Cooperative Mixed Strategy for Service Selection in Service Oriented Architecture

Transcription

1 Cooperatve Mxed Strategy for Servce Selecton n Servce Orented Archtecture Yn Shen, Student Meber, IEEE, and Yushun Fan Abstract In Servce Orented Archtecture (SOA), servce brokers could fnd any servce provders whch offer sae functon wth dfferent qualty of servce (QoS). Under ths condton, users ay encounter dffculty to decde how to choose fro the canddates to obtan optal servce qualty. Ths paper tackles the Servce Selecton Proble (SSP) of te-senstve servces usng the theory of gaes creatvely. Pure strateges proposed by current studes are proved to be proper to ths proble because the decson conflcts aong the users result n poor perforance. A novel Cooperatve Mxed Strategy () wth good coputablty s developed n ths paper to solve such nconstant-su non-cooperatve n-person dynac gae. Unlke related researches, offers users an optzed probablty ass functon nstead of a deternstc decson to select a proper provder fro the canddates. Therefore t s able to elnate the fluctuaton of queue length, and rase the overall perforance of SOA sgnfcantly. Furtherore, the stablty and equlbru of are proved by sulatons. W I. INTRODUCTION ITH ts technque neutralty, loosely coupled structure, and locaton transparency, Servce Orented Archtecture (SOA) provdes prosng paradgs for effectve enterprse-scale syste developent and ntegraton []-[3]. There are three operatons n standard SOA,.e., publshng, fndng, and bndng []. Servce-fndng operaton produces an portant effect on perforance of the archtecture because t offers unque opportunty that the three partes ncludng user (requester, or bulder), broker, and provder collaborate together. In recent years, ore and ore scholars of servce fndng turned ther ephass fro the proble of servce-atchng to another specal branch, Servce Selecton Proble (SSP), to deal wth the ncreasng selecton proble of servce provders [5]-[5]. SSP s defned as the proble that servce requester has to ake the optal decson of selectng the servce provder fro ultple canddates dscovered by broker. The canddates can provde servces wth slar or substtutable functons but dfferent qualty of servce (QoS) [6]. SSP s Manuscrpt receved March 6, 7. Ths work was supported n part by the Natonal H Technology Research and Developent Progra of Chna under Grant 6AAZ5, and by the Natonal Natural Scence Foundaton of Chna under Grant 667. Yn Shen s wth the Autoaton Departent of Tsnghua Unversty, Bejng, Chna; and Chengdu Electroechancal College, Chengdu 63 (phone: ; fax: ; e-al: sy@eee.org). Yushun Fan s wth the Autoaton Departent of Tsnghua Unversty, Bejng, Chna (e-al: fanyus@tsnghua.edu.cn). lnked closely wth servce coposton stuaton where user takes a selecton of varous servces coposed together to for a busness process [5]-[7]. It s wdely accepted that the user shall fetch QoS nforaton fro the QoS broker as a decson-akng bass, and then select fro the canddates accordng to hs preferences [6]-[9],[]-[5]. Many strateges based on optzaton theory have been developed to help users to gan optal perforance by selectng correct provders. The ost studed ethods are dynac prograng [6][7], Genetc Algorth [][9], nteractve ethod [], fuzzy logc [][], and other perforance evaluaton approaches []-[5], etc. All these ethodologes are based on optzaton theory, assung that a servce user need not consder the possble strateges of other users when akng hs decson. Ths assupton s only correct n a lted collecton of SSP n whch perforance of provders are assued to be ndependent wth the users strateges. However, the assupton s ncorrect for the te-senstve servces that need the watng and servce te to be optzed. The watng te has to be extended, for exaple, when a provder wth capacty constrant s overloaded by frequent selecton. Unlke the tradtonal schedulng proble [6][7], there s no adnstrator n the peer to peer world of SOA, who dentfes or pleents the selecton strategy. Thus the actual watng te s deterned not only by the provder s perforance, but also by the users selecton strateges. In ths case, a servce user has to copete wth other users n choosng hs best strategy. Thus SSP becoes a dynac gae theory proble aong the ultple users who are allowed to ake ther decsons