Computer Networks 53 (2009) 1062–1072

Contents lists available at ScienceDirect

Computer Networks

journal homepage: www.elsevier .com/locate /comnet

Cross-layer optimized wireless multicast for layered media q

Ákos Kovács *, István Gódor 1

Ericsson Research, Traffic Analysis and Network Performance Laboratory H-1300 Budapest 3, P.O. Box 107, Hungary

a r t i c l e i n f o

Article history:Received 8 January 2007Received in revised form 20 November 2008Accepted 16 December 2008Available online 27 December 2008

Responsible Editor: Qian Zhang

Keywords:WirelessMulticastLayered mediaOptimizationCross-layer

1389-1286/$ - see front matter � 2009 Elsevier B.Vdoi:10.1016/j.comnet.2008.12.008

q This work is part of theFP6/ IST Project M-Pipe anEuropean Commission.

* Corresponding author. Tel.: +36 14377064; fax:E-mail addresses: Akos.Kovacs@ericsson.com (Á

dor@ericsson.com (I. Gódor).1 Tel.: +36 1 437 5237.

a b s t r a c t

This paper discusses the maximization of user-perceived quality of layered media in wire-less networks. We present a system model and an optimization framework for such net-works assuming: (i) the data rates are nested, (ii) the media is multicast, (iii) layeredand (iv) the media layer information is available in lower network layers based on cross-layer information. In order to show the general applicability of the proposal, different max-imization targets and algorithmic solutions are presented.

� 2009 Elsevier B.V. All rights reserved.

1. Introduction

In this paper we show how the user-perceived qualityof media can be maximized in wireless networks. It is as-sumed that (i) a single, shared channel is used for multicasttraffic in a cell, (ii) multiple data rates are available usingdifferent channel coding and modulation and (iii) termi-nals closer to the Access Point (AP) can receive multicastusing both higher and lower data rate transmission, whileterminals farther from the AP can receive lower data ratetransmission only. This latter property is referred to asnested data rates (see Fig. 1). It is also assumed that themedia is multicast, layered and the media layer informa-tion is available in lower network layers based on cross-layer information. We present a system model and anoptimization framework in order to provide maximumuser-perceived media quality in such networks. The goalcan be achieved by sending the media layers at the optimal

. All rights reserved.

d is co-funded by the

+36 1437 7767.. Kovács), Istvan.Go-

data rate. The presented model is based on, but not re-stricted to IEEE 802.11 Wireless Local Area Networks(WLANs). In the optimization framework, different maxi-mization targets and algorithmic solutions are presented.

The user-perceived quality of the media depends on thereceived number of layers and the quality properties of thelayered media. The layered media to be multicast typicallyconsists of a base and a number of additional enhancementlayers. Each additional layer provides progressive qualityimprovement. The larger the number of received layers isthe higher the perceived quality becomes. There are avail-able codecs providing such layered media stream like Ad-vanced Audio Coding – Bit-Sliced Arithmetic Coding(AAC-BSAC) [1] and Scalable Speech Audio Codec (SSAC)[2] for audio and H.264/MPEG-4 Advanced Video Coding(AVC) [3] and its extension, the Scalable Video Coding(SVC) [4,5] for video. These media codecs provide FineGrade Scaling (FGS). With FGS continuously increasingquality can be provided resulting in dozens or even hun-dreds of media layers in case of audio [2] and practicallymore than 30 in case of video.2

2 E.g, taking only two levels of spatial scalability, five of temporal andthree of SNR results 30 layers.

Border of Receving Area of 11 Mbps (Rate#4), inside also available: 1, 2 and 5.5 Mbps

AP Border of Receving Area of 5.5 Mbps (Rate#3), inside also available: 1 and 2 Mbps

Border of Receving Area of 2 Mbps (Rate#2), inside also available: 1 Mbps

Border of Receving Area of 1 Mbps (Rate#1)

Fig. 1. Illustration of nested data rates of WLAN 802.11b networks.

Á. Kovács, I. Gódor / Computer Networks 53 (2009) 1062–1072 1063

In order to take advantage of the media’s scalability, themedia layer information has to be available not only at theapplication layer, but also at lower network layers usingcross-layer information forwarding techniques. E.g., theLayer Independent Descriptor (LID) concept [6] proposesthe Differentiated Services Code Point (DSCP) in InternetProtocol (IP) packet headers to indicate the importance ofthe packets to achieve graceful quality control.

The (cross-layer based) quality control is more impor-tant in wireless access networks, where users may perceivedifferent channel quality due to their distance from the APand wireless phenomena (like multipath fading, etc.), inother words receiving capability of the users. Therefore,the received media quality should be optimized per user.This optimization task raises difficulties for current WLANstandards. Today in WLAN, only one data rate is used formulticast delivery, which data rate is mostly the lowestand provided by the most robust modulation. This tech-nique does not differentiate the users, but ensures thatall users can receive all of the data. The data rate of thetransmission is fixed and the possibility of changing thedata rate per packet is not used for multicast, becausethere are no quality feedback and processing functions de-fined. This insufficiency of the current delivery methodlimits deploying efficient multicast schemes in wireless ac-cess networks like WLAN.

The existing multicast approaches [7] do not take theadvantage of nested data rates existing, e.g., in WLAN net-works, because these approaches were developed origi-nally for wireline technology. In networks having nesteddata rates, all lower data rates are available for a user inconjuction with potential additional higher data rates.(See Fig. 1 for illustration in case of WLAN 802.11b net-works.) In such networks if the receiving capability of theusers is known (based on some reporting technique), theplayblack quality of the layered media can further be im-proved and tailored to the individual receiving capabilityof the users.

In this paper a system model is presented for multicastoptimization and a proposal is given to complete the cur-rent WLAN standard according to this model. Note, thatthe proposed solutions can be implemented in any othernetwork providing nested data rates. Besides the systemmodel, three possible optimization targets and solutionsare presented. Since the exact solution of two targets isNP-complete, heuristics are also given. We show that theheuristics are good enough for the most common case ofmany layers.

The rest of the paper is organized as follows. First, anoverview of the related work is given. Then a general sys-

tem model for multicast optimization is presented in Sec-tion 3, which allows the operators to maximize the user-perceived quality of media delivery meanwhile optimizingthe channel usage of wireless network. The system modelis introduced through an example of IEEE 802.11 WLAN[8,9] networks. Separately from the system, different tar-gets of the optimization task and algorithmic solutionsare presented in Section 5. The performance of the pro-posed solutions is analyzed in Section 6. Finally Section 7summarizes the paper.

