Effcent Bandwdth Management n Broadband Wreless Access Systems Usng CAC-based Dynamc Prcng Bader Al-Manthar, Ndal Nasser 2, Najah Abu Al 3, Hossam Hassanen Telecommuncatons Research Laboratory School of Computng, Queen s Unversty ngston, ON, Canada 7L 3N6 {manthar, hossam}@cs.queensu.ca 2 Department of Computng & Informaton Scence Unversty of Guelph Guelph, ON, Canada NG 2W nasser@cs.uoguelph.ca 3 College of Informaton Technology UAE Unversty, Al-An P.O. 7555, UAE najah@uaeu.ac.ae Abstract Whle the demand for moble broadband wreless servces contnues to ncrease, rado resources reman scarce. Even wth the substantal ncrease n the supported bandwdth n next generaton Broadband Wreless Access Systems (BWASs), t s expected that these systems wll severely suffer from congeston due to the rapd ncrease n demand of bandwdth ntensve applcatons. Wthout effcent bandwdth management and congeston control schemes, network operators may not be able to meet the ncreasng demand of users for multmeda servces, and hence they may suffer mmense amount of revenue loss. In ths paper, we propose an admsson-level bandwdth management scheme consstng of Call Admsson Control (CAC) and dynamc prcng. The man am of our proposed scheme s to provde monetary ncentves to users to use the wreless resources effcently and ratonally, hence, allowng effcent bandwdth management at the admsson level. By dynamcally determnng the prces of unts of bandwdth, the proposed scheme can guarantee that the arrval rates to the system are less than or equal to the optmal ones computed dynamcally, hence, guaranteeng a congeston-free system. Smulaton results show the effectveness and strengths of our proposed approach. eywords call admsson control, congeston, dynamc prcng. I. INTRODUCTION The success of emergng next generaton Broadband Wreless Access Systems (BWASs) such as 3.5G Hgh Speed Downlnk Packet Access (HSDPA) [] and WMAX [2] wll depend, among other factors, on ther ablty to manage ther shared wreless resources n the most effcent way. Ths s a complex task due to the expected ncrease n demand for multmeda servces that have dverse and very hgh bandwdth requrements. Therefore, bandwdth management s crucal for the success of such communcaton systems. To support as many users as possble whle satsfyng ther bandwdth requrements, network operators typcally employ Call Admsson Control (CAC), whch s an admsson-level bandwdth management strategy. By lmtng the number of admtted users calls n the system, CAC can guarantee that the packet-level Qualty of Servce (QoS) (e.g., packet delay, average throughput, etc) of ongong calls wll not get degraded as a result of new ncomng ones. CAC s very effcent n mprovng the packet-level QoS of ongong calls especally durng congeston perods. However, t may not be as effcent n mprovng the admsson-level QoS (e.g., call blockng probabltes). Ths s because CAC, by tself cannot avod congeston due to the fact that t does not provde ncentves to the users to use the shared wreless system resources ratonally and effcently. Therefore, the call blockng probabltes can reach hgh levels durng congeston perods. To overcome ths problem, there has been some research work recently on ntegratng admsson-level dynamc prcng wth CAC n order to control call request arrvals to the system through monetary ncentves [3], [4] and [5]. Hence, mantanng the call-level QoS at the desred thresholds. In admsson-level dynamc prcng, the prce for a unt of tme or bandwdth s determned when the user ntates a call request before she s admtted to the system. The prce n ths case s fxed for the call duraton. Ths prce s dynamcally determned accordng to the network load. Dynamc prcng can competently promote ratonal and effcent use of the shared wreless resources by nfluencng the users behavors. Ths s because t can dscourage prce-senstve users from over usng the wreless network when t s congested and encourage them on the other hand to ncrease ther demand when the network s underutlzed. Dynamc prcng s, therefore, a promsng soluton to traffc control problems, whch can help allevate the problem of congeston and provde effcent bandwdth management. In addton, t can ensure economc effcency snce t ensures that the wreless resources are gven to those who value them the most. Furthermore, dynamc prcng s cost-effectve and t can generate hgher revenues. Several CAC schemes wth dynamc prcng have been proposed n the lterature [3], [4] and [5]. The scheme n [3] dynamcally computes the optmal prce so that the prceaffected call arrval rates mze the summaton of users Ths research s supported by the government of the Sultanate of Oman, Natural Scences and Engneerng Research Councl of Canada and by a grant from Bell Unversty Labs. 978--4244-243-9/08/$25.00 2008 IEEE 484