ndependently. Fro the vewpont of gae theory, all the current ethodologes are pure strateges. But t s clear that n an n-person non-cooperatve gae, the saddle pont of pure strategy ay not exst []. In ths case, any pure strategy, no atter how effcent t sees to be, wll result n the proble of "wnner takes all", because ost of the users wll pursue the provder evaluated to be the best, whle the other provders have probablty uch hgher to be dle. Moreover, the delay between decson and pleentaton of selecton strategy n SOA s expected to be prolonged because SOA recoends one-te transsson process usng text data to establsh servce connecton, nstead of tradtonal Reote Procedure Call (RPC). If users stll apply any pure strateges, ncoplete nforaton durng ths perod wll exacerbate the provders perforance further /7/$5./ 7 IEEE 5

2 Nevertheless, the copettveness n all ndustral sectors s pushng servce users to strve for a further optzaton of delvery te. Therefore the optzed soluton of SSP s recognzed as a key technque for te-senstve servces. The rest of the paper s organzed as follows: Frst, secton II forulates the proble, and analyze the decreasng of servce level caused by tradtonal pure strategy. Next, secton III presents Cooperatve Mxed Strategy () and descrbes the approach of optal soluton n detal. And then, perforance, stablty, and equlbru of are sulated n secton IV. Fnally, secton V concludes the paper. II. PROBLEM DEFINITION AND PURE STRATEGY A. Defnton of Te-Senstve SSP The selecton proble for te-senstve servces s defned as follows: A knd of servce requests arrve at the SOA envronent, followng the Poson dstrbuton wth a ean arrval rate. The user of each requests choose one provder fro the parallel canddates { M } = to coplete the request. After akng such decson, the user takes te T p to coplete preparaton procedures lke bndng provder, transttng data, etc. And then, the servce te of the requests on the selected provder M s a negatve exponental dstrbuted varable wth ean μ. Paraeters T p,, and μ, =,, are assued to be constant whle the queue te of the requests vary durng the stochastc process. The objectve of SSP for each user s to apply an optal strategy to select servce provder fro the canddates, whch wll yeld a nzed expected value of the total te T, ncludng watng te and servce te. An deal strategy shall have the followng features: () Optzaton under cooperaton stuaton. Those who pleent the strategy, the obedent, wll reach the optal expectaton of total te. () Punshent for the dsobedent. The dsobedent that dsrupts the cooperaton by adaptng hs own strategy other than the publc one, has to wat even longer than expected. The frst crteron s reasonable obvously, whereas the second one ensures the stablty and equlbru of the ethod. Ths crteron can correct the devaton fro the stable attractor of the syste, and keep the syste steady around ts optal state because ost partes have to obey the publc strategy to avod extra punshent. B. Pure Strategy and Decson Conflct When selectng provder at te t, the request J acqures the queue length L and servce level paraeter μ fro QoS broker. Unable to obtan other users strateges for reference, a reasonable pure strategy of J s to choose the provder by whch J would expect to be copleted n the shortest duraton,.e., n T = (L + ) / μ () Wthout loss of generalty, let us assue that J chooses provder M. So we have L /μ L /μ, =,3,..., at t, accordng to the objectve (). Now we want to know the lower bound of the postve real nuber t, wth whch the above nequalty keeps true n the nterval (t-t, t], whereas s false n (t-t-, t-t) for arbtrary sall postve real nuber. We notce that at least one of the followng events ust have occurred n (t-t, t]: () M copletes a certan task, or () a partcular task arrves at a provder other than M to queue up. Fro Lttle s forula[9], the expected value of t whch akes one of the above events occur should be no less than ether / or /μ. Therefore the lower bound of t should be larger than / snce generally we have >μ. Hence M keeps ts poston of the best provder n the te nterval (t-/, t]. Servce requesters arrvng durng ths nterval should have ade the sae choce to bnd M, as J does at t. Accordng to the prncple of Frst-Coe-Frst- Serve (FCFS), they wll be served