AN ADMISSION CONTROL SCHEME FOR PROPORTIONAL DIFFERENTIATED SERVICES ENABLED INTERNET SERVERS USING SUPPORT VECTOR REGRESSION

Transcription

1 AN ADMISSION CONTROL SCHEME FOR PROPORTIONAL DIFFERENTIATED SERVICES ENABLED INTERNET SERVERS USING SUPPORT VECTOR REGRESSION CHENN-JUNG HUANG, YI-TA CHUANG and CHIH-LUN CHENG Insttute of Learnng Technoogy, Natona Huaen Teachers Coege, Huaen, Tawan 97 Abstract: Unpredctabe response tme s a common probem n contemporary web servers. Long response deays substantay cut company revenues owng to the arge number of aborted e-commerce transactons. Ths work presents an admsson contro mode and a traffc scheduer scheme of the web server under a proportona dfferentated servce, embeddng a tme seres predctor to estmate the traffc oad of the cent n the next measurement tme perod. The expermenta resuts ndcate that the proposed modes can effectvey reaze proportona deay dfferentaton servce n mutcass Web servers. Keywords: Proportona dfferentated servce, tme seres predcton, fuzzy ogc systems, sef-smarty, support vector regresson. INTRODUCTION Wth the ncreasngy popuarty of the Internet wordwde, the use of web servers to advertse and se merchandse n busness s sgnfcanty ncreasng. Tradtona web servers provde servce for cent requests usng frst-come-frst-serve (FCFS) servce mode. FCFS ntroduces unpredctabe response tmes for the cents when ncomng traffc s bursty. Customers may become frustrated by a ong response tme and end the network connecton wth the web servers wthout fnshng the transactons wth the enterprses, eadng to oss of revenue for the busnesses and unsatsfactory quaty of servce (QoS). Athough QoS provson n network transmsson, such as Integrated Servces (IntServ) and Dfferentated Servces (DffServ), s currenty of nterest among researchers, network ayer QoS guarantees aone mght not be abe to offer cents perceptbe servces when the servers are overoaded by unexpectedy sgnfcant rses n the number of connectons. Recent works have consdered the ssues of prortzed processng n web servers. Eggert and Hedemann [Eggert and Hedemann, 999] attempted to provde QoS servce at the appcaton eve by spttng cent requests nto two casses as n the network ayer, and restrctng the process poo sze and response transmsson rate for varous prorty groups. Bhatt and Fredrch [Bhatt and Fredrch, 999] deveoped tered servce eves and overoad management mode, and mpemented admsson contro mechansm by bockng ow-prorty tasks when the number of hgh prorty obs exceeded some predefned threshod. Both admsson contro modes foow conventona fxed bandwdth eased ne scheme that satsfes the requrement of bursty workoad. Athough the servce quaty of hgh prorty tasks s guaranteed, some precous bandwdth s st unused under the average oad. Vasou and Lutfyya [Vasou and Lutfyya, ] proposed a QoS archtecture that permts the number of requests for dfferent prorty groups to be dynamcay atered accordng to the performance of the hgh prorty group durng the runtme, the servce quaty of ow prorty tasks s st somewhat degraded. Chen et a. [Chen and Mohapatra, ] proposed so-caed servce dfferentatng Internet servers to manage the QoS servce provded by the network ayer. The servers aso provde sgnfcanty better servces to hgh prorty tasks usng prortzed schedung and task assgnment approaches. However, the starvaton effect of the ow prorty task group s st sgnfcant owng to the monopozaton of the resource aocated to the hgh prorty tasks. To enhance the performance of the ow prorty tasks, Lee et a. [Lee et a, 4] presented an admsson contro agorthm that permts proportona deay dfferentated servce (PDDS) [Dovros et a, ] at appcaton eve. Under PDDS, cent requests are ntay cassfed to dfferent prorty groups as n [Chen and Mohapatra, ], and the servces acheved by casses are proportona to ther ratos set n the servce contracts. Cents can then be charged accordng to ther maxmum average watng tme QoS requrement. Kanoda and Knghty [Kanoda and Knghty, 3] deveoped the atency-targeted mutcass admsson contro agorthm, whch empoys measurements of requests and servce atences to manage each cass QoS. Lee et a and Kanoda & Knghty [Lee et a, 4; Kanoda and Knghty, 3] mght resove the starvaton ssue for the ow prorty tasks as reveaed by Chen & Mohapatra [Chen and Mohapatra, ]. However, compared wth the approach usng HTTP tags and HTML nks n [Rtter et a, ], the agorthm of Lee et a [Lee et a, 4] requres cents to feed two expct parameters, maxmum arrva