utltes (.e., socal welfare of the system), where the user utlty s assumed to be a functon of the call blockng probabltes, whch are n turn a functon of the arrval rates. However, the scheme s desgned to prevent network congeston only, where a flat rate prcng s assumed when the network s underutlzed. Therefore, users are not gven any ncentves to ncrease ther usage of the network when t s underutlzed, whch results n resource wastage, and hence potental revenue loss. In addton, the scheme lacks the support of comprehensve QoS snce t assumes that all calls requre the same amount of resources, whch makes t unsutable for next generaton BWASs. In [4], a CAC-based dynamc prcng scheme s proposed. In ths scheme users are dvded nto two types, prorty users and conventonal users. When the network s underutlzed, all users n ths scheme are consdered conventonal users and are placed n the conventonal queue awatng admsson, where they are charged a flat rate. Durng congeston perods, a dynamc prce s computed and the users are gven the opton to choose between beng prorty users, where they are charged a hgher dynamc prce and are placed n the prorty queue to be served faster or beng conventonal users, where they are charged a flat rate and they are served slower. The dynamc prce s determned so that the mum number of users that the network can accommodate, and yet conform to the delay the users can spend n the queue awatng admsson [4]. However, even though the scheme consders the delay the users experence n the admsson queues, t does not take nto account the new call blockng and handoff call droppng probabltes. Ths may not be practcal snce wreless network operators have a lmt on the number of calls they can block, whch s usually determned by regulatons. In addton, smlar to the scheme n [3], ths scheme assumes that calls requre the same amount of resources. Hence, t may not be unsutable for Next generaton BWASs. Moreover, the scheme s desgned to prevent congeston only. Therefore, t does not provde ncentves to users to ncrease ther demand for the network servces when the network s underutlzed. The CAC-based dynamc prcng scheme n [5] ams at reducng congeston and mzng revenues n wreless cellular networks. The scheme consders the effects of prces on call arrvals, retrals (.e., requestng the same servce agan after beng blocked) and substtutons among servces (.e., substtutng a servce for another after beng blocked). Usng some assumptons about the new and handoff call arrval rates, the scheme dynamcally determnes the prces of network servces as to encourage or dscourage the arrval rates to the system n order to reserve some bandwdth for arrvng handoff or hgher revenue-generatng users. Even though the scheme consders dfferent QoS classes, t assumes that users wthn each class request the same amount of bandwdth. Ths s stll not practcal n next generaton BWASs snce n these systems, each class can have a number of servces each requestng a certan amount of bandwdth (e.g., audo streamng and vdeo streamng n the streamng class). In addton, the scheme s complex and requres many calculatons to determne the prce. Moreover, all the schemes n [3], [4] and [5] are based on certan assumptons about users demand models and cannot, therefore, be generalzed to work wth dfferent demand models wthout affectng the way prces are computed. Ths lmts ther scalablty, snce dfferent network operators mght have dfferent demand models dependng on ther subscrbers. The schemes n [6], [7], [8], [9] and [0] apply dynamc prcng at admsson level wthout usng CAC. These schemes, therefore, cannot acheve optmzed call-level QoS. Therefore, there s a need for a CAC-based dynamc prcng scheme that s able to support dfferent QoS classes wth dfferent users havng dfferent bandwdth requrements, work wth varous demand