real-timesupportforhccafunctioninieee802.11enetworks:a ... · devices which support multimedia...

SECURITY AND COMMUNICATION NETWORKSSecurity Comm. Networks 00: 1–20 (2009)Published online in Wiley InterScience(www.interscience.wiley.com) DOI: 10.1002/sec.0000

Real-time support for HCCA function in IEEE 802.11e networks: aperformance evaluation†

Gabriele Cecchetti, Anna Lina Ruscelli

ReTiS Lab., Scuola Superiore S. Anna, Pisa, Italy

Summary

The IEEE 802.11 standard for wireless networks has been recently enhanced with the IEEE 802.11eamendment which introduces Quality of Service support. It provides differentiation mechanisms atthe Medium Access Control layer, using two additional access functions: the Enhanced DistributedChannel Access (EDCA) function and the HCF Controlled Channel Access (HCCA) function. Only theHCCA mechanism is suitable for serving traffic streams with real-time requirements such as multimediaapplications and Voice Over IP. The IEEE 802.11e standard does not specify a mandatory HCCAscheduling algorithm, but it offers a reference scheduler as the guideline in the resources schedulingdesign.In this paper we analyze four HCCA alternative schedulers to the reference one. They offer real-timeguarantees proposing different solutions to the request of QoS and real-time support expressed by theincreasing diffusion of multimedia applications. A performance evaluation is conducted to show the maindifferences between the considered schedulers, including the reference one.The results show that under several scenarios there is not a unique best scheduler, but there exists avariety of solutions depending on the specified requirements. The conclusions of the paper offer someguidelines in the choice of the scheduler tailored for a particular scenario of interest.Copyright c⃝ 2009 John Wiley & Sons, Ltd.

KEY WORDS: quality of service; real-time guarantees; scheduling algorithms; performance evaluation;wireless lan

1. Introduction

In wireless communications the Quality of Service(QoS) and real-time guarantees provisioning areimportant issues due to the diffusion of mobiledevices which support multimedia applications

∗Correspondence to: Gabriele Cecchetti, Retis Lab., ScuolaSuperiore S. Anna, Pisa, Italy.E-mail: [email protected]†This paper is an extended version of “PerformanceEvaluation of Real-Time Schedulers for HCCA Functionin IEEE 802.11e Wireless Networks”, presented at ACMQ2SWinet, Vancouver, Canada, October 2008.

such as Voice over IP (VoIP), multimedia video,and videoconferencing. Furthermore,Constant BitRate (CBR) and Variable Bit Rate (VBR) trafficwith different requirements need differentiatedQoS levels with real-time guarantees.

On the other hand the space and time-varyingcharacteristics of the wireless channel [1] affectthe network performance in terms of Signal toInterference plus Noise Ratio (SINR) and BitError Rate (BER). Thus it is not possible toassure hard QoS constraints in terms of exactvalues of network and application parameters,

Copyright c⃝ 2009 John Wiley & Sons, Ltd.

Prepared using secauth.cls [Version: 2008/03/18 v1.00]

2 G. CECCHETTI, A. L. RUSCELLI

including delay, delay jitter, packet loss ratio,and throughput. However the characteristics ofmultimedia applications allow us to consider softrequirements expressed by admitted intervals forthe parameters’ values. In fact, for multimediaapplications, missing deadlines implies only adegradation of the received QoS and notcatastrophic events. Therefore we can considersoft-real time constraints with the exception thatwe cannot employ the classical methodologies usedin static real-time systems because they cannothandle the dynamic characteristics of both thewireless medium and the multimedia applications.

Even if IEEE 802.11 [2] is the recognizedstandard for Wireless Local Area Networks(WLAN), its earlier releases were designedto provide only best effort services thus ithas been recently enhanced with the IEEE802.11e amendment in order to include the QoSguarantees.

The recent standard has introduced a dif-ferentiation mechanism at the Medium AccessControl (MAC) layer, using two additional accessfunctions: the Enhanced Distributed ChannelAccess (EDCA) function and the HCF ControlledChannel Access (HCCA) function. The EDCAfunction is based on distributed control andenables prioritized channel access, while the latterrequires centralized scheduling and allows theapplications to negotiate parameterized serviceguarantees. In particular, since only the HCCAmechanism is suitable for respecting real-timeconstraints, we consider only this MAC function.

However, the IEEE 802.11e standard does notspecify a mandatory HCCA scheduling algorithm,but it offers a reference scheduler, which iscompatible with the use of link adaptation,and respects a minimum set of performancerequirements. In particular it periodically assignsa fixed transmission time interval to all themanaged Traffic Streams (TSs).

Some studies evaluated the new standardthrough analytical techniques and simulations anddemonstrated that the HCCA improves the QoSsupport particularly for CBR traffic. However,it performs poorly with VBR traffic, since itassigns fixed transmission parameters thereforeit cannot accommodate the traffic variability. Asa consequence several researchers have suggestedalternative scheduling algorithms to the referenceone, in order to improve its QoS provisioning for

VBR traffic, but they do not consider the real-timeconstraints specifically.In this paper we present a performance

evaluation of five schedulers tailored for thereal-time guarantees support over IEEE 802.11eHCCA networks, in terms of the admissioncontrol, the resource utilization efficiency and theaccess delay. This is neither a fully comprehensiveevaluation nor a survey of the existent schedulingalgorithms for IEEE 802.11e WLANs, but theanalysis of some known algorithms that, to thebest of our knowledge, can provide real-timefeatures for WiFi networks.The rest of the paper is organized as follows: in

section 2 we describe the IEEE 802.11 standard,in section 3 we summarize some schedulingalgorithms for HCCA function, in section 4 weillustrate the considered schedulers including thereference one and in Section 6 we evaluate theirperformance. Finally, the conclusions are drawnin Section 7.

2. IEEE 802.11 MAC Protocols

This section provides a description of the IEEE802.11b and IEEE 802.11e MAC protocols.

2.1. IEEE 802.11b MAC Protocol

The IEEE 802.11b MAC supports two datapacket transmission modes. One mandatory, theDistributed Coordination Function (DCF) and oneoptional, the Point Coordination Function (PCF).DCF deals on distributed medium access

based on Carrier Sense Multiple Access/CollisionAvoidance (CSMA/CA), where each node com-petes with the other ones for access to themedium. The collisions are avoided througha random backoff procedure which ensuresdifferent waiting transmission times to thestations. Considering that all stations havethe same channel access parameters, the samemedium access priority, and there is no streamsdifferentiation but the medium access is managedonly by means of a random procedure, DCF doesnot provide QoS support but supplies only besteffort service.The basic feature of PCF is the polling

mechanism, managed by the Access Point (AP),which allows only polled stations to transmit.During the Contention Free Period (CFP), theAP polls its associated stations according to the

Copyright c⃝ 2009 John Wiley & Sons, Ltd.Prepared using secauth.cls

Security Comm. Networks 00: 1–20 (2009)DOI: 10.1002/sec

SCN-SI-010 3

polling list, usually in a round-robin manner.If the CFP terminates before all stations havebeen polled, the polling list will be resumed atthe next CFP from the previous stopping point.Since the polling list is fixed and each polledstation (STA) occupies the medium until it hasdata to send, independently from the presence ofpollable STAs with more stringent QoS and real-time requirements, PCF is not suitable to provideany form of QoS support.

2.2. The IEEE 802.11e MAC Protocol

The IEEE 802.11e standard has introduced aservice differentiation mechanism at the MAClayer. It uses the two additional access functions,EDCA and HCCA, multiplexed by the new HybridCoordination Function (HCF).The services differentiation enables the protocol

to distinguish the packets with different servicerequirements, overcoming the limits of the besteffort model and delivering data with differentQoS levels and negotiated services. The designapproach has been to enhance the preexistentMAC functions, DCF and PCF, through the newmentioned ones, which are based on the samecorresponding medium access methods. Moreover,in order to ensure compatibility with legacydevices, the standard allows the coexistence ofDCF and PCF with EDCA and HCCA.

2.2.1. EDCA

The EDCA function enhances the distributedcontrol of DCF by means of a prioritizedchannel access which ensures different prioritizedQoS levels. The incoming traffic in a STAis classified in four Access Categories (AC),corresponding to four different service levels.Each AC has its own transmission queue andits own set of channel access parameters, usedfor the packets classification. The most importantones are Contention Window, which sets thebackoff interval, and Transmission Opportunity(TXOP) which is the maximum duration of anode transmission. The access to the medium isregulated by two contention phases. The first one,internal to each STA, is won by the AC with thesmallest backoff time. This value is used by eachSTA in the external contention for the wirelessmedium. So nodes with higher priority can accessthe channel earlier than the other ones.