rate and maxmum average watng tme, nto the server to aunch the admsson contro mechansm. However, ths scheme s mpractca n reaty. Furthermore, Lee et a and Kanoda and Knghty [Lee et a, 4; Kanoda and Knghty, 3] modeed the aggregate request rate from a cents as a Posson process, whch s nconsstent wth the evdence shown n [Artt and Jn, ] that the traffc from Word Wde Web (WWW) transfers has sef-smarty. Thus, the vadty of the performance evauaton reports gven n [Lee et a, 4; Kanoda and Knghty, 3] s dsputed. Ths work presents two admsson contro modes to enabe PDDS at appcaton eve. Each proposed mode predcts the tota maxmum arrva rate and maxmum I. J. of SIMULATION Vo. 6 No and 4 ISSN x onne, prnt

2 average watng tme of each prorty task group for the next measurement perod accordng to the arrva rate of each cass durng the current and the ast three measurement perods. The admsson contro modes can then appy the predcted vaues to determne the next cent for servce from one of the queues mantaned for each prorty task group. Sgnfcanty, the system automatcay derves the two above-mentoned parameters, thus resovng the mpractca probem of specfyng the parameters by the cents as seen n [Lee et a, 4]. Moreover, Artt and Jn [Artt and Jn, ] revea that the WWW traffc possesses the characterstc of sef-smarty, and sef-smar tme seres are predctabe. Meanwhe, the works of [Abraham, 3; Dhyan et a, 3; Bonno et a, 3; Antono et a, ] aso exhbt the effectveness of predctng the number of web server access. Therefore, ths work attempts to utze the support vector regresson (SVR) technque to reaze the predcton mechansm, and compares ths scheme wth another we-known machne earnng technque, fuzzy ogc system, whch s renowned by ts mathematca framework to dea wth rea word mprecson, and permts decson-makng by estmated vaues under ncompete or uncertan nformaton. The SVR s specfed to mpement the proposed modes because of ts superor performance n many appcatons, such as tme seres predcton [Van Geste et a, ], Internet traffc predcton, ca cassfcaton for AT&T s natura daog system, mut-user detecton and sgna recovery for a code dvson mutpe access (CDMA) system [Chen et a, ; Gong and Kuh, 999; Haffner et a, 3; Kuh, ; Hasegawa et at, ]. VLSI chps aso provde many soutons that enabe the SVR to be hardware-computed. Hgh-speed ow cost SVR chps have been ntroduced recenty, makng mpementaton of SVR usng hardware feasbe [Anguta et a, 999]. The rest of ths paper s organzed as foows. Secton ntroduces the proposed admsson contro modes. Secton 3 then descrbes the predcton agorthms adopted n the admsson contro modes, namey the fuzzy ogc system and support vector regresson technques. Secton 4 ntroduces the smuaton resuts, whch compare the proposed agorthms wth FCFS servce mode and wth two representatve tme seres predctors n the terature. Concusons are fnay drawn n Secton 5.. ADAPTIVE ADMISSION CONTROL MODELS WWW traffc has been observed to have sef-smarty [Artt and Jn, ]. Lang [Lang, ] found that WWW traffc s predctabe through sef-smar tme-seres. Therefore, ths work ncorporate a predcton agorthm nto the proposed admsson contro mode to estmate the rato of the expected average watng tme between dfferent casses, and to nvestgate whether the servce contract of each cass s voated. Sef-smarty s brefy defned and characterzed beow.. Sef-Smarty Let = ( X, t =,,,L ) X be a statonary stochastc t process. If the average of the seres X s computed over non-overappng bocks of sze m, an m-aggregated statonary tme seres ( m ) ( ( m X = X ) k, k =,,,L) s obtaned as foows: m X ( m) = km+ Xk =, m N m, () ( m) when the varances and the autocorreatons of X and X satsfy the foowng reaton: ( m ) Var ( ) ( X ) Var X = ( ),.5 < H <, () H m r ( )( ) = rx ( ),, (3) X m where H denotes the Hurst parameter. It s sad that X s exacty sef-smar. In addton, X s asymptotcay sef-smar f the foowng reaton s satsfed: ( m ) Var ( ) ( X ) X ~ ( ),. 