Quality Driven Web Services Composition

Transcription

1 Quaity Driven Web Services Composition Liangzhao Zeng University of New South Waes Sydney, Austraia Jayant Kaagnanam IBM TJ Watson Research Center New York, USA Bouaem Benataah University of New South Waes Sydney, Austraia Maron Dumas Queensand University of Technoogy Brisbane, Austraia Quan Z Sheng University of New South Waes Sydney, Austraia qsheng@cseunsweduau ABSTRACT The process-driven composition of Web services is emerging as a promising approach to integrate business appications within and across organizationa boundaries In this approach, individua Web services are federated into composite Web services whose business ogic is expressed as a process mode The tasks of this process mode are essentiay invocations to functionaities offered by the underying component services Usuay, severa component services are abe to execute a given task, athough with different eves of pricing and quaity In this paper, we advocate that the seection of component services shoud be carried out during the execution of a composite service, rather than at design-time In addition, this seection shoud consider mutipe criteria (eg, price, duration, reiabiity), and it shoud take into account goba constraints and preferences set by the user (eg, budget constraints) Accordingy, the paper proposes a goba panning approach to optimay seect component services during the execution of a composite service Service seection is formuated as an optimization probem which can be soved using efficient inear programming methods Experimenta resuts show that this goba panning approach outperforms approaches in which the component services are seected individuay for each task in a composite service Categories and Subect Descriptors H35 [Information Systems]: Web-based services Genera Terms Management, Performance Keywords Web services, QoS, Service Composition 1 INTRODUCTION Web services technoogies are emerging as a powerfu vehice for organizations that need to integrate their appications Copyright is hed by the author/owner(s) WWW2003, May 20 24, 2003, Budapest, Hungary ACM /03/0005 within and across organizationa boundaries In particuar, the process-based composition of Web services is gaining a considerabe momentum as an approach for the effective integration of distributed, heterogeneous, and autonomous appications [1] In this approach, appications are encapsuated as Web services and the ogic of their interactions is expressed as a process mode This approach provides an attractive aternative to hand-coding the interactions between appications using genera-purpose programming anguages A Web service is a sef-described appication that uses standard Internet technoogies to interact with other Web services An exampe of a Web service is a SOAP-based interface to pace bids in an auction house Once depoyed, services can be aggregated into composite services An exampe of a composite service woud be a Trave Panner system that aggregates mutipe component services for fight booking, trave insurance, accommodation booking, car renta, and itinerary panning, which are executed sequentiay or concurrenty The process mode underying a composite service identifies the functionaities required by the services to be composed (ie, the tasks of the composite service) and their interactions (eg, contro-fow, data-fow, and transactiona dependencies) Component services that are abe to provide the required functionaities are then associated to the individua tasks of the composite services and invoked during each execution of the composite service The number of services providing a given functionaity may be arge and constanty changing Consequenty, approaches where the deveopment of composite services requires the identification at design-time of the exact services to be composed are inappropriate The runtime seection of component services during the execution of a composite service has been put forward as an approach to address this issue [2, 6, 11] The idea is that component services are seected by the composite service execution engine based on a set of criteria However, previous approaches in this area have not identified a set of criteria (other than price and appication-specific criteria) for seecting Web services In addition, existing service seection approaches adopt a oca seection strategy, meaning that they assign a component service to an individua tasks, one at a time As a resut, these approaches are not abe to hande goba user con- 411