2.2.2. HCCA

HCCA provides a centralized polling scheme toenhance PCF with QoS support through a trafficstreams classification, based on stations requests,into eight TSs. Each TS can be uni-directional(uplink or downlink) or bi-directional (both ofthem) and corresponds to a specific servicelevel identified by particular values of protocolparameters. The QoS-aware Hybrid Coordinator(HC), usually located at the QoS Access Point(QAP), manages the access to the medium.

In order to be included in the polling list ofthe HC, a QoS Station (QSTA) uses the QoSmanagement frame Add Traffic Stream (ADDTS)to send to the QAP a QoS reservation requestfor each of its TSs. TS parameters are collectedin a Traffic Specification (TSPEC), negotiatedbetween QSTA and QAP. Mandatory fields areillustrated in table I.

Table I. Traffic specification mandatory fields.

TSPEC parameters Symbol

Mean data rate (bps) RNominal SDU size (bytes) LMinimum PHY rate (bps) ΓDelay bound (s) DMaximum service interval (s) MSI

The negotiation of TSPECs parameters duringthe reservation request allows HCCA to guaranteeTSs a parameterized QoS access to the medium.

Moreover, during this phase HC performs theadmission control, checking if the acceptance ofthe requesting TS can compromise the serviceguarantees of the already admitted TSs. If the TSis admitted, QAP sends to the QSTA a positiveacknowledgement which contains also the servicestart time for the frames of the considered TS.As result of this negotiation HC has to alsocompute, as aggregation of the QSTA TSPECs,the following transmission parameters sent to theQSTA at the polling time:

Service Interval (SI) : the time interval be-tween two successive polls of the node. It isa submultiple of the 802.11e beacon intervalduration

Transmission Opportunity (TXOP) : thenode transmission duration based on themean application data rates of its TSs.




SI and TXOP are the basic parameters used bythe scheduling algorithms to manage the access tothe medium and their choice is the key in the QoSprovisioning.When there are admitted QSTAs, the QAP

listens to the medium and, if it is idle for aPCF Interframe Space (PIFS), it gains controlof the channel. The QAP, during the ControlledAccess Phase(CAP), polls a single QSTA atturn, according to the polling list generatedby the scheduler considering the QoS and real-time requirements. It is necessary to distinguishbetween downlink TXOP, during which the QAPsends bursts of QoS Data to QSTA and uplinkTXOP, that starts when the polled QSTA takesthe medium control. If the polled QSTA does nothave packets for the considered TS, i.e. the TS isnot backlogged, or if the head-of-line packet doesnot fit into the remaining TXOP duration, theQSTA sends a QoS CF-Null frame to the QAP.This situation is shown in fig. 1.

3. Related Works

Many research studies evaluated the performanceof the reference IEEE 802.11e scheduler employinganalytical techniques and simulations [3, 4, 5].They demonstrated the usefulness of the proposedmechanism, but they also highlighted its limits.In fact, scheduling each QSTA with a fixed coupleof transmission parameters, SI and TXOP, allowsto provide some relevant results in terms of real-time and QoS support only for CBR traffic,which shows stable and invariant values of itsparameters. But in the case of VBR traffic, (likeVoIP, videostreaming etc.), a fixed scheduler givesa non-optimal resources utilization since it doesnot react to the network and traffic variability.The same conclusions are valid for the AdmissionControl which is missing of the necessary flexilitythat gives a more stringent condition with theadmission of less TSs than possible, wasting theavailable resources.Several alternative scheduling algorithms have

been proposed to improve the QoS provisioningof IEEE 802.11e networks, in particular in thecase of VBR traffic, [6, 7, 8, 3]. Neverthelessthey do not consider real-time constraints andfew works evaluate the real-time issues of thereference scheduler, i.e. its ability to respecttiming constraints expressed in term of TSsdeadlines, and propose possible solutions.

Looking at the proposed real-time schedulingalgorithms for IEEE 802.11e networks, to thebest of our knowledge, they are characterizedby the different mechanisms used to providetemporal guarantees. Table II illustrates suchmechanisms adopted by some algorithms. The

Table II. Scheduling mechanisms.

Scheduling mechanisms Algorithms

Queue length estimation FHCF, ARROWFeedback based mechanism FBDSTimed token based mechanism WTTPDeadline-based scheduling SETT-EDD, RTH,

WCBS

algorithms named in the table using italic fontwill be described more accurately and comparedin section 4.One component of the packet delay is the queue

delay, due to the lateness in the transmissionqueue delivering. A resources assignment tailoredto the network traffic can reduce the waiting timeexperienced by the packets in the transmissionqueues.The algorithms can employ different types of

information, depending on the use of a theoreticalmodel or the actual values of the queues length.In the latter case, a better resource reservationis possible with respect to it being based onestimated queue length where the assigned TXOPcould be major (waste time), minor (delayedpacket) or equal (ideal case) with respect to theactual needed transmission time.In ARROW [9] the next TXOP is computed

considering the actual buffered TSs data at thebeginning of the polling and communicated to theQAP through the QS header field. Moreover, sinceeach new packet could be delayed at least oneSI, the scheduler ensures that the deadline is notexceeded and the delay requirements respectedbounding the MSI. However the increasing pollingoverhead, due to the fact that the MSI upperlimit could be less than half of the standard,is compensated by a more accurate TXOPcomputation. Knowing the queue length for eachTS, the scheduler can manage differently theTSs with different requirements. Finally thenext QSTA to be polled is chosen by meansof an Earliest Due Date (EDD) [10] deadlinesscheduling.The Application-Aware Adaptive HCCA Sche-

duler [11], derived from ARROW, distinguishes



SCN-SI-010 5

Figure 1. Sample HCCA exchange sequence.

uplink and downlink schedulers, whereas the Ear-liest Deadline First (EDF) algorithm [12] adaptsthe polling order to the stations’ requirements,expressed by means of the computed deadlines.The uplink scheduler assigns each QSTA aminimum and a maximum SI, modified to followthe application and network conditions and todigest the buffered traffic, respecting the QoSrequirements. They are computed considering themean TSPEC values and depending of periodic oraperiodic traffic.

A feedback mechanism can be used to updatethe value of transmission parameters to thenetwork variability. All the QSTAs and theirtransmission queues are regarded as a systemwhose balance is perturbed by new incomingTSs. The Feedback Based Dynamic Scheduler(FBDS) [13] behaves as a closed loop controllerwhich restores this balance by bandwidthrecovering, limiting the maximum delay. It assignsdynamically, by means of a feedback mechanism,the TXOP according to queue length estimationat beginning of the new CAP phase througha discrete time model, while SI remains fixed.Moreover it compensates the errors producedby channel perturbations not previewed by theestimation algorithm using the actual queuelength information sent by each QSTA.

In [7] the authors propose the SchedulingEstimated Transmission Time - Earliest Due Date(SETT-EDD) algorithm, which is based on theuse of the deadline concept. The algorithm usesvariable TXOP and SI but it maintains thesame admission control as the reference scheduler.Both the TXOPs and the SIs values are limited

between the minimum and a maximum SI sothat transmissions can be done at least at PHYrate thus respecting the deadlines. The pollingorder is managed by the real-time algorithmDelay-Earliest Due Date (Delay-EDD) [14], whichpolls the QSTA according to the non decreasingdeadlines.

4. The Analyzed Real-Time HCCA Schedulers

In this section we analyze with more detail theschedulers chosen for our performance evaluation.We begin describing the limits of the IEEE802.11e HCCA reference scheduler in terms ofreal-time provided guarantees. Then we describewith more detail the schedulers chosen for ourperformance evaluation as tailored for the real-time provisioning for HCCA WLANs.The scheduling algorithm and the admission

control test of each scheduler are analyzed. Inparticular the admission control test represents asufficient feasibility condition for the schedulingalgorithm and it allows for determination ofthe set of streams which can transmit withintheir timing constraints. When a new TS asksthe QAP for the right to be transmitted, thescheduler checks if the new traffic stream can beadmitted to the medium without jeopardizing theguarantees of already admitted streams and if itwould require more capacity than the system canprovide. If the TS cannot be admitted the QSTA isnotified of insufficient available capacity, otherwisethe scheduler updates the total used bandwidthand allocates data structures to perform the TSscheduling.




4.1. The Reference Scheduler

The reference scheduler proposed by the IEEE802.11e standard suggests how to compute themain protocol parameters, SI and TXOP, suitableto meet the requirements globally expressed byeach QSTA. Different values for each specificQSTA are computed only for TXOP, whereas SI iscomputed as a unique value for all non-AP QSTAswith admitted streams.