models and compute the dynamc prces n a smple way In ths paper, we propose an admsson-level bandwdth management scheme that conssts of a CAC component and a dynamc prcng component. The proposed scheme ams at effcently managng the bandwdth of BWASs n order to smultaneously satsfy the bandwdth requrements of users, mze the utlzaton of BWASs and prevent congeston. By dynamcally computng the prces of unts of bandwdth, our scheme s able to force the arrval rates of call requests to the system towards the optmal ones as determned by the CAC component. The proposed scheme mproves the scheme n [3] n the followng aspects. Frst, unlke the scheme n [3], our proposed scheme employs dynamc prcng durng all network condtons (.e., whether the network s underutlzed or congested). Ths way, our scheme s able to mze the utlzaton of BWASs when these networks are underutlzed as well as preventng congeston when they are over utlzed. Second, our scheme supports multple QoS classes wth calls havng multple bandwdth requrements, whch makes t more sutable for BWASs. Fnally, congeston prces n the scheme n [3] are computed based on certan assumptons about the users utltes not based on the amount of avalable network bandwdth. The scheme, therefore, s ncapable of capturng the dynamcs of the network (.e., changes n avalable bandwdth) and the varyng bandwdth requrements of users. Ths explans the nablty of the scheme to acheve zero call blockng probabltes despte of assumng accurate users demand model [3]. On the contrary, congeston prces n our scheme are computed dynamcally based on the amount of avalable bandwdth n the network. As a result, our scheme s shown to acheve zero call blockng probabltes n case accurate users demand model s assumed. The rest of ths paper s organzed as follows. Secton II provdes an overvew of the proposed scheme. Secton III presents a descrpton of the proposed scheme. Performance results are presented n Secton IV. Fnally, conclusons and future work are dscussed n Secton V. 485
II. OVERVIEW OF BANDWIDTH MANAGEMENT SCHEME In ths secton, we provde an overvew of our proposed bandwdth management scheme. We assume that there are QoS classes, where class has hgher prorty than class +, and +. We consder that class calls request b unts of bandwdth. Our scheme conssts of two components namely, the CAC component and the dynamc prcng component. Our scheme works as follows. The CAC component contnuously montors the amount of avalable bandwdth (.e., unutlzed bandwdth). When the amount of avalable bandwdth changes due to call completon or new admtted calls, the CAC component then computes the optmal arrval rate for each QoS class n order to mze the utlzaton of the new avalable bandwdth n the system and acheve certan farness levels between QoS classes. The actual arrval rates for the QoS classes are, however, dfferent from the optmal rates determned by the CAC component. In ths case, the dynamc prcng component dynamcally determnes the prces of unts of bandwdth for each class based on the users demands n order to force the actual arrval rates to be less than or equal to the optmal ones. The dynamc prces are computed ndependently from the optmal arrval rates, hence, smplfyng the mplementaton of our scheme and provdng the network operators the flexblty to use any CAC and users demands functons wthout affectng the computaton of prces. It should be noted that n ths paper, we focus on new calls only. Ths s because handoff calls are not affected by dynamc prcng when chargng at admsson level, snce they have been already charged at the cell where the calls have been ntated. Therefore, the network operator cannot nfluence the behavor of handoff users by changng the prce. In ths case, the network operator can use a form of Guard Channel schemes n whch a certan amount of bandwdth s exclusvely reserved for handoff calls n order to mantan the handoff call droppng blockng probablty below a certan threshold []. III. DESCRIPTION OF THE BANDWIDTH MANAGEMENT SCHEME COMPONENTS In ths secton, we descrbe each component of our proposed bandwdth management scheme. We make the followng defntons. Let: N : number of admtted calls n class. N = N : total number of admtted calls. = B: total bandwdth of the system B free ( t ) : total avalable bandwdth at tme