2 straints and preferences For exampe, the overa duration of the composite service execution shoud be minimized, or a given budget constraint shoud be satisfied In this paper, we present quaity-driven approach to seect component services during the execution of a composite service The saient features of our approach are: The rest of the paper is organized as foows Section 2 presents the service composition mode and defines some key concepts used throughout the paper Section 3 defines the service quaity criteria used for service seection and expains how the vaues of these quaity criteria can be computed for a given service Section 4 formuates the goba service seection probem and describes a inear programming method to efficienty sove it Section 5 presents a prototype impementation of the proposed approach, as we as a set of experiments comparing the goba panning approach with the oca seection approach Finay, Section 6 discusses reated work, and Section 7 draws some concusions 2 WEB SERVICE COMPOSITION MODEL In this section, we wi present some basic concepts in Web service composition first, then give some definitions on composite service execution panning 21 Composite services and communities A composite Web service is an umbrea structure aggregating mutipe other eementary and composite Web services, which interact with each other according to a process mode Foowing our previous work [2], we choose to specify the process mode of a composite service as a statechart [12] The choice of statecharts for specifying composite Web services is motivated by two main reasons: (i) statecharts have a we-defined semantics; and (ii) they offer the basic fow constructs found in contemporary process modeing anguages (ie, sequence, conditiona branching, structured oops, concurrent threads, and inter-thread synchronization) The first characteristic faciitates the appication of forma manipuation techniques to statechart modes, whie the second characteristic ensures that the service composition mechanisms deveoped in the context of statecharts, can be adapted to other process modeing anguages ike, for exampe, those that are being designed by Web services standardization efforts (eg, BPEL4WS, WSCI, BPML) 1 A statechart is made up of states and transitions In the proposed composition framework, the transitions of a statechart are abeed with events, conditions, and assignment operations over process variabes States can be basic or compound Basic states are abeed with invocations to Web A Web services quaity mode We propose an extensibe muti-dimensiona Web services quaity mode statecharts within them Specificay, compound states come services operations Compound states contain one or severa The dimensions of this mode characterize non-functiona properties that are inherent to Web services in a singe statechart within it whereas an AND-state contains in two favors: OR and AND states An OR-state contains genera: execution price, execution duration, reputation, reiabiity, and avaiabiity tended to be executed concurrenty Accordingy, OR-states severa statecharts (separated by dashed ines) which are in- are used as a decomposition mechanism for moduarity purposes, whie AND-states are used to express concurrency: Quaity-driven service seection In order to overcome the imitations of oca service seection outined above, they encode a fork/oin pair The initia state of a statechart is denoted by a fied circe, whie the fina state is we propose a goba panning approach In this approach, quaity constraints and preferences are assigned denoted by two concentric circes: one fied and the other to composite services rather than to individua tasks unfied within a composite service Service seection PSfrag is then repacementsa simpified statechart W specifying a Trave Panner formuated as an optimization probem and a inear composite t i Web Service is depicted in Figure 1 In this composite f programming method is used to compute optima service execution pans for composite services Experi- t t service, a search for attractions is performed in parae c with a fight and an accommodation booking After menta resuts show that the proposed service seection strategy significanty outperforms oca seection t d strategies these searching and booking operations are competed, the distance from the hote to the accommodation is computed, and either a car or a bike renta service is invoked Note that when two transitions stem from the same state (in this case the state t 4), they denote a conditiona branching, and the transitions shoud therefore be abeed with disoint conditions W s1 W s2 W s3 W s4 W s5 (W e1 ) (W e2 ) t 1 t 2 FightTicketBooking AttractionSearching t 3 t 4 t 5 HoteBooking DrivingTimeCacuation BikeRenta CarRenta Legend State And state Transition Initia State Fina State Figure 1: Statechart of a composite service Trave Panner A basic state 2 of a statechart describing a composite service can be abeed with an