelecommunicationsNCSR “Demokritos”

Ag. Paraskevi, Greecemkafetz@iit.demokritos.gr

Kimon KontovasilisInstitute of Informatics &

TelecommunicationsNCSR “Demokritos”

Ag. Paraskevi, Greecekkont@iit.demokritos.gr

Lambros SarakisInstitute of Informatics &

TelecommunicationsNCSR “Demokritos”

Ag. Paraskevi, Greecesarakis@iit.demokritos.gr

ABSTRACTEfficient coordination and management of radio resourcesin heterogeneous wireless networks is a key requirement forthe realization of the Beyond 3G vision. In this context,radio resource optimization algorithms as well as infrastruc-tures that allow for their performance evaluation are of greatvalue. Towards these directions, the goal of the paper istwofold: first to present a testbed that enables integrationof geographically distributed modules simulating differentradio access technologies with independent realizations ofinter-system optimization functions, and second to presentthe integration of an algorithm for admission control in IEEE802.11 networks into the described platform. The algorithm,which is based on the notion of Effective Capacity and pro-vides statistical overflow probability-related QoS guarantees,is validated using both a simulator in a stand-alone fashionand a simulator attached to the presented testbed. Theintegration of the algorithm on the testbed validates thetestbed’s capability to serve as a facility for optimizationstudies in simulated composite radio networks.

Categories and Subject DescriptorsI.6.7 [Simulation and Modeling]: Simulation SupportSystems; C.2.1 [Computer-Communication Networks]:Network Architecture and Design; C.4 [Performance ofSystems]: Modeling techniques; G.3 [Probability andStatistics]: Queueing theory

General TermsAlgorithms, Performance, Design, Experimentation, The-ory, Verification

KeywordsDistributed Virtual Testbed, Event-based middleware, Ad-

Permission to make digital or hard copies of all or part of this work forpersonal or classroom use is granted without fee provided that copies arenot made or distributed for profit or commercial advantage and that copiesbear this notice and the full citation on the first page. To copy otherwise, torepublish, to post on servers or to redistribute to lists, requires prior specificpermission and/or a fee.SIMUTools 2009, Rome, Italy.Copyright 2009 ICST, ISBN 978-963-9799-9799-45-5.

mission control, IEEE 802.11, QoS, Effective Capacity

1. INTRODUCTIONThe convergence of circuit-switched and packet-switched

networks and, at the same time, the increase of availabledata rates of wireless links has led to the rapid evolutionof the so-called Beyond 3G networks integrating heteroge-neous wireless access technologies. In such environments,terminals are equipped with multiple radio interfaces andare capable of attaching to different wireless access networkseither alternatively or concurrently. Furthermore, serviceswith strict Quality of Service (QoS) requirements can beallocated to different networks (either at the session estab-lishment phase or during the session progress) so as to fulfillcost and QoS restrictions.

In this regard, efficient coordination and management sche-mes must exist to allocate available resources among the dif-ferent networks in an optimized manner. This is the idea ofinter-system optimization addressed by joint (or common,or multi-access) Radio Resource Management (RRM) sche-mes [9]. What is defined by joint RRM (JRRM) is a set ofalgorithms usually running on a centralized network entitythat take decisions on the optimal access network selectionor the reallocation of resources between different wirelessaccess networks.

Many inter-system optimization schemes have been pro-posed in the literature for initial access selection, joint ad-mission control and resource scheduling, inter-system han-dovers and load sharing for heterogeneous networking en-vironments. Validation of such algorithms under close-to-reality conditions requires platforms that are able to inte-grate (possibly) disparate simulators (and even emulatorsand physical testbeds) of wireless technologies and, at thesame time, provide for hooks that allow for smooth inte-gration of individual implementations of optimization algo-rithms. Another requirement for such a platform is the easeof usability: the end-user should be able to identify availablesimulators and algorithms, construct simulation models andsubmit them for execution, monitor the simulation statusand get simulation results.

Many experimental testbeds exist for wireless network re-search (e.g., WHYNET [14], Emulab [16], Panlab [5]) ad-dressing simulated, emulated, physical and hybrid systems.However, although these represent significant efforts towardsproviding versatile experimentation platforms, they mostly

target at providing large-scale experimental infrastructuresthat can be used to evaluate new end-to-end approaches fornext generation networking environments (like Panlab) or atproviding platforms that interconnect well-established sim-ulators (like ns-2) with a special focus on supporting cross-layer optimization studies (like WHYNET). In this regard,they constitute testbeds in which smooth integration of newsimulators and JRRM algorithms can be difficult, as vitalparts of the infrastructure need to be modified.

Motivated by the need for an open and flexible testbedthat allows for simulations of different radio access tech-nologies and, at the same time, provides facilities for testingoptimization algorithms, the UNITE project [13] aims to de-liver a simulator coordination platform (Virtual DistributedTestbed - VDT) that interconnects possibly physically iso-lated software simulators and independent implementationsof optimization algorithms. Each one of the interconnectedsimulators may correspond to a different radio technology,constructing, in this way, a VDT for simulating heteroge-neous wireless networks found in a Beyond 3G setting. Byanalyzing these requirements, it becomes apparent that afederated simulation system is a natural choice for such anexperimental testbed.

Distributed systems organized in a federated fashion havereceived a lot of attention during the last years (see, forexample, [11]) and the High Level Architecture (HLA) hasbeen specified as a standardized way for the interconnectionand coordination of federates [6]. However, the HLA im-plies strict rules that the federates must adhere to in orderfor them to communicate. For this reason, a customizedmechanism for simulator coordination including data dis-tribution and time management has been adopted for theUNITE VDT.

In this paper we describe the proposed simulator coor-dination platform, illustrate its application for integratingadmission control algorithms and provide results from simu-lations governed by a particular admission control algorithmfor IEEE 802.11, so as to validate the architecture and thealgorithm integration (extending in this way the work pre-sented in [15]).

Towards this end, we present the overall architecture ofthe platform, we discuss the functionality of the key buildingblocks and present the events through which communicationand synchronization of the different simulation clusters areachieved. The main strengths of this architecture include a)support for distributed simulations of composite networkswhere different radio technologies are simulated in disparatesimulators, b) integration of virtual and physical clusters,and c) integration of inter-system (and cross-layer) optimiza-tion algorithms. The design of the VDT is modular in thesense that new wireless network simulators and realizationsof joint admission control (and handover) algorithms can beattached to the testbed (expanding in this way the overalltestbed’s capabilities) without major modifications.