t. B free ( t ) can be computed as follows: N B = B b () free j = j= where bj ( t ) s the bandwdth assgned to user j n class at tme t. λ : arrval rate of new calls to class at tme t. Therefore, the total demand requested by class at tme t s equal to λ b, where b s the bandwdth request of class. λ : mum total arrval rate to the system (.e., of all QoS classes). λ can be easly computed from observed hstorcal data of the network operator or t could be set to the total number of subscrbers. p ( t ) : prce n terms of unts of money per unt of bandwdth for class servces. A : percentage of users who have suffcent Wllngness to Pay (WTP) to make call requests to class. Clearly, A s a A = f p [0,], where functon of the prce (.e., ( ) p f ( A ) = ). A can be constructed from the system s hstory by observng the users responses to changes n the prce. It s reasonable to assume that A s monotoncally decreasng functon of the prce. That s, when the prce ncreases, A ether remans the same or t decreases. It should be noted that the computaton of A s a pure economc topc that s outsde the scope of ths paper. However, we utlze a well-known demand functon n secton IV to model A, although our scheme can work wth any functon for A as explaned next. The man objectve of our CAC component s to fnd the optmal arrval rate for each QoS class such that the utlzaton of avalable bandwdth s mzed. To acheve ths objectve, the CAC component wll solve the followng optmzaton problem every tme t detects a change n the avalable bandwdth: Objectve: λ b { λ, λ2,..., λ } = Subject to: λ b Bfree, = = N λ λ, and bj + λ b / B υ,, (2) j= where the frst constrant ensures that the demand of all QoS classes does not exceed the total avalable bandwdth (.e., supply). The second constrant ensures that the resultng total The monetary value users are wllng to pay for a certan servce. 486
arrval rate to the system (.e., = λ ) s realstc and does not exceed the mum arrval rate, whch s lmted by the number of subscrbers. The last constrant s used to ensure farness among QoS classes by restrctng that the share of bandwdth for each class (.e., the bandwdth of admtted calls + bandwdth of new calls) does not exceed a predefned value ( υ ) determned by the network operator. For example, to acheve absolute farness (.e., an equal bandwdth share) between QoS classes, υ should be set to /. Besdes ensurng farness, the second constrant can be used to promote certan servces or ncrease revenues by assgnng more bandwdth to QoS classes that are expected to yeld hgher revenues (e.g., hgher prorty classes). It should be noted that the objectve functon and the constrants n (2) do not nclude the call blockng probabltes. Ths s because our prcng component, as descrbed below, can guarantee to force the actual arrval rates to be less than or equal to the optmal ones computed n (2). Hence, the system s guaranteed to be congeston-free. In addton, the objectve functon and the constrants n (2) are lnear. Hence, the optmal arrval rates { λ, λ2,..., λ } can be found usng Lnear Programmng (LP) technques. The actual arrval rates to the system are, however, dfferent from the optmal arrval rates (.e., { λ, λ2,..., λ } { λ, λ2,..., λ }. Therefore, the dynamc prcng component wll adjust the prces of unts of bandwdth for each QoS class such that the actual arrval rates are less than or equal the optmal ones computed n (2) (.e., { λ, λ2,..., λ } { λ, λ2,..., λ } as follows. We know from the arrval rate to class at tme t (.e., λ ) that t consttutes the followng rato of the total users that could request the servce λ λ From (3) we know that the rato of users that have suffcent λ ( ) WTP to make call requests to class s at least t (there λ could be other users who have suffcent WTP, but choose not to make such requests at ths tme). However, the optmal rato should equal to λ λ λ λ Therefore, when, the prce of class servces λ λ wll be adjusted so that (3) (4) λ A = f ( p ) =,, (5) λ There are two cases, and hence two mplcatons of prce λ λ settng. The frst case s when >, accordng to (5), λ λ the prce should be ncreased so that λ λ λ A = =. In ths case, f A s accurate n λ λ λ modelng the users WTP, then the rato of ncomng users who have suffcent WTP to make call requests s guaranteed to equal the optmal rato. The