4.1.1. Reference Scheduler Algorithm

The goal to assure that each QSTA transmitsits TSs during an interval tailored to itsrequirements is solved considering the mean valuesof transmission parameters.The SI is suggested to be less than the beacon

interval, ensuring that the QSTAs will be polledat least one time during the beacon duration.It has also to be less than the minimum of theMaximum SI (MSIi) of each QSTAi. This meansthat the QSTAs are polled within an interval lessthan the minimum of those requested for eachof their TSs and this assures no deadline misses.For that which concerns the MSIi, the standardsuggests to compute it as the ratio between theDelay Bound and the possible number of retries.The standard states that the number of retriesmay be chosen to meet a particular probabilityof dropping a packet when it exceeds its DelayBound.The TXOP is globally assigned to a QSTA

and not to its single TSs. It is computed usingthe SI previously obtained and the mandatorynegotiated TSPEC parameters. First of all thescheduler computes the max number Ni of maxsizeMAC Service Data Unit (MSDU) transmittedat the mean data rate Ri:

Ni =

⌈

SI ⋅Ri

Li

⌉

Then it computes TXOP as the maximumbetween the time to transmit Ni MSDU withnominal size Li and the time to transmit one maxsized MSDU (Mi, i.e. 2304 bytes) at the PhysicalTransmission Rate (Ri), which is equal to theminimum PHY rate Γi of the QSTAi:

TXOPi = max

(

Ni ⋅ Li

Γi

,Mi

Γi

)

+O

This choice is conservative, since it preserves themaximum time to transmit, at the minimum PHY

rate, the maximum number of bytes that canarrive during SI. Note that TXOPi takes intoaccount the transmission overhead (O) in timeunits: such overhead includes interframe spaces,ACKs and CF-Polls.SI and TXOPi are fixed values, based on “worst

case” conditions, and they are recomputed only ifa new TS arrives with a MSIi greater than thepreexistent ones. This means that all the QSTAsare polled with the same period, SI, and that thedifferent TSs of a station i are served with thesame computation time, TXOPi.

4.1.2. Reference Scheduler Admission Control

The admission control test suggested for theAdmission Control Unit (ACU) of the referencescheduler depends on the values of SI and TXOPcomputed for each TS as described before. Itassures that the ratio between the TXOP and theSI of the new admitted TS, added to the sumof the ratio of TXOPi and SI of each alreadyadmitted TSi, must not exceed the portion ofbandwidth reserved to the HCCA function:

TXOPk+1

SI+

k∑

i=0

TXOPi

SI≤

T − TCP

T(1)

where k is the number of already admittedstreams, k + 1 is the index of the newly arrivingstream, T is the beacon interval and TCP is thetime used for EDCA traffic.

4.2. Fair HCF

Fair HCF (FHCF) [15] aims to improve thefairness of both CBR and VBR traffic and thedelay performances assigning variable TXOPs bymeans of the estimation of the uplink TSs queueslength.The mathematical model proposed for the

queues shows the relationship between pollinginterval and queuing delays and it is used toestimate the global packet delay. More specifically,it distinguishes between the packet queuing delayQ and the waiting time delay W. The formeris due to the delay in the queue, influenced bythe variations in packet size and data rate, whilethe latter is defined as the interval between thepacket arrival time and the QSTA polling time.The authors distinguish the case when the queuesare empty at the end of the TXOP, especially for



SCN-SI-010 7

CBR traffic, and the case where the queues arenot empty, which is more realistic for real wirelessnetworks. In the first situation, there is a furtherdistinction if the packet arrives before the pollingtime of its QSTA (the packet has to wait for thepolling QSTA):

di(t) =

( i−1∑

j=1

Tj − t+qiMi

Reff

+�i

Reff

)

or if the packet arrives during (the packet hasto wait for the transmission of previous queuedpackets) or after (the packet has to wait for thenext SI):

di(t) =

(

SI − t+

i−1∑

j=1

Tj +�i

Reff

(

t−

i−1∑

j=1

Tj

))

where Tj is the allocated TXOP for TSj, qi is thenumber of packets in the queue, Mi is the MSDUsize, Reff is the effective data throughput, �i isthe application data rate. In this case the delayis bounded by the SI. Instead, if the queues arenot empty at the TXOP end, the packets can bequeued later than the next SI since the delaysare cumulative. In the best case the packet willbe transmitted in the next SI but in general, ifthere is highly variable traffic, the packet delaycan become really unpredictable. This delay isexpressed by:

Di = maxt

{di(t)} = SI − Ti +qeiMi

�i

where qei is the queue length at the end of theTXOP Ti.These equations suggest two different ways to

control the maximum delay: increase the TXOPTi or reduce the SI but this implies an increasingnumber of polling and an increasing overhead.FHCF adopts the alternative method to reducethe delay through the control of the queue lengthbefore the polling time. To deal with these delaysthe authors designed FHCF with two schedulers:the QAP scheduler and the node scheduler.The QAP scheduler estimates the varying queue

length for each QSTA at the beginning of the nextSI, qesti , and compares this value with the idealone, qideali . In particular,

qesti =�i(SI − tei )

Mi

+ qei

where tei is the time when qei is evaluated. Sincesending rate and packet size can change, thisestimation can not be accurate, so the QAPscheduler corrects the qei computation using itsexpected value. Then it computes the additionalrequired time testi (positive or negative) for eachTS of a QSTA and reallocates the correspondingTXOP duration according to the number ofadditional packets in the queue. Moreover, itevaluates the actual available time after theallocation of all the TXOPs in one SI in the idealcase and, if it is not sufficient, it decreases fairlyall the assigned times by a percentage of tei . Thisallows management of the traffic variations.The node scheduler, located in each QSTA, just

after the CF-Poll reception, can redistribute theadditional time of the TXOP, which is alwaysglobally allocated to the QSTA, among its TSs.It executes the same computation than the QAPscheduler but more accurately since each QSTAknows exactly its TSs queues size at the beginningof the polling and it is able to estimate its queuelength at the end of TXOP and the requested TSadditional time. According to its allocated TXOP,it evaluates the remaining time T ′ that can be re-allocated considering the number of packets Ni

to transmit in the TSi and the time required totransmit a packet, computed according to its QoSrequirements:

T ′ = T −

p∑

i=1

Ni ⋅

(

Mi

Reff

+ 2SIFS +ACK

)

.

FHCF uses the same admission control test ofthe reference scheduler (eq. 1).

4.3. Wireless Timed Token Protocol

Wireless Timed Token Protocol (WTTP) [16]is based on the Timed Token Protocol (TTP)[17], a token passing MAC protocol for ring-based networks. The token is used to manage acircular list of nodes in a round-robin manner.Each node refers downlink and uplink TSs. Onespecial node represents contention traffic usingcontention-based schemas as EDCA and DCF.

4.3.1. WTTP Scheduler Algorithm

The scheduler visits each node of the list fora time called sojourn time and either schedulesTXOP or refrains from generating CAP according




to the node TSs, (if the node represents contentiontraffic). The nodes are inserted in the round robinlist only if they actually have traffic to be served:

∙ downlink TSs are added and removedwhenever they become backlogged or idle;

∙ uplink TSs are added and removed if eachQSTA piggybacks the backlog of its TSson outgoing data messages (standard IEEE802.11e feature). However such nodes areput back in the round robin list at least eachminimum SI.

The token circulation is ruled by the TargetToken Revolution Time (TTRT), a protocolparameter selected as reference round durationwhich corresponds to the SI: its value is computedby the QAP according to the TSPEC valuesnegotiated by the QSTA during the admissioncontrol phase. The authors set the value of TTRTto half the smallest delay bound, since the roundrobin duration is bounded by 2 ⋅ TTRT :

TTRT =1

2mini{Di}.

This maximum limit makes the scheduler moreconservative in terms of maximum delay tolerableby the QSTAs, implying that some QSTAs can bepolled more frequently than necessary, increasingthe system overhead.The sojourn time is given by either one or both

the following two components:

∙ synchronous bandwidth: a fixed time Hi

computed as a percentage of TTRT and usedto transmit frames with HCCA rate basedguarantees

∙ asynchronous bandwidth: a variable and notreserved portion of TTRT used to transmitthe remaining not guaranteed traffic.

In other words, each node i has a Hi >= 0 anda Token Rotation Timer (TRTi), initially set toTTRT, which counts down the time from thelast server visit to obtain the maximum fairshare of asynchronous bandwidth that node i canexploit. When a node is served, the asynchronousbandwidth is computed as follows:

ai =

{

0 TRTi < 0,min{TTRT −Hi, TRTi} TRTi ≥ 0.

In this way the asynchronous TSs are not reservedany capacity: they may transmit frames only if

the token arrives earlier than expected, i.e., beforea TTRT time has elapsed from the last tokenvisit. The early arrival of the token usually occurswhen the synchronous TSs consumed less thanthe reserved capacity during the previous tokenrevolution. The assignment of the sojour time,which is substantially the TXOP, is different bythe other approach. In fact in the case of QoStrafficHi is computed as a fixed fraction of TTRT:this means that the time destined to the QoS TSsis not recomputed and, in case of unused resources,the remaining time is not assigned to other QSTAsbut to the non-QoS traffic of the polled QSTA.Since the contention traffic is managed by theEDCA function this implies that the QSTA hasto change the medium access mode adopting thecontention method and, when it has exhausted therecovered time, it has to go back to the HCCA.Obviously, this increases the system overhead.