To illustrate the flexibility and potential of the UNITEVDT framework, the paper discusses a recently developedalgorithm for admission control, designed for providing tail-related QoS guarantees in IEEE 802.11 WLANs [7]. Thetheory underlying the admission control captures the ser-vice rate variability in 802.11 stations through the notion ofEffective Capacity. The resulting algorithm can be used forcontroling the admission of new stations in the WLAN, ornew traffic flows over already active stations. In the second

Figure 1: The architecture of the simulator coordi-nation testbed.

case, the algorithm could be used as a trigger for a verti-cal handover to another network (of the same technology, oreven a different one, in case the station has multimodal ca-pability). The fact that the admission control scheme couldbe integrated in the VDT with relatively minor effort, bymaking use of VDT constructs devised independently fromthe algorithm’s conception, provides reasonable evidence forthe potential of the simulator coordination platform.

The rest of the paper is organized as follows: Section 2 de-scribes the architecture of the simulator coordination plat-form, while Section 3 presents the main functional blocks(as well as the events that are used for the communicationbetween remote entities) that are required to support thetestbed’s goals. Section 4 presents the admission control al-gorithm for IEEE 802.11 networks and provides model val-idation and performance evaluation results using the ns-2simulator in a stand-alone fashion. The integration of thealgorithm into the simulator coordination platform, in whichsimulation of the IEEE 802.11 technology is addressed as afederate service managed remotely, and its evaluation in thissetting are presented in Section 5. Finally, Section 6 con-cludes the paper.

2. ARCHITECTURE OF THE SIMULATORCOORDINATION PLATFORM

The aim of the described platform is to integrate indepen-dent physically/geographically distributed simulation clus-ters and realizations of optimization algorithms so as to al-low for inter-system optimization studies in heterogeneousnetworking environments. Towards this end, two main func-tional requirements need to be addressed. The first relatesto the derivation of a simulation facility for composite wire-less networks by interconnecting existing simulators of di-verse Radio Access Technologies (RATs) running on differ-ent platforms. The second has to do with the incorporationof facilities for JRRM functions. These requirements canbe realized by the functional architecture depicted in Fig-ure 1. This architecture relies on the identification of fourconceptual levels for the simulator coordination testbed andthe definition of discrete building blocks residing at theselevels. At the topmost level, the end-user interface providesaccess to the Testbed Controller, which is responsible for

connecting the VDT platform with the external world. Us-ing this interface and the facilities provided by the TestbedController (e.g., identification of available simulators andoptimization algorithms, exposure of simulator parametersand list of supported events, etc.), end-users compile simu-lation plans and submit them for execution (plans are cre-ated through a graphical tool called VDT Editor). Addi-tional functionalities of this component include managementof user authentications and permissions, retrieval of pastsimulation plans from the repository, validation of simula-tion plans, scheduling of the simulation instance execution,monitoring of the simulation state and storage of simula-tion plans to the repository for future use (the repository isimplemented as part of the UNITE Data Base–UDB).

When a simulation is eligible for execution the TestbedController hands over control to the Central Controller. Thelatter is comprised of three modules: a) the UDB/Repositorymodule responsible for storing simulation results and simula-tion plans, b) the Statistics module that performs statisticalprocessing on the data of the UDB, and c) the Scenario Man-ager that undertakes functions like terminal management,service and traffic stream management and simulator clus-ters time management. The Scenario Manager is the corecomponent of the simulation platform and its functionalitywill be discussed in the next section.

At the lowest level of the VDT architecture, there are anumber of components (federates) responsible for simulat-ing radio technologies and implementing JRRM algorithms.These components are attached to the testbed by formingVDT modules (RAT VDT modules and UNITE RRM –URRM– VDT modules, respectively). These modules arecomprised of three entities: a) the actual simulator or opti-mization algorithm (in general the module/cluster service),b) the Federated Gateway (FG) which manages the simula-tion cluster or the execution of the algorithm and translatescluster-specific messages to VDT messages and vice versa,and c) the VDT module API which provides the interface(functions and parameter definitions), through which VDTmodules communicate with the components of the CentralController.

The cluster service calls specific functions from the FG tosend VDT Events. In the other direction, a specific functionfrom the cluster service is called each time a VDT Eventis received from the VDT Central Controller, so that theevent can be processed. The communication framework issupported by an event-based middleware implemented usingSOAP messages over HTTP. The control of the events andtheir routing towards the distributed modules registered forthem is undertaken by the Central Controller.

The VDT module is the basic building block of the VDTarchitecture and it can be used for any function that needsto be implemented in a separate hardware/software envi-ronment. In general, the VDT module can be consideredas an abstract implementation of a desired function servinga specific purpose. In this regard, the notion of the VDTmodule can perfectly apply as well to the simulation controlfunctions of the Central Controller. Additionally, the fed-erated service of a VDT module can also correspond to anemulated or physical networking testbed.

3. FUNCTIONS AND EVENTS FOR SIMU-LATION CONFIGURATION AND EXE-CUTION

Simulation configuration and execution control is under-taken by the Scenario Manager, which is in charge of log-ically integrating VDT modules corresponding to simula-tors of different radio technologies and realizations of inter-system optimization algorithms, thus presenting to the end-user a virtual distributed simulator of a composite networkconsisting of heterogeneous segments. It takes input fromthe Testbed Controller in the form of simulation plans editedby the end-users. Each simulation plan contains informationabout the simulators and the optimization algorithms thatwill be used and their configuration parameters, the numberof terminals that will participate in the simulation, the con-nectivity capability of these terminals with respect to thechosen RATs, the service types, QoS requirements and traf-fic configuration parameters, and the preferred RAT for ser-vice initiation. The information contained in the simulationplans corresponds to configuration parameters that do notchange during the simulation run-time (e.g., it is assumedthat session arrival events are independent of the state ofthe simulation).