second case s when λ λ <, the prce should be lowered such that λ λ λ λ λ A =. In ths case, the prce s lowered λ λ λ so that enough users have suffcent WTP to make call requests. It should be noted that users wth suffcent WTP may not make call requests at the tme dependng on ther preferences. Usng our scheme they are, however, encouraged to make such calls due to lower prces. In ths case, the ncomng arrval rate s guaranteed to be less than or equal to the optmal rato. Based on the above dscusson and from (5), the dynamc prcng component wll set the new prces as follows λ p(t) = f,, λ As t can be seen, the prce equaton n (6) s very smple to compute and s ndependent from the objectve functon n (2). Such ndependence allows the network operator to use any objectve functon n (2) wthout affectng the computatons of prces and vce versa. In addton, based on the aforementoned dscusson, the actual arrval rates are guaranteed to be less than or equal to the optmal ones computed n (2). Hence, usng our prcng scheme, the system s guaranteed to be congeston-free. IV. PERFORMANCE EVALUATION In ths secton, we evaluate the performance of our proposed scheme by means of call-level dynamc dscrete event smulaton. We test our scheme on Hgh-Speed Downlnk Packet Access (HSDPA) system. HSDPA s a 3.5G wreless system that has been ntroduced by the 3 rd Generaton Partnershp Project (3GPP) as an extenson to the 3G cellular (6) 487
system Unversal Moble Telecommuncaton System (UMTS). More nformaton on HSDPA can be found n []. A. Smulaton Model For smplcty, we smulate a sngle-cell scenaro. We gnore handoff call, snce they are not affected by dynamc prcng as aforementoned. The base staton s located at the center of the cell. Therefore, only one base staton s nvolved n allocatng the rado resources. The cell radus s m and the base staton s transmsson power s 38 dbm. Two QoS classes are consdered (.e. =2). For demonstraton purposes, we let class and class 2 calls request a bandwdth of 28 bps and 64 bps, respectvely. In addton, we set υ n (2) to / n order to acheve an equal share of bandwdth between the two classes. Actual arrval rates to the system are normally tme vared, and therefore, we adopt a 24 hour model for the arrval rates. In ths model, the day s dvded nto 24 hours startng at mdnght, where dfferent arrval rates are assgned to dfferent hours of the day based on observaton of the call arrvals n a typcal busness day [2], [3]. It s observed n [3] that the peak hours (mum call arrvals) occur around :00 AM and 6:00 PM. In our smulaton, each hour of the day s smulated by 400 s and the performance results are collected at end of the each smulated hour. Call arrvals are modeled by a Posson process where the mean total arrval rates to the system for each hour of the day are shown n Fgure. The total arrval rate to the system s equally dvded between the two classes. The arrval rates n Fgure consttute the actual arrval rates before dynamc prcng s mplemented. When dynamc prcng s mplemented, the actual arrval rates wll depend on the prces. In ths case, durng congeston perods, our prcng component guarantees that the actual arrval rate wll match the optmal one as dscussed n Secton III. On the other hand, when the network s underutlzed, whch occurs n early mornng hours (00:00-05:00 AM) and at nght (2:00-24:00 PM), our prcng component guarantees to provde ncentves to users to use the network servces whle smultaneously preventng congeston. However, as dscussed n Secton III, not every user who has a suffcent WTP to make a call at a certan tme s wllng to make such a call at that tme. In ths case, the arrval rate to the system may stay at ts low level or t may ncrease up to the optmal one dependng on the preferences of users. To evaluate such a case when the network s underutlzed, we test our proposed scheme wth 0% ncrease n the arrval rate (.e., the actual arrval rate stays at ts low value and does not ncrease as a result of lower prces) and wth a 0%, 30% and 50% ncrease of the optmal arrval rate, respectvely (.e., 0%, 30% and 50% of the users who have suffcent WTP to make call requests as a result of reducng the prces wll make such calls, respectvely). The call duraton of each call s modeled by an exponental dstrbuton wth a mean value of 30 s. Users are unformly dstrbuted n the cell. Pedestran A (Ped A) envronment s used n our smulaton, whch s recommended by 3GPP. Moble users n Ped A envronment move at a fxed speed of 3 m/hr. We adopt the same channel model as n [4]. The smulaton tme step s one tme frame, whch s 2 ms n HSDPA []. Other smulaton parameters are lsted n Table I. Arrval Rate (Connecton/Sec) 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0. 