invocation to either of the foowing: An eementary Web service, ie, a service which does not transparenty rey on other Web services A composite Web service aggregating severa other services A Web service community, ie, a coection of Web services with a common functionaity athough different non-functiona properties (eg, with different providers, different QoS parameters, reputation, etc) The concept of Web service community addresses the issue of composing a arge and changing coection of Web services Service communities provide descriptions of a desired functionaity (eg, fight booking) without referring to any actua service (eg, Qantas fight booking Web service) 1 See 2 Aso caed task in remainder of this paper t 6 t 7 t 8 412

3 The set of members of a community can be fixed when the community is created, or it can be determined through a registration mechanism, thereby aowing service providers to oin, quit, and reinstate the community at any time When a community receives a request to execute an operation, this request is deegated to one of its current members The choice of the deegatee is done at execution time based on the parameters of the request, the characteristics of the members, the history of past executions, and the status of ongoing executions Sections 3 and 4 dea with the seection of deegatees during the execution of a composite service whose states are abeed with invocations to communities 22 Execution paths and pans In this section, we define two concepts used in the rest of the paper: execution path and execution pan Definition 1 (Execution path) An execution path of a statechart is a sequence of states [t 1, t 2, t n], such that t 1 is the initia state, t n is the fina state, and for every state t i (1 < i < n), the foowing hods: PSfrag repacements t i is a direct successor of one of the states in [t 1,,t i 1] t i t f t i is not a direct successor of any of the states in t a1 [t i+1,,t n] t a2 There is no state t in [t 1,, t i 1] such that t and t i beong to two aternative branches of the statechart If t i is the initia state of one of the concurrent regions of an AND-state AST, then for every other concurrent region C in AST, one of the initia states of C appears in [t 1,, t i 1, t i+1,, t n] In other words, when an AND-state is entered, a its concurrent branches are executed This definition reies on the concept of direct successor of a state A basic state t b is a direct successor of another basic state t a if there is a sequence of adacent transitions 3 going from t a to t b without traversing any other basic state In other words, the first transition in the sequence stems from t a, the ast transition eads to t b, and a intermediate transitions stem from and ead to either compound, initia, or fina states, but are not incident to a basic state It is straightforward to see that an acycic statechart has a finite number of execution paths To simpify the presentation, we initiay assume that a the statecharts that we dea with are acycic If a statechart contains cyces, a technique for unfoding it into an acycic statechart needs to be appied beforehand The method used to unfod the cyces of a statechart is to examine the ogs of past executions in order to determine the average number of times that each cyce is taken The states appearing between the beginning and the end of the cyce are then coned as many times as the cyce is taken in average Detais about this unfoding process are omitted for space reasons Under the assumption that the underying statechart is acycic, it is possibe to represent an execution path of this statechart as a Directed Acycic Graph (DAG) as foows W s1 W s2 W s3 W s4 W s5 Definition 2 (DAG representation of an execution path) Given an execution path [t 1, t 2, t n] of a statechart ST, the DAG representation of this execution path is a graph obtained as foows: The DAG has one node for each task t 1, t 2, t n The DAG contains an edge from task t i to task t iff t is a direct successor of t i in the statechart ST If a statechart diagram contains conditiona branchings, it has mutipe execution paths Each execution path represents a sequence of tasks to compete a composite service execution Figure 2 gives an exampe of a statechart s execution paths In this exampe, since there is one conditiona branching after task t 4, there are two paths, caed W e1 and W e2 respectivey In execution path W e1, task t 6 is executed after task t 5, whie in execution path W e2, task t 7 is executed after task t 5 t 1 t 1 t 2 t 3 t 2 t 3 t 4 Execution Path 1 t 4 t 6 t 5 t 5 Execution Path 2 (W e1 ) (W e2 ) Figure 2: DAG representation of the execution paths of the statechart of Figure 1 As stated before, the basic states of a statechart