before J. Thus, when J starts to queue at te t+t p, t wll fnd that the queue s longer than ts expectaton, and ts actual expected value of total te becoes: ET j [L + + ( μ )*n(t p, /)] / μ The expected value s clearly nferor to (), thus s not the desred soluton. Moreover, the ncreasng trend of expected watng te s not the worst aspect of pure strategy, as shown n the followng sulaton. C. Perodc Fluctuaton Produced by Pure Strategy In the sulaton, T p and are set to. and. {μ } of provders are selected randoly to be 9,, 9, 7,, 6,,,, and 5 respectvely., requests are generated and assgned to the provders usng pure strategy descrbed n [3]. A knd of "order", fluctuaton of the queue length, eerges durng the rando experent. Fgure shows the queue length sequence of M 3 as an exaple. Queue Length Te (n) Fg.. Queue length fluctuates perodcally under pure strategy because of the slar decsons aong users. In fact, t s the hoogenety of users selecton crtera that results n the slar decsons. Ths spontaneous behavor of order not only prolongs the average total te, but also causes load balance and neffcency aong servce provders. The decson conflct s nevtable when pure strategy s appled. Unfortunately, better pure strategy ntensfes the 53

3 fluctuatons of servce copleton te further. In order to solve ths proble, a stochastc xed strategy s developed n the followng secton to optze the servce perforance. III. COOPERATIVE MIXED STRATEGY FOR SERVICE SELECTION PROBLEM A. Mxed Strategy and Its Payoff Functon A xed strategy eans J does not drectly choose the provder whch has nu expected value of total te, but choose a provder M wth probablty ~ p, =,,...,, satsfyng ~ = p = Obvously any pure strategy s n realty a specal xed strategy: ~ p ; ~ ' = p =, f '. Let p be the probablty that M s chosen by the users at te t, =,,... It should be notced that, on the one hand, ~ p s a saple of p. They are closely related wth each other, and the dstrbuton of p s deterned by varous ~ p saples. But on the other hand, f the users of J and other requests are consdered as dfferent players n a gae, p and ~ p usually have dfferent atheatcal characterstcs because J does not have to adopt the sae strategy wth other users. ~ { p} = and { p} = construct a xed strategy stuaton of the dynac gae. Now we offer the payoff functon of the stuaton. We notced that p( t) = { p = s a stochastc process, and the queue length L( t) = { L = s also a stochastc process dependent on both t and p. Due to the preparaton te T p, the payoff functon of J should be E(T (t+t p ) L (t), p ) under the condtons that J has been nfored the value of L(t) and p(t). For ease of notaton, we use ET (t+t p ) nstead of the above condtonal expected value as long as t can not cause any confuson. We have: ET (t+t p ) = ax [L + + (p μ )*T p, ] / μ () Therefore, the payoff of J,.e., the expected value of total te, can be represented by (3). ET = ET t T ~ = ( + p ) p (3) Equaton (3) forulates the objectve of te-senstve SSP. Generally speakng, () s ncoputable for J because the dstrbuton { p s unavalable for J. Nevertheless, = when all the users hold a knd of cooperatve atttude,.e., applyng the approach descrbed n the next secton, they can not only forecast the ean (), but also optze ther expected payoff (3). Moreover, ths knd of cooperatve atttude s defntely not an extravagant hope even under the peer to peer world of SOA because, as we wll show later n secton IV, users holdng uncooperatve atttude wll ake ther own payoff reduced. B. Cooperatve Mxed Strategy Accordng to the above analyss, the proposed Cooperatve Mxed Strategy () approach for SSP can be suarzed as follows steps: () Servce provders regster to servce broker; () QoS broker accesses the provders to get ther QoS nforaton updated, partcularly the paraeters μ and the current queue length L, n an event-drven or a perod anner; () When a request eerges to the archtecture, the user fetches updated QoS nforaton fro the QoS broker frst; (v) Then the user calculates the probablty p correspondng wth each provder M, wth the ethod descrbed n the next secton; (v) Fnally the requester uses a based roulette wheel approach based on the above dstrbuton to deterne whch provder shall be chosen. To pleent roulette wheel selecton, a real