4.3.2. WTTP Admission Control

Since TTRT is the average inter-service time fora node, and 2 ⋅ TTRT is an upper bound theadmission control test for the WTTP algorithmis the following:

∑

nodei

Hi + � ≤ TTRT. (2)

Therefore a node that has a synchronousbandwidth equal to Hi is in fact entitled to anaverage rate equal to Hi/TTRT times the channelspeed, and has a bounded medium access time.The term � is an overhead due to the time requiredto regain the control of the medium and to starta new CAP after a DCF/EDCA phase.

4.4. Real-Time HCCA

The Real-Time HCCA (RTH) algorithm [18] isdesigned to provide real-time support in HCCAassuring the traffic streams a fixed amount ofcapacity during a fixed period.

4.4.1. RTH Scheduler Algorithm

The periodic scheduler, based on EDF algorithmplus Stack Resource Policy (SRP) [19], takes intoaccount the non-preemptability of frame transmis-sions, that are considered critical sections. Thescheduler activity is split into offline activityat stream lifetime timescale, which performs the



SCN-SI-010 9

more complex activity, and online activity at theframe transmission timescale.Since the scheduling parameters are computed

offline, the online activity consists only of readingthe next entry [i,ti,TXOPi], composed by theindex of the next QSTA which can access themedium, the polling time ti and the duration ofits transmission. Conversely, the offline activityattends to the admission control and to thetimetable computation. The admission controlphase takes into account the requirements ofthe new and admitted QSTAs and provides theparameters used in the timetable computation.When the MAC sublayer Management Entity

has admitted a QSTA it translates a set of theTSPEC mandatory parameters [Ri, Ni, Di, Γi]into that used in the enforcement procedure andnotifies the QSTA of successful operation. EachTS is characterized through two parameters, thecapacity Ci and the period Ti, derived from theTSPEC set. The capacity Ci is computed as thetime needed to send at rate Ri how many NominalSDU can be held during a period Ti, and it canbe expressed as follows:

Ci =

⌈

Ri ⋅ Ti

Ni

⌉

⋅ tNi

where tNiis the Nominal transmission time. The

period Ti is:

Ti =

⎧

⎨

⎩

Di Di <

⌈

(Ni

Ri

⌉

,⌊

Ri

Ni

⋅Di

⌋

⋅Ni

Ri

otℎerwise.

4.4.2. RTH Admission Control

In order to admit a TS the QAP verifiesthe schedulability test introduced in the multi-programmed environment by SRP which includesthe blocking time due to the non-preemptabilityof TS transmission. In the case of HCCA, theminimum critical section bi for a TSi is equalto the nominal size SDU transmission time tNi

including the poll time for uplink TSs, tPi, i.e.,

bi = tNi+ tPi

.When TSi is in a critical section and is

scheduled instead of the highest priority TSj , TSiis said to block TSj . So, the blocking time for aTSi is the maximum critical section durations ofTSs with a period longer than TSi:

Bi = maxj>i

{bj}.

The schedulability analysis produces the followingsufficient condition to determine the set of nschedulable TSs:

Bi

Ti

+∑

j≤i

Cj + �j ⋅ tPj

Tj

⩽ 1 ∀i : 1 ⩽ i ⩽ n. (3)

where �j is the maximum number of times thatthe QAP has to poll TSj , during Tj. Moreover,larger versions of Bi and bi, Bi and bi, which arethe maximum quantities that the equality holdsin the equation, are used in order to reduce theMAC overhead.

4.5. Wireless Capacity Based Scheduler

The Wireless Capacity Based Scheduler (WCBS)for HCCA [20], suitable for serving Soft Real-Timeapplications, is derived from the basic ideas ofthe Constant Bandwidth Server [21], a soft real-time scheduling algorithm for real-time operatingsystems. The original algorithm is based on EDFand provides a mechanism to serve multimediaapplications with soft real-time constraints alongwith hard real-time applications served by hardreal-time algorithms.In WCBS, the basic idea is that WiFi networks

handle traffic streams instead of tasks: some ofthem require temporal guarantees, while othersare just best effort traffic streams. The nature ofthe wireless medium does not allow transmissionswith hard real-time requirements, so here wedo not need a different real-time algorithm forthem, different to the operating systems. The besteffort traffic is served during the contention phase,while the traffic streams requiring QoS guaranteeswill be served during the contention-free phaseaccording to the HCCA protocol.

4.5.1. WCBS Static and Dynamic Parameters

WCBS uses static and dynamic parameters to rulethe transmitted packet scheduling.The scheduler assigns to each TSi an ordered

pair of static parameters:

Qi : the budget, i.e., the maximum transmissiontime which can be assigned during a period;

Pi : the service interval of the TSi.

In particular, Qi is the maximum capacity,expressed in time units, that a stream i canconsume in its period Pi. These parameters are




computed during the admission control phase andtheir values are based on the TSPECi. They donot change during normal conditions. The ratioUi = Qi/Pi is denoted as the factor utilization ofthe stream, i.e., the TSi bandwidth.During the scheduling, each TSi is characterized

by the following dynamic parameters, whichrepresent the actual stream status:

ci : the current capacity, i.e., the remainingtime that can be assigned to TSi during thenext TXOP;

di : the absolute deadline before the budgettransmission time has to finish;

pi : the next time an uplink TSi will be polledwhen it has no more data to transfer or ithas exhausted its TXOP;

state : the current state of the stream can be onebetween transmitting, active, polling, idle.

The scheduling process marks each TS witha status according to its condition and itsparameters’ values. It is used to determine thesubsequent temporal evolution and it can be:

transmitting if the TS is transmitting packets;

active if the TS is in the transmitting queuebecause it has a packet to send and ci > 0.Moreover if it is an uplink stream it is thenext TS which will be polled;

idle if the TS is a downlink stream that hasno packet to transmit or has exhausted itscapacity;

polling if the TS is an uplink stream and it is inthe polling queue because it has packets tosend and ci > 0 but is still too early to bepolled.

Note that only one stream at time can be in thetransmitting state.

4.5.2. WCBS Admission Control

The admission control test of WCBS is:

N∑

i=0

Qi

Pi

≤T − TCP

T(4)

where Pi is computed as the maximum SI and Qi

is determined by means of a weighted function f of

TSi = EDF_extractTXOPi = ci

ci <= min_cap.

No

TSi.state = Transmitting

TSi.state = PollingPOLL_enqueue(TSi)

Yes

TSi.dir ?TSi.state = Idle DL UL

di < now ||c i > (di – now)ui

ci = Qi

di = now + Pi

Yes No

TSi.state = ActiveEDF_enqueue(TSi)

ci = Qi

di += Pi

pi = di

di < now

di = now + P i

Yes

TSi

Yes

TSi rejectedNo

pi <= now

Yes

U + ui <= MaxU

U += Uici = Qi , di = Pi

UL: pi = di

Qi,Pi,ui = Calc(TSi)

new data ?

Yes

No

j = head(EDF_queue)

j == i

Yes

NoNo

Figure 2. WCBS scheduling algorithm.

Qmin and Qmax, which evaluates the minimumand maximum budget needed to transmit duringa period Ti, respectively, Nominal SDUs at themean data rate and Maximum SDUs at the peakdate rate. Qmin and Qmax are expressed as:

Qmin :=

⌈

Ri ⋅ Pi

Li

⌉

⋅ tN , Qmax :=

⌈

Λi ⋅ Ti

MLi

⌉

⋅ tN

where Pi is the period, Ri is the mean data rate,Λi is the peak data rate, Li is the nominal SDUsize, MLi is the maximum SDU size for the itℎ

TSPEC, and tN is the Nominal transmission timecomputed as:

tN = tDATA + tSIFS + tACK + tSIFS

= Li ⋅ ΓDATA + ℎACK ⋅ ΓACK + 2 ⋅ tSIFS .

4.5.3. WCBS Scheduling Algorithm

After the admission control phase, the temporalevolution of the scheduler is as follows (see Fig.2).

1. For each new admitted TSi, at thebeginning: ci = Qi, di = now + Pi, state =active, where now is the current time.

2. Whenever a TSi is active its transmissionrequest is enqueued in an EDF queue.



SCN-SI-010 11

3. The streams are served in EDF order: thescheduler extracts the next TSi to servefrom the top of EDF queue, and it sets: TSi

state to transmitting , TXOP = ci, thenit decreases the capacity ci by the effectivetransmission time.

4. When a TSi finishes transmitting, thenext pending transmission, if any remains,is served using its current capacity anddeadline.