In the following we briefly describe the functions of theScenario Manager and the operations of the VDT modulesthat are responsible for the simulation of radio technologiesand the implementation of joint admission control (in gen-eral initial access selection) algorithms. The discussion ofthe functions is complemented by a set of events that havebeen used to implement the corresponding functionality. De-tailed description of the functions of the Scenario Managerand the events needed for the integration of VDT modulesimplementing handover algorithms can be found in [15].

3.1 Terminal and Service ManagementThe VDT platform addresses two types of terminals: single-

mode terminals which are connected to and managed by onlyone simulator and multi-mode terminals that have connec-tivity to several simulators (each one simulating the radiotechnology of a given interface). At the system level (com-posite simulator level) every mobile terminal is allocateda unique global identifier (ID). For multimodal terminals,however, there is an additional need to identify the mobilein all systems it is supposed to have access to. This intro-duces a second level of mobile indexation inside RAT VDTmodules. The mapping between the global mobile ID andthe simulator-specific ID is based on the concept that (vir-tual) multi-mode terminals are identified by many instances,each one of which applies to a different simulator. The globalIDs are provided by the VDT Editor and are managed bythe Scenario Manager in a centralized manner while the lo-cal ones are assigned by the corresponding simulators uponterminal registration in the simulated radio access system.

The VDT testbed supports several types of services, all ofwhich comply with a common service description. Servicescorrespond to sessions initiated between mobile nodes in dif-ferent RATs and fixed nodes (i.e., nodes that are supposedto belong to a wired infrastructure outside the wireless sim-ulators). The description of the VDT services and the con-figuration of the associated sessions are given in a combinedstructure (in the form of records provided by the VDT Ed-itor) including service description (identified by service ID

and corresponding parameters), service QoS requirements(QoS identifier and parameters), service binding (source anddestination terminal IDs), service start (and possibly stop)time and description for the session initiation (inter-sessionidle time distribution and parameters) and session durationprocesses (distribution and parameters). Examples of ser-vices include Voice over IP, HTTP, Near Real Time Video,FTP (provided they are supported by the simulators) as wellas more abstract services corresponding to CBR and Poissontraffic models. Services are characterized by QoS require-ments (e.g., minimum rate, tolerable packet loss/delay, etc.)and are bound to source and destination terminals identifiedby unique global IDs. Session arrivals are managed by theservice start time and intersession idle time, while sessiontermination takes place in accordance to session duration(and possibly a service stop time).

When a new session (traffic stream) is setup, it is assigneda unique ID. Several sessions (applications) can run on amobile using any of the available radio access technologies.The mapping of the sessions to source/destination mobileterminals and corresponding RAT VDT modules is handledby the Scenario Manager.

3.1.1 Supporting EventsWhen a new mobile must be added to a simulator, an Ad-

dMobile event is sent by the Scenario Manager. Since theinsertion of a new mobile has always to do with imminentservice activation, the characteristics of the service (includ-ing QoS requirements) are incorporated in the body of themessage. The newly added mobile will just appear in thesimulator as an idle node; it will not yet produce or receiveany traffic.

Service configuration is performed by the SendStreamTo-

Mobile event sent by the Scenario Manager to the RATVDT module. Traffic activation and deactivation, on theother hand, takes place through the StartTxToMobile andStopTxToMobile events (the StartTxToMobile event is notcombined with the SendStreamToMobile event in order toavoid multiple configurations of the same service in casesof frequent handovers). The StartTxToMobile is sent bythe Scenario Manager to the RAT VDT module, indicatingthat a specific mobile should start transmitting traffic ac-cording to the profile that had been previously configured.The StopTxToMobile event indicates that a specific mobileshould stop transmitting traffic. An additional RemoveMo-bile event has been specified to indicate that a specific mo-bile should be completely removed from a simulator in orderto free resources. All of the previously described events arecomplemented by appropriate replies.

3.2 Initial Configuration and Time Manage-ment

Each VDT module has specific configuration parametersthat must be initialized before the simulation starts. ForRAT VDT modules, these configuration parameters are sep-arated in two types. The first includes parameters that arespecific to the simulated RAT (e.g., physical and MAC layerparameters) and the second includes parameters (commonfor all simulators) for adding background traffic to the simu-lator (in this case the number of terminals and their serviceID are provided). For URRM VDT modules, on the otherhand, the parameters are algorithm-dependent.

Time management in the VDT platform is achieved by

a simple yet efficient scheme that guarantees that all con-nected simulators are synchronized at specified points intime (i.e., they have simulated up to a given time). Withthe aim of facilitating the validation of inter-system opti-mization algorithms, these points in time have been selectedto correspond to instances when inter-system optimizationfunctions are executed. This means that for the cases ofinitial network selection and inter-system handover oper-ations the simulators need to be synchronized both whennew sessions are setup and when handover algorithms areexecuted. The requirement to address synchronization atsession setup points results in an “irregular” set of synchro-nization points (in the sense that the intervals between syn-chronization points are not constant). Handover optimiza-tion algorithms, on the other hand, can be reasonably as-sumed to be executed on a regular basis (e.g., every fewseconds). If, however, this assumption is not acceptable itis the responsibility of the URRM VDT module to predictthe time in future it will require service by the time managerand communicate this information to the Scenario Manager.This prediction can be assisted from the fact that handoverdecisions are usually not based on single snapshots of thesystem state but rather on observations taken during sometime period.

3.2.1 Supporting EventsVDT module initialization parameters are provided by the

Scenario Manager through the ProcessConfiguration event(this is the only event sent by the Scenario Manager, theformat of which is simulator-dependent). After the initialconfiguration of the simulators, the Scenario Manager sendsa Start event to all RAT VDT modules to start the simula-tion.

Time synchronization in the VDT platform is achievedvia the ProcessTime and ProcessTimeReply events. Theformer is sent by the Scenario Manager to the simulatorsand allows them to simulate for the specified time interval(i.e., simulation step delineated by the start time and stoptime parameters). The stop time can coincide with the nexttime in future that the optimization algorithms will be ex-ecuted. The ProcessTimeReply event is sent as a reply tothe Scenario Manager, as soon as the specified interval hasbeen simulated and update of measurements in the UDBhas taken place. At the end of the simulation duration, aStop event is sent to all VDT modules, indicating that thesimulation is completed.