describing a composite service can be abeed with invocations to communities If this is the case, actua Web services (ie, members of communities) need to be seected during the execution of the composite service Hence, it is possibe to execute a path in very different ways by aocating different Web services to the states in the path The concept of execution pan defined beow captures the various ways of executing a given execution path Definition 3 (Execution pan) A set of pairs p = {< t 1, s i1 >, < t 2, s i2 >,, < t N, s in >} is an execution pan of an execution path W e iff: {t 1, t 2, t N} is the set of tasks in W e t 7 For each 2-tupe < t, s i > in p, service s i is assigned the execution of task t t 8 t 8 3 Two transitions are adacent if the target state of one is the source state of the other 413

4 3 WEB SERVICE QUALITY MODEL In a Web environment, mutipe Web services may provide simiar functionaities with different non-functiona property vaues (eg, different prices) In the composition mode presented in the previous section, such Web services wi typicay be grouped together in a singe community To differentiate the members of a community during service seection, their non-functiona properties need to be considered For this purpose, we adopt a Web services quaity mode based on a set of quaity criteria (ie non-functiona properties) that are transversa to a Web services, for exampe, their pricing and reiabiity Athough the adopted quaity mode has a imited number of criteria (for the sake of iustration), it is extensibe: new criteria can be added without fundamentay atering the service seection techniques buit on top of the mode In particuar, it is possibe to extend the quaity mode to integrate non-functiona service characteristics such as those proposed in [22], or to integrate service QoS metrics such as those proposed by [25] In this section, we first present the quaity criteria in the context of eementary services, before turning our attention to composite services For each criterion, we provide a definition, indicate its granuarity (ie, whether it is defined for an entire service or for individua service operations), and provide rues to compute its vaue for a given service 31 Quaity Criteria for Eementary Services We consider five generic quaity criteria for eementary services: (1) execution price, (2)execution duration, (3) reputation, (4) reiabiity, and (5) avaiabiity Execution price Given an operation op of a service s, the execution price q price(s, op) is the amount of money that a service requester has to pay for executing the operation op Web service providers either directy advertise the execution price of their operations, or they provide means to enquire about it Execution duration Given an operation op of a service s, the execution duration q du (s, op) measures the expected deay in seconds between the moment when a request is sent and the moment when the resuts are received The execution duration is computed using the expression q du (s, op) = T process(s, op) + T trans(s, op), meaning that the execution duration is the sum of the processing time T process(s, op) and the transmission time T trans(s, op) Services advertise their processing time or provide methods to enquire about it The transmission time is estimated based on past executions of the service operations, ie, T trans(s, op) = ni=1 T i (s,op), where T n i(s, op) is a past observation of the transmission time, and n is the number of execution times observed in the past Reiabiity The reiabiity q re (s) of a service s is the probabiity that a request is correcty responded within a the maximum expected time frame (which is pubished in the Web service description) Reiabiity is a technica measure reated to hardware and/or software configuration of Web services and the network connections between the service requesters and providers The vaue of the reiabiity is computed from historica data about past invocations using the Tabe 1: Aggregation functions for computing the QoS of execution pans Criteria Aggregation function Price Q price (p) = N i=1 q price(s i, op i ) Duration Q du (p) = CP A(q du (s 1, op 1 ),, q du (s N, op N )) Reputation Q rep(p) = 1 N N i=1 qrep(s i) Reiabiity Q re (p) = Π N i=1 (eq re(s i ) z i ) Avaiabiity Q av(p) = Π N i=1 (eqav (s i) z i ) expression q re (s) = N c(s)/k, where N c(s) is the number of times that the service s has been successfuy deivered within the maximum expected time frame, and and K is the tota number of invocations Avaiabiity The avaiabiity q av(s) of a service s is the probabiity that the service is accessibe The vaue of the avaiabiity of a service s is computed using the foowing expression q av(s) = T a(s)/θ, where T a is the tota amount