22 45th Hawa Internatonal Conference on System Scences Effcent QoS Aggregaton n Servce Value etworks Steffen Haak Research Center for Informaton Technology (FZI) haak@fz.de Benjamn Blau SAP AG benjamn.blau@sap.com Abstract In recent years, the trend towards standardzaton, smplfcaton and modularzaton n the servce sector has fostered the rase of Servce Value etworks where provders and consumers jontly co-create value. Wth many dfferent competng servces avalable, the user experence, whch s captured by the non-functonal Qualty-of-Servce (QoS) attrbutes, s an mportant compettve factor. QoS computaton for complex Web servces,.e. the aggregaton of QoS factors from atomc servces, s essental for an automated an optmzed Web servce selecton process. However, the computatonal complexty of QoS aggregaton has often been dsregarded n the respectve feld of research, whereas computatonal effcency s nevtable for the applcaton of optmzaton approaches n on-lne scenaros. The threefold contrbuton of ths paper conssts of an elaboraton on the computatonal complexty of aggregatng QoS, an approxmaton scheme that allows for a computatonal effcent optmzaton and a broad analytcal and smulaton-based evaluaton of ths approach.. Introducton For a long tme, the development and consumpton of Web servces have been complex actvtes exclusvely performed by techncal experts. Wth the rase of lght-weght technologes such as RESTful archtectures [, 2] and fat free message exchange formats such as JSO [3], a trend towards radcal smplfcaton of servce creaton and consumpton s observable. Ths trend fosters ecosystems of prosumers that collaboratvely develop and consume dynamc servces n networked envronments,.e. Servce Value etworks (SVs) [4]. SVs are networked economes consttuted of dstrbuted servce provders and consumers. In these envronments, servces are dynamcally composed nto value-added complex servces fulfllng the need of a long-tal of ndvdual customers [5]. One of the key characterstcs of servces n general s ther ntangble nature. More precsely, value creaton through servces s domnated by ntangble elements [6]. The consumer of a servce experences the performance or actvty whch embodes the man porton of created value [7, 8]. From a techncal perspectve, such ntangble elements experenced by the customer are emboded by the Qualty-of-Servce (QoS) dmenson. When dynamcally composng dstrbuted servces nto complex servces, the computaton of the QoS s a central task n SVs. Especally the computatonal complexty of ths task has often been gnored n the respectve feld of research, whereas computatonal effcency s nevtable for the applcaton n on-lne scenaros. Such an on-lne scenaro could be a real-tme servce ntegraton platform, allowng for an autonomc composton of complex servces out of atomc servces based on customer reuests. Typcally the customer specfes hs functonal needs by askng for a set of so called canddate pools, whch cluster functonally euvalent servce alternatves. An example reuest could be for a complex servce consstng of storage, bllng and payment. The challenge s not only to fnd techncally feasble solutons, moreover fndng the optmal composton from a QoS perspectve s a cumbersome task, reurng en effcent aggregaton of QoS values from the sngle servces. Tacklng the challenge of aggregatng QoS n the context of servce composton n SVs, the contrbuton of the work at hand s threefold: Layng the groundwork and dentfyng the current research gap, we () elaborate the computatonal complexty of aggregatng QoS dependng on feasble attrbute types and assess possble optmzaton approaches. We () provde an optmzaton approach based on an approxmaton scheme that restores computatonal effcency n case of attrbute types wth P-hard complexty regardng ther aggregaton. Fnally, we () evaluate our approach analytcally and smulatonbased regardng the error that arses from the approxmaton n dfferent determnstc and stochastc scenaros. The remander of ths work s structured as follows. Secton 2 covers the related work n ths area. We then present n Secton 3 the foundatons and challenges of aggregatng QoS by descrbng typcal classes of QoS attrbutes and common approaches for optmzng 978--7695-4525-7/2 $26. 22 IEEE DOI.9/HICSS.22.237 52