3.3 Management of Simulators’ ResultsEach simulator is responsible for producing simulation re-

sults for the time interval it was instructed to simulate. Af-ter the end of the simulation step and before sending thepermission for time advancement, each RAT VDT moduleupdates the UDB with new measurements. These resultsare of interest to both the Statistics entity of the CentralController and the URRM VDT modules.

3.3.1 Supporting EventsSimulator results for the most recent time interval are

propagated to the UDB through the UpdateValues event.In response to this event, the UDB module sends an Up-

dateValuesReply event, after the reception of which theRAT VDT module sends permission for time advancementthrough the ProcessTimeReply event.

To achieve proper communication between the URRMand the UDB modules two events have been specified. TheGetSystemInformation event is sent from the URRM to theUDB, requesting “system level” simulation results for thewhole composite network (e.g., total load, number of con-nected terminals, simulator configuration parameters, ter-minal and session mapping to different RAT VDT modules,etc.), while the GetMobileInformation event requests de-tailed information (e.g., throughput, packet error rate, sig-nal strength, etc.) for specific mobiles.

3.4 Management of Admission Control VDTmodules

When an admission control VDT module is included ina simulation, the default session allocation strategy, accord-ing to which sessions are allocated to different networks onthe basis of preferences defined by the end-user, is bypassedand the allocation of new sessions to simulators takes placeaccording to the decision of the algorithm. In such cases,the Scenario Manager takes the necessary actions to call theadmission control algorithm at appropriate time instancesand implement the decision.

3.4.1 Supporting EventsAt session initiation times an AdmissionRequest event is

sent by the Scenario Manager to the admission control mod-ule to let it decide on the admission of a new service inthe system. Apart from the terminal ID, service’s descrip-tion and QoS requirements, the message also includes a listof simulated networks that are candidates for accepting thenew service (this list may be utilized by joint admission con-trol algorithms). After the execution of the algorithm, anAdmissionReply event is sent to the Scenario Manager, in-dicating the acceptance or rejection of the service. In theformer case, the network in which the service will start isindicated.

4. EFF. CAPACITY-BASED ADMISSIONCONTROL FOR IEEE 802.11 WLANS

In this section we briefly describe a recently developedAdmission Control (AC) algorithm for IEEE 802.11 DCFenvironments. The algorithm is based on the notion of Ef-fective Capacity and can be used to provide tail-related QoSguarantees. The ability to enforce tail-related QoS guar-antees is a considerable merit of the algorithm given thatvirtually all other results for estimating and/or enforcingQoS for IEEE 802.11 are limited to mean-value quantities(e.g., throughput, mean packet delay, mean queue length).We limit ourselves to only the material necessary for un-derstanding and for applying the AC scheme. For details,proofs and further implications, the reader is referred to [7].Section 5 will address the details of incorporating this ad-mission control scheme within VDT.

4.1 Theoretical foundationsEffective Bandwidth theory offers a linkage between source

characteristics, system resources (server capacity and buffersize) and QoS tail-constraints. It was developed by a greatnumber of contributions from various researchers (see [8]for a survey in the field). The notion of Effective Band-width encapsulates the details of bandwidth time-varyingbursty sources in a single function, the Effective Bandwidth

function, that can be used to express the minimum con-stant server capacity required to satisfy a given overflowprobability-related constraint.

The theory was originally developed for queueing systemswith constant server capacity. When the server’s capacity istime-varying independently from the input, the theory canbe generalized, by defining an Effective Capacity function tocapture the server’s burstiness. Although this generalizationhas been studied for some years (see, e.g., [3, 4]) it did notattract much attention until recently when the importanceof wireless systems grew considerably. This is because mostsuch systems feature a variable service rate and the notionof Effective Capacity is ideal for modeling such settings. Inthe literature, application of Effective Capacity to wirelesssystems has focused on the modeling of rate fluctuations atthe physical layer [17, 18]. Instead, here and in [7] Effec-tive Capacity theory is employed to Medium Access Control(MAC) layer modeling.

For a quick review of the Effective Bandwidth/Capacitytheory, consider a single-server queue with time-varying ca-pacity, where the capacity fluctuations are independent fromthe input. Let the input traffic that produces an amount ofdata V (t) within the time window (0, t] feed the queue anddenote by C(t) the amount of data that can be served withinthe time window (0, t]. Assuming stationary and ergodic in-crements for the input and output processes and some ad-ditional mild technical conditions1, the probability that thequeue size Q exceeds a certain threshold x has at all timesan asymptotic exponential upper bound of rate θ, viz.,

aB(θ) ≤ aC(−θ) ⇐⇒ limx→∞

− log Pr{Q > x}x

≥ θ, (1)

where

aB(s) ,1

slimt→∞

1

tlogE[esV (t)], ∀s > 0,

and

aC(s) ,1

slimt→∞

1

tlogE[esC(t)], ∀s < 0,

stand for the Effective Bandwidth function and the EffectiveCapacity function, respectively.

Assume now that the queue has a finite size x, and thatwe want to provide stochastic QoS by limiting the overflowprobability to a value ≤ e−ε. By using the tail percentilePr{Q > x} of the respective infinite queue as a proxy for theoverflow probability, (1) suggests that the QoS is respected(asymptotically, for large ε and x, maintaining a finite ratio)if and only if

aB(θ?) ≤ aC(−θ?), for θ? = ε/x. (2)

Application of the admission control in (2) is eased furtherby noting that by definition, the Effective Bandwidth func-tion is additive, i.e., the Effective Bandwidth of a super-position of N independent traffic flows, i = 1, . . . , N (eachpossibly of different traffic profile) is simply the sum of the

Effective Bandwidths a(i)B (θ) of the constituent flows, namely

aB(θ) =

NXi=1

a(i)B (θ), ∀θ. (3)

As an example, consider the case where the AC scheme isapplied for deciding if a terminal with two running services

1Always satisfied in the setting addressed in this paper.

is admissible or not. Assume that the first service gener-ates constant rate traffic stream with bit rate Rcbr and thatthe second service generates a Poisson traffic stream witha constant packet size D and a mean bit rate Rpois. Then

a(1)B (θ) = Rcbr and a

(2)B (θ) = Rpois e

θD−1Dθ

and the EffectiveBandwidth of the aggregate traffic is governed by (3) withN = 2. Note that if the two services have different QoS re-quirements, ε1 and ε2, then the test in (2) should be appliedwith θ? = maxi=1,2 εi/x in order to satisfy both in a sharedbuffer.