5. When a TSi with ci < min capacity† isserved, if TSi is an uplink stream, itsstate becomes polling , it is inserted in thepolling queue, and the following quantitiesare set: ci = Qi, di = di + Pi, pi = di. If itsdeadline is still expired, i.e., di < now, thenit is postponed to another period by now,i.e., di = now + Pi. In this way, using thisrecharging mechanism, a TSi does not haveto wait for a deadline expiration to rechargeits capacity and then it is ready earlier totransmit again.

6. A TSi remains in the polling state untilpi ≤ now, then it is extracted from thepolling queue, it becomes active and it isinserted in EDF queue.

7. When an idle downlink TSi is servedbecause it has new data to transmit, ifci ≥ (di − now)Ui the scheduler rechargesthe stream capacity to the maximum value,ci = Qi, and it generates a new deadline bya period from now: ci = Qi, di = now + Pi.Then it becomes active and it is inserted inEDF queue.

8. If there are no active streams a ContentionPeriod is started.

5. Scheduling Algorithms Properties Analysis

In this section we briefly analyze the selectedalgorithms from a theoretical and mathematicalpoint of view. This general dissertation doesnot have the aim to provide a completeanalytical study but a theoretical validation ofthe performance evaluation which is the goal of

†min capacity is the minimum capacity needed to transmitan SDU, and eventually CF-Poll for an uplink TSi.

the present work. So we have chosen to highlightthe temporal isolation property, which assures aprotection mechanism in the transmission, and thecomputational complexity, which evaluates themathematical efficiency of the selected algorithms.

5.1. Temporal Isolation

The use of the standard TXOP parameter in orderto assign a maximum transmission time to eachadmitted QSTA is the key in the provisioningof negotiated QoS levels. Moreover the TXOPparameter is functional to provide the temporalisolation property to the scheduling algorithms.In particular for WCBS the temporal isolation

property is derived from the analogue oneof the Constant Bandwidth Server algorithm.The introduction of the budget for eachtraffic stream provides a solution to executionoverruns, which happen when a TS asks to betransmitted more than expected, jeopardizing thetemporal guarantees of other streams. The budgetassignment performs a bandwidth reservation:each stream is assigned a fixed capacity, andwhen it requires additional time, its deadlineis postponed. In this case such streams mightexperience a delay, whereas the guarantees assuredfor the other streams remain unaffected.The same conclusions are valid for FHCF

and RTH. Moreover, RTH considers the blockingtime derived by the granularity of the TSstransmission that implies a minimum length ofassured transmission duration.From this point of view WTTP is more

conservative, again. In fact its QoS TXOPs arefixed and the unused bandwidth is recovered forthe contention traffic. Thus, if a QSTA consumesless than expected the non-QoS traffic will be sent,but this does not jeopardize the QoS traffic thathas been already sent.

5.2. Computational Complexity

The evaluation of the computational complexityof the considered algorithms highlights theirefficiency in term of computational resourceutilization.WCBS can be split into two phases, each having

a particular task and a related complexity. Thefirst is the insertion of the admitted TS in theright place within the EDF queue, following adeadline-based order: the operation of inserting an




element into an ordered queue of n elements, then admitted streams, has a complexity O(log n).Furthermore, if we consider the worst case where nTSs are admitted at the same time, the complexitygrows to O(n ⋅ log n). The second phase, whichinvolves the extraction of the stream from thequeue, has a complexity of O(1).In RTH the admission control test has an

O(n) computational complexity. Admitted TSsare scheduled following the timetable computationbased on EDF-order of periods Ti, for i = 1..n. Itscomputational complexity is due to the selectionof the TS to be transmitted between that withunfulfilled capacity, so it is equal to O(n) infunction of number n of admitted TSs. Instead thecomputational complexity of the online scheduleractivity is O(1) in terms of the number of QSTAs.The same considerations can be applied to the

reference and the FHCF schedulers. In fact, afterthe admission control test, which is the same forboth, and the computation of the transmissionparameters they do not apply any particularalgorithm in the polling list sorting but acceptthat which is performed during the negotiationphase and expressed in terms of QoS requirements.Since the sorting algorithm is not specified, wecan infer that it can be characterized, in general,by an O(n) complexity due to the comparisonof requested QoS levels, while the extraction ofan element from the polling list has an O(1)complexity.The WTTP computational complexity is

bounded by the algorithm which enforces thepolling order. In fact the Timed Token Protocolmanages the token circulation between the nodesimposing the sequence of polled QSTAs. Sothe QAP does not have to perform any otherselection or ordination activity. The simplicitygives a O(1) complexity. This is the drawbackbetween the adoption of a simple solution with apredetermined ordination algorithm, which limitsits computational complexity, and the lack in theflexibility.

6. Performance Analysis

In this section we present some results ofthe algorithms performance evaluation obtainedthrough simulation. We first describe the si-mulation settings and the used traffic model.Then we discuss the results about admissioncontrol analysis, the null rate experienced over

the medium, the unreserved capacity available forcontention traffic and the mean access delay.As shown by the obtained results an optimal

algorithm suitable to meet all the differentrequirements as a unique solution does notexist. In fact each requirement involves differentparameters sometimes in opposition. Tuning someof them to obtain the desired scheduler behaviorfrom a particular point of view as specified byparticular values of network parameters can implythe degradation of other performance. Thereforethe choice of the global optimal scheduler isbounded by the drawback of opposite trends ofthe parameters and is obtained through the useof a specific algorithm tailored for the desiredbehavior.

6.1. Simulation Settings

We used the physical layer parameters specified bythe High Rate Direct Sequence Spread Spectrum(HR-DSSS) (Table III).

Table III. MAC/PHY simulation parameters

Parameter Value Parameter Value

SIFS 10 �s PHY header 192 �sPIFS 30 �s Data rate 11 Mb/sDIFS 50 �s Basic rate 1 Mb/sSlotTime 20 �s Bit error rate 0 b/s

MAC level fragmentation, multirate support,RTS/CTS protection mechanism are disabledand we assume that all nodes can directlycommunicate with each other, without the hiddennode problem. We use the HCCA implementationdescribed in [22] for the ns-2 network simulator[23], as a framework to implement the proposedalgorithm.The analysis has been carried out using the

method of independent replications. Specificallywe ran independent replications of 600 secondseach with 100 seconds warm-up periods untilthe 95% confidence interval is reached for eachperformance measure. Confidence intervals are notdrawn whenever negligible.

6.2. Traffic Model

We use two types of uplink (UL) traffic streamsrequiring QoS guarantees: VoIP and video. TheVoIP traffic is simulated using a VoIP generatormodule for ns-2 described in [24].



SCN-SI-010 13

The VoIP streams of packets are modeled asan ON/OFF source: during the ON (talkspurt)periods the traffic is CBR with parameters thatdepend on the encoding scheme; during theOFF (silence) periods no packets are generated.Talkspurt and silence periods are distributedaccording to the Weibull distribution [25] thatmodels a one-to-one conversation: �ON=1.423s,kON=0.824s, �OFF=0.899s, kOFF=1.089s (whichyields E[ON ]=1.58s, E[OFF ]=0.87s). The em-ployed encoding schemes are G.711, G.723.1 andG.729A [26] with the parameters as shown inTable IV. For both encoding schemes we set theTSPEC delay bound to the packet interarrivaltime (period) and the mean data rate to the peakrate during talkspurts.

Table IV. VoIP encoding schemes

Codec G711 G723.1 G729A

Frame size (s) 80 30 10Period (s) 0.02 0.0455 0.02Sample per packet 2 1 2Payload size (B) 160 30 20IP/UDP/RTPHeader size (B) 40 40 40SDU size (B) 200 70 60Data rate (b/s) 80000 12320 24000

The video stream traffic is generated using pre-encoded MPEG4 trace files from the Internetarchive of traces [27]. An MPEG4 encoderproduces streams of variable size frames at fixedintervals [28]. They are chosen to represent avideconference session (LectureHQ-Reisslein tracefile) and a video streamed over the network(Jurassic Park High Quality trace file). TheTSPEC parameters are shown in table V.

Table V. Traffic parameters for video streams

video stream VideoConf. VideoStr.

Mean frame size (B) 660 3800Max frame size (B) 11386 16745Period (s) 0.033333 0.040Mean data rate (b/s) 157712 770000Peak data rate (b/s) 2732640 3300000

Best effort data traffic is transmitted usinglegacy DCF. Stations with data traffic operatein asymptotic conditions, i.e., they always havea frame to transmit. The packet length of datatraffic is constant and equal to 1500 bytes.