QoS-aware servce compostons by means of Graph algorthms and lnear programmng technues. Secton 4 contans our man contrbuton, a computatonal effcent approxmaton for aggregatng multplcatve attrbutes along wth a broad evaluaton of the approxmaton followng n Secton 5. The paper s concluded n Secton 6 wth remarks on the practcal mplcatons and an outlook on our next steps wthn ths research. 2. Related Work The challenge of aggregatng QoS attrbutes of atomc servces n varous process patterns s a wellnvestgated feld n BPM and servce composton research. In the context of automatc servce composton dfferent approaches have been proposed n recent lterature. Mostly backward channg s appled to derve sutable compostons startng from a central objectve [9, ]. In most of the cases such models are based on formal descrpton languages focusng on functonal servce characterstcs as proposed by the W3C recommendaton SAWSDL, OWL-S and WSMO. Other approaches such as [] also ncorporatng the servce lfecycle aspect and tme dependences. The man scope of all of these approaches s dealng wth servce functonalty as the only crtera for composton and largely gnores other non-functonal or QoS propertes. Focusng on mathematcal models for aggregatng ualty attrbutes of sngle servces nto complex servces dependng on multple types of process patterns are nvestgated n [2-5]. Although ths stream of research consders dfferent types of attrbutes and the mplcatons for a correspondng aggregaton algorthm, computatonal complexty and desred effcency s not n scope of ther nvestgaton. Another stream of research targets a more comprehensve soluton for managng functonal and non-functonal (.e. QoS) servce characterstcs across the entre lfecycle based on a model of atomc and composte servces [6, 7]. Although outlned approaches also focus on automaton and on-lne computaton of QoS aggregaton, complexty aspects and effcent algorthm desgn s not n the focus. [8] propose an lnear programmng (LP) approach that enables an automated Web servce composton whle maxmzng the user experence whch s modeled as QoS dependent utlty functon. The authors also descrbe how dfferent types of QoS values can be aggregated n such an optmzaton scenaro. They argue that probablty values lke the relablty or avalablty of a servce can be aggregated usng the followng functon: e In order to nclude the aggregaton nto a LP formulaton, the logarthm functon s appled. However, ths seems not to be reasonable, as the result s an addton of the QoS attrbutes, rather than a multplcaton. 3. Foundaton & Challenges We dentfy fve dfferent types of attrbute types that are categorzed by ther aggregaton functon (cf. Table ). Table : Classes of QoS attrbutes Type Aggregaton Functon Addtve Multplcatve Average Mn mn (, 2,...,) Max max (,,..., ) 2 Common representatves for the addtve attrbutes are the response tme or the costs of a servce. Typcal attrbutes that are aggregated n a multplcatve way are the avalablty or relablty of a servce, whch are expressble as probablty values. Reputaton or user ranks can be aggregated usng the average functon. Representatves for the mn or max functon are the throughput of a servce or the encrypton level. In the wnner determnaton problem, as t occurs n multattrbutve auctons [9] or n the complex servce aucton [2], the welfare or utlty maxmzng confguraton or path has to be calculated. In optmzaton scenaros for QoS-aware composton of Web servces [8] or custom cloud servces [2], the challenge s to fnd the optmal complex servce confguratons wth respect to user-defned QoS preferences. There are two dfferent approaches to tackle ths problem. In a more graph orented approach, as t s sutable for fndng the best offer n a Servce Value etwork, graph algorthms can be used to fnd the best offer by calculatng the optmal path through the 53