5 < H < H Var, (4) m r r, m. (5) ( )( ) ( ) X m X The autocorreatons gven n Eqs. (3) and (5) te that the degree of varabty or burstness s dentca at dfferent tme scaes for sef-smar stochastc process, and the autocorreaton does not drop to zero as m. Ths s n contrast to the characterstc of the stochastc processes used n typca data modes: r X ( m )( ), m. (6) As for the varances gven n Eqs. () and (4), they decrease more sowy than when m. m As the study n [Artt and Jn, ] showed that the sef-smar traffc pattern generated by Web browsers fts very we to a Pareto-type dstrbuton, our smuaton mode w thus assume the packet nterraca tmes for each prorty task group to be ndependent and dentcay dstrbuted accordng to the nfnte-varance Pareto dstrbuton wth shape parameter α and cut-off parameter k: F f t f () t α = k () t = P ( T t ) k t ( ) = F( t) = k α + α α, t > k,(7) k =, t, t. (8). Proportona Deay Dfferentated Servce The fundamenta prncpe of proportona deay dfferentated servce (PDDS) deveoped by Dovros et a. [Dovros et a, ] states that hgher-cass requests w receve better performance than the ower cass requests. Specfcay, N > servce casses are assumed, where the prorty of each cass s set n a decreasng order, and the average watng tme s then set n nverse proporton to the prorty for each cass: D P =,, N, (9) D P where D and P denote the average watng tme and prorty for cass, respectvey. That s, the average I. J. of SIMULATION Vo. 6 No and 5 ISSN x onne, prnt

3 watng tme s shorter for the prorty task groups that pay hgher usage costs. The proposed admsson contro mode forecasts the average watng tme of each cass for next measurement perod, and utzes Eq. (9) to choose the cass cent wth the argest gap n the rato of the average watng tme to receve the next servce from the server. Sgnfcanty, a new cent requestng for servce s paced n the cent s cass queue f the average watng tme for that cass does not exceed a predefned threshod. Each cass has ts own average watng tme threshod vaue, and t s adustabe durng operaton accordng to the number of the cents who eave the cass queue wthout servce owng to ong watng tme..3 Parameters Requred for Admsson Contro Mode The cent s asked to suppy some essenta nformaton before beng admtted nto the servce. Each cent has two optons, ts cass and the maxmum average watng tme that t can endure. The specfcaton for the cent s cass s smpe and consstent wth the approach taken n the dfferentated servce enabed at the network ayer, whe the provson of the maxmum average watng tme drecty refects the customer s requrement. The correspondng admsson contro modes are deveoped based on the two parameter specfcatons..3. Usng cent cass as the parameter As depcted n Fg., each cass cent watng for the servce s paced n the correspondng cass queue, whch s managed usng the FCFS servce mode. Accordng to conservaton aw [Boch et a, 998], f the average arrva rate s λ for a cent of cass durng the next measurement perod, then the average watng tme for cass, gven by D, shoud be: N = N λ D = λ D, () = where D represents the average watng tme for the aggregate traffc servced by a work-conservng FCFS server. The average watng tme for cass durng the next measurement perod can be derved from Eqs. (9) and () as foows: N λ D = D =, N N λ. () P P = Tme seres predctor Watng tme record Cass buffer Cent request Admsson contro unt Cass buffer Scheduer Cent accepts servce Cass N buffer Decne cent request Fgure : Admsson contro mode wth a parameter of each cent s cass specfcaton. Snce the average arrva rate s dffcut to obtan for each MAX D P. Correspondngy, ncomng cass cent s cass durng the next measurement perod n Eq. (), a tme P seres predctor s ncorporated nto the proposed admsson contro mode to foresee the average arrva rate and permtted to use the server f the foowng reatonshp s therefore resove the ssue of the unreasonabe request for fufed: N the arrva rate specfed by each cent as presented by Lee ˆ λ D MAX et a [Lee et a, 4]. = D P Cass cents are assumed to have the owest prorty, N ˆ λ P, () and the maxmum watng tme for each cass cent s P MAX assumed as D. Accordng to Eq. (9), the maxmum = P watng tme acceptabe to a cass cent s gven by where λˆ represents the average arrva rate for cass aggregate traffc foreseen by the tme seres predctor. I. J. of SIMULATION Vo. 6 No and 6 ISSN x onne, prnt