2. Related work

The existing multicast approaches [7] come from thewireline technology. In Single Stream Adaptive Approach(SSAA), a single version of the media is transmitted forall users. This approach is typical in today’s wireless net-works. In Replicated Streams Approach (RSA), differentquality versions of the media are sent out by the media ser-ver and users having higher bandwidth sign up for a higherquality of the media. Each version forms an independentstream and multicast group. This method is bandwidthwasting because of multiple and parallel transmissions ofthe media. To overcome the drawback of RSA, the LayeredStreams Approach (LSA) can be used, where different lay-ers of the media are sent to different multicast groups.With receiver-driven controlling methods [10–12], eachuser can decide to which groups (layers) they sign up. Thisapproach eliminates the redundant transmission, mean-while optimizes the perceived quality at the expense ofsome application layer overhead [13]. RSA and LSA is appli-cable in wireless networks as well. However, since currentWLAN standards do not support information on receivingcapability, both approaches have to use the most robusttransmission rate, i.e. the slowest, to reach all users inthe cell. Hence, it is possible that not all quality versions(RSA) or media layers (LSA) fit in the channel capacitydue to the higher bandwidth demand. Nevertheless LSAutilizes the layered structure of the media in a more effi-cient way, running out of capacity is a risk for both solu-tions. Another way how RSA and LSA can be adopted towireless networks is proposed in [13]. Different data ratescan be considered as individual access networks. That is, alllayers and quality versions has to be sent to all access net-works. This solution still transmits redundant data andoverload the network. Instead of a receiver-driven methodwe propose a centralized. We aim to optimize the overallperformance of the system regarding the user-perceivedquality. A simple way to realize such an optimization con-cept could be to use a table lookup mechanism [14]. From a

1064 Á. Kovács, I. Gódor / Computer Networks 53 (2009) 1062–1072

limited set of predefined rate – layer combinations, thesystem chooses a pair considered to be optimal. Such asolution has two main drawbacks. On the one hand the re-quired bandwidth of the layers is media dependent, thusthe lookup table should be made for each media stream.On the other hand this mechanism is static, i.e., it doesnot consider the movement of receivers thus the changingof their receiving capabilities. A dynamic method can beachieved by the optimized distribution of the media layersif the receiving capabilities of users and the nested datarates are considered.

An important parameter of the optimization process isthe available bandwidth of the shared medium. The avail-able bandwidth depends on the theoretically free channelcapacity, and the actual channel condition. Decreasedchannel capacity reduces the amount of data that can betransmitted at a given data rate. The most trivial way todetect free capacity is to send probing packet sequence,but there are more sophisticated methods as well. In [15]a mechanism is proposed that uses the mobile nodes toestimate link characteristics based on the wireless linkquality and contention status of the whole WLAN. Theauthors of [16] propose a non-intrusive link bandwidthmodel based on Signal to Noise Ratio (SNR) to estimate freecapacity. Back-Propagation Neural Network and BayesianInterference methods are used in this model. In [17] amodel is established to analyze the throughput of theWLAN channel. Giuseppe in [18] proposes a simple analyt-ical model to compute the saturation throughput perfor-mance in presence of a finite number of terminals and inthe assumption of ideal channel conditions. Further ade-quate techniques for predicting the free channel capacitycan be found in [19–21], thus the calculation is out of thescope of this paper.

The channel condition affects the number of availabledata rates. In case of bad channel condition it is possiblethat at most the slowest rate is available only. There aremany techniques for measuring the channel condition[22–27]. Basically they are Frame Error Ratio (FER), SNR,and retry based methods. These techniques are practicallyused for choosing that modulation for transmission, whichprovides the fastest data rate along with acceptable net-work parameters like frame error rate and signal-to-noiseratio, thus are often referred to as rate control schemes.Auto Rate Fallback (ARF) [28] was one of the first rateadaptation algorithms. It was designed to optimize theapplication layer throughput. Each sender attempts touse a higher transmission rate after a fixed number of suc-cessful transmissions (given by a threshold) at a given rateand switches back to a lower rate after one or two consec-utive failures. In Adaptive ARF (AARF) proposed in [29], thethreshold is continuously changed at runtime to better re-flect to channel conditions. Receiver Based Auto Rate(RBAR) [30] also aims to optimize the application through-put and mandates the use of the Request to Send (RTS)/Clear to Send (CTS) mechanism. The receiver of the RTSframe calculates the transmission rate to be used by theupcoming data frame transmission based on the SNR ofthe received RTS frame and on a set of thresholds calcu-lated with an a priori wireless channel model. Although,RBAR has little practical interest since it cannot be imple-

mented in existing WLAN standards, it gives a theoreticallygood reference model. The goal of MiSer [31] is to optimizethe local power consumption instead of applicationthroughput. A set of optimal rate/transmission power pairsis calculated offline with a specific channel model. At run-time, the wireless nodes execute table lookups to deter-mine current optimal transmission rates. There existsalso proposals on rate control for especially multimediatraffic. All solutions aim to keep the latency of transmissionlow, since real-time media delivery is very delay sensitive.They propose enhanced Automatic Repeat Request (ARQ)schemes and improved Forward Error Correction (FEC)algorithms to guarantee limited delay. Such solutions canbe found in [14,32,33]. To detect the available data rateswe propose to use triggered and unsolicited messages bythe users. The decision algorithm that triggers these mes-sages can be any applicable rate control scheme. Hereby,our solution can be a completion of the methods above,since we address to maximize the user-perceived quality.

In case of the basic IEEE 802.11 specification of WLAN,all packets has statistically the same chance to access theradio channel provided by the Distributed CoordinationFunction (DCF). Based on the performance of DCF (e.g.[34]) and its possible improvements (e.g. [35]), a better ac-cess mechanism called Enhanced Distributed Access Chan-nel (EDCA) is proposed in IEEE 802.11e [36]. EDCA providesfour Access Categories (AC) with different priorities. Ingeneral, we can say that AC3 has the highest and AC0 hasthe lowest chance to access the channel. In [37], theauthors propose that the packets of the media with differ-ent importance (they consider 5 partitions of the mediaprovided by the H.264 video codec [3]) should be mappedto different ACs, thereby, providing better performance ofthe video transmission. This idea can be incorporated intoour proposed system model. Based on [37], the priorityclass of each layer of the media can be set similarly tothe data rates (see Section 4.2).

3. System model for multicast optimization

In this section, first we describe our system architec-ture. Next, the optimization goals are introduced, while fi-nally we show how this optimization approach can bemapped onto our WLAN architecture resulting an opti-mized system model.