network. [2] propose the use of the Djkstra algorthm [22] to compute the shortest path wthn polynomal tme complexty, however, ths can only be acheved under the assumpton of addtve and monotone aggregaton operatons, as Djkstra reles on the Bellman property [23]. Ths however hnders the ncluson of any QoS attrbute other than the ones that fall nto the class of addtve aggregaton functons (cf. Table ). Algorthms, that calculate the optmal path n networks wthout the Bellman property, are known to be P-hard. In order to capture other non-bellman attrbute classes, an extenson to Djkstra s proposed n [24], yet the P-hard complexty s thereby traded for exponental space complexty. [8] and [2] user lnear programmng and lnear nteger programmng technues to compute the optmal servce composton. Dependng on the problem nstantaton, these problems can stll be computed n polynomal tme. The objectve functon, however, s reured to be of lnear form. For maxmzng or mnmzng an addtve QoS attrbute, the objectve functon typcally has the followng form, where x,, x n denote bnary decson varables. max x Whle there are lnearzaton methods for both mn and max functon avalable, multplcatve attrbutes cannot be ncluded nto a lnear objectve functon of such an optmzaton problem. 4. Proposed Soluton In order to cope wth the challenges we presented n the precedng secton, a lnear functon approxmatng the multplcaton result s desrable. We thereby can make use of some neat propertes of the attrbutes as they occur n our doman of nterest. QoS attrbutes that reure the multplcaton operaton are commonly probabltes or ratos lke the average avalablty, falure rate, etc. Common to these values s there doman whch les n the nterval [, ] and ther proxmty to the value of one. Ths s especally the case w.r.t. the avalablty of servces or the probablty of success of servce components, whch s the nverse of ther falure rate. In almost all cases these values are above 99%. onetheless, these values have to be aggregated, as there s a huge dfference between an aggregated value of 99.5% and 99.9%. Seen as avalablty over the perod of one year, ths slght change would account for an addtonal outage tme of 35 hours. In Web servce scenaros where downtme s extreme costly, e.g. a bankng servce, these addtonal hours can uckly cost up to mllons, thus a credble aggregated value becomes nevtable. When composng a servce out of several servces or servce components, we therefore need to aggregate all the ndvdual QoS attrbutes to allow us to make a statement on the resultng overall performance of the complex servce composton. The exact aggregaton value for probablty values,, (assumng stochastc ndependence of the attrbute values of the dfferent servce components) can be calculated as follows: Ths aggregaton functon, beng non-lnear, however s not effcent computable n optmzaton scenaros wth many dfferent servce composton alternatves. In those scenaros lnear programmng (LP) technues or graph algorthms are appled. However they reure lnearty of the objectve functon for the LP approach or Bellman optmalty respectvely. In ether case, solutons exst to lnearze the problem or to cope wth the state dependency wthn graph algorthms lke Djkstra [22], yet these solutons come at the cost of space complexty, thus not reducng the overall complexty, whch s known to be P-hard. 4.. Approxmaton For achevng an effcent computaton,.e. a soluton lyng wthn polynomal tme complexty, a lnear approxmaton can be of great beneft. Recapturng the typcal characterstcs of the attrbutes that are aggregated by multplcaton, the followng lnear approxmaton functon offers computatonal effcency whle smultaneously only yeldng a small error, dependng on the concrete problem nstantaton: a α a denotes the aggregated value, the number of attrbutes, are the attrbute values and parameter α can be used to parameterze the functon accordng to the doman of attrbute values, whch allows to reduce the expected error of the approxmaton functon. Values for α that acheve a mnmum expected error typcally le wthn the nterval [, ]. The functon can be further smplfed by pullng out the constants from the summaton term: a α +α Other than the exact aggregaton, ths approxmaton s lnear and thus can be ncluded n any 54
LP formulaton or dynamc programmng approach that reures the Bellman property. 5. Evaluaton We evaluate the proposed approxmaton analytcally and by runnng dfferent smulatons. For measurng the error the followng error functon s defned: a e, a Based on ths error functon, we conducted varous senstvty analyses wth dfferent parameter settngs. For reducng the mathematcal complexty and a better understandng of the reader, we set α to a value of one for some parts of the evaluaton, as we can draw the same conclusons wthout losng generalty. We then obtan a smplfed aggregaton functon wth α eualng one: a + 5.. Determnstc Values For a frst analytc error estmaton we assume the attrbute values to be determnstc and parameter α to be one. Thus the error can