4 When the server s ready to servce the next cent, the scheduer as shown on the rght of Fg. appes the foowng equaton to decde whch cass cent shoud be chosen for servce: k =arg W P max, (3) N W P where W denotes the watng tme for the cent at the front of the cass queue, and P represents the prorty of cass. As the cents of some casses can toerate onger watng tme, such as best-effort traffc, the proposed agorthm can pop up an nteractve daog box to ask cents f they wsh to wat onger when the server s overoaded. The usage cost and the prorty for the cents s reduced f they are wng to wat onger. Ths approach can ower the number of customers n the hgher-cass queues under a bursty workoad, and thus mantan the strngent QoS requrement for hgher cass cents. The agorthm for the admsson contro mode s now summarzed wth each cent s cass as the parameter as foows.. When a cass cent arrves, utze the tme seres predctor to predct the average arrva rate of cass aggregate traffc durng next measurement perod.. Appy Eq. () to compute the average watng tme of cass durng the next measurement wndow. 3. Empoy Eq. () to determne whether the ncomng cent s admtted to use the server. 4. If admtted, pace the cent at the end of the cass queue. If Eq. () s not satsfed, but the cent s wng to wat onger, then search for the frst ower cass that satsfes the requrement of Eq. (). If found, pace the cent nto the correspondng cass queue..3. Usng maxmum watng tme as the parameter The proposed admsson contro mode aso permts the cent to specfy the maxmum watng tme as reveaed n Fg.. Notaby, the proposed mode requres a cassfer to derve the ncomng cent s cass, as dspayed n Fg.. MAX Let D represent the maxmum watng tme for cass cents wth the owest prorty, and the maxmum watng tme requested by the ncomng cent s set to ω. Equaton (9) ndcates that the ongest watng tme that MAX cass cents can bear s gven by D P ; the P cassfer can then appy the foowng equaton to determne the ncomng cent s cass: = arg mn ω P N max,. (4) MAX D P Notaby, the above equaton s apped to ocate the hghest prorty task group wth a maxmum watng tme requrement onger than that s acceptabe to the cent. The agorthm for the admsson contro mode as shown n Fg. can be summarzed as foows:. Empoy the cassfer to cassfy the ncomng cent accordng to Eq. (4).. Appy the tme seres predctor to predct the average arrva rate of the ncomng cent s cass,, durng the foowng measurement perod. 3. Use Eq. () to cacuate the average watng tme of cass durng the foowng measurement wndow. 4. Empoy Eq. () to determne whether the ncomng cent s admtted to use the server. 5. If admtted, pace the cent at the end of the cass queue. If Eq. () s not satsfed, but the cent s wng to wat onger, then search for the frst ower cass that satsfes the requrement of Eq. (). If found, then pace the cent nto the correspondng cass queue. 3. TIME SERIES PREDICTOR The fuzzy ogc technque has been used to sove severa connecton admsson contro n ATM and wreess networks and tme seres predcton probems effcenty n the terature [Lang, ]. We thus try to appy fuzzy ogc controer concept to predct maxmum arrva rate and maxmum watng tme as shown n the scheme presented n the prevous secton. 3. Fuzzy Logc Predctor As n [Ren and Ramamurthy, ], we use the average arrva rate for each cass durng the current and the ast four measurement perods λ ( t 3), λ ( t - ), λ ( t -), and λ () t to predct the average arrva rate durng next measurement perod λ ˆ( t +). Fg. 3 shows the correspondng fuzzy ogc tme seres predctor. The basc functons of the components empoyed n the predctor are descrbed as foows. Fuzzfer: The fuzzfer performs the fuzzfcaton functon that converts crsp nput data nto sutabe ngustc vaues that are needed n the nference engne. Fuzzy rue base: The fuzzy rue base s composed of a set of ngustc contro rues and the attendant contro goas. Inference Engne: The nference engne smuates human decson-makng based on the fuzzy contro rues and the reated nput ngustc parameters. The max-mn nference method s used to assocate the outputs of the nferenta rues [Buckey and Esam, ], as descrbed ater n ths subsecton. Defuzzfer: The defuzzfer acqures the aggregated ngustc vaues from the nferred fuzzy contro acton and generates a non-fuzzy contro output, the foreseen average arrva rate of each cass durng next measurement perod. The Mamdan defuzzfcaton method s empoyed n ths paper to compute the centrod of membershp functon for the aggregated output, where the area under the graph of membershp functon for the aggregated output s dvded nto two equa subareas [Buckey and Esam, ]. I. J. of SIMULATION Vo. 6 No and 7 ISSN x onne, prnt