As with Effective Bandwidths, the form of the EffectiveCapacity function depends on the details of the server pro-cess. When the latter is of an On/Off type, alternatingbetween On periods at peak server rate r and Off periods atzero rate, the Effective Capacity function can be explicitlydetermined [7] as

aC(θ) = u(θ)/θ, ∀θ < 0, (4)

where u(θ) is the unique negative solution of

log γon(rθ − u) + log γoff(−u) = 0. (5)

The functions γon(·) and γoff(·) are the moment generatorscorresponding to the distributions of the On and Off so-journs, respectively.

If one is interested in applying the AC test (2), ratherthan in evaluating the Effective Capacity function itself, itis not necessary to numerically solve (5). Using the mono-tonicity of the related functions [7], it can be proven thatthe inequality in (2) is equivalent to

log γon

`−rθ? + θ?aB(θ?)

´+ log γoff

`θ?aB(θ?)

´≤ 0. (6)

This condition simplifies greatly the computational aspectsof the AC scheme.

4.2 The Eff. Capacity of IEEE 802.11 mobilestations

Because of the CSMA/CA access algorithm used by theIEEE 802.11 protocol, a mobile station behaves as an On/Offserver [7]. The server is On, at rate equal to the channelbit rate r, when transmitting successfully the payload of apacket. In all other states of the CSMA/CA protocol (ter-minal backing-off, colliding with other terminals, or doingoverhead operations before or after a successful transmis-sion, e.g., RTS/CTS or ACK), the server is Off.

On the basis of these observations, the moment generatorfunction of the On period is

γon(ω) = E[eωPr ], (7)

where P is the payload size of the packet being transmit-ted. For packets of constant payload, the On period reducessimply to random variable of constant length.

The moment generator function of the Off period reads

γoff(ω) = eωtover“Bo + (1−Bo)eωtslotγbo(ω)

”. (8)

This last equation reflects that, if the backoff counter drawnat stage 0 (immediately after a successful transmission) iszero (an event of probability Bo), then the Off period sim-ply lasts the deterministic time tover, required for the over-heads before and after the successful transmission. In thecomplementary event, with probability 1 − Bo, the Off pe-riod additionally includes the deterministic time slot tslot

required for initially decrementing the backoff counter by

one, plus the time spent by the terminal in backoff mode.The moment generator for this backoff time is

γbo(ω) =go`γs(ω)

´−Bo

γs(ω)(1−Bo)

×»m−1Xl=0

“(1− p)plelωtcoll

lYj=1

gj`γs(ω)

´”+

(1− p)(peωtcoll)mQmj=1 gj

`γs(ω)

´1− pgm

`γs(ω)

´eωtcoll

–.

(9)

In (9), gi(z) stands for the generator function of the backoffcounter at the ith backoff stage. The equation reflects thefact that, beyond stage m, the backoff windows maintain thesame distribution. Eq. (9) allows for general backoff windowdistributions. For the special uniform backoff window dis-tribution described in the IEEE 802.11 standard,

gj(z) =

Wj−1Xl=0

1

Wjzl =

1

Wj

zWj − 1

z − 1,

where

Wj = 2min{j,m}Wo, j ≥ 0

and Wo−1 denotes the upper margin of the backoff windowat the 0th stage. For any backoff window distributions, wealways have Bo = go(0).

The quantity p in (9) denotes the probability of a collisionseen by a packet being transmitted on the channel (condi-tional collision probability). The value of p is obtained bysolving the relations

1− p = (1− τ)n−1, (10)

and

τ =1

1 + (1− p)`Wo

1−Bo − 1 +Pm−1i=1 piW i + pmWm

1−p

´ , (11)

between the conditional collision probability p and the trans-mission probability τ [1]. In (10) and (11), n is the numberof competing stations and Wi, i = 1 . . .m is the mean back-off window at the ith stage. Eq. (11) assumes saturationconditions, where every mobile station has always a packetto send.

Finally, the function γs(ω) appearing in (9) denotes themoment generator function of the time needed for the re-duction by one of the backoff counter viz.,

γs(ω) =Pcolleωtcoll + Pemptye

ωtslot

+ Psucc(1−Bo)γon(ω)eωtover

1−Boγon(ω)eωtovereωtslot ,

where

Psucc = (n− 1)τ(1− τ)n−2, Pempty = (1− τ)n−1,

Pcoll = 1− Psucc − Pempty,(12)

are the probabilities with which a successful transmission,an empty slot and a collision, respectively, are observed bya station backing-off (which observes n−1 other independentstations) and where tcoll represents the deterministic time fordetecting a collision. Eq. (12) again relies on the assumptionof saturation conditions.

All quantities involved in the On/Off server characteriza-tion just described, either depend on the probabilities p and

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1x 10!3

0

0.5

1

1.5

2

2.5

3

3.5

4

4.5x 105

QoS exponent ! (1/bit)

Effe

ctiv

e C

apac

ity E

c ( !

) (bi

t/sec

)2 terminals5 terminals10 terminals20 terminals

Increasingnumber of stations

Figure 2: Curves of aC(−θ) vs θ, for different valuesof the number of stations n.

τ (each of which can be expressed in terms of the otherthrough (10)), or are directly available as parameters spec-ified by the standard. Therefore, it is straightforward tocompute values of the moment generators of the On andOff periods (through (7) and (8), respectively), towards, ei-ther the computation of Effective Capacity values (through(4) and (5)), or testing for the AC through (6).

Note that, since the Effective Capacity model of the IEEE802.11 station employs the assumption that all other com-peting stations are saturated, the Effective Capacity of thestation does not depend on the details of traffic throughthese competing stations; the only information required isthe total number of stations n. Besides simplicity, in mostcases the saturation assumption provides a conservative up-per bound to the overflow probability (equivalently, a conser-vatively strict AC test), which becomes all the more accuratewhen the IEEE 802.11 WLAN approaches congestion.