6.3. Admission Control Analysis

We evaluated the number of admitted stationsusing the considered schedulers under differentscenarios: CBR traffic only, VBR traffic only andmixed traffic. This approach shows the schedulersbehavior under different traffic conditions. Ananalytical comparison of the different admissioncontrol formulas confirms and explains thesimulation results.Fig. 3 shows the number of admitted G.729A

TSs as a function of admitted G.711 streams.Since they are CBR TSs and their codecs have

0

2

4

6

8

10

12

14

16

0 2 4 6 8 10 12 14

Num

ber

of G

.729

A V

oIP

UL

stre

ams

Number of G.711 VoIP UL streams

ReferenceFHCFWTTP

RTHWCBS

Figure 3. Admission Control: number of admittedVoIP G.729A vs. G.711 UL TSs.

the same SI, the number of admitted TSs issubstantially similar for each scheduler except forWTTP which underutilizes the medium. The poorperformance of WTTP is due to the fact that, asexplained previously, in the case of QoS traffic thesojour time Hi is computed as a fixed fractionof TTRT and the unused resources are destinedto the non-QoS traffic of the polled QSTA. Thissimulation shows that the considered schedulers,except WTTP, behave similarly in the case ofCBR traffic.In Fig. 4 we analyze the number of admitted

G.723.1 TSs as a function of admitted G.711streams. Since G.723.1 can have two differentrates, here we show that the reference schedulerand FHCF perform worse. This is due to thecodecs having different periods (20 ms and 45.5ms for G.711 and G.723.1, respectively) whilethese schedulers poll the stations at the smallestinterarrival period (thus 20ms). This is often morethan needed and the computed TXOP for TSsis overestimated. RTH and WCBS have betterperformance because they are derived by EDF so




0

5

10

15

20

25

30

35

40

0 2 4 6 8 10 12 14

Num

ber

of G

.723

.1 V

oIP

UL

stre

ams

Number of G.711 VoIP UL streams

ReferenceFHCFWTTP

RTHWCBS

Figure 4. Admission Control: number of admittedVoIP G.723.1 vs. G.711 UL TSs.

they allow a more precise estimation of the neededcomputation time for variable traffic. They assigndifferent SI and TXOP for each TSs. The WTTPadmission control performs even worse than thereference scheduler because it takes into accountonly an a priori evaluation of the lowest DelayBound of considered TSs. Such a parameter is themost strict requirement with respect to the otheradmission control tests. This simulation highlightsthat the reference scheduler lacks in the flexibilityrequired for the variable traffic.

Finally, Fig. 5 shows the number of admittedTSs under different scenarios involving two dif-ferent kinds of VBR streams, the videoconferenceand the videostreaming, as a function of thenumber of admitted VoIP G.711 and G.729A TSs.These TSs have four different periods: 33ms, 40ms,

0

1

2

3

4

5

6

7

8

0 2 4 6 8 10

Num

ber

of G

.711

G72

3 V

oIP

UL

stre

ams

Number of Videconference and Videostream UL streams

ReferenceFHCFWTTP

RTHWCBS

Figure 5. Admission Control: number of admittedVoIP G.711 and G723.1 vs. VC and VS UL TSs.

20ms and 45.5ms. In this scenario the schedulers

able to deal with such different TSPECs arewinning. In fact WCBS and RTH perform betterthan the others, and in particular WCBS admitsmore TSs because it computes a smaller TXOPand is able to recover the unused bandwidthwith more flexibility. On the other hand, thereference scheduler cannot efficiently manage TSswith different TSPECs, because it polls TSs withdifferent periods more frequently than needed,setting the scheduling duration to the smallest TSperiod and assigning an overestimated TXOP tothe TSs. Similarly, WTTP performs even worsebecause of its pessimistic admission control test.

To make an analytical evaluation of thedifferent admission control tests, now we put theirequations in the same form and use the samenotation:

Refer. :TCP

T+

k∑

i=0

TXOPi

SI≤ 1 (5)

FHCF :TCP

T+

k∑

i=0

TXOPi

SI≤ 1 (6)

WTTP :�

TTRT+

∑

nodeiHi + �

TTRT≤ 1 (7)

RTH :TCP

T+

Bi

Ti

+∑

j≤i

TXOPj

Ti

≤ 1 (8)

WCBS :TCP

T+

k∑

i=0

TXOPi

Ti

≤ 1 (9)

where the TXOPi comprises the tP i polling timefor uplink streams and the RTH admission controltest takes into account the contention period. Notethat:

∙ eq. 5 and eq. 6 are the same by definition;∙ eq. 7 is composed of one term representingthe asynchronous traffic and one termrepresenting the synchronous traffic. Ifwe set � so that �/TTRT = TCP /T , wehave to compare only the second termwhich represents the HCCA portion of thetraffic. Because TTRT = 1

2mini{Di} and,

in general, the SI is chosen to not makethe mini{Di} expired, we can approximateas TTRT = 1

2⋅ SI. Since this quantity is

smaller than SI, the ratio of WTTP isgreater than the ratio of the referencescheduler, so the index of the sum can reacha smaller value, admitting less QSTAs;



SCN-SI-010 15

∙ eq. 8 and eq. 9 differ only for the term Bi/Ti

which takes into account the critical sectioninvolved in SRP and has a small valuebecause usually Ti ≫ Bi: for this reasonWCBS and RTH always admit almost thesame number of TSs;

∙ eq. 9 differs from eq. 5 only by the secondterm; because in general Ti ≥ SI so the ratioTXOPi/Ti ≤ TXOPi/SI and therefore theindex of the sum can reach a higher value.

6.4. Efficiency Analysis

This section is dedicated to analyzing theschedulers’ efficiency under different scenarios.The scheduler efficiency is evaluated as a measureof how well it utilizes the network resources. Suchevaluation is done through the analysis of the nullrate and the polling interval experienced duringthe polling of the QSTAs and considering thethroughput left to DCF and EDCF TSs.The null rate is defined as the number of Null

packets received by the QAP after it has senta CF-Poll frame: this happens when a QSTAhas no packets to transmit. Evaluating the nullrate, we can check if the polling time computationis suitable for the considered traffic or that theQAP is polling the QSTAs more frequently thannecessary, increasing the system’s overhead.We consider a scenario with four VoIP G.711

uplink TSs and four VoIP G.723.1 uplink TSs.In Fig. 6 we show the null rate value for theanalyzed schedulers. With G.711 TSs the null rate

Nul

l rat

e (1

/s)

G711 G723.1

ReferenceFHCFWTTP

RTHWCBS

0

5

10

15

20

25

30

Figure 6. Null rate of UL VoIP TSs.

is almost the same for the reference, RTH andWCBS because they have the same SI and thesame computed TXOP and this type of traffic

does not need more scheduling flexibility. Thenull rate is higher for FHCF and WTTP, becauseboth try to empty the queue node polling moreaggressively. With G.723.1 TSs the reference,WTTP and FHCF schedulers show a higher nullrate because they use the shortest polling periodof 20 ms instead of 45.5 ms. WCBS and RTHpoll the TSs only at their packet arrival, so theirpolling mechanism is more efficient. In particular,FHCF performs worse with respect to the otherschedulers since it adapts the TXOP to the trafficvariations but maintains the same SI. This meansthat even if the scheduler exhausts the queuedpacket until the next SI, which is a good choice, itpolls the QSTA when it does not have any packetto send. Note that Null messages are only due tothe silence periods, and they are unavoidable.

In fig. 7 we show the polling intervals for ascenario with a mix of VoIP and videoconferenceTSs. The reference and the FHCF schedulerscompute the same polling interval, even if thereare TSs with different interarrival time. Insteadthe WCBS and RTH calculate two different valueswhich allow polling the QSTAs when they havepackets to send. The particular behavior of WTTPis due to the fact that it initially computes for theconsidered QSTA a fixed polling interval TTRTi

equal to TTRT, and only during the transmissiondoes it try to follow the traffic variability bychanging this value. But this adaptation isperformed slowly as shown in the figure, due toits poor reactivity. This simulation highlights the

0.010

0.015

0.020

0.025

0.030

0.035

2 4 6 8 10 12 14

Pol

l int

erva

l (s)

Number of QSTA

ReferenceFHCFWTTP

RTH VoIPRTH VC

WCBS VoIPWCBS VC

Figure 7. Poll interval of mix of VoIP andvideoconference UL TSs.

scheduling efficiency in adapting or not adaptingthe SI parameter to the traffic variability. The




choice of the polling interval affects also thesystem overhead.Now we analyze the unreserved capacity that

the schedulers leave for contention-based traffic,scheduled by DCF or EDCF functions.In fig. 8 we show the unreserved (E)DCF

capacity as a function of the increasing numberof QSTAs with VoIP and VBR traffic. Obviously,

0M

1M

2M

3M

4M

5M

6M

2 4 6 8 10 12 14

Unr

eser

ved

(E)D

CF

Cap

acity

Thr

ough

put (

bps)

Number of QSTA

ReferenceFHCFWTTP

RTHWCBS

Figure 8. Unreserved DCF capacity throughput withVoIP and VBR traffic.

as the number of QSTAs increases the capacityleft for the best effort traffic becomes poorsince the available resources are used to servethe QoS traffic. Therefore the schedulers presentdifferent levels of efficiency. The reference andWTTP schedulers perform worse since the QSTAsTXOPs (Hi for WTTP) are fixed. In particularWTTP computes TXOP and SI considering ana priori evaluation of the lowest Delay Bound,as mentioned in the admission control analysis,so its admission control is more stringent, theresource management is less efficient and thesystem overhead is higher. These factors impact inthe poor resources left to the non-QSTAs. FHCF,which adapts the TXOPs to the traffic variabilityconsidering the transmission queues length, showsa better behavior. It efficiently manages theavailable resources, leaving more capacity tothe best effort traffic. The flexibility of RTHand WCBS in the TXOP and SI computationallows a more efficient admission control and lesspolling overhead that’s why they show the bestperformance, managing efficiently the availableQoS capacity leaving more resources for the non-QSTAs.It is interesting to analyze the schedulers’

behavior with or without ideal channel conditions.