be calculated straghtforward wthout the use of stochastc theory. In case all attrbute values are eual,.e. 2 n, we obtan: a e, + Ths euaton can be smplfed to: a + e, The resultng error as a functon of and s depcted n Fgure. One can see that the error ncreases wth and decreases wth. Fgure : Determnstc error e (n %, all eual) However, n realty, not all values are eual. One way to measure ths effect n a determnstc manner, s by spreadng the attrbute values eually n the nterval [ ε, + ε], wth ε + ε We can derve the followng error functon, whch shows the nfluence of the values beng spread apart by two tmes ε. 2 a e, +ε + ε 2 2 +ε 2 2 2 2 Fgure 2 shows that the error of the approxmaton actually frst decreases wth epslon, untl a value for ε s reached, where the approxmaton evaluates to the exact same result as the exact value of the product. After ths pont, the error ncreases agan. The exact locaton of ths root depends on the number of aggregated attrbutes and ther mean value. The locaton of the root forms a valley where a decreasng value of corresponds to an ncreasng value of ε (Fgure 3). From a practcal pont of vew, ths s an advantageous property, as realstc values commonly fall nto ths valley. 55
.25 5 of the approxmaton, whch s only gven for the range bl bu :.2.5..5.5.4.3..2.3.4.5 epslon 2 a ( ) e S,S b b b b n bl bl U L U L 3 n bl By smplfyng the above euaton to a ( ) e S,S b U b + b 3 L L.2..4.3.2...2.3.4.5 epslon one can see that the error s actually ndependent on the upper bound b U, yet reures a lmt analyss where s close to b L. Fgure 4 shows the error n dependency on b L and. In most cases, t s not much hgher than the error wthout usng a scorng functon, yet wth a close proxmty of to b L, the error rases towards nfnty. Ths mples that the proposed approxmaton functon s not sutable, where the exact value of the aggregaton functon s very close to the lower bound of the scorng functon. However, from a practcal pont of vew, ths case can be consdered very unlkely...2.3.4.5 epslon Fgure 2: Determnstc error n %,.995 So far, we analyzed the error from an absolute pont of vew. In some scenaros, e.g. mult-crtera optmzaton problems, a utlty or scorng functon that transforms the aggregated value nfluences the conseuent error n the applcaton. A typcal scorng functon defnes upper (b U ) and lower bounds (b L ) for the aggregated attrbute value, wth a lnear utlty progresson n between these bounds: bl bu L bu bl > bu < S b b Ths transformaton n fact zooms nto the area of nterest, scalng up small dfferences n the attrbute value, thus also rasng the expected error. Usng the same smplfcaton as above, we can derve the error U 5.2. Stochastc Values In the last subsecton we presented evaluaton results that were relyng on determnstc values for the QoS attrbutes that have to be aggregated. In realty, however, we are confronted wth values that orgn from some stochastc or random process wth an underlyng probablty dstrbuton. I.e. not all are eual, nor are the spread evenly on a gven nterval. However, these determnstc consderatons allowed for some much smpler analytc conclusons, whch become far more complex n the stochastc case. As the more sophstcated analytc consderatons reure a numercal ntegral evaluaton wth an ncreasng complexty n the number of varables, we also rely on Monte Carlo smulatons wthn ths subsecton. In the center of nterest s agan the error arsng from usng the approxmaton nstead of the exact multplcaton functon. In the stochastc case, ths error s not a determnstc value but a probablty functon. The expected value of ths functon thereby resembles a comparable result to the determnstc error value. 56
Fgure 3: Determnstc error n % (5) Fgure 4: Error wth scorng functon n % () 57