5 Tme seres predctor Watng tme record Cassfer Cass buffer Cent request Admsson contro unt Cass buffer Scheduer Cent accepts servce Cass N buffer λ(t 3) Decne cent request Fgure : Admsson contro mode wth the parameter of the maxmum watng tme. where m denotes the mean, σ represents the varance. λ(t-) λ(t-) λ(t) λ ˆ( t +) Fuzzfer Defuzzfer Inference engne Fuzzy rue base Fgure 3: The fuzzy ogc based tme seres predctor. Fg. 4 shows the mappng of four nputs of the fuzzfer and the output parameter of the nference engne nto some approprate ngustc or membershp vaues, whch are expressed by the vaues wthn the range of and. Three membershp functons for each of four nputs and the output are gven n Fg. 4, where the ngustc varabes ow, medum and hgh gve the measure of the average arrva rate for each cass. Note that the foowng Gaussan membershp functon s chosen for the antecedents and the consequent: ( ( )) ( ) exp λ t m µ =, =,,,3 λ t,(5) σ Fgure 4: Membershp functon for the antecedents and the consequent. The nput and output fuzzy sets are correated to estabsh the nferenta rues of the fuzzy ogc tme seres predctor. Note that three fuzzy sets are used for each antecedent, so the number of fuzzy rues s 3 4 =8. By way of ustraton, each fuzzy rue can be nterpreted as: Fuzzy rue R : IF λ ( t 3) s A and λ ( t - ) s B and λ( t ) s C and λ () t s D, THEN ˆ λ ( t +) s E. (6) The nference engne then umps to the foowng concuson for fuzzy rue R : I. J. of SIMULATION Vo. 6 No and 8 ISSN x onne, prnt

6 ' ' ' ' ( A B C D ) ( A B C D E ) ' E = o, (7) ' ' ' where A, B, C and D stand for the ' membershp grades of four nputs obtaned from fuzzy rue R, respectvey, and the expresson nsde the second parenthess denote the smpfed representaton for Eq. (6). Fg. 5 ustrates the reasonng procedure for a two-rue Mamdan fuzzy nference system. Note that the composton of mnmum and maxmum operatons, whch corresponds the o operator n Eq. (7), s empoyed n the evauaton of the fuzzy rues. The µ µ µ µ non-fuzzy output of the defuzzfer can then be expressed by the foowng agebrac expresson: ( ( ) ) ( ) µ ˆ A λa ˆ λa d ˆ λa ˆ λ t + = µ ˆ λ d ˆ λ, (8) where ( ) A λ A A ( ) µ ˆ denotes the membershp functon of the aggregated output λˆ. A Mn µ A A µ λ(t 3) λ(t ) λ(t ) λ(t) ˆ λ ( t + ) Mn µ µ µ µ Max λ(t 3) λ(t ) λ(t ) λ(t) µ ˆ λ ( t + ) ˆ λ ( t + ) λˆ A Fgure 5: The reasonng procedure for Mamdan defuzzfcaton method 3. Support Vector Regresson Approach Support vector regresson (SVR) has recenty ganed popuarty owng to ts many attractve features and emnent emprca performance [Vapnk, 995]. The maor dfference between the SVR and tradtona regresson technques s that the SVR empoys the structura rsk mnmzaton (SRM) approach, rather than the emprca rsk mnmzaton (ERM) approach typcay adopted n statstca earnng. The SRM attempts to mnmze an upper threshod on the generazaton rather than mnmze the tranng error, and s expected to perform better than the tradtona ERM approach. Furthermore, the SVR s a convex optmzaton, whch guarantees that the oca mnmzaton s the unque mnmzaton. To sove a nonnear regresson or functona approxmaton probem, the SVR nonneary maps the nput space nto a hgh-dmensona feature space usng an approprate kerne representaton, such as poynomas and rada bass functons wth Gaussan kernes. Ths approach s utzed to bud a near regresson hyperpane n the feature space, whch s nonnear n the orgna nput space. The parameters can then be derved by sovng a quadratc programmng probem wth near equaty and nequaty constrants [Vapnk, 995]. A tranng data set n D = {( x, y ) R R, =,..., } comprsng par tranng data ( x, ), = y, s gven. The nput x,... terms are n-dmensona vectors, and the system response y terms are contnuous vaues. The SVR attempts to approxmate the foowng functon usng data set D: N (, w) w ( x) b f x = ϕ +, (9) = where b denotes the bas term, and the w terms represent the subects of earnng. Furthermore, a mappng z = Φ( x) s seected n advance to map nput vectors x nto a hgher-dmensona feature space F, whch s ϕ x. spanned by a set of fxed functons ( ) By defnng a near oss functon wth the foowng ε-nsenstvty zone as ustrated n Fg. 6: f y f ( x, w) ε y f ( x, w) = ε y f ( x, w) ε otherwse,() The w terms n Eq. (9) can be estmated by mnmzng the rsk: C R = w + y f ( x, w), () ε = where C denotes a user-chosen penaty parameter that determnes the trade-off between the tranng error and VC dmenson of the SVR mode. Sgnfcanty, the VC dmenson s a scaar vaue that measures the capacty of a I. J. of SIMULATION Vo. 6 No and 9 ISSN x onne, prnt