of time (in seconds) in which service s is avaiabe during the ast θ seconds (θ is a constant set by an administrator of the service community) The vaue of θ may vary depending on a particuar appication For exampe, in appications where services are more frequenty accessed (eg, stock exchange), a sma vaue of θ gives a more accurate approximation for the avaiabiity of services If the service is ess frequenty accessed (eg, onine bookstore), using a arger θ vaue is more appropriate Here, we assume that Web services send notifications to the system about their running states (ie, avaiabe, unavaiabe) Reputation The reputation q rep(s) of a service s is a measure of its trustworthiness It mainy depends on end user s experiences of using the service s Different end users may have different opinions on the same service The vaue of the reputation is defined as the average ranking given by to the service by end users, ni=1 R ie, q rep = i, where R n i is the end user s ranking on a service s reputation, n is the number of times the service has been graded Usuay, the end users are given a range to rank Web services, for exampe, in Amazoncom, the range is [0, 5] Given the above quaity criteria, the quaity vector of a service s is defined as foows: q(s) = (q price(s), q du (s), q av(s), q re(s), q rep(s)) (1) Note that the method for computing the vaue of the quaity criteria is not unique The goba panning mode presented Section 4 is independent of these methods 32 Quaity Criteria for Composite Services The above quaity criteria are aso appied to evauate the QoS of composite services Tabe 1 provides aggregation functions for the computation of the QoS of a composite service CS when executed using pan p = {< t 1, s i1 >, < t 2, s i2 >,, < t N, s in >} A brief expanation of each criterion s aggregation function foows: 414

5 g repacements t a1 t a2 ti t f Execution price: The execution price Q price(p) of an execution pan p is a sum of every service s i s execution price q price(s i, op i) Execution duration: The execution duration Q du (p) of an execution pan p is computed using the Critica Path Agorithm (CPA) [23] Specificay, the CPA is appied to the the execution path of execution pan p, seen as a proect digraph The critica path of a proect digraph is a path from the initia state to the fina state which has the ongest tota sum of weights abeing its nodes In the case at hand, the weights abeing the nodes correspond to the maximum expected execution durations A task that beongs to the critica path is a critica task, whie a service that beongs to the critica path is a critica service Figure 3 provides an exampe of critica path In this exampe, the proect digraph represents execution path W e1 and its execution pan p, where p={ < t 2, s 23 >, < t 3, s 38 >, < t 4, s 45 >, < t 5, s 59 >, < t 6, s 62 > } Each task s execution duration is given in the proect digraph There are two proect paths in this proect digraph, where proect path 1 is < t 2, t 5, t 6 > and proect path 2 is < t 3, t 4, t 5, t 7 > The tota execution time of proect path 1 (proect path 2) is 37 seconds (62 seconds) Since proect path 2 s tota execution duration is onger than that of proect path 1, the critica path for the proect digraph is proect path 2 Thus, the execution pan s tota execution duration is 62 seconds Task t 3, t 4, t 5 and t 7 are critica tasks Services s 38, s 45, s 59 and s 62 are critica services s 13 s 28 s23, 20 Seconds s t 2 35 t 6 s62, 15 Seconds s 49 t 1 t t8 s 52 5 W s1 W s2 t s59, 2 Second 3 t 4 W s3 W s4 W s5 s38, 25 Seconds s45, 20 Second (W e1 ) (W e2 ) Legend: critica path of proect digraph critica task critica service Figure 3: Critica Path Reputation: The reputation Q rep(p) of an execution pan p is the average of each service s i s reputation q rep(s i) in the execution pan p Reiabiity: Reiabiity Q re (p) of an execution pan p is a product of e q re(s i ) z i In the aggregation function, z i is equa to 1 if service s i is a critica service in the execution pan p, or 0 otherwise If z i = 0, ie, service s i is not a critica service, then e q re(s i ) z i = 1, and hence, the reiabiity of service s i wi not affect the vaue of execution pan s reiabiity Avaiabiity: The avaiabiity Q av(p) of an execution pan p is a product of e qav(s i) z i, where q av(s i) is service s i s avaiabiity Using above aggregation functions, the quaity vector of a composite service s execution pan is defined as: Q(p) = (Q price(p), Q du (p), Q av(p), Q re(p), Q rep(p)) (2) 4 GLOBAL SERVICE SELECTION As mentioned before, in existing approaches, the seection of component service to execute a task is determined independenty to other tasks of composite services [2, 11, 6] More precisey, in our previous