Ths expected error, whch s based on ndependently dstrbuted ualty values, can be calculated by buldng the ntegral over all random varables. We thereby have to multply the densty functons of all random varables (.e. the actual probablty of obtanng a gven value) tmes the error that results of ths random varable nstantaton. Wth f ( ) beng the probablty densty functon of varable, we obtan the followng formula: a ( ) E e, a f ( ) f ( ) d d... For two varables (2) whch are dstrbuted unformly,.e. ~ 2 ~ U(a,b), we obtan: a a 2 ( ) a E e, b b 2+ + b a 2 2 2 d d 2 Fgure 5 shows a numercal evaluaton of the nterval n dependence on a and b. The worst case expected error n ths fgure wth a.8 and b.9 evaluates to 3.62%, whereas a typcal settng wth a dstrbuton between a.99 and b.999 evaluates to a low expected error value of.3%. Analogously one can estmate the expected error for other dstrbutons and dfferent numbers of varables. However the computatonal complexty for evaluatng the ntegral strongly ncreases wth. Fgure 5: Stochastc error n % 2, ~ 2 ~ U(a,b) Table 2 gves an overvew on the expected error for the unform dstrbuton, for two to fve varables and dfferent parameters for the unform dstrbuton (a, b). Table 2: Stochastc error ( ~ U(a,b)) a b Error n % 2.9.99.348 2.99.999.3 3.9.99.85 3.99.999.9 4.9.99 2.257 4.99.999.9 5.9.99 3.9 5.99.999.3 In Table 3 the expected error for two and three normally dstrbuted varables s shown, wth dfferent parameters for the mean value and varance of the normal dstrbuton. All error values are calculated by means of numercal ntegraton. Table 3: Stochastc error ( ~ (μ, )) μ σ 2 Error n % 2.9..979 2.9..238 2.99..698 2.99..4 2.999..666 2.999..6 3.9. 5.56 3.9. 3.988 3.99..296 3.99.5.37 3.999..26 3.999..282 As a last evaluaton, we conducted a Monte Carlo smulaton study to expose the relatonshp of mean value, varance and parameter α on the expected error. All smulatons are based on the average of, runs. As shown n Fgure 8, the error ncreases slghtly wth the varance and decreases wth the mean value. The depcted fgures represent the approxmaton error for fve and ffteen aggregated varables respectvely. Another smulaton run, depcted n Fgure 6, shows the nfluence of a growng Servce Value etwork,.e. a growng number of aggregaton varables for a typcal settng wth a mean value of 99.5% and a varance of.. In ths realstc scenaro we can see, that the proposed approxmaton easly scales up to 3, meanng that we have 3 dfferent Web servce canddate pools, each of them possbly havng many dfferent concrete Web servce nstance alternatves. If 58
Error n % 4 3.5 3 2.5 2.5.5 Mean Error 2 3 4 5 Fgure 6: Stochastc error n % ~ (.995,.) each canddate pool conssted of dfferent choces, one would have to optmze over 3 composton alternatves, effcently tractable usng our proposed approxmaton scheme, yet only yeldng an expected error value of.27 %. As stated n Secton 4., the approxmaton can be further mproved by adjustng the parameter α. In Fgure 7, one can see that parameter α can be effectvely used to reduce the error even for low mean values, where a lower value of α s more approprate when the QoS attrbutes are low as well. 6. Concluson In the precedng sectons we descrbed the mportance of QoS aggregaton n SVs, Web and Cloud servce compostons. It became clear, that n both wnner determnaton and other optmzaton approaches, computatonal effcency s an mportant matter. Yet, current lterature does not provde answers for effcently aggregatng multplcatve QoS attrbutes. We therefore proposed an approxmaton functon, whch s very sutable for hgh probablty values close to one, as they are usually found for the avalablty and other probabltes n the context of Web or Cloud servces. A detaled evaluaton explaned dfferent nfluences on the expected error, lke the mean value, the varance or the ntroducton of a scorng functon. In addton, we presented how parameter α wthn the approxmaton functon can be used to acheve even more precse results. Approaches as descrbed n [8, 2, 25] can be of great beneft, as they enable a customzed servce confguraton that maxmzes the user experence by optmzng the resultng QoS accordng to the customer preferences. However, n any on-lne scenaro, these approaches are ether lmted to small problem szes or restrcted to addtve QoS attrbutes, both resultng n sub optmal servce allocatons. Even n offlne Mean Error 6 Error n % 5 4 3 2.9.9.92.93.6.5 Mean Value.94.95.96.97.98.9.8.7 Alpha.99 Fgure 7: Stochastc error n %, ~ (μ,.), 59
scenaros, where computatonal tme s less mportant, computatonal effcency can be of great beneft, as computatonal tme s costly or the model complexty can be ncreased n other aspect due to the greater effcency of QoS aggregaton. In ths context our approach s strkng as t provdes a precse approxmaton scheme whch elmnates the exponental problem complexty to polynomal tme. Hence, computatonal effort can be reduced from mnutes to mllseconds, whch boosts wnner determnaton and effcent allocaton n SVs. Our current approach s lmted to values close to Error n %.25.2.5..5 Mean Error 5.2.5..5.9.9.92.93.8. Mean Value.94.95.96.97.98.2.4.6 Varance.99 Error n % 5 4 3 2 Mean Error 5 4 3 2.9.9.92.93.8. Mean Value.94.95.96.97.98.2.4.6 Varance.99 Fgure 8: Stochastc error n %, ~ (μ, 2 ), 5 (above) and 5 (below) 52