0 0 3 6 9 2 5 8 2 24 Hour of the Day Fgure. Arrval Rates n a Typcal Busness Day TABLE I SIMULATION PARAMETERS Smulaton tme per hour 400s Base Staton Transmsson power 38 dbm Antenna gan 7 db Base Staton buffer sze 30 MB Shadowng Lognormal dstrbuton Intra-cell nterference 30 dbm Inter-cell nterference -70 dbm B. Demand Model As aforementoned, our prcng scheme s general and can work wth any demand model. To test our scheme, however, we utlze the followng well-known demand model [6], [5]: ( ) c p A = f p = a e (7) where a ( t ) s the demand shft constant for class users at tme t and c ( t ) s the prce elastcty of demand (.e., the change n demand for a certan product or servce due to a change n ts prce). The reason for usng ths partcular demand model s that t can support dfferent QoS classes and dfferent user behavors by consderng ther prce elastcty of demand and ther demand shft constants, whch can assume dfferent values for dfferent tmes of the day. To ensure that A = f ( p ) [0,], we set a ( t ) to. In addton, for demonstraton purposes, we set c ( t ) to and 2 for classes 488
and 2, respectvely. These two values are chosen so that class users are less responsve to prce changes than those of class 2. Ths way, class users are charged hgher than class 2 users because class users have hgher prorty 2. It should be noted that the actual values of c ( t ) should be determned by market studes on real demand behavors for the dfferent users. C. Smulaton Results In ths secton, we compare the performance of our Bandwdth Management Scheme denoted by (BMS+x%, where x %= 0%, 0%, 30%, 50% ncrease n user calls when the network s underutlzed as a result of lower prces as dscussed n the prevous secton) wth a Conventonal CAC scheme denoted by (CCAC). In CCAC, no dynamc prcng s mplemented. Instead, users are charged fxed prces and call requests are always accepted as long as there s enough bandwdth to support them. In ths case, we fx the prces to 0.35 and 0.7 unts of money per unts of bandwdth for classes and 2, respectvely. These two values are chosen so that at least %70 of users have suffcent WTP to make call requests accordng to the demand model n (7). In practce, fxed prces are determned so that the majorty of people have suffcent WTP to make calls, whch s one of the man causes of congeston. To demonstrate the ablty of our scheme to mze the utlzaton of the system whle preventng congeston and ncreasng revenues, t suffces to show: Percentage of bandwdth utlzaton: the percentage of the utlzed bandwdth to the total bandwdth. Call blockng probablty: the probablty that a call s blocked due to nsuffcent bandwdth to meet ts requrements. Percentage of bandwdth share: the percentage of used bandwdth for each class to the total utlzed bandwdth. Ths metrc s used to test our farness measure n (2). Revenue: the amount of money earned durng the day. Fgure 2 shows the percentage of bandwdth utlzaton for our scheme and the CCAC scheme. The fgure shows that our scheme can sgnfcantly ncrease the bandwdth utlzaton of the system as more users (.e., 0%, 30% and 50%) decde to make call requests as a result of lower prces durng off-peak hours. In case users are not effected by lowerng the prces (.e., case wth 0% ncrease), the bandwdth utlzaton of our scheme s the same as CCAC, whch s expected snce our scheme s dstngushed by ts ablty to ncrease the usage of the network when t s underutlzed. We remark, however, that snce most users are prce-senstve, they wll try to make ther calls when the prces are lower. Hence, the case of 0% can rarely occur n practce. Therefore, usng our scheme, the network operator can ncrease the usage of the network when t s underutlzed, hence, ncreasng ts revenues. In addton, our scheme can effcently prevent network congeston, and hence achevng 0% blockng probabltes as shown n Fgure 3. The reason for ths s that our scheme