7 set of functons [Vapnk, 995]. Equaton () can be further derved as the foowng constraned optmzaton probem: C R( w, ξ, ξ ) = w + ξ + ξ, () = = subect to constrants: T y w x b ε + ξ T w x + b y ε + ξ, (3) ξ, ξ where ξ and ξ denote the respectve measurements above and beow the zone wth the radus ε n Vapnk s oss functon as gven n Eq. (). e ( x) ( α ) k( x,x) b f = α +, (6) = where b denotes the optma bas. 4. PERFORMANCE EVALUATION A seres of smuatons was performed to measure the performance and behavora specfcs of the proposed admsson contro modes. Ony the admsson contro mode utzng the maxmum watng tme as the parameter s mpemented n ths study, snce t s very smar to the agorthm that utzes each cent s cass specfcaton as the parameter except n the presentaton of the parameters. The performance metrcs of most nterest ncude the throughput of the admtted cents from the premum cass, the percentage of nfrngement of QoS requrement for admtted cents from each prorty task group and the acheved watng tme rato for dfferent casses of requests. 4. Smuaton Scenaro ε Fgure 6: ε-nsenstvty oss functon. ε y f(x,w) Schökopf et a. deveoped a modfcaton of orgna Vapnk s SVR agorthm, caed ν-svr, and camed that t can automatcay mnmze the radus ε [Schökopf et a, ]. Lagrange mutper methods can be empoyed to demonstrate that the constraned optmzaton probem n Eqs.() and (3) maxmzes the souton of the foowng equaton: W ( α, α ) = ( α α ) y ( α α )( α α ) k( x x ) =,(, = 4) under constrants: ( α α ) = = C α, α, =,...,, (5) ( + ) α α C ν = where ( α, α ) denotes one of Lagrange mutper pars; C represents a reguarzaton constant specfed a pror; ν s a constant greater than or equa to zero, and k(x x ) denotes normay a Gaussan kerne or poynoma kerne. The best nonnear regresson hyperfuncton s then represented as: An event-drven smuator s deveoped to examne the admsson contro modes proposed n ths work. Athough prevous works report that the WWW traffc pattern s sef-smar characterstc, the precse generaton of representatve sef-smar WWW traffc for performance evauaton remans an open probem. Varous offered oads on the Web server were therefore smuated, mantanng fxed targeted watng tme ratos by utzng the rea trace based on the Web server ogs for the 998 Word Cup Soccer web ste. The trace ncudes requests for 5 unque fes, and the average request sze s 7KB. The access ogs provde the request tmestamp, cent ID, obect URL, servce status and repy sze of each request. Tabe shows the smuaton parameters. To smpfy the study durng the experments, Ths work ony consdered two prorty eves, premum and basc cents, because the maor concern of the smuatons s to examne the effectveness of the proposed admsson contro modes. The watng tme dfferentaton of the two cass cents s. The maxmum watng tmes of a cents are drawn unformy between and seconds for the agorthm usng the maxmum watng tme as the parameter. The servce tmes of a requests are exponentay dstrbuted wth a mean equa to 8ms. The prorty of each ncomng request s assgned randomy, and the number of hgh prorty tasks and ow prorty tasks s amost the same. 4. Smuaton Resuts A seres of smuatons was conducted for the proposed admsson contro modes embedded wth the two tme seres predctors,.e. support vector regresson (SVRAC) and fuzzy ogc system (FLAC). The smuaton resuts were compared wth those of the frst-come-frst-served servce mode (FCFS). I. J. of SIMULATION Vo. 6 No and ISSN x onne, prnt

8 Parameter Prorty Leve Measurement perod seconds Dsk seekng overhead. ms. Dsk bandwdth Network bandwdth Mbps Mbps Maxmum server process number Maxmum queue ength Watng tme dfferentatons Tabe : Smuaton parameters Vaue Fgures 7 9 dspay the throughputs of admtted cents from the premum and basc casses for three admsson contro modes. Tabe sts the throughput ratos of the three agorthms. Tabe reveas that the average throughput of the admtted cents from premum cass n SVRAC s sgnfcanty better than that n FLAC, whe FCFS does not dstngush between the two prorty task groups as expected. Meanwhe, athough a three agorthms attan about the same performance under ow workoads, Fgs. 7 9 ndcate that the machne earnng technques, such as support vector regresson and fuzzy ogc system, effectvey predct sef-smar tme seres for admsson contro modes under hgh traffc oad. Scheme Average throughput Throughput rato (%) Premum cass Basc cass Premum cass Basc cass SVRAC FLAC FCFS Tabe : Comparson of throughputs for two prorty task groups Premum cass Basc cass 9 8 Throughput (reqs/sec) Smuaton tme (sec) Fgure 7. Throughput of the admtted cents from two casses n the SVRAC scheme. Premum cass Basc cass Throughput (reqs/sec) Smuaton tme (sec) Fgure 8. Throughput of the admtted cents from two casses n the FLAC scheme. I. J. of SIMULATION Vo. 6 No and ISSN x onne, prnt