Our system model is presented through a typical WLANscenario, multicasts layered media, like it is depicted inFig. 2. The system architecture consists of a Multicast Back-bone (MB) connecting the Access Network and the MediaServer via a Media Gateway (MG). The Access Networkconsists of Access Points (APs) and the user stations. Forsigning up to or leaving a specific multicast group, InternetGroup Management Protocol (IGMP), e.g., IGMPv2 can beused [38].

In this architecture, the MG is the last node, which hasdetailed information about the layered structure of themedia (based on the cross-layer information available inIP packets [6]). The MG is ‘‘directly connected” to theAPs, which set the data rate of the packets. That is, theMG is a proper place to control the channel utilizationand run an optimizing algorithm.

AP

MediaGateway

Media Server

LayeredMedia

MulticastBackbone

AP

UserStation

UserStation

UserStation

UserStation

Fig. 2. The network architecture.

Table 1Overview on wireless standards.

Standards Supports rateswitching

Available data rates(Mbps)

802.11a/g(WLAN)

Yes 6–54 in eight steps

802.11b (WLAN) Yes 1–11 in four steps802.16 (WiMax) Yes 7.47–104.5 in 13 stepsHIPERLAN-1 No 23.5HIPERLAN-2 Yes 6–54 in seven stepsHIPERMAN Yes 14–63 in six steps

Á. Kovács, I. Gódor / Computer Networks 53 (2009) 1062–1072 1065

There are a number of different wireless standards thatare conceptually similar to WLAN, shown in Table 1. ExceptHIPERLAN-1, all of them supports data rate switching,since all can switch between modulation schemes duringdata transmission even per packet. That is, all technologiescan support our system model for the proposed multicastoptimization without major modifications.3

The goal of the optimization presented in the followingis to maximize the user-perceived quality of multicast lay-ered media meanwhile minimizing the channel usage inwireless networks. The goal can be achieved by sendingthe media layers at the optimal data rate.4 The proposedexploitation of the nested data rates provides more effectiveuse of the wireless resources if the receiving capabilities ofthe users are different. In that case, faster data rates canbe used for some media layers resulting less occupation ofthe radio channel. We propose a system model based onthe LSA approach. Unlike LSA, our method is not receiver dri-ven but centralized. That is why it is not necessary to usedifferent multicast groups for different layers of the media.A particular media stream is considered as a single flowand the distribution of consisted layers are controlled bythe system. In the following we provide all network func-tions that is necessary for the optimization.

3 Note, that in case of Super G mode 104 Mbps is available for 802.11g,too. In 802.16 WiMax, the data rates depend on the channel bandwidth.

4 As shown in [13], layered media allows to forward only those layerstowards the users, which can actually be used by them.

A short summary of the proposed system model is asfollows.

(1) The MG collects information about the receivingcapabilities of the users per AP.

� The users send their receiving capability informa-

tion directly to the MG.� The MG collects the receiving capabilities of the

users per AP for each media stream, separately.

(2) The MG computes the optimal data rate of the media

layers for each AP based on the receiving capabilitiesof the users and the layer information of the media.

� We assume that the cross-layer information

about the layer of the packets can be carried inthe IP header [6], which means that the MG canbe a proper network element for processing thisinformation.

(3) The MG tells the APs which data rate to use for theforwarded packets based on the result of theoptimization.

� The MG forwards the packets to the appropriate

APs.

(4) The optimization proceeds at Step 1.

4. Realization of the system model

In this section we show how the introduced WLANarchitecture should be modified and what novel featuresshould be added in order to support the layered multicastoptimization. These features include the collection ofreceiving capabilities, data rate setting and error protection.

4.1. Information on receiving capability

The WLAN standards define various modulation typesfor data delivery, resulting in different data rates. To opti-mize the user-perceived quality, the exact knowledge ofthe current receiving capability of the users per AP isneeded, i.e., how many users are able to receive at eachdata rate. As in current WLAN standards no receiving capa-bility reporting is used (in case of multicast traffic), some

Table 2IGMPv2 receiving capability update message format.

1066 Á. Kovács, I. Gódor / Computer Networks 53 (2009) 1062–1072

novel extensions in the architecture must be considered.After introducing the technical background of how to de-tect the receiving capability, a protocol extension is pro-posed to signal this information. Then the way isdescribed how the MG collects this information. Finallythe setting of the data rates is introduced.

4.1.1. Detection of receiving capabilityIn WLAN, the Physical Layer Convergence Protocol

(PLCP) header of each packet is sent out at the lowest datarate. This ensures that all users can detect the header. Thedata rate used for the payload is indicated by the PayloadSignaling Field (PSF) of the header. Thus each user candetermine the highest perceptible data rate of acceptablequality [22–27].

If the radio conditions are stable or change slowly, theuser could send a receiving capability update report whenhis/her receiving capability is changed. If the radio condi-tions change rapidly, then a periodic receiving capabilityupdate could result in less signaling overhead comparedto the continuous updates. However, if the receiving capa-bility information is not up-to-date, then the efficiency ofthe optimization can be degraded.

4.1.2. Signaling of receiving capabilityReceiving capability signaling can be handled by

IGMPv2. For such purpose a new message type should beused, the IGMPv2 Receiving Capability Update message(IGMP-RCU). Table 2 shows the proposed IGMP-RCU mes-sage format.

The meaning of the fields is as follows. The Type field5

identifies the IGMP-RCU message. The Max Response Timefield should be set to 0 by the users. The Group Address con-tains the multicast group address to which the user isreporting the receiving capability update. All nodes signedup to this address ignore the IGMP-RCU messages exceptthe MG. The Old Rate is the highest data rate that the usercould receive before the update and 0 can be reserved fornew entry users. Since there can be multi-standard capabledevices in the network, the identifier of a rate should implic-itly signal the standard applied by the user, thus there is noneed to define signaling to tell the MG what kind of standardis used in a given AP. The New Rate is the data rate that theuser can receive hereafter.

5 The value of the type field needs an Internet Assigned NumbersAuthority (IANA) registration.

Some further notes must be considered: (i) the MGcould collect the user information in a receiving capabilitytable. If Old Rate is 0, then the MG adds the new user to thetable. If New Rate is 0, the MG deletes this user from thetable.6 (ii) In order to avoid the loss of IGMP-RCU messages,an IGMP-like error protection (doubled messages) can beconsidered. As the lowest data rate is the most robust anderror-tolerant, it can be used to protect IGMP-RCU messagesagainst accidental bit errors.