optmally determnes the prces of unts of bandwdth as to encourage enough users to make call requests, hence, ensurng that the system s never congested. Ths fgure confrms the superorty of our scheme compared to CCAC scheme where users are not provded any ncentves to regulate ther usage of the network. Such scheme can result n very hgh blockng probabltes durng peak hours leadng to user unsatsfacton and potental revenue loss. For nstance, Fgure 3 shows that at peak hours (e.g., AM), the blockng probablty of CCAC can reach up to.4%. Table II shows the percentage of bandwdth share for each class. As aforementoned, we set υ to / n our objectve functon so that each class gets an equal share of bandwdth. The table shows that our scheme acheves better bandwdth share than CCAC. The reason for the unfar bandwdth share n CCAC s that accordng to our traffc model, the actual arrval rate (before dynamc prcng s mplemented) s equally dvded between the two classes and snce class users request double the amount of bandwdth compared to class 2 users, ths results n hgher bandwdth share for class. Our scheme, on the other hand, determnes the dynamc prces of unts of bandwdth so that the arrval rates for the QoS classes acheve the mum possble bandwdth utlzaton whle mantanng a certan farness level (.e., absolute farness n ths case). Hence, our scheme acheves better farness as shown n Table II. An nterestng result revealed from Table II s that even though we set υ to / n our scheme, the bandwdth share of class s stll hgher than that of class 2. Ths s due manly to the hgh bandwdth requests of class users n off-peak hours where users are not affected by our dynamc prces. Ths s clearly shown n Fgure 4, whch depcts the bandwdth share of class n each hour of the day. In peak hour perods, users are more affected by dynamc prces, and hence ther call requests arrval rates equal the optmal computed ones, hence, achevng / bandwdth share. In fact, as more users decde to make call requests as a result of lowerng the prces, the bandwdth share of both classes approaches / because as aforementoned, prces are desgned to acheve an equal share of bandwdth n our experments. Ths explans the ncreased farness of our scheme n Fgure 4 as more users tend to make call requests durng off-peak hours. Table III shows the total revenue collected throughout the day for our scheme and CCAC. Our scheme clearly outperforms CCAC n terms of revenues due to chargng users hgher prces durng peak hours. In addton, as more users decde to make call requests, more revenues can be collected. The revenue collected from class users s hgher than that from class 2 users because the formers pay hgher prces for class servces n addton to requestng double the amount of bandwdth. It should be noted that more revenue can be earned f more bandwdth s assgned to class (.e., f class s assgned more than / bandwdth share). Therefore, the 489
farness constrant n (2) can be used also to ncrease revenues by assgnng more bandwdth to classes that are expected to yeld hgher revenues. V. CONCLUSIONS AND FUTURE WOR Wreless network servces have encountered an enormous demand n the past few years. Ths trend s expected to contnue n the future due to the emergence of new broadband wreless technologes that are able to support hgh data rates, hence, offerng a wde range of multmeda servces. Due to scarcty of rado resources, wreless bandwdth must be managed n a way that mzes the effcency of the wreless network and meet the requrements of both network operators and users. In ths paper, a novel admsson-level bandwdth management scheme s proposed for next generaton broadband wreless access systems. The proposed scheme conssts of two components namely, the Call Admsson Control (CAC) component and the dynamc prcng component. Our scheme ams at provdng monetary ncentves to users to use the wreless bandwdth effcently and ratonally. By dynamcally computng the prces of wreless servces accordng to the network load, our scheme s able to prevent congeston whle ncreasng the utlzaton of the network. Our scheme s smple to compute and can work wth any CAC and dynamc prcng functons due to the separaton of the CAC functon and dynamc prce computaton. Dynamc prcng, however, can guarantee to prevent congeston only f the users