Thus we considered the presence of a uniformchannel error. In fig. 9 we compare the unreserved(E)DCF capacity throughput when there are 10QSTAs transmitting using VoIP codec G.729A. Asexpected, the schedulers’ performance decreases.In fact all the considered schedulers do notimplement any error recovery strategy, and thecomputed TXOPs, even if they are different forthe various TSs, are not recomputed after theadmission control and negotiation phase. Thus theschedulers are not able to recover the channelerror. In particular, only WTTP has a minorloss in the unreserved capacity since it polls theQSTAs more aggressively, “compensating” thenon-ideal conditions due to the channel error. Sothe QSTAs are able to react to the losses bytaking advantage of the more frequent polling toretransmit the corrupted packets.

Unr

eser

ved

(E)D

CF

Cap

acity

Thr

ough

put (

bps)

Schedulers without and with uniform error rate of 20%

0.0M

5.0M

1.0M

1.5M

2.0M

2.5M

3.0M

3.5M

REF FHCF WTTP RTH WCBS

Figure 9. Unreserved (E)DCF capacity throughputwhen 10 QSTAs are transmitting using VoIP G.729A

without or with uniform error rate.

Finally, we consider the QSTAs throughputwith the different schedulers. Fig. 10 shows thethroughput of the last admitted QSTA when thereare 10 admitted stations with G.729A traffic, inthe presence or the absence of channel errors.Since this is not a saturation condition andthe traffic is CBR, the QSTA has almost thesame throughput with the different schedulers.In fact we know that they perform similarly incase of CBR traffic. Only WTTP exhibits lowerperformance due to the more frequent polling.When introducing the uniform channel error, asinferred from fig. 9, the QSTA experiences a dropin the throughput. The behavior of one singleQSTA highlights the previous considerations



SCN-SI-010 17

about the schedulers, that are missing an errorrecovery policy.

Thr

ough

put o

f QS

TA

#10

(bps

)

Schedulers without and with uniform error rate of 20%

0.0K

5.0K

1.0K

1.5K

2.0K

2.5K

3.0K


Figure 10. Throughput of QSTA #10 with VoIPG.729A TS without or with uniform error rate.

Since the schedulers do not implement any errorcompensation mechanism it is not meaningfulto analyze their behavior with non-homogeneouschannel error. In fact as with uniform errorthe performance drops down, we can easilyinduce that the inhomogeneity in the channelconditions impacts severly in the schedulers’activity. Furthermore, no particular scheduler issuitable to avoid this effect.

6.5. Delay Analysis

In this section we investigate the access delaydefined as the time elapsed from the packetreaching the MAC layer to that of the packet beingsuccessfully acknowledged.Fig. 11 shows that, in case of CBR traffic like

that produced by G.729A codec, all the analyzedschedulers have almost the same mean delay,as expected. The only exception is WTTP: thelower polling interval assigns the QSTAs a sooneropportunity to transmit the TSs, reducing theexperienced delay. Note that the delay value isreported only for admitted QSTA. Looking at theCumulative Distribution Function (CDF) of theaccess delay shown in fig. 12, we have a clearconfirmation that the results about the meandelay are correct. Even this analysis shows thatWTTP has higher probability to experience asmaller value of access delay while the othersperform almost with the same probability forthe same values with a little advantage for thereference scheduler.

0.010

0.011

0.012

0.013

0.014

0.015

0.016

0.017

0.018

0.019

0.020

2 4 6 8 10 12 14 16 18 20

Mea

n de

lay

(s)

Number of QSTA

ReferenceFHCFWTTP

RTHWCBS

Figure 11. Mean delay of G.729A uplink TSs.

0.00

0.20

0.40

0.60

0.80

1.00

0.005 0.010 0.015 0.020 0.025 0.030

CD

F

Access delay with 8 QSTAs (s)

ReferenceFHCFWTTP

RTHWCBS

Figure 12. CDF of access delay of 8 QSTAstransmitting VoIP G.729A TSs.

In Fig. 13 we consider a scenario with anincreasing number of upload videoconference TSs.We note that the EDF-based schedulers produce

0.010

0.015

0.020

0.025

0.030

0.035

0.040

0.045

0.050

0.055

0.060

2 4 6 8 10 12 14

Mea

n de

lay

(s)

Number of QSTA

ReferenceFHCFWTTP

RTHWCBS

Figure 13. Mean delay of VC uplink TSs.

a mean delay greater than the others. In fact




the EDF algorithm executes a new sorting foreach CAP phase, while the other schedulersmaintain a fixed order of TSs. WCBS performseven worse because it also postpones the currentdeadline when there is not enough capacityto transmit. This motivates its shown highvariability in the mean delay too. The referenceand FHCF schedulers perform similarly becauseFHCF changes only the TXOP parameter whilekeeping SI fixed. WTTP has better behavior sinceit assigns the TTRT, corresponding to the SI, avalue smaller than the minimum tolerable delay,ensuring that all the QSTAs respect their timingrequirements. So WTTP is more conservative inminimizing the experienced delay at the cost ofan increased system overhead, due to the reducedpolling interval, as we explained in the previoussection.Observing the CDF of the access delay

experienced by a QSTA in the same scenario,when 8 QSTAs are transmitting we see that bothRTH and WCBS cause the VBR traffic to havea lower probability keeping low access delay thanthe other schedulers. FHCF and WTTP are thebest for this kind of traffic.

0.00

0.10

0.20

0.30

0.40

0.50

0.60

0.70

0.80

0.90

1.00

0.00 0.05 0.10 0.15 0.20

CD

F


ReferenceFHCFWTTP

RTHWCBS

Figure 14. CDF of access delay of 8 QSTAstransmitting VC TSs.

Now we want to take into account how theaccess delay changes in the presence of a uniformerror rate over the wireless channel. In thisscenario every packet has a probability of 0.25 tobe lost.In fig. 15 the mean access delay experienced by

each scheduler under this scenario is shown. Notethat now every scheduler is affected by the lossof packets. While in fig. 13 the mean values spanthe range from 0.01 s to 0.06 s, now this range

is increases to span the interval from 0.02 s to0.25 s. In particular WCBS and RTH mean accessdelays increase more than 3 times: this is greaterthan the Delay Bound admitted for that TSPEC.In such a case the packet will be discarded by thereceiving client. Only FCHF and WTTP still havean acceptable value, even if it is increased.

0.000

0.050

0.100

0.150

0.200

0.250

0.300

2 4 6 8 10 12 14

Mea

n de

lay

(s)

Number of QSTA

Reference w/EFHCF w/EWTTP w/E

RTH w/EWCBS w/E

Figure 15. Mean delay of VC uplink TSs in thepresence of uniform error rate over the wireless

channel.

The effect produced by the uniform error rateon the access delay is visible even if we look at itscumulative distribution function shown in fig. 16.Comparing this figure with fig. 14 we note thatthe slope of each probability curve is decreased,meaning that every scheduler has less probabilityto keep the access delay under a specific value.

0.00

0.10

0.20

0.30

0.40

0.50

0.60

0.70

0.80

0.90

1.00

0.00 0.05 0.10 0.15 0.20

CD

F


ReferenceFHCFWTTP

RTHWCBS

Figure 16. CDF of access delay of 8 QSTAstransmitting VC TSs in the presence of uniform error

rate over the wireless channel.

As explained in the throughput analysis, sincethe evaluated schedulers do not use any type



SCN-SI-010 19

of error recovery, the channel error directlyimpacts the experienced performance, withoutany reaction by the schedulers. All the schedulersshow very poor performance with a considerableincrease in the mean delay. The distinctionsbetween the different algorithms are the same asin the case of no error. Thus FHCF and WTTPperform better, whereas the poor ones are theEDF-based, RTH and WCBS. In particular thelatter shows the same high variability than in theideal conditions, due to the postponed deadlines.