4.1.3. Initial probing phaseBefore the distribution of the stream, the MG needs to

know the initial receiving capability of the users. There-fore, the MG should probe all the data rates for a shortpacket train (e.g., the first packets of the media). This InitialProbing Phase (IPP) allows the receivers to detect their ex-act receiving capability. During IPP all layers are transmit-ted at all possible data rates (allowed by the standard andthe available channel capacity). All users must reply withan IGMP-RCU message after the reception of the first pack-ets. In this message, there is 0 for the Old Rate indicatingthat a new user has joined the group.

The MG indicates the users the beginning and the end ofthe IPP by an IGMP-RCU message with the Max ResponseTime set to the length of the IPP. This message is sent tothe appropriate multicast group. At the end of the IPP,the MG starts sending each layer enforcing such modula-tion that results in the appropriate data rate.

4.1.4. Secondary probesIt is possible that some data rates are not used during

the transmission (e.g., no user can receive them in theIPP or a new user has joined the system). Periodic Second-ary Probes are performed in order to trigger the users tosend IGMP-RCU messages to reveal whether they can re-ceive at the unused data rates. Note, that both Initial andSecondary Probes are used to collect information on thechannel condition, that is to detect available data rates inthe system.

4.2. Data rate setting

Each data rate is associated with a Virtual Local AreaNetwork (VLAN) [39] between the APs and the MG. The

6 Note, that the table registers only the total number of users belongingto each data rate.

Á. Kovács, I. Gódor / Computer Networks 53 (2009) 1062–1072 1067

MG puts the media packets only to the proper VLAN. Withthis method there is no need to maintain detailed per-media information about the layer distribution in theAPs. The only new feature for an AP is to be able to modu-late the packets according to the VLAN ID.

4.3. Error protection

In case of wireless transmission, error protection is al-ways an issue. The protection can be applied either perframe or per layer.

One way how a per frame protection can be carried outis the use of Forward Error Correction (FEC). This adds ex-tra data overhead to the protected packets. Another way isthe application of Unequal Error Protection (UEP). UEP pro-tects stronger those frames which are more important forthe decoder. Note that per frame protection does not influ-ence the overall system behaviour we described. Mean-while the system sets the appropriate modulationscheme for each layer, within the layers the individualframes can be protected as well.

UEP can also be applied per layer. That is, a layer (whichis more important for the decoding process) can be pro-tected more by transmitting it on a more robust data rate.For example, the base layer can be sent at 1 Mbps in case ofa WLAN 802.11b network and not on the selected optimaldata rate.

5. Optimization targets and solutions

This section focuses on the optimization process calcu-lating the optimal data rate for each layer of the media.Optimality means that the channel is efficiently utilizedand the best possible quality is provided for all usersaccording to a given optimization target. First we give pos-sible optimization alternatives, then later in this sectionwe define their corresponding targets. Finally, we presentthe solution of each alternative and if necessary, scalablesolutions are shown as well.

5.1. Optimization targets

The targets of the optimization task can be differentdepending on whether user location has influence on thereceived quality or not. Thus we define two main targetsand a third, hybrid one. In case of all targets, there is a con-straint on the channel occupation7 for the multicast media.The three targets are as follows.

� Maximize the average: provides the optimization of theaverage user-perceived quality depending on the placeof the users. A service based on this target can bethought of as an optimized best-effort multimedia ser-vice: the user has to be at the right place to get the right

7 The bandwidth is the product of the applied data rate and the channeloccupation time, so the channel occupation is the real limiting factor incase of multi-rate wireless networks.

quality. That is, the perceived quality of an individualuser may be decreased in order to increase the overallquality in average. For this target we show an optimalalgorithm called MaxAvg in Section 5.2.

� Maximize per rate: provides the optimization of theperceived quality of individaul users independently oftheir place. Each user gets the best possible qualitywithout decreasing the perceived quality of othersbeing at less advantageous places. That is, the medialayers are sent at the lowest data rate allowed bythe channel occupation constraint. For this target weshow an optimal algorithm called MaxRate in Section5.3.

� Hybrid: provides a minimum quality level that everyonereceives and optimizes the remaining channel ‘‘capac-ity” among users in good radio condition. That is, theabove two maximization targets are mixed. For this tar-get we show an optimal algorithm called Hybrid in Sec-tion 5.4.

The formulization of the above optimization targets areas follows.

Let L be the total number of layers the media consists ofand R be the number of data rates available in the network.

Let ½l1; l2; . . . ; lR� denote a layer distribution among datarates 1; . . . ;R. That is, li denotes the number of layers to-tally received at rate i. Note, that physically only li � li�1

number of layers are sent at rate i. (See Table 3 forillustration.)

Let ½n1;n2; . . . ;nR� denote the user distribution amongdata rates 1; . . . ;R. That is, ni denotes the number of usersreceiving data at the highest rate i. (That is, sending a layerat rate i affects the perceived quality of Nr ¼

PRr¼inr users

total.)The target of the targets is to find the optimal ½̂l1; l̂2 . . . l̂R�

layer distribution so that (the ‘‘required quality”)

1.PR

r¼1nrQ l̂rin case of MaxAvg,

2.PR

r¼1Ql̂rin case of MaxRate,

3. Ql̂kþPR

r¼kþ1nrQ l̂rin case of Hybrid (if k is the first

highest rate received by users, i.e., ni ¼ 08i < k)

is maximal and the channel constraint

XR

r¼1

Pl̂rk¼l̂r�1þ1

bk

Sr6 C ð1Þ

is satisfied where Qli is the total quality provided bylayers up to i, bi is the required bandwidth of layer i,Si is the speed of data rate i and C is the channel occu-pation constraint. (Note that l0 is considered as 0.) SeeSection 6 and Fig. 3 for examples of quality and band-width curves.The number of layers ðliÞ to be received at a given data

rate (i) determines which layers the AP has to send oneach modulation scheme. Table 3 shows the relation be-tween the received and sent layers at each data rate inan illustrative example of IEEE 802.11b and 10 medialayers.

0%

20%

40%

60%

80%

100%

0% 20% 40% 60% 80% 100% Sent Layers

ytilauQ

aideM

deviecreP

0%

20%

40%

60%

80%

100%

0% 20% 40% 60% 80% 100% Sent Layers

htdiwdna

Bderiuqe

R

Typical SVC with Movie #1 SVC with Movie #2 Extremity

Fig. 3. Quality and bandwidth as a function of the number of layers.