demand models are accurate. Therefore, we are currently nvestgatng the effect of naccurate demand models on the system performance and how such naccuracy can be effcently dealt wth. REFERENCES [] 3GPP TS 25.308, Hgh Speed Downlnk Packet Access (HSDPA); Overall Descrpton, Release 5, March 2003. [2] IEEE 802.6 Workng Group, IEEE 802.6-2005e Standard for Local and Metropoltan Area Networks: Ar nterface for fxed broadband wreless access systems amendment for physcal and medum access control layers for combned fxed and moble operaton n lcensed bands, December 2005. [3] J. Hou, J. Yang and S. Papavassllou, Integraton of Prcng wth Call Admsson Control to Meet QoS Requrements n Cellular Networks, IEEE Transactons on Parallel and Dstrbuted Systems, vol. 3, no. 9, pp. 898-90, September 2002. [4] S. Yaparoj and F.C. Harmantzs, Congeston Prcng wth Alternatves for Moble Networks, Proceedngs of the IEEE Wreless Communcatons and Networkng Conference (WCNC), Atlanta, U.S.A, vol. 4, pp. 67-676, March 2004. [5] S.L. Hew and L. B. Whte, Optmal Integrated Call Admsson Control and Congeston Prcng wth Handoffs and Prce- Affected Arrvals, Proceedngs of the Asan-Pacfc Conference on Communcatons (APCC), Perth, Australa, pp. 396-400. October 2005. [6] S. Mandal, D. Saha and A. Mahant, A Technque to Support Congeston Prcng Strategy for Dfferentated Cellular Moble Servces, Proceedngs of the IEEE Global Telecommuncatons Conference (GLOBECOM), St. Lous, U.S.A, vol. 6, pp. 3388-3392, December 2005. [7] S. Yaparoj and F.C. Harmantzs, Aucton-based Congeston Prcng for Wreless Data Servces, Proceedngs of the IEEE Internatonal Conference on Communcatons (ICC), Istanbul, Turkey, pp. 059-064, June 2006. [8] S. Mandal, D. Saha and M. Chatterjee, Prcng Wreless Network Servces Usng Smart Market Models, Proceedngs of the IEEE Consumer Communcatons and Networkng Conference (CCNC), Las Vegas, U.S.A., vol., pp. 574-578, January 2006. [9] S. Mandal, D. Saha and M. Chatterjee, Dynamc Prce Dscoverng Models for Dfferentated Wreless Servces, Journal of Communcatons, vol., no. 5, pp. 50-56, August 2006. [0] E. Vterbo and C.F. Chassern, Dynamc Prcng for Connecton-Orented Servces n Wreless Networks, Proceedngs of the IEEE Internatonal Symposum on Personal, Indoor and Moble Rado Communcatons (PIMRC), San Dego, U.S.A, vol., pp. 68-72, September 200. [] D. Hong and S.S. Rappaport, Traffc Model and Performance Analyss for Cellular Moble Rado Telephone Systems wth Prortzed and None-prortzed Handoff Procedures, IEEE Transactons of Vehcular Technology, vol. 35, no. 3, pp. 77-92, August 986. [2] D. Lu and Y. Zhang, A Self-Learnng Adaptve Crtc Approach for Call Admsson Control n Wreless Cellular Networks, Proceedngs of the IEEE Internatonal Conference on Communcatons (ICC), Anchorage, U.S.A., vol. 3, pp.853-857, May 2003. [3] R. L. Freeman, Telecommuncaton System Engneerng, 3rd edton, Wley, 996. [4] B. Al-Manthar, N. Nasser and H. Hassanen, Packet Schedulng n 3.5G Hgh-Speed Downlnk Packet Access Networks: Breadth and Depth, IEEE Network Magazne, vol. 2, no., pp. 4-46, January 2007. [5] E. D. Ftkov-Norrs, A hanfar, Congeston prcng n Cellular Networks, A Moblty Model wth a Provder-Orented Approach, Proceedngs of the IEEE Internatonal Conference on 3G Moble Communcaton Technologes, London, U, pp. 63-67, March 200. TABLE II PERCENTAGE OF BANDWIDTH SHARE Scheme Class Class 2 BMS+0% 59.956% 40.044% BMS+0% 57.04% 42.896% BMS+30% 53.420% 46.580% BMS+50% 5.536% 48.464% CCAC 69.344% 30.656% 490
70 60 50 TABLE III TOTAL EARNED REVENUE DURING THE DAY (UNITS OF MONEY) Scheme Total Revenue Class Class 2 BMS+0% 542 x0 4 353 x0 4 89 x0 4 BMS+0% 562 x0 4 4 x0 4 5 x0 4 BMS+30% 59 x0 4 387 x0 4 204 x0 4 BMS+50% 627 x0 4 402 x0 4 225 x0 4 CCAC 523x0 4 355 x0 4 68 x0 4 BMS+0% BMS+0% BMS+30% BMS+50% CCAC Capacty Share (%) 75 70 65 60 55 50 45 40 35 30 0 3 6 9 2 5 8 2 24 Hour of the Day BMS+0% BMS+0% BMS+30% BMS+50% CCAC Fgure 4. Percentage of Bandwdth Share for Class at Dfferent Hours of the Day Utlzaton (%) 40 30 20 0 0 0 3 6 9 2 5 8 2 24 Hour Of the Day Fgure 2. Percentage of Bandwdth Utlzaton at Dfferent Hours of the Day 2 0 Blockng Probablty (%) 8 6 4 2 BMS+0% BMS+0% BMS+30% BMS+50% CCAC 0 0 3 6 9 2 5 8 2 24 Hour of the Day Fgure 3. Call Blockng Probablty at Dfferent Hours of the Day 49