7. Conclusions

In this paper we have compared five differentschedulers for HCCA IEEE 802.11e networks,namely, reference scheduler, FHCF, WTTP, RTHand WCBS. We have described their parameters,admission control tests, and their temporalevolution.In order to evaluate how suitable they are to

support traffic streams requiring soft real-timeguarantees, we have analyzed their characteristics.In particular the admission control phase has beentested under different scenarios both analyticallyand through simulation. Then the efficient use ofthe medium has been studied considering the nullrate and the polling interval produced with mixedtraffic streams. Our results have been confirmedby evaluating the unreserved capacity availablefor contention-based traffic. While the admissioncontrol and the efficiency tests have shown thatEDF-based schedulers perform better than theothers, the delay analysis with VBR scenarios hasillustrated that RTH and WCBS result in greateraccess delays. The drawback of the reference,FHCF and WTTP schedulers is essentially dueto restrictive admission control test, fixed SI, andoverestimated TXOP.There is not a clear winning strategy among

the five proposed schedulers: the scheduler forHCCA function has to be chosen according tothe scenario. However EDF-based algorithms suchas RTH and WCBS seem to be more suitablewhen applications require temporal guarantees.In particular, WCBS has a simpler design,lower computational complexity, slightly betterefficiency than RTH, and it has the feature topostpone the deadlines if required: the increaseddelay experienced by the TS transmission couldbe easily reduced using part of EDCA function totransmit further packets or by inserting WCBS in

a more general hierarchical scheduler for differenttraffic types.

Summarizing the obtained results we canextract some guidelines useful in the choice ofthe scheduler tailored for a particular scenarioof interest. In table VI we classify the behaviorof the analyzed schedulers, with respect todifferent parameters, in four levels of goodness(–,-,+,++). We consider the performance of thealgorithms in term of computational complexity(O()), admission control quality(AC), efficiencyand overhead (Eff./Ov.), type of traffic: CBR,mixed of CRB (M-CBR), VBR, mixed traffic(MIX) , and finally in term of reactivity to non-ideal channel conditions (Ch.Err).

Table VI. Guidelines for choosing the right scheduler.


A.C. – – – – + ++Eff./Ov. – – – – + ++O() + – + – ++

CBR + + – ++ ++M-CBR – + – ++ ++VBR – + ++ – –MIX – + – – +Ch.Err. – + ++ – – –

Acknowledgement

This work has been supported by the FRESCOREuropean project (Contract n. 034026).

References

1. Y. L. Boudec and T. Thiran, “A short tutorialon network calculus I: Fondamental bounds incommunication networks,” in IEEE ISCAS, 2000, pp.IV–93–IV–96.

2. IEEE802.11, “Wireless LAN medium access control(MAC) and physical layer (PHY) specification,”IEEE, Piscataway, NJ, 2007.

3. S. Tsao, “Extending earliest due-date schedulingalgorithms for wireless networks with locationdependent errors,” in Proc. IEEE VTC, Boston, MA,Sept. 2000.

4. A. Grilo and M. Nunes, “Performance evaluationof IEEE 802.11e,” in Proc. PIMRC, vol. 1, Lisboa,Portugal, Sept. 2002, pp. 511 – 517.

5. J. Cowling and S. Selvakennedy, “A detailedinvestigation of the IEEE 802.11e HCF referencescheduler for VBR traffic,” in Proc. ICCCN, Chicago,US, Oct 2004.

6. H. Fattah and C. Leung, “An overview of schedulingalgorithms in wireless multimedia networks,” IEEEWireless, vol. 9, no. 5, pp. 76–83, Oct. 2002.




7. A. Grilo, M. Macedo, and M. Nunes, “A schedulingalgorithm for QoS support in IEEE 802.11e networks,”IEEE Wireless Communications, vol. 10, no. 3, pp.36–43, June 2003.

8. S. Lu, V. Bharghavan, and R. Srikant, “Fairscheduling in wireless packet networks,” IEEE/ACMTrans. Net., vol. 7, no. 4, pp. 473 – 489, Aug. 1999.

9. D. Skyrianoglou, N. Passas, and A. K. Salkintzis,“Arrow: An efficient traffic scheduling algorithmfor ieee 802.11e hcca,” IEEE Trans. WirelessCommunications, vol. 5, no. 12, pp. 3558 – 3567, Dec.2006.

10. J. R. Jackson, “Scheduling a production lineto minimize maximum tardiness,” University ofCalifornia, Los Angeles, CA, Research Report 43,1955, management Science Research Project.

11. I. Inanc, F. Keceli, and E. Ayanoglu, “An adaptivemultimedia qos scheduler for ieee 802.11e wirelesslans,” in Proc. IEEE ICC, Istanbul, Turkey, June2006.

12. C. Liu and J. Layland, “Scheduling algorithm formultiprogramming in a hard-real-time environment,”Journal of ACM, vol. 20, no. 1, Jan. 1973.

13. G. Boggia, P. Camarda, L. G. Grieco, and S. Mascolo,“Feedback-based control for providing real-timeservices with the ieee 802.11e mac,” IEEE/ACMTrans. Net., vol. 2, no. 15, pp. 323 – 333, April 2007.

14. D. Ferrari and D. Verma, “A scheme for real-timechannel establishment in wide-area networks,” IEEEJournal on Selected Areas in Communications, vol. 8,no. 3, pp. 368–379, April 1990.

15. P. Ansel, Q. Ni, and T. Turletti, “FHCF: A simple andefficient scheduling scheme for IEEE 802.11e wirelesslan,” Mobile Networks and Applications, vol. 11, no. 3,pp. 391–403, June 2006.

16. C. Cicconetti, L. Lenzini, E. Mingozzi, and G. Stea,“An efficient cross layer scheduler for multimediatraffic in wireless local area networks with IEEE802.11e HCCA,” ACM Mob. Comput. and Commun.Reviews, vol. 11, no. 3, pp. 31–46, 2007.

17. R. Grow, “A timed-token protocol for local areanetworks,” in Proc. Electro/82, ser. Token AccessProtocols, I. Electronic Conventions, Ed., no. Paper17/3, El Segundo, CA, May 1982.

18. C. Cicconetti, L. Lenzini, E. Mingozzi, and G. Stea,“Design and performance analysis of the real-timehcca scheduler for ieee 802.11e wlans,” Comput.Netw., vol. 51, no. 9, pp. 2311–2325, 2007.

19. T. P. Baker, “Stack-based scheduling for real-timeprocesses,” Real-Time Syst., vol. 3, no. 1, pp. 67–99,1991.

20. G. Cecchetti, A. L. Ruscelli, and F. Checconi, “W-cbs: A scheduling algorithm for supporting qos in ieee802.11e,” in Proc. ACM-IEEE QSHINE, Vancouver,Canada, Aug. 2007.

21. L. Abeni and G. Buttazzo, “Integrating multimediaapplications in hard real-time sistems,” in IEEE RealTime-Systems Symposium, Dec. 1998, pp. 4–13.

22. C. Cicconetti, L. Lenzini, E. Mingozzi, and G. Stea,“A software architecture for simulating IEEE 802.11eHCCA,” in Proc. 3rd IPS MoMe, Warsaw, Poland,Mar 2005.

23. Network Simulator 2,http://www.isi.edu/nsnam/ns/.

24. A. Bacioccola, C. Cicconetti, and G. Stea, “User-levelperformance evaluation of voip using ns-2,” in Proc.NSTools, 2007.

25. P. T. Brady, “A model for generating on-off speechpattern in two way conversation,” Bell System

Technical Journal, vol. 48, pp. 2445–2472, 1969.26. Traffic Analysis for Voice over IP. Cisco Press,

November 2001.27. http://traces.eas.asu.edu/, 2005.28. F. H. P. Fitzek and M. Reisslein, “Mpeg4 and

h.263 video traces for network performance,” IEEENetwork, vol. 15, no. 6, pp. 40–54, November 2001.

Authors’ Biographies

Gabriele Cecchetti is a researcher at the ScuolaSuperiore Sant’Anna University, Pisa, Italy. Hiscurrent research activities span several areas,including: design and performance evaluation ofscheduling algorithms for QoS provisioning in wirelesscomputer networks with soft real-time constraints,design and evaluation of cross-layer frameworkarchitecture for providing QoS in heterogeneouswireless networks, and wireless link emulation.

Anna Lina Ruscelli received the MS degreein Electronic Engineering from the University ofPerugia, Italy, in 2005. She is currently a PhDstudent in Innovative Technologies of Informationand Communications Engineering and Robotics atScuola Superiore Sant’Anna, Pisa, Italy. Her researchinterests include Quality of Service support overheterogeneous networks, IEEE 802.11e networks, andcross-layer design.



real-timesupportforhccafunctioninieee802.11enetworks:a ... · devices which support multimedia...

Documents