1068 Á. Kovács, I. Gódor / Computer Networks 53 (2009) 1062–1072

5.2. Algorithm MaxAvg

The goal of this algorithm is to maximize the average ofthe user-perceived quality depending on the place of theusers. This is achieved by finding the proper modulationand data rate for each layer. The exact solution can be for-mulated as a combination with repetition, where we selectthe proper number of layers to be received at each datarate. Since the problem is NP-complete (see Appendix Afor proof), the presented exact solution is applicable onlyto a small number of media layers and data rates in prac-tice. Therefore, we present a scalable solution to solvethe problem in polynomial time.

5.2.1. Scalable solutionThe main idea of the (MaxAvg Scalable (MAS)) can be

summarized as follows: (1) start from an empty networkand (2) fill the network with the following two opera-tions: (i) add a new layer to the highest data rate or(ii) move an existing layer to a lower data rate. Mean-while satisfying the channel occupation constraint, themost effective operation should be applied in each stepof the algorithm based on the gain function. If layer lis added or moved to data rate r (i.e., l ¼ lR þ 1 in caseof add to r = R and l ¼ lr þ 1 in case of move), then thegain function Gl;r is as follows:

Gl;r ¼Nr � DQ l

DBl;r; ð2Þ

where Nr is the number of users affected by the operation(see Section 5.1 for details), DQ l is the increase of the qual-ity per affected user, DBl;r is the increase of the total chan-nel occupation caused by the operation.

Table 3Relation between received and sent layers, an example.

Data rate # Receivedlayers

Selected layers byalgorithm

Sent layers in practice(by AP)

1 Mbps l1 ¼ 1 1 12 Mbps l2 ¼ 3 1, 2, 3 2, 35.5 Mbps l3 ¼ 8 1–8 4–811 Mbps l4 ¼ 10 1–10 9, 10

The evaluation of a step is based on the quality and thebandwidth function of the media. See typical curves inFig. 3. The quality function is derived from the rate distor-tion function showing the distortion of the media as afunction of the received amount of data. That is, the higherthe amount of data (equally to the number of layers in ourcase) is received, the lower the distortion becomes and thehigher the perceived quality is. The bandwidth of a layercan be derived from the bandwidth function showing therequired bandwidth in function of the number of layerssent. The channel occupation of a layer is the product ofthe bandwidth of the layer and the data rate at whichthe layer is sent out.

In sum, that operation should be chosen at each step ofthe algorithm, which gives the most quality increase withthe least channel occupation time increase.

At the expense of computation time, the probability offinding optimal solution can be increased. In that case,not only maxl;rfGl;rg is chosen, but the next k best possiblechoices, as well. This mechanism forks the main thread ofthe algorithm into k alternative direction in each step.After the main thread is finished, the alternative threadsare checked, as well. The final solution is chosen from themain and the alternative solutions according to the maintarget defined in Section 5.1.

The proposed MAS algorithm can be summarized asfollows:

1. (Initialization) Start from an empty network, i.e.,li ¼ 08i ¼ 1; . . . ;R.

2. (Main thread) Fill the network with layers:

(a) (Evaluation) Evaluate and order each possibleoperation (allowed by the channel occupation con-straint) based on the gain function.(b) (Alternatives) Store k alternative direction andcontinue the main thread with the best direction(if channel is not full or not all layers are added tothe network) by going to Step 2a.

3. (Alternative Threads) Check the alternative solutions:

(a) (Start an Alternative) Start from the next (previ-ously stored) alternative direction (i.e., a partiallyfilled network).

Á. Kovács, I. Gódor / Computer Networks 53 (2009) 1062–1072 1069

(b) (Evaluate and Fill) Evaluate each possible opera-tion (allowed by the channel occupation constraint)based on the gain function. Continue this threadwith the best direction (if channel is not full or notall layers are added to the network) by repeating thisstep.(c) (Iteration) Continue with a new thread by goingto Step 3a.

4. (Final Solution) Pick the best solution from the solu-tions of the main and the alternative threads basedon the objective.

5.3. Algorithm MaxRate

The solution of the problem is quite straightforward.We have to find the maximal number of layers to be re-ceived at each rate. The data rates are chosen one afterthe other (started with the lowest one) and the most layersare allocated to each data rate (determined by the channeloccupation constraint) independently from how many usercan receive the individual data rates. The number of layersto be received at a given data rate determines which layersthe AP has to send on each modulation scheme similarly tothe case of MaxAvg, see Table 3 for illustration.

Of course, the unused channel capacity decreases whileheading from a lower data rate to a higher one. The com-plexity of the solution is a linear function of the numberof layers to be sent.

5.4. Algorithm Hybrid

The goal of this algorithm is to maximize the quality levelthat everyone receives and to optimize the remaining chan-nel capacity among users in good radio condition by maxi-mizing the average of the user-perceived qualitydepending on the place of the users. Hybrid can be formu-lated as a combination of the algorithms MaxAvg and Max-Rate. The number of layers received at the lowest data rate isdetermined by MaxRate. The distribution of the remaininglayers is determined by MaxAvg. Thus the Hybrid algorithmis also NP-complete and needs scalable solution as well.

5.4.1. Scalable solutionAs in the case of the exact solution, a scalable solution

(Hybrid-S) can be given as a similar combination of Max-Rate and MAS. The main difference between Hybrid-Sand MAS is the reduced state-space where the solutionshould be found: there are less layers and data rates in aparticular example.

8 The whole channel available for the media.

6. Performance analysis

This section summarizes the results of the performanceanalysis. First, the generation of the problem instances isdescribed based on the network and media parametersapplied in the analysis. Then the comparison of the exactand heuristic solutions is presented. Finally the complexityanalysis of the algorithms is given.

Since the Hybrid algorithm can be considered as a spe-cial case of the MaxAvg algorithm, the analysis focuses

only on MaxAvg and MAH algorithms. MAH was used withk = 0 (only the main thread of the algorithm is used) andk = 2 settings (two alternative threads are used in each stepof the algorithm, as well). Since the MaxRate algorithmprovides optimal solution in polynomial time, MaxRate isnot involved in the analysis.

6.1. The network and media parameters applied in theanalysis

The performance of the algorithms was investigated incase of IEEE 802.11b WLAN networks connected to the sys-tem realization presented in Section 3. In this case thenumber of data rates was R = 4. In order to keep the com-plexity and the running time of the exact solutions compa-rable with the heuristics, we applied 10 media layers(L = 10). The users can be grouped according to the maxi-mal data rate that they can receive. In the tests, the num-ber of the users in each of these groups was varied from0 to 4 resulting 624 (meaningful) test network scenarios.