4.3 Standalone validationThe Effective Capacity model has been validated against

stand-alone ns-2 simulation results, under various forms oftraffic load and number of competing terminals. For detailssee [7]. Here we limit ourselves in two results, in the inter-est of further highlighting the concepts already discussed.In both of the results the system parameter values corre-spond to Frequency Hopping Spread Spectrum (FHSS) PHYlayer [12], a configuration which, although somewhat out-dated, was intentionally chosen the same as the one in [1],for comparison purposes. Also, in all cases, the payload size(see (7)) was chosen constant and equal to 1/4 of the stan-dard’s maximum PDU (i.e., P = 8184 bits).

Figure 2 illustrates the accuracy of the Effective Capac-ity, by comparing graphs of the function (dashed, dottedand solid lines), computed with the use of the saturation-based analytic model, against simulation results (marks).The model results are consistent with known properties ofthe Effective Capacity function (monotonicity, limθ→0 aC(θ)equal to the mean service rate [7]) and depend on the num-ber of terminals in the network. As intuitively expected,for a given QoS parameter value θ the Effective Capacitydecreases as the network size increases.

! "! #! $! %! &!! &"! &#! &$!&!

!"

&!!&

&!!

'()()*+,(-.*/-*0123)45*

,6)789,:*07,+1+/9/4;

*&!*4)7</-195*199*514(714).*)=2)04*4>)*,+5)76).*********************

*

*

?/<(914/,-*@,/55,-*47188/2

A-19;4/219*B,.)9**@,/55,-*47188/2

?/<(914/,-*CDE*47188/2

A-19;4/219*B,.)9**CDE*47188/2

Figure 3: Modeling and simulation results of queuetail probabilites for 10 stations, one unsaturated, fortwo types of load.

In the simulation runs used for producing Figure 2, thevalues of the Effective Capacity function were indirectlymeasured, by feeding a “tagged” IEEE 802.11 station withtraffic of known profile, sampling the probability with whichthe station’s buffer exceeded a given threshold and exploit-ing the linkage (see (1)) between the Effective bandwidth ofthe input traffic, the Effective Capacity of the server andthe probability tail just mentioned. All terminals, besidesthe tagged one, were operating under saturation conditions.

The remarkable match between theory and simulationsvalidates the model and indicates its suitability for estimat-ing queue tails or, equivalently, for taking AC decisions inIEEE 802.11 WLANs under heavy load. This fact has beenreinforced by further extensive comparisons with simulationresults, of which one we now present.

Figure 3 depicts curves of the queue tail probabilities ver-sus the tail threshold (in semilog scale) for a network with10 stations, of which 9 are saturated. The queue of the un-saturated station has been observed under two kinds of traf-fic load, CBR and Poisson, both featuring the same meanrate of 79.84 kbps. The slope θ of the queue tail for themodel-derived curve in each loading case was determinedaccording to the theory, namely as θ = max{s : aB(s) ≤aC(−s)}, i.e., as the unique solution of aB(s) = aC(−s).

As shown in the figure, the simulation-derived queue tailsdecays exponentially and the decay rates agree well with themodel results. It may also be observed that the model cap-tures the dependence between the slope of decay and thedetails of the station’s input traffic (expressed through therespective Effective Bandwidth function). The decay rate isfaster when the unsaturated station is loaded with CBR traf-fic than with Poisson traffic of the same mean value, due tothe smoother nature of the former. The difference in offsetbetween model and simulation curves is attributed to theinherent shortcoming of the asymptotic theory in locatingthe point where the queue tail starts to decay exponentially.Despite this blemish, the Effective Capacity approach en-capsulates in a single function all complicated details of theIEEE 802.11 MAC layer and provides good QoS estimates.

5. INTEGRATION AND EVALUATION OFTHE AC SCHEME ON VDT

To address the integration of the AC scheme in the VDT,we now briefly summarize the basic steps of the scheme.When appropriate, we distinguish between tests for admis-sion of a new service on an already active IEEE 802.11 sta-tion (NS-test in the sequel) and tests for admitting a newlyarrived station (NT-test).

For both types of admission test, the first step consistsin calculating the appropriate value θ? and the EffectiveBandwidth of the input traffic aB(θ?). As already discussed(see the paragraph following eq. (3)), θ? is determined bythe strictest QoS requirement, among those of the servicesrunning on the station under test, and the station’s buffercapacity x, namely

θ? = maxi=1,...,s

{− logPi}/x. (13)

In case of an NS-test, (13) includes both the existing servicesand the newly arriving one, while for NT-tests the equationincludes the QoS specifications for all services to be initiallyengaged, should the terminal be admitted. Once θ? is avail-able, the Effective Bandwidth aB(θ?) is computed throughan application of (3). Note that the VDT uses service IDsto communicate service-related information between mod-ules. These IDs are mapped to the respective traffic andQoS descriptors internally, when needed.

The last step in applying the AC test is checking the in-equality in (6). Towards this, the total number of stations nin the WLAN is retrieved (augmented by one to accountfor the arrival of the new station, in case of an NT-test),equations (10) and (11) are solved, the probability values tothe left of (12) are computed and, finally, the test in (6) isapplied by invoking (9), (8) and (7).

Note that, in case of NT-tests, normally the test in (6)should be applied to all preexisting stations as well, becausethe increased number of stations n affects their Effective Ca-pacity as well (see Figure 2). In light of the mostly conserva-tive nature of the saturation-based model, these additionaltests are ommited in practice.

Table 1 compiles the input parameters for the AC schemeand the VDT entities from which these parameters are re-trieved. The first row in this table refers to the values of theIEEE 802.11 MAC parameters (DIFS, SIFS, backoff windowparameters, etc.) that pertain to the network into which ad-mission is requested. These parameters are required for cal-culating the constant values (including time constants liketover, tcoll, etc.) used by the equations in Subsection 4.2. Allnetwork-wide (as opposed to station-specific) parameters inthe table are distinguished by a mark, and a similar distinc-tion is made for parameters whose values don’t change withtime.

5.1 Communication between VDT modulesThe exchange of messages required to implement session

initiation with joint admission control is illustrated in Fig-ure 4. In this figure (which, for completeness, provides alsothe message exchange for session termination and mobile re-moval) it is assumed that the admission control algorithmcan operate on more than one network (although in thispaper only one is used) and that a normal ProcessTime/ProcessTimeReply round has been performed just prior tothe arrival of a new session.