nuber r s randoly generated over the range (,] frst. And then provder M s selected f the condton p < k= k r p () k= k s et. Here p s defned as. Snce { p } k= k s a onotoncally ncreasng sequence = of, and = p = k k, one and only one M wll be chosen by the roulette wheel. Usng ths stochastc saplng ethod, offers opportunty, ore or less, to each provder. It s clear that the opportunty ght not be dvded equally aong the provders because better provder should be ore lkely to be selected. The next secton tres to fnd out the optal dstrbuton. C. Optal Soluton of Now we assue that all the users hold cooperatve atttude,.e., ~ p =p (The uncooperatve case wll be studed n secton IV.). Then (3) turns to the nonlnear prograng proble (5) wth varables t) = { p, accordng to (). p( = L + μ T p n ET = λt p ax p + p =, μ λtp (5) p = st =. p =,,..., Proble (5) can be verfed to be a convex prograng proble wth varables and + constrants. Because the local nu of any convex prograng proble s also a global optal soluton, (5) has good coputablty. Now we dscuss ts analytc soluton. a = μ Let, and K={ b L + T μ <, p b = =,,..., λt =,,...,}. The followng reversble transfor (6) turns the prevous proble (5) to a nonlnear prograng proble (7): 5

4 p p ' = p + b b b ' = b n EL st. J = = p ' = = K K K K a ax[( p ' + b ') b ',] p ' n(b,) K, = b ' K, = =,,..., When b ', the optal soluton s: K p ' =. ( = b b K K ' ) ', n When ' <, = b K, proble (7) equals to the lnear prograng proble () wth varables p ' and + constrants: n B p ' = =,,..., ab ' B B p ' = b ' =,,..., ab ' < B st a (). p = = b ' K = ', p ' n(b,) =,,..., By applyng the reverse transforaton of (6) to the above optal solutons, we can obtan the optal probablty ass functon p(t) necessary for fnally. Therefore, has good coputablty snce ts optal soluton can be ganed n at ost polynoal te O() [], and ts coputaton te wll be uch less than dynac prograng. IV. SIMULATIONS A. Servce Perforance Iproveent wth Sulaton of actual envronent was undertaken to verfy the effcency of. All servce users, servce broker, QoS broker, and servce provders are sulated on the Intel Pentu IV.7-GHz usng VB language. The paraeters of the experent are selected as follows. One knd of servce requests eerge accordng to Posson process wth ntensty =. provders of the servce regster to both the servce broker and QoS broker. The te that provders coplete a request follows negatve exponental dstrbuton wth randoly selected paraeters μ = 9,, 9, 7,, 6,,,, and 5 respectvely. Durng the experent,, requests are assgned to the provders usng approach. A coparson experent s conducted by Selecton Method Usng PROMETHEE () proposed n [3]. The crtera of s forulated as (), and the paraeters are confgured sae as that of. Total te and queung length of each request are recorded, and the statstcs of both and are lsted n Table I. (6) (7) TABLE I COMPARISON BETWEEN AND Ite / (%) average total te (n) % axu total te (n) % average queue length % axu queue length % It s obvous fro Table I that outperfors tradtonal pure strateges lke snce averagely shortens % of the total te and queue length. Therefore servce users can obtan uch better servce level when the capablty of provders s not changed at all. Queue length sequences of three typcal servce provders are shown n Fg. -Fg. respectvely to copare the perforance of and. Queue Length L t 6 6 Te t (n) Fg.. Queue length sequence of provder M. M has the best servce rate μ =9 aong the provders. The sequence produced by has shorter queue and saller apltude than that by. Queue Length L t 6 6 Te t (n) Fg. 3. Queue length sequence of M 3, the provder wth edu paraeter μ 3=9. Fg. -Fg. ndcate that cooperatve xed strategy can not only reduce the average queue length sgnfcantly, but also balance the workload of provders wth varous capablty. 55