The efficiency of the heuristics depends on the appliedquality and bandwidth functions. Fig. 3 illustrates theuser-perceived media quality and the bandwidth require-ment of the media as function of the number of receivedlayers. (Note that we assumed no packet loss in the net-work, so if a user can receive a layer, then he/she will re-ceive all packets of the layer.)

The typical curve of the quality can be derived from thetypical Rate Distortion curve which has an decreasing log-arithmic shape [40,41]. In the theory, the distortion is larg-est if no data is received and equals to zero if the coding islossless. In the quality context, this means that the play-back quality is the best if the user (coder) receives all lay-ers of the media and the quality is zero if no layers arereceived. That is, the quality curve has an increasing loga-rithmic shape. The typical quality curve and the corre-sponding typical bandwidth curve are depicted bytriangles in Fig. 3. Besides the typical curves, the character-istics of Scalable Video Coding (SVC) [4,5] are given for twomovies. Movie #1 represents a dynamic movie trailer andis depicted by circles. Movie #2 represents a documentaryand is depicted by squares. Finally, an extreme characteris-tic is represented by solid line.

The performance of the algorithms was analyzed notonly with the above media types, but with fictitious, virtualones. Therefore, additional quality and bandwidth curveswere generated as transition between the typical curvesand the extreme curves. Then a virtual media can be cre-ated by associating such a quality curve with such a band-width curve. In sum 91 media streams were tested.

The absolute bandwidth of the media streams was setin such a way that all layers are available for all users ifthe channel occupation constraint is 100%.8 In order toanalyze the effects of background traffic, the channel occu-pation constraint was varied from 0 to 100% in 10% stepsresulting 11 versions. So in total we got 624,624 test cases.

Table 4Comparison of the results of the exact and heuristic solutions.

Algorithms Solutions compared to MaxAvg (%) Ratio of optimal solutions (%)

Average Deviation

MAH (k = 0) 99.7605 0.2290 89.41MAH (k = 1) 99.9375 0.0699 95.34MAH (k = 2) 99.9385 0.0690 95.45

1070 Á. Kovács, I. Gódor / Computer Networks 53 (2009) 1062–1072

6.2. Numerical results

The results are summarized in Table 4. Since MaxAvgprovides optimal solutions by definition, it was used as areference for MAH. The first column indicates the twocases we investigated. The second column shows the aver-age user-perceived media quality compared to the optimalgiven by MaxAvg. The third column shows the standarddeviation from the average. The fourth column shows thepercentage of test cases in which the heuristic algorithmscould find the optimal solution.

The results show that the heuristic is very effectiveespecially in the case when alternative threads (whichcan be practically considered as a look-ahead meachnism)are used. MAH with k = 2 can reproduce the optimal solu-tion in more than 95% of the test cases. Moreover, the aver-age solution reaches 99.94% of the optimal quality. Thestandard deviation of MAH is practically negligible, 0.07%.

6.3. Complexity analysis

Beside the performance analysis presented above, thecomplexity of the heuristic compared to the exact solutionis another important characteristic.

The complexity of the exact solution (MaxAvg) is prac-tically the complexity of a combination with repetition(this is the number of layer distributions to be checkedduring the optimization process) where we want to chooseR elements from L + 1 possible ones: OE ¼ O ðLþRÞ!

R!L!

� �.

The ‘‘worst case complexity” of the main thread of theheuristic can be calculated as follows. In the worst case, anew layer is added to the highest data rate and this layeris moved to a lower data rate R� 1 times. At most L layercan be added to the solution, so there areð1þ ðR� 1ÞÞL ¼ RL steps of the main thread of MAH. Ineach step, R possible direction (possible layer distribution)should be checked (1 add operation and R� 1 move oper-ation), so the complexity of the MAH with k = 0 is:O0 ¼ OðR2LÞ.

In case of alternative threads, k alternative runnings areneeded per each step of the main algorithm. The complex-ity of the alternative runnings can be estimated from aboveby the complexity of the main thread, i.e., O0. Thereby, thecomplexity of the main algorithm per step isRþ kO0 ¼ Rþ kR2L. After simplifications, the complexityof MAH with k > 0 can be formulated as: Ok ¼ OðkR3L2Þ.Note that value of k is practically a constant multiplicatorin the complexity, so it does not modify the exact theoret-ical complexity. However, this multiplication could beimportant in real-time network environments, so the slightimprovement of the result (from k = 1 to k = 2) shown in

Table 4 is almost negligible compared to the increase ofthe running time.

It can be seen that MAH with k = 0 is the fastest solutionand has good performance including the quality of the re-sults. If smaller standard deviation of the solutions isneeded, then we propose MAH with k = 1 or k = 2 resultingbetter quality and acceptable running time. For example, incase of IEEE 802.11g WLAN networks with eight data rates(meaning L = 10 and R = 8 parameters), MAH with k = 1 isstill at least two-times faster than MaxAvg.

7. Conclusion

In the paper, multicast delivery of layered media isinvestigated in case of wireless networks. We proposed asystem model for optimization based on cross-layer infor-mation. In this model, the properties of the layered mediacan be utilized in order to maximize the user-perceivedmedia quality meanwhile efficiently utilizing the availablechannel capacity.

The paper describes a centralized optimization processin wireless access networks, where the Media Gateway col-lects the receiving capability information of the users,determines the optimal data rate for each layer and in-forms the Access Points which data rate to use for eachlayer.

For the calculation of the optimal data rate, we proposeboth exact and effective scalable algorithms, which incor-porate knowledge about the quality properties of the lay-ered media and the receiving capabilities of the users.

The proposed techniques and methods have low com-plexity and can be implemented in a relatively simpleway in real network environment.

Acknowledgements

The authors would like to thank Vera Vértesi, SándorGy}ori and the anonymous referees for their valuable com-ments and help.

Appendix A. Proof of NP-completeness of the ExactMaxAvg Algorithm

The MaxAvg algorithm can be formalized as follows. Gi-ven the number of data rates (R P 0), the user distributionðn1; . . . ;nR P 0Þ, the speed of the data rates ðS1; . . . ; SR P 0Þ,the channel occupation constraint (C P 0), the number ofthe media’s layers (L P 0) and the required quality(Q P 0). The task is to check if a layer distribution

½̂l1; l̂2 . . . l̂R� exists for whichPR

r¼1

Pl̂rk¼̂lr�1þ1

bk

Sr6 C (note that

Á. Kovács, I. Gódor / Computer Networks 53 (2009) 1062–1072 1071

l0 ¼ 0 by definition) andPR

r¼1nrQ l̂rP Q , where bi is the re-

quired bandwidth of layer i and Q li is the total quality pro-vided by layers up to i.