work [2], service seection is done at each service community ocay The seection of a service is based on a seection poicy invoving parameters of the request, the characteristics of the members, the history of past executions, and the status of ongoing executions Athough service seection can be ocay optimized, the goba quaity constraints may not be satisfied For exampe, a goba constraint such as composite services execution price is ess than 500 doars can not be enforced In this section, we present a goba panning based approach for Web services seection We first present an approach of seecting an optima execution pan for a composite service, then present a nove inear programming based method for optima execution pan seection 41 Seecting an Optima Execution Pan The basic idea of goba panning is the same as query optimization in database management systems Severa pans are identified and the optima pan is seected The foregoing discussion makes it cear that a statechart has mutipe execution paths and each execution path has its own set of execution pans if the statechart contains conditiona branchings In this subsection, we assume that the statechart does not contain any conditiona branchings and has ony one execution path We wi discuss the case where a statechart has mutipe execution paths in Section 42 We aso assume that for each task t, there is a set of candidate Web services S that are avaiabe to which task t can be assigned Associated with each Web service s i is a quaity vector (see equation 1) Based on the avaiabe Web services, by seecting a Web service for each task in an execution path, the goba panner wi generate a set of execution pans P : P = {p 1, p 2,, p n} (3) n is the number of execution pans After a set of execution pans is generated, the system needs to seect an optima execution pan When seecting the execution pan, instead of computing the quaity vector of a particuar Web service, each execution pan s goba service quaity vector needs to be computed The seection of execution pan uses Mutipe Attribute Decision Making (MADM)[16] approach Once the quaity vector for each execution pan is derived, by accumuating a the execution pans quaity vectors, we obtain matrix Q, where each row represents an execution pan s quaity vector Q 1,1 Q 1,2 Q 1,5 Q 2,1 Q 2,2 Q 2,5 Q = (4) Q n,1 Q n,2 Q n,5 415

6 A Simpe Additive Weighting (SAW) [4] technique is used to seect an optima execution pan Basicay, there are two phases in appying SAW: 411 Scaing Phase Some of the criteria used coud be negative, ie, the higher the vaue is, the ower the quaity is This incudes criteria such as execution time and execution price Other criteria are positive criteria, ie, the higher the vaue is, the higher the quaity is For negative criteria, vaues are scaed according to Equation 5 For positive criteria, vaues are scaed according to Equation 6 V i, = V i, = { Q max Q i, Q min 1 if { Qi, Q min Q min if 0 if 1 if = 0 0 = 0 = 1, 2 (5) = 3, 4, 5 (6) In the above equations, is maxima vaue of a quaity criterion in matrix Q, ie, = Max(Q i,), 1 i n is minima vaue of a quaity criterion in matrix Q, ie, Q min = Min(Q i,), 1 i n In fact, we can compute and Q min without generating a possibe execution pans For exampe, in order to compute the maximum execution price (ie, price) of a the execution pans, we seect the most expensive Web service for each task and sum up a these execution prices to compute price In order to compute the minimum execution duration (ie, Q min du ) of a the execution pans, we seect the Web service that has shortest execution duration for each task and use CPA to compute Q min du The computation cost of and Q min is poynomia After the scaing phase, we obtain the foowing matrix Q : V 1,1 V 1,2 V 1,5 Q V 2,1 V 2,2 V 2,5 = V n,1 V n,2 V n,5 412 Weighting Phase The foowing formua is used to compute the overa quaity score for each execution pan: Score(p i) = 5 (V i, W ) (7) =1 where W [0, 1] and 5 =1 W = 1 W represents the weight of each criterion In (7), end users can give their preference on QoS (ie, baance the impact of the different criteria) to seect a desired execution pan by adusting the vaue of W The goba panner wi choose the execution path which has the maxima vaue of Score(p i) (ie, max(score(p i))) If there are more than one execution pans which have the same maxima vaue of Score(p i), then an execution pan wi be seected from them randomy 42 Handing Mutipe Execution Paths In Section 41, we assume that the statechart ony has one execution path In this subsection, we discuss the case where statecharts have mutipe execution paths Assume that a statechart has n execution paths For each execution path, an optima execution pan can be seected So, the goba panner