9 Premum cass Basc cass 7 6 Throughput (reqs/sec) Smuaton tme (sec) Fgure 9. Throughput of the admtted cents from two casses n the FCFS scheme. Tabe 3 sts the percentage of the cents that nfrnge ther QoS requrement (maxmum watng tme), whe Fgs. show the number of decned cents requests owng to the voaton of maxmum watng tme requrement. The anaytca resuts revea that the SVRAC mode provdes the best servce for the admtted premum cass cents. Athough the other two agorthms mght accept more basc-cass cents for servce than the SVRAC agorthm, the hgher percentage of voatons of QoS requrements for the admtted cents of premum cass as sted n Tabe 3 s hghy undesrabe for the reazaton of Internet servers provdng proportona dfferentated servces. Scheme Reect rato (%) Premum cass Basc cass SVRAC FLAC FCFS Tabe 3: Percentage of nfrngement QoS requrement for two prorty task groups Premum cass Basc cass 7 6 Reected requests (reqs/sec) Smuaton tme (sec) Fgure. Reected cents requests from two casses n the SVRAC scheme. I. J. of SIMULATION Vo. 6 No and ISSN x onne, prnt

10 Premum cass Basc cass 6 5 Reected requests (reqs/sec) Smuaton tme (sec) Fgure. Reected cents requests from two casses n the FLAC scheme. Premum cass Basc cass Reected requests (reqs/sec) Smuaton tme (sec) Fgure. Reected cents requests from two casses n the FCFS scheme. Fgures 3 5 show the average watng tmes of admtted cents of premum and basc casses for the three admsson contro modes. Tabe 4 sts the watng tme ratos for the three agorthms. Tabe 4 reveas that the rato of SVRAC s sghty coser to the preset rato of, than that of FLAC, whe FCFS does not dscrmnate between the two prorty task groups. The throughput of the admtted premum cass cents s sgnfcanty hgher than those of the basc cass as reveaed n Fgs. 7 to, but the dfferences between the support vector regresson and fuzzy ogc system n the average watng tme rato for the cents of two casses were found to be nsgnfcant. Scheme Rato SVRAC.3537 FLAC FCFS Tabe 4: The watng tme rato for the three schemes I. J. of SIMULATION Vo. 6 No and 3 ISSN x onne, prnt

11 Premum cass Basc cass Average watng tme (sec) Smuaton tme (sec) Fgure 3. Average watng tme of two casses n the SVRAC scheme. Premum cass Basc cass Average watng tme (sec) Smuaton tme (sec) Fgure 4. Average watng tme of two casses n the FLAC scheme. Premum cass Basc cass 3 Average watng tme (sec) Smuaton tme (sec) Fgure 5. Average watng tme of two casses n the FCFS scheme. I. J. of SIMULATION Vo. 6 No and 4 ISSN x onne, prnt

12 5. CONCLUSION Ths work presents two adaptve admsson contro modes to provde a proportona deay dfferentated servces from an Internet server. Two dfferent tme seres predctors, namey support vector regresson and fuzzy ogc, are embedded n the admsson contro modes to estmate the traffc oad of the cent n the next measurement perod. The predcton s needed to determne whether the cent can be accepted n the admsson contro, and the forecast s promsng because a sef-smar tme seres s predctabe. Smuaton resuts demonstrate that the mpementaton of tme seres predcton agorthm wth support vector regresson (SVR) s sgnfcanty better then fuzzy ogc system and frst-come-frst-serve servce mode when the performance metrcs of the throughput and the rato of reected cents from premum cass are compared. Meanwhe, the average watng tme rato of the cents from the two casses for both the SVR agorthm and the fuzzy ogc system s aso mantaned wthn a reasonabe range of the predetermned rato. In the subsequent research, we w not ony ncorporate other ntegent toos, ncudng neuro-fuzzy and genetc agorthms, nto the proposed admsson contro mode, but aso expore some ensembe earnng method to combne the resuts of ndvdua machne earnng technques so that the accuracy of predcton for the arrva rate of the aggregate traffc can be further enhanced. REFERENCES Abraham A. 3, Busness ntegence from web usage mnng, Journa of Informaton & Knowedge Management, vo., no. 4, pp Anguta, D., Bon, A. and Rdea, S. 999, Learnng agorthm for nonnear support vector machnes suted for dgta VLSI, Eectroncs Letters, vo. 35, no. 6, pp Antono, G., Casazza, G., D Lucca, G., D Penta, M. and Mero, E., Predctng Web ste access: an appcaton of tme seres, The 3rd Internatona Workshop on Web Ste Evouton, vo., pp Artt, M. F. and Jn, T., A workoad characterzaton study of the 998 word cup web ste, IEEE network, pp Barford, P. and Crovea, M. E. 998, Generatng representatve Web workoads for network