The problem can be reformulated by defining: Ai ¼ 1=Si,Di ¼ ni�1 � ni and BR ¼ nR. Then we are looking for aNi ¼ li � li�1 P 0 such that

PRi¼1Ni 6 L,

PRi¼1 ¼ AiNi P B

andPR

i¼1 ¼ DiNi 6 Q . Further simplification is to letAi ¼ Di, C = Q and L = R.

Thus a problem, Modified Knapsack Problem (MKP) canbe defined as:

Input : A1; . . . ;AR P 0

Q P 0

Task : check if exists N1; . . . ;NR 2 N such thatXR

i¼1Ni 6 R

XR

i¼1AiNi ¼ Q :

Claim 1. The input of the MKP can be constructed from theinput of the Exact Cover by 3-sets (X3C) problem in polyno-mial time.

Proof. In X3C, there is a base set I ¼ f1; . . . ; tg with somesub-sets fXi : i 2 Ig of three elements ðjXij ¼ 3Þ. The taskis to check if a single-layer cover exists, i.e., there is J # Ifor which [i2JXi ¼ f1; . . . ; tg and Xi \ Xj ¼ ;ði; j 2 JÞ. Thuslet the input of the X3C be given: system of setsXiði 2 IÞ; jXij ¼ 3 over the base set.

Based on the above, the input of MKP can be con-structed as:

R :¼ jIj;r :¼ 1þ R;

Ai :¼X

j2Xirj�1 ði 2 IÞ;

Q :¼Xt

j¼1

rj�1: �

Definition 2. Let N1; . . . ;NR 2 N and define fjðj 2 f1; . . . ; tgÞcovering numbers as: fj ¼

PRi¼1Ninfj2Xig, where nfj2Xig ¼ 0 if

j R Xi and nfj2Xig ¼ 1 if j 2 Xi are indicators.

Lemma 3. The sumPR

i¼1AiNi and series ff1; . . . ; ftg uniquelydetermine each other.

Corollary 4.PR

i¼1AiNi ¼ Q () ff1; . . . ; ftg ¼ f1; . . . ;1g.

Proof. Transpose equationPR

i¼1AiNi as follows:PR

i¼1AiNi ¼PR

i¼1NiP

j2Xirj�1

� �¼PR

i¼1NiPt

j¼1nfj2Xigrj�1 ¼

Ptj¼1rj�1 PR

i¼1nfj2XigNi

� �¼Pt

j¼1rj�1fj ¼ P. By definition,

at most R sets can contain an element j, thusfj < r ð8j ¼ 1; . . . ; tÞ, so fj covering numbers are uniquelyrepresented in base r. The primary condition P = Q can bemet if and only if fj ¼ 18j ¼ 1; . . . ; t. h

Claim 5. X3C has solution with the above input if and only ifMKP has solution with the above constructed input.

Proof. Suppose that X3C has solution. Then the solution ofMKP is given by Nj ¼ 1 if j 2 J and Nj ¼ 0 if j R J.

Suppose the opposite that MKP has N1; . . . ;NR solution.Then

PRi¼1AiNi ¼ Q ) ff1; . . . ; ftg ¼ f1; . . . ;1g ) N1; . . . ;NR

6 1) the solution of X3C is J ¼ fj 2 I : Nj ¼ 1g. h

Theorem 6. MKP is NP-complete.

Proof. Claim 1 says that X3C can be reduced to MKP inpolynomial time and Claim 5 says that MKP has solutionif and only if X3C has solution. Therefore, if MKP can besolved in polynomial time, then X3C also could be solved.Since X3C is known to be NP-complete [42], MKP is alsoNP-complete. h

References

[1] Performance and functionality of existing MPEG-4 technology in thecontext of CFI on scalable speech and audio coding, ISO/IEC JTC 1,Call for Information (CfI) SC 29, WG 11/M11657, 2005.

[2] Terms of reference for embedded VBR (EV) speech codecstandardization (draft v4.3 + revisions of 2nd of august 2005),annex Q09.A in’TD 167 R1 (PLEN/16), 2005.

[3] Advanced video coding for generic audiovisual services, ITU-T andISO/IEC JTC 1, ITU-T Recommendation H.264 and ISO/IEC 14496-10(MPEG-4 AVC), version 3, 55-PLEN, 2005.

[4] H. Schwarz, D. Marpe, T. Schierl, T. Wiegand, Combined scalabilitysupport for the scalable extension of H.264/AVC, in: Proceedings ofthe IEEE International Conference on Multimedia and Expo,Amsterdam, The Netherlands, 2005, p. 4.

[5] H. Schwarz, D. Marpe, T. Wiegand, Basic concepts for supportingspatial and SNR scalability in the scalable H.264/MPEG4-AVCextension, in: Proceedings of the IEEE International Workshop onSystems, Signals and Image Processing, Chalkida, Greece, 2005.

[6] A. Takács, A. Kovács, F. Kalleitner, H. Brand, Forward information – ageneral approach for scalable audiovisual service delivery, in:Proceedings of the ISWCS 2005, Siena, Italy, 2005, pp. 158–162.

[7] H. Yousefi’zadeh, H. Jafarkhani, A. Habibi, Layered media multicastcontrol (LMMC): rate allocation and partitioning, IEEE/ACMTransactions on Networking 13 (3) (2005) 540–553.

[8] The IEEE 802.11b-1999 standard, IEEE, 802.11a, 1999. <http://standards.ieee.org/getieee802/download/802.11b-1999.pdf>.

[9] The IEEE 802.11g-2003 standard, IEEE, 802.11g, 2003. <http://standards.ieee.org/getieee802/download/802.11g-2003.pdf>.

[10] S. McCanne, V. Jacobson, Receiver-driven layered multicast, in: ACMSIGCOMM, New York, 1996, pp. 117–130.

[11] L. Vicisano, J. Crowcroft, L. Rizzo, TCP-like congestion control forlayered multicast data transfer, in: IEEE INFOCOM (3), Stanford, CA,1998, pp. 996–1003.

[12] J. Byers et al, FLID-DL: congestion control for layered multicast, IEEEJournal on Selected Areas in Communications 20 (8) (2002) 1558–1570.

[13] D. Wu, T. Hou, Y.Q. Zhang, Scalable video coding and transport overbroadband wireless networks, Proceedings of the IEEE Special Issueon Multi-Dimensional Broadband Wireless Technologies andApplications 89 (1) (2001) 6–20.