5 Queue Length L t 6 6 Te t (n) Fg.. Queue length sequence of M 5. Wth μ 5=, M 5 s the least copetent aong the provders. It s dle under n ost of the te, but suddenly facng a backlog of tasks at about t=.5 and t=7.5. When users applyng, M 5 s fully utlzed over the te horzon, and the sudden rse n the queue s elnated. B. Stablty of Stochastc Process Durng the experent descrbed above, we can easly track the rando countng process N (t) that represents the total nuber of requests assgned to provder M n a te nterval between and t. Results ndcate that N (t) s an ndependent ncreent process under, whch does not hold under. N 3 (t) s shown n Fg. 5 as a saple. Te t (n) 6 6 Total nuber of requests assgned to M 3, N 3 (t ) Fg. 5. Countng process N 3(t) under and. The ncreent N 3(t+ )- N 3(t) s te-ndependent and proportonal to under, whereas t s te-dependent under. Moreover, snce N ()=, N (t) can be treated as a stochastc Posson ncreent process that s covarance statonary. On the contrary, output of pure strategy lke s hard to be forecasted or controlled due to ts lack of stablty. Fnally, stablty of the strateges s easured quanttatvely. We calculate the varance of nterarrval tes,.e., te ntervals between successve occurrences of assgnng requests to M. Fg. 6 shows the varances of all the provders on a logarthc coordnate syste. Fg. 6 ndcates that wth saller varance, the output process of s uch ore stable than that of. C. Non-cooperatve Case and qulbru The global optzed soluton calculated n Secton III aybe not the equlbru of the n-person non-cooperatve gae accordng to gae theory. In fact, the optal soluton ganed fro (7) aybe not n lne wth the party s own best nterests. Hence t s very necessary to dscuss the non-cooperatve equlbru of the dynac gae n ths case. Varance of nterarrval te Servce provers Fg. 6. Varance of nterarrval te of the requests assgned to provders. The varance under s about one order of agntude saller than that under. Assung that q proporton of servce users are dsobedent, they hold uncooperatve atttude and adopt pure strategy. After obtanng other user s strateges by (7)-(), the dsobedent selects the best provder wth probablty. Obvously, when q=, the non-cooperaton case degrades to cooperaton case. We execute a seres of experents to sulate the syste perforance when q changes. The paraeters of the experents are confgured sae as that n the prevous secton. Fg. 7 shows average total te recorded n experents correspondng wth varous values of q. The data fro Table I whch llustrates the perforance of pure strategy are also drawn as a constant functon n Fg. 7 for coparson. Average total te (n) The dsobedent under The obedent under...6. Proporton of the dsobedent, q Fg. 7. Increasng trend of total te when the proporton of the dsobedent, q, ncreases. It s worth notcng that the watng te of the obedent s always less than that of the dsobedent. It can be seen fro Fg. 7 that () average total te ncreases onotoncally when the nuber of the dsobedent ncreases; and, () for each proporton q, total te of the users adoptng uncooperatve strategy s, contrary to what ght be expected, 56

6 longer than the others. The phenoenon s caused by strategy conflct aong the dsobedent. Therefore, uncooperatve atttude does not gve the dsobedent extra payoff, but reduces ther payoff. Ths statstcal result encourages ratonal users to adapt nstead of any pure strategy to obtan optal servce. The above features ake the deal ethod snce t satsfes the crtera defned n secton II. In addton, t can be concluded also fro Fg. 7 that donates pure strateges provded the proporton q of the dsobedent does not exceed 75%. Snce the stuatons wth the dsobedent over 75% are detrental to all partes, the syste wll return to the -based stuaton quckly under the users jont efforts. Therefore, we can expect a convergng trend of users strateges towards, whch ensures the -based stuaton the equlbru pont of the gae. V. CONCLUSION Global copetton ncreases pressure on servce users to have ther requests copleted faster than ther copettors. For such te-senstve requests, applyng tradtonal pure strateges to select proper provder would result n poor perforance, as llustrated n secton II of ths paper. Usng cooperatve xed strategy (), ths paper overcoes the servce selecton proble of te-senstve servces. The sulaton results have proved that not only shortens the average watng te and queue length, but also balances the load aong provders, elnates the sudden ncrease n the queue, and reduces the volatlty of syste perforance. A sgnfcant contrbuton of ths paper s that the servce selecton proble s treated for the frst te as an n-person gae nstead of an optzaton proble. successfully gudes the users to cooperatve copetton that benefts