and server performance evauaton, ACM SIGMETRICS Performance Evauaton Revew, vo. 6, no., pp Bhatt, N. and Fredrch, R. 999, Web server support for tered servces, IEEE Network, vo. 3, no. 5, pp Boch, G., Grener, S., de Meer, H. and Trved, K. S. 998, Queueng networks and Markov chans: modeng and performance evauaton wth computer scence appcatons. New York: John Wey & Sons, Inc. Bonno, D., Corno, F. and Squero, G.. 3, Dynamc predcton of Web requests The 3 Congress on Evoutonary Computaton, vo. 3, pp Buckey, J. and Esam, E., An ntroducton to fuzzy ogc and fuzzy sets (advances n soft computng). Physca Verag. Chen, S.; Samngan, A.K. and Hanzo, L., Support vector machne mutuser recever for DS-CDMA sgnas n mutpath channes, IEEE Transactons on Neura Networks, vo., no. 3, pp Chen, X. and Mohapatra, P., Performance evauaton of servce dfferentatng Internet servers, IEEE Trans. Computers, vo. 5, no., pp Dhyan, D., Bhowmck, S. and Ng, W.-K. 3, Modeng and predctng a Web page accesses usng Markov processes, 4th Internatona Workshop on Database and Expert Systems Appcatons, pp Dovros, C., Stads, D. and Ramanathan, P., Proportona dfferentated servces: deay dfferentaton and packet schedung, IEEE/ACM Trans. Networkng, vo., no., pp. -6. Eggert, L. and Hedemann, J. 999, Appcaton-eve dfferentated servces for web servers, Word Wde Web Journa, vo. 3, no. 3, pp Gong, X. and Kuh, A. 999, Support vector machne for mutuser detecton n CDMA communcatons, The Thrty-Thrd Asomar Conference on Sgnas, Systems, and Computers, vo., pp Haffner, P., Tur, G., and Wrght, J. H. 3, Optmzng SVMs for compex ca cassfcaton, 3 IEEE Internatona Conference on Acoustcs, Speech, and Sgna Processng, vo., pp. I-63 I-635. Hasegawa, M., Wu, G. and Mzuno, M., Appcatons of nonnear predcton methods to the Internet traffc, The IEEE Internatona Symposum on Crcuts and Systems, vo., pp. III-69 III-7. Kanoda, V. and Knghty, E. W. 3, Ensurng atency targets n mutcass Web servers, IEEE Trans. Parae and Dstrbuted Systems, vo. 4, no., pp Kuh A., Adaptve kerne methods for CDMA systems, IEEE Internatona Jont Conference on Neura Networks, vo. 4, pp Lee, S. C., Lu, J. C. and Yau, D. K. 4, A proportona-deay DffServ-enabed Web server: admsson contro and dynamc adaptaton, IEEE Trans. Parae and Dstrbuted Systems, vo. 5, no. 5, pp Lang, Q., Ad hoc wreess network traffc-sef-smarty and forecastng, IEEE Communcaton Letters, vo. 6, no. 7, pp Paxson, V. and Foyd, S. 997, Why we don't know how to smuate the Internet, 997 Wnter Smuaton Conference, pp Ren, Q. and Ramamurthy, G., A rea-tme dynamc I. J. of SIMULATION Vo. 6 No and 5 ISSN x onne, prnt

13 connecton admsson controer based on traffc modeng, measurement, and fuzzy ogc contro, IEEE J. Seected Areas n Comm., vo. 8, no., pp Rtter, H., Pastoors, T. and Wehre, K., DffServ n the web: dfferent approaches for enabng better servces n the Word Wde Web, Jont Conf. of Broadband Communcatons, Hgh Performance Networkng and Performance of Communcaton Networks, pp Schökopf, B., Smoa, A., Wamson, R. and Bartett, P. L., New support vector agorthms, Neura Computaton, vo., pp Van Geste, T., Suykens, J.A.K. and Baestaens, D.-E., Lambrechts, A., Lanckret, G.., Vandaee, B., De Moor, B. and Vandewae, J., Fnanca tme seres predcton usng east squares support vector machnes wthn the evdence framework, IEEE Transactons on Neura Networks, vo., no. 4, pp. 89 8,. Vapnk, V. 995, The nature of statstca earnng theory. New York: Sprnger-Verag. Vasou, N. and Lutfyya, H., Managng a dfferentated quaty of servce n a Word Wde Web server, IEEE Inter. Symp. Integrated Network Management, vo. VII, pp AUTHOR BIOGRAPHIES Chenn-Jung Huang was born n Huaen, Tawan, n 96. He receved a B. Sc. degree n Eectrca Engneerng from Natona Tawan Unversty, Tawan and an M. S. degree n Computer Scence from Unversty of Southern Caforna, Los Angees, n 984 and 987. He receved a Ph. D degree n Eectrca Engneerng from Natona Sun Yat-Sen Unversty, Tawan, n. He s currenty an Assocate Professor at the Insttute of Learnng Technoogy, Natona Huaen Unversty of Educaton, Tawan. Hs research nterests ncude computer communcaton networks, data mnng and machne earnng appcatons. Y-Ta Chuang and Chh-Lun Cheng are pursung a Master s degree at the Insttute of Learnng Technoogy, Natona Huaen Unversty of Educaton, Tawan. Ther research nterests ncude computer communcaton networks, data mnng and appcatons of machne earnng technques. I. J. of SIMULATION Vo. 6 No and 6 ISSN x onne, prnt