[14] I. Haratcherev et al, Automatic IEEE 802.11 rate control forstreaming applications, Wireless Communications and MobileComputing (5) (2005) 421–437.

[15] W. Xiuchao, A.L. Ananda, Link characteristics estimation for IEEE802.11 DCF based WLAN, in: IEEE LCN 2004, 2004.

[16] L.C.J. Zhang, I. Marsic, Models for non-intrusive estimation ofwireless link bandwidth, in: Personal Wireless CommunicationConference, Venice, Italy, 2003.

[17] Y. Tay, K. Chua, A capacity analysis for the IEEE 802.11 MAC protocol,Wireless Networks 7 (2001) 159–171.

[18] G. Bianchi, IEEE 80211 – saturation throughput analysis, IEEECommunications Letters 2 (12) (1998) 318–320.

[19] I. Haratcherev, J. Taal, K. Langendoen, R. Lagendijk, H. Sips,Optimized video streaming over 802.11 by cross-layer signaling,IEEE Communications Magazine 44 (1) (2006) 115–121.

[20] M.U. Demircin, P. van Beek, Bandwidth estimation and robust videostreaming over 802.11e wireless LANs, in: Proceedings of the IEEE

1072 Á. Kovács, I. Gódor / Computer Networks 53 (2009) 1062–1072

International Conference Multimedia and Expo (ICME), Amsterdam,The Netherlands, 2005, pp. 1250–1253.

[21] R. Prasad, M. Murray, C. Dovrolis, K. Claffy, Bandwidth estimation:metrics, measurement techniques, and tools, IEEE Network 17 (6)(2003) 27–35.

[22] B.E. Braswell, J.C. McEachen, Modeling data rate agility in the IEEE802.11a WLAN protocol, in: Proceedings of the OPNETWORK,Washington, DC, August, 2001, pp. 118–133.

[23] K. Balachandran, S.R. Kadaba, N. S, Channel quality estimation andrate adaptation for cellular mobile radio, IEEE J-SAC 17 (7) (1999)1244–1256.

[24] I. Haratcherev et al., Hybrid rate control for IEEE 802.11, in:Proceedings of the ACM International Workshop MobilityManagement and Wireless Access Protocols, Philadelphia, PA,2004, pp. 10–18.

[25] J. del Prado Pavon, S. Choi, Link adaptation strategy for IEEE 802.11WLAN via received signal strength measurement, in: Proceedings ofthe IEEE ICC’03, vol. 2, Anchorage, AK, 2003, pp. 1108–1113.

[26] Z. He, J. Cai, C.W. Chen, Joint source channel rate-distortion analysisfor adaptive mode selection and rate control in wireless videocoding, IEEE Transactions Circuits and Systems for Video Technology12 (6) (2002) 511–523.

[27] R. Chandramouli, K. Subbalakshmi, N. Ranganathan, Stochasticchannel-adaptive rate control for wireless video transmission,Pattern Recognition Letters 25 (7) (2004) 793–806. Special issueon Video Objects:Representation, Creation, Coding, Transmission,Manipulation and Retrieval.

[28] A. Kamerman, L. Monteban, WaveLAN-II: a high-performancewireless LAN for the unlicensed band, Bell Lab Technical Journal(1997) 118–133.

[29] M. Lacage et al., IEEE 802.11 rate adaptation: a practical approach,in: ACM MSWiM 2004, 2004, pp. 161–175.

[30] G. Holland et al., A rate-adaptive MAC protocol for multi-hopwireless networks, in: ACM MOBICOM 2001, 2001.

[31] D. Qiao et al., An optimal low-energy transmission strategy for IEEE802.11a/h, in: ACM MOBICOM 2003, 2003, pp. 161–175.

[32] A. Majumdar et al., Multicast and unicast real-time video streamingover wireless LANs, IEEE Transactions on Circuits and Systems forVideo Technology 12 (6) (2002) 524–534.

[33] I. Kozintsev, J. McVeigh, Improving last-hop multicast streamingvideo over 802.11, in: BROADNETS 2004, 2004.

[34] Y. Zheng, K. Lu, D. Wu, Y. Fang, Performance analysis of IEEE 802.11DCF in binary symmetric channels, in: Proceedings of the of IEEEGlobecom 2005, St. Louis, MO, USA, 2005, pp. 3144– 3148.

[35] J. Zhaoy, Q. Zhang, Z. Guo, W. Zhu, Throughput and QoS optimizationin IEEE 802.11 WLAN, in: 3GWireless2002+WAS2002, San Fransisco,USA, 2002.

[36] Wireless LAN medium access control (MAC) enhancements forquality of service (QoS), IEEE, 802.11e draft 8.0, 2004.

[37] A. Ksentini, M. Naimi, Toward an improvement of H.264 videotransmission over IEEE 802.11e through a cross-layer architecture,IEEE Communications Magazine 44 (1) (2006) 107–114.

[38] W. Fenner, Internet group management protocol, version 2, RFC2236, November 1997.

[39] The IEEE 802.1q-2003 standard – Virtual LANs, IEEE, 802.1Q (2003).<http://standards.ieee.org/getieee802/download/802.1Q-2003.pdf>.

[40] T.M. Cover, J.A. Thomas, Elements of Information Theory, John Wiley& Sons Inc, 1991.

[41] M. Dai, D. Loguinov, H. Radha, Statistical analysis and distortionmodeling of MPEG-4 FGS, in: Proceedings of the IEEE InternationalConference on Image Processing (ICIP 2003), Barcelona, Spain, 2003,pp. 301–304.

[42] L. Rónyai, G. Ivanyos, R. Szabó, Algoritmusok, Typotex, 1999.

Ákos Kovács received M.Sc. degree in com-puter science at the Budapest University ofTechnology and Economics (BUTE) in 2005.Since then he has been with the Departmentof Telecommunications and Media Informat-ics, where he is currently pursuing his Ph.D.degree, and a part-time research fellow atTrafficLab, Ericsson TelecommunicationsHungary.

István Gódor received both his M.Sc. andPh.D. degree in electrical engineering fromBudapest University of Technology and Eco-nomics, Budapest, Hungary in 2000 and 2005,respectively. He is a research fellow at Erics-son Research, Traffic Analysis and NetworkPerformance Laboratory of Ericsson Hungary.His research interests include network design,combinatorial optimization and cross-layeroptimization. He is a member of IEEE and theScientific Association for Infocommunicationsof Hungary.