one. It wll be nterestng to fnd out, f smlar or totally dfferent lnear approaches can be found for other domans. From a practcal pont of vew, an evaluaton n an appled optmzaton scenaro s planned, where the error of the approxmaton can actually be monetzed and compared to the costs of an exact soluton. 7. References [] C. Pautasso, O. Zmmermann, and F. Leymann, "Restful web servces vs. bg'web servces: makng the rght archtectural decson," n Proceedng of the 7th Internatonal World Wde Web Conference, 28, pp. 85-84. [2] L. Rchardson and S. Ruby, RESTful web servces: O'Relly Meda, 27. [3] D. Crockford, "JSO: The fat-free alternatve to XML," 26. [4] B. Blau, J. Kramer, T. Conte, and C. Van Dnther, "Servce value networks," n Proceedngs of the IEEE Conference on Commerce and Enterprse Computng, 29. CEC '9, 29, pp. 94-2. [5] J. Kraemer, T. Conte, B. Blau, C. Van Dnther, and C. Wenhardt, "Servce value networks: Unleashng the combnatoral power of servce mashups." [6] T. Menl and B. Blau, "Web servce dervatves," n Proceedngs of the 8th Internatonal World Wde Web Conference, 29, pp. 27-28. [7] C. H. Lovelock and J. Wrtz, Servces marketng: Prentce Hall, 99. [8] A. M. Rushton and D. J. Carson, "The marketng of servces: managng the ntangbles," European Journal of Marketng, vol. 23, pp. 23-44, 989. [9] F. Lécué and A. Léger, "A formal model for semantc web servce composton," The Semantc Web-ISWC 26, pp. 385-398, 26. [] E. Srn, B. Parsa, D. Wu, J. Hendler, and D. au, "HT plannng for web servce composton usng SHOP2," Web Semantcs: Scence, Servces and Agents on the World Wde Web, vol., pp. 377-396, 24. [] D. Berard, D. Calvanese, G. De Gacomo, M. Lenzern, and M. Mecella, "Automatc servce composton based on behavoral descrptons," Internatonal Journal of Cooperatve Informaton Systems, vol. 4, pp. 333-376, 25. [2] M. B. Blake and D. J. Cummngs, "Workflow composton of servce level agreements," n Proceedngs of the Internatonal Conference on Servces Computng, 27, pp. 38-45. [3] M. C. Jaeger, G. Rojec-Goldmann, and G. Mühl, "Qos aggregaton for web servce composton usng workflow patterns," n Proceedngs of the 8th IEEE Internatonal Enterprse Dstrbuted Object Computng Conference, 24. [4] T. Unger, F. Leymann, S. Mauchart, and T. Schebler, "Aggregaton of servce level agreements n the context of busness processes," n Proceedngs of the 2th Internatonal IEEE Enterprse Dstrbuted Object Computng Conference, 28, pp. 43-52. [5] R. Knapper, B. Blau, S. Speser, T. Conte, C. Wenhardt, Servce Contract Automaton Proceedngs of the 6th Amercas Conference on Informaton Systems (AMCIS), 2. [6] A. Ludwg and B. Franczyk, "COSMA An Approach for Managng SLAs n Composte Servces," Servce- Orented Computng ICSOC 28, pp. 626-632, 28. [7] V. Muthusamy, H. A. Jacobsen, T. Chau, A. Chan, and P. Coulthard, "SLA-drven busness process management n SOA," n Proceedngs of the 29 Conference of the Center for Advanced Studes on Collaboratve Research, 29, pp. 86-. [8] L. Zeng, B. Benatallah, M. Dumas, J. Kalagnanam, and Q. Z. Sheng, "Qualty drven web servces composton," n Proceedngs of the 2th nternatonal conference on World Wde Web, 23, pp. 4-42. [9] M. Bchler and J. Kalagnanam, "Confgurable offers and wnner determnaton n mult-attrbute auctons," European Journal of Operatonal Research, vol. 6, pp. 38-394, 25. [2] B. Blau, C. Wenhardt, C. van Dnther, T. Conte, and Y. Xu, "How to Coordnate Value Generaton n Servce etworks A Mechansm Desgn Approach," Busness & Informaton Systems Engneerng, vol., p., 2. [2] S. Haak and S. Grmm, "Towards Custom Cloud Servces," n Proceedngs of the 8th Extended Semantc Web Conference, ESWC 2, Heraklon, Crete, Greece, May 29-June 2, 2, 2. [22] E. W. Djkstra, "A note on two problems n connexon wth graphs," umersche mathematk, vol., pp. 269-27, 959. [23] R. E. Bellman, Dynamc programmng: Dover Pubns, 23. [24] B. S. Blau, "Coordnaton n servce value networks : a mechansm desgn approach", PhD thess, 29. [25] B. Blau, "Coordnaton n Servce Value etworks," Management, vol. 5, pp. 398-43, 29. 52