all of the. Further steps nclude classfcaton of gaes related to selecton proble, as well as applcatons of ore gae theory to ths doan. Another contrbuton of ths paper s that ntroduces a knd of ndeternate ethodology nto servce selecton proble. Such rando approach has features ncludng versltude of the real world, effcency of coputaton, and stablty to be controlled. The ethod n ths paper s applcable only to the Posson process queue odel M/M/, so the general queue odel, G/G/, needs further study. ACKNOWLEDGMENT Yn Shen would lke to thank Sen Zeng, We Tan, John He, Jan Zhou, Supng Jang, Xaoyan Bn, and Chuanzhen Zang for ther coents whch led to proveents of ths paper. [] M. P. Papazoglou, Servce-orented coputng: concepts, characterstcs and drectons, n IEEE Proceedngs of the Fourth Internatonal Conference on Web Inforaton Systes Engneerng, Netherlands, Dec. 3, pp. 3-. [3] W.T. Tsa, C. Fan, Y. Chen, R. Paul, J.-Y. Chung, Archtecture classfcaton for SOA-based applcatons, n Proceedngs of the Nnth IEEE Internatonal Syposu on Object and Coponent-Orented Real-Te Dstrbuted Coputng, Korea, Apr. 6, pp.. [] D. Booth, H. Haas, F. McCabe, E. Newcoer, M. Chapon, C. Ferrs, D Orchard. (, February ). Web servces archtecture W3C workng group note [Onlne]. Avalable: ws-arch/ [5] L.-J. Zhang, B. L, C. Tan, H. Chang, On deand web servces-based busness process, IEEE Internatonal Conference on Systes, Man and Cybernetcs, vol., Oct. 3, pp [6] F. L, F.-C. Yang, S. Su. On dstrbuted servce selecton for QoS drven servce coposton, n E-coerce and Web Technologes Proceedngs Lecture Notes n Coputer Scence, Poland, vol., Sep. 6, pp. 73. [7] T. Yu, K.-J. Ln. Servce selecton algorths for coposng coplex servces wth ultple QoS constrants, Servce-Orented Coputng ICSOC 5 Lecture Notes n Coputer Scence, Netherlands, vol. 36, Dec. 5, pp 3-3. [] H.-C. Wang, C.-S. Lee, T.-H. Ho, Cobnng subjectve and objectve QoS factors for personalzed web servce selecton. Expert Systes wth Applcatons, vol. 3, no., pp. 57-5, Feb. 7. [9] S.-L. Lu, Y.-X. Lu, N. Jng, G.-F. Tang, Y. Tang. A dynac web servce selecton strategy wth QoS global optzaton based on ult-objectve genetc algorth, n Grd and Cooperatve Coputng - GCC 5 Lecture Notes n Coputer Scence, Bejng, vol. 3795, Nov. 5, pp. 9. [] M. Couzz, B. Pernc. Negotaton support for web servce selecton, n Technologes for E-Servces Lecture Notes n Coputer Scence, Toronto, vol. 33, Aug., pp [] Y. Zhang, S.-S. Zhang, S.-Q. Han, A new ethodology of QoS evaluaton and servce selecton for ubqutous coputng, n Wreless Algorths, Systes, and Applcatons Lecture Notes n Coputer Scence, X an, vol. 3, Aug. 6, pp [] M. Ouzzan, A. Bouguettaya, Effcent access to web servces, IEEE Internet Coputng, vol., no., pp. 3-, Mar.. [3] Y.-J. Seo, H.-Y. Jeong, Y.-J. Song, Best web servce selecton based on the decson akng between QoS Crtera of servce, n Ebedded Software and Systes - ICESS 5 Lecture Notes n Coputer Scence, X an, vol. 3, Dec. 5, pp. -9. [] M. C. Huebscher, J. A. McCann, Usng real-te dependablty n adaptve servce selecton, n IEEE Proceedngs of the Jont Internatonal Conference on Autonoc and Autonoous Systes and Internatonal Conference on Networkng and Servces, Taht, Oct. 5, pp. 76. [5] L.-H. Vu, M. Hauswrth, K. Aberer, QoS-based servce selecton and rankng wth trust and reputaton anageent, n On the Move to Meanngful Internet Systes Proceedngs Lecture Notes n Coputer Scence, vol. 376, Oct. 5, pp [6] P. A. D. Castro, M. G. Pres, H. A. Canlargo, O. M. Jr. E. R. R. Kato, Genetc learnng of fuzzy rules appled to sequencng proble of FMS, IEEE Internatonal Conference on Systes, Man and Cybernetcs, Cyprus, vol. 5, Oct., pp [7] Y.-W.; Zhong, J.-Y. Yang, A hybrd genetc algorth wth laarckan ndvdual learnng for tasks schedulng, IEEE Internatonal Conference on Systes, Man and Cybernetcs, vol., Oct., pp [] J.-H. Wang, The Theory of Gaes. Oxford: Oxford Unversty Press, 9, pp. 7. [9] D. Gross, C. M. Harrs, Fundaentals of Queung Theory. New York: John Wley & Sons, 95, pp. 79. [] N. Kararkar, A new polynoal-te algorth for lnearprograng, Algorthca, vol., no., 9, pp REFERENCES [] Z. Stojanovc, A. Dahanayake, H. Sol. Modelng and desgn of servce-orented archtecture, IEEE Internatonal Conference on Systes, Man and Cybernetcs, vol. 5, Oct., pp