has n seected execution pans Since each seected optima execution pan ony covers a subset of the entire statechart, then the goba panner needs to aggregate these n execution pans into an overa execution pan to covers a the tasks in the statechart This overa execution pan wi be used to execute the statechart For exampe, for trave panner statechart W (see Figure 1), there are two execution paths W e1 and W e2 The optima execution pans p 1 and p 2 of these two execution paths are seected From the Figure 2, it can be seen that both execution paths W e1 and W e2 are subsets of W Thus neither p 1 nor p 2 covers a tasks in W Since the goba panner conducts panning before the execution time, it does not know which execution path wi eventuay be used for the composite service Therefore it needs to aggregate p 1 and p 2 into an overa execution pan which covers a the tasks in W Assume that statechart W has k tasks (ie, t 1, t 2,, t k ) and n execution paths (ie, W e1, W e2,, W en) Thus, for each execution path, the goba panner seects an optima execution pan Consequenty, we obtain n optima execution pans (ie, p 1, p 2,, p n) for these execution paths The goba panner adopts the foowing approach to aggregate mutipe execution pans into an overa execution pan 1 Given a task t i, if t i ony beongs to one execution path (eg, W e), then the goba panner seects W e s execution pan p to execute the task t i We denote this as t i(p ) For exampe, in trip panning statechart, task t 7 (ie,carrenta ) ony beongs to execution path W e2 In this case, W e2 s execution pan p 2 is used to execute t 7, ie, t 7(p 2) 2 Given a task t i, if t i beongs to more than one execution paths (eg, W e, W e+1,, W em), then there is a set of execution pans (ie, p, p +1,, p m) that can be used to execute W si In this case, the goba panner needs to seect one of the execution pans from p, p +1,, p m The seection can be done by identifying the hot path for task t i Here, the hot path of a task t i is defined as the execution path that has been most frequenty used to execute task t i in the past For exampe, in trave panner statechart, task t 3 (FightTicketBooking) beongs to both execution path W e1 and W e2 Assume that the statechart W has been used to execute the composite service for 25 times Aso assume that, in 20 times the execution of the composite service foows the execution path W e1; whie in 5 times, the execution of the composite service foows the execution path W e2 This indicates that execution path W e1 has been more frequenty used to execute task t 3 (ie, W e1 is the hot path for t 3) Thus, W e1 s execution pan p 1 is used to execute t 3, ie, t 3(p 1) The system keeps composite service execution traces in an execution history [10] This aows the goba panner to identify hot path for each task 43 Linear Programming Soution The approach of seecting an optima execution pan given in the previous section requires the generation of a possibe 416

7 execution pans Assume that there are N tasks in a statechart and there are M potentia Web services for each task The tota number of execution pans is M N The computation cost of seecting an optima execution pan is O(M N ) Such an approach is impractica for arge scae composite services, where both the number of tasks in the composite services and the number of candidate Web services in communities are arge For exampe, assume that a composite service has one execution path and 10 tasks, and for each task, there are 10 candidate Web services Then the tota number of execution pans is It is very costy to generate a these pans and seect an optima one In this subsection, we present a method based on inear programming (LP) [15], which can be used to seect an optima execution pan without generating a the possibe execution pans There are three inputs in LP: variabes, an obective function and constraints on the variabes, where both the obective function and constraints must be inear LP attempts to maximize or minimize the vaue of the obective function by adusting the vaues of variabes based on the constraints The output of LP is the maximum (or minimum) vaue of the obective function as we as the vaues of variabes at this maximum or minimum point In order to use LP to seect an optima execution pan, we mode the seection of an optima execution pan as an LP probem The variabes of the LP probem are y i representing the participation of service s i in the seected execution pan The vaue of each variabe y i is 1 if service s i is in the seected pan, 0 otherwise The obective function of the LP probem, which is based on equations 5,6, and 7, is: Max ( 2 =1 ( Q max Q i, W ) + 5 =3 ( Qi, W )) where W [0, 1] and 5 =1 W = 1 W is the weight assigned to quaity criteria In the