Table 1: Input parameters for the AC scheme(‘*’=for entire WLAN,‘†’=static configuration).

Input parameters From where retrievedin VDT

IEEE 802.11 MAC UDB*,†parameters

Traffic & QoS Defined in VDT Editor,descriptors provided by Scenario Manager

x UDB†n UDB*

Figure 4: Messages exchange between modules forsession initiation with AC.

When a session is initiated for a non-existent station, anew mobile is added to each simulator. Assuming that thenew service can be supported in both networks an Admis-

sionRequest event is sent to the joint admission controlURRM module. The latter retrieves information from theUDB in order to take decision. The decision is then propa-gated to the Scenario Manager, which takes care of config-uring the traffic profile for the new station and of preparingtraffic activation (which will actually take place after thenext ProcessTime event).

After the traffic configuration is over, the simulators areinstructed to continue with normal operation and the Sce-nario Manager takes care of new session arrivals and timeadvancement. For each active session, the Scenario Managerkeeps track of the sessions’ duration, at the end of which itsends appropriate StopTxToMobile commands. When allservices for a station have been ended, it can be safely re-moved from the simulators (with the RemoveMobile event)in order to free computational and memory resources.

The implementation of the algorithm is addressed in theURRM entity. The algorithm runs through the Admission-

Control() function. The IEEE 802.11 MAC parametersand the number of stations already present in the network(see Table 1) are requested and obtained from UDB thoughevents GetSystemInformation and GetSystemInformation-

Reply, respectively, while the buffer size of the station sub-jected to the AC is requested and obtained from UDB viathe GetMobileInformation and GetMobileInformationRe-

ply events (see Figure 5). The input traffic descriptors

GetSystemInformationReply

AdmissionRequest & GetMobileInformationReply

!

!

!

!

!

!!!!!"##$!

!

!

!

!

!

AdmissionRequestReply!

!

!

!!!!!!!!!!!!!!!!!"#$!

!

!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!

!!$%&'(%)&!"#$!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!

*&+,,'(!-).(&'!!!!

!

!

!!!!!!/,,0!1+2!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!

!

!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!

/,,0!1+234'3%5! /,,0!6+7+('%8!

9!#,!%)&:'2+;.!

!%)&:':'+;.!

<=6>&);+%)3!

7&#?+?';'%').!

=3:'..'#2!6#2%&#;!(&'%)&'#2!

*&+,,'(!3).(&'7%#&!

"#$!3).(&'7%#&!!

$%&'(%)&!

"#$!

"#$!)@7!

1A,,)&!

%5&).5#;3!

Figure 5: Flow chart representation of the AC algo-rithm.

and the overflow probability-related QoS requirements areobtained through the AdmissionRequest from the ScenarioManager. The URRM computes the required Effective Band-width functions on the basis of the traffic descriptors. Fi-nally, according to the AC criterion, i.e., inequality (6), theURRM decides whether to accept the new service or notand this decision is disseminated to the Scenario Managerthrough the AdmissionRequestReply event.

5.2 Scenarios for the AC algorithm evaluationThe integration of the AC algorithm into the VDT was

validated through distributed simulations involving the VDTCentral Controller (including for the purposes of the teststhe Scenario Manager and the UDB), an IEEE 802.11 simu-lator (called Pythagor) [10] that was extended for interfacingto the VDT, a URRM module implementing the AC, and thegraphical application for the VDT Editor. The VDT CentralController (running on Linux) was hosted in a server locatedin Lisbon, Portugal, while the other components (runningon Windows XP) were hosted on different PCs located inAthens, Greece. The communication between the differentmodules was over the standard Internet.

In all the simulation tests considered in this subsection,the Pythagor was set to simulate a legacy IEEE 802.11 net-work with a rate of 1 Mbps. This configuration, althoughdated, was intentionally chosen to enable comparison withthe standalone tests of Subsection 4.3, as well as many othertests in the literature (notable [1]). The first test addresseda setting according to which a number of sessions were ini-tiated on mobile stations connected to an IEEE 802.11, andthe AC algorithm was called at session initiation instances todecide whether the sessions could be admitted in the WLAN.The AC algorithm was taking decisions on the basis of thenumber of terminals that were present in the WLAN andon the QoS requirements of the new sessions. New sessionswere allocated on previously inactive stations, and their du-ration was appropriately set to ensure that sessions neverswitched-off for the duration of the simulation. The sec-ond test involved the same session generation process as thefirst one, but assessed the performance of the system in the

Table 2: Simulation results provided by testbed’sPythagor VDT.

] Type 2 stations Overflow Avg. queue+ 8 Type 1 stations probability length

×10−2 (in pkts)1 0.11 1.03142 0.53 1.76703 4.20 3.84204 7.34 4.78935 9.61 5.9315

absence of AC.The simulation setup assumed two types of sessions with

different Traffic and QoS descriptors. Type 1 correspondedto CBR traffic with a mean rate of 100 kbps and with noQoS requirement (P1 = 1). Type 2 was characterized bya Poisson traffic pattern with a mean rate of 60 kbps, andP2 = 10−2. A common buffer threshold was assumed for allinvolved stations, equal to x = 20 packets. Also, a constantpacket (payload) size P = 8184 bits was used throughout.

According to the simulated scenario, eight Type 1 ses-sions (allocated to an equal number of stations) asked tojoin in the WLAN, each one after the other (with constantinter-arrival time of one second). The AC algorithm wascalled every time a new session was to be added to the sys-tem. By virtue of the null QoS requirement (resulting ina value of θ? = 0), the AC criterion for Type 1 reducedsimply to the stability condition (i.e, mean input rate <mean server rate). Because of the fact that the mean rate(100 kbps) of each Type 1 session was less than the hostingstation’s saturation throughput (mean server rate, see [1, 2,7]) for eight competing stations, the eight sessions succes-sively passed the admission criterion and became accepted.

After the sequence just described, five sessions of Type 2also demanded for service successively, but now the time be-tween session arrivals was longer (800 simulation steps, eachone second long) to allow for reliably assessing metrics rel-evant to the QoS performance. The Type 2 sessions weresubjected to two tests: In the first one, the AC algorithmwas kept active and its outcome was that only up to nine ses-sions in total could be admitted (i.e., the preexisting Type 1sessions plus the first Type 2 session), as inclusion of moresessions violated the Type 2 QoS requirement. The correct-ness of the AC decision was validated by sampling overflowsbeyond the threshold at the buffer of the station hostingthe session admitted last. Average queue lengths were alsosampled, as an additional indicator of congestion. The timeaverages for these single-run results appear in the first rowof Table 2. As it can be observed, the value of the sampledoverflow probability is below the Type 2 QoS specification,validating the AC decision.

The second test was targeted at validating the AC algo-rithm’s decision to accommodate only one Type 2 session.The test was realized by deactivating the AC and letting allarriving Type 2 sessions be admitted. The buffers of thehosting stations were sampled as previously (for at least 800steps) towards obtaining performance metrics. The relevantresults appear in the second to last rows within Table 2. Itcan be seen that the AC algorithm acted conservatively tosome (minor) extent, as one more session could be admittedon the basis of the results in the table. As commented in

Subsection 4.3, some deviation from exact answers is due tothe asymptotic nature and the saturation assumption un-derlying the scheme. Overall, however, the AC algorithmguarded against QoS violation quite effectively, as can be ob-served by the growth of overflow probabilities as the numberof sessions increases.

6. CONCLUSIONSThe paper presented a simulator coordination testbed that

can be used for optimization studies in composite wirelessnetworks and illustrated how a recently developed algorithmfor admission control in IEEE 802.11 networks can be inte-grated in an effective manner. The functionality of the keybuilding blocks of this testbed was discussed and the eventsthat were used to interconnect and synchronize geographi-cally distributed simulation clusters and realizations of ad-mission control algorithms were presented. The descriptionof the admission control algorithm was complemented byits validation on a stand-alone simulator as well as on thepresented simulator coordination platform. The results val-idated both the capability of the platform to efficiently inte-grate simulators and implementations of optimization func-tions, and the capability of the admission control algorithmto provide statistical QoS guarantees for the mobile node’sbuffer size.

7. ACKNOWLEDGMENTSThe work in this paper has been partly undertaken in the

context of the FP6 European research project UNITE (Vir-tual Distributed Testbed for Optimization and Coexistenceof Heterogeneous Systems), Contract No. IST-4-STREP-026906. The authors acknowledge contributions of their col-leagues from the UNITE consortium.

8. REFERENCES[1] G. Bianchi. “Performance Analysis of the IEEE 802.11

Distributed Coordination Function”. IEEE JSAC,18(3):535–547, 2000.

[2] G. Bianchi and I. Tinnirello. “Remarks on IEEE802.11 DCF Performance Analysis”. IEEE Commun.Lett., 9(8):765–767, 2005.

[3] C. Chang and J. Thomas. “Effective Bandwidth inHigh-Speed Digital Networks”. IEEE JSAC,13(6):1091–1100, 1995.

[4] C. Chang and T. Zajic. “Effective bandwiths ofdeparture processes from queues with time varyingcapacities”. In Proc. IEEE INFOCOM, pages1001–1009, 1995.

[5] A. Gavras, H. Bruggemann, D. Witaszek, K. Sunell,and J. Jimenez. “Pan European Laboratory for NextGeneration Networks and Services”. In Proc. IEEETRIDENTCOM’06, Mar. 2006.

[6] IEEE 1516-2000, “Standard for Modeling andSimulation (M&S) High Level Architecture(HLA)—Framework and rules”, Sept. 2000.

[7] E. Kafetzakis, K. Kontovasilis, and I. Stavrakakis. “ANovel Effective Capacity-Based Framework forProviding Statistical QoS Guarantees in IEEE 802.11WLANs”. NCSR “Demokritos” Technical Report,available online at www.iit.demokritos.gr/nel/

papers/DEMO-effcap-wlan.pdf. Submitted forpublication, 2007.

[8] F. Kelly. Notes on effective bandwidths. In F. Kelly,S. Zachary, and I. Zeidins, editors, StochasticNetworks: Theory and Applications, volume 4, pages141–168. Oxford University Press, 1996.

[9] P. Magnusson, J. Lundsjo, J. Sachs, and P. Wallentin.“Radio resource management distribution in a beyond3G multi-radio access architecture”. In Proc. IEEEGLOBECOM, pages 3472–3477, Dec. 2004.

[10] The “Pythagor” simulation tool. Available athttp://www.icsd.aegean.gr/telecom/pythagor.

[11] G. F. Riley, M. H. Ammar, R. M. Fujimoto, A. Park,K. Perumalla, and D. Xu. “A federated approach todistributed network simulation”. ACM Transactionson Modeling and Computer Simulation, 14(2):116–148,Apr. 2004.

[12] “IEEE Standard for Wireless LAN Medium AccessControl (MAC) and Physical Layer (PHY)Specifications, Nov. 1997. P802.11”.

[13] The UNITE IST-FP6 project.http://www.ist-unite.org/.

[14] M. Varshney, Z. Xu, S. Mohan, Y. Yang, D. Xu, andR. Bagrodia. “WHYNET: a framework for in-situevaluation of heterogeneous mobile wireless systems”.In Proc. WINTECH’07, pages 35–42, 2007.

[15] D. Vassis, L. Sarakis, G. Kormentzas, andC. Verikoukis. “A Distributed Testbed forPerformance Evaluation of Inter-system OptimizationSchemes in Heterogeneous Wireless Networks”. InProc. 5th International Workshop on Next GenerationNetworking Middleware (NGNM), pages 109–122,Sept. 2008.

[16] B. White, J. Lepreau, L. Stoller, R. Ricci,S. Guruprasad, M. Newbold, M. Hibler, C. Barb, andA. Joglekar. “An integrated experimental environmentfor distributed systems and networks”. In Proc. of theFifth Symposium on Operating Systems Design andImplementation, pages 255–270, Boston, MA, Dec.2002. USENIX Association.

[17] D. Wu and R. Negi. “Effective Capacity: A WirelessLink Model for Support of Quality of Service”. IEEETrans. Wireless Commun., 2(4):630–643, 2003.

[18] X. Zhang, J. Tang, H. Chen, S. Ci, and M. Guizani.“Cross-Layer-Based Modeling for Quality of ServiceGuarantees in Mobile Wireless Networks”. IEEECommun. Mag., 44(1):100–106, 2006.