foowing subsections, we discuss the constraints on the variabes of the LP probem 431 Constraints on Execution Duration and Execution Price In this subsection, we consider constraints on the execution duration and the execution price of an execution pan Assume that A is the set of a tasks (ie, basic states) of the statechart For each task t, there is a set of Web services S that can be assigned to this task, but on the end, for each task t, ony one Web service shoud be seected Given that y i (y i = 0 or 1) denotes the participation of Web service s i in the seected pan, this atter fact is captured by the foowing constraints: i S y i = 1, A (9) For exampe, there are 100 potentia Web services that can execute task, since ony one of them wi be seected to execute the task, then we have 100 i=1 yi = 1 Assume that variabe x represents the eariest start time of task t, variabe p represents the execution duration for task, and variabe p i represents the execution duration for task t by service s i We use the notation t t k to denote that task t k is task t s direct successor task We (8) have the foowing constraints: i S p i y i = p, A (10) x k (p + x ) 0, t t k,, k A (11) Q du (x + p ) 0, A (12) Constraint 10 indicates that the execution duration of a given task t is equa to the execution duration of one of the Web services in A Constraint 11 captures the fact that if task t k is a direct successor of task t, the execution of task t k must start after task t has been competed Constraint 12 indicates that the execution of a composite service is competed ony when a its tasks are competed Assume that z i is an integer variabe that has vaue 1 or 0: 1 indicates that Web service s i is a critica service and 0 indicates otherwise We have the foowing constraint on execution pan s execution duration Q du : Q du = p iz i (13) A i S For execution price, assume that variabe c i represents the execution price of Web service s i, then we have the foowing constraint on tota execution price of composition service: Q price = c i y i (14) A S An aternative of constraint 14 is as foows: A S c iy i B, B > 0 (15) where B is the budget constraint given by the user This constraint indicates that the entire composite service s execution price can not be greater than B By introducing a budget constraint the above probem needs to be expicity soved as an integer programming probem This probem is a specia case of the knapsack probem and hence it is NP-hard [19] Notice that constraints on other criteria can be easiy incorporated into LP if the aggregation function is a inear function For exampe, assume that variabe r i represents the reputation of Web service s i, we can have the foowing constraint on the execution pan s reputation: Q rep = r iy i (16) A S 432 Constraints on Reiabiity and Avaiabiity In this subsection, we consider constraints on criteria where the aggregation function is not a inear function Among the criteria that are used to seect Web services, both the avaiabiity s and the reiabiity s aggregation functions are non-inear (see Tabe 1) We can inearize them using a ogarithm function as shown beow Assume that variabe a i represents the reiabiity of Web service s i Since z i indicates whether Web service s i is a critica service or not, the reiabiity of execution pan is: Q re = Π A S e ai zi 417

8 By appying the ogarithm function n, we obtain: n(q re ) = n e ai zi A S Since A zi = 1 and zi = 0 or 1, we obtain: n(q re ) = a iz i A S Let Q re = n(q re ), we have the foowing constraint on the execution pan s reiabiity: Q re = a iz i (17) A S Simiary, assuming that b i represents the avaiabiity of the Web service s i, the foowing constraint is introduced: Q av = b iz i (18) A i S where Q av = n(q av) Criteria that can be introduced into the LP probem are not imited to what we defined in Section 3 Other criteria can be added once the aggregation functions are given 433 Constraints on Uncertainty of Execution Duration In the previous sections, we assume that the execution durations p i of Web services are deterministic In reaity, the execution duration of a Web service s i may be uncertain For exampe, Service s may advertise that the execution duration is 5 seconds, but the actua execution duration may be 45, 46, or 52 seconds To address this issue, we mode each p i using a norma distribution Variabe p i has therefore the foowing probabiity function: f(x) = 1 [ exp 1 2πσ 2 ( x µ ) 2 ], < x < σ where the mean µ and the std deviation σ are given by: σ 2 = 1 n 1 µ = 1 n n x i (19) i=1 n (x i µ) 2 (20) i=1 By appying formuas 19 and 20 to the execution ogs of a Web service s i, we can obtain µ i and σ i for this Web service Since i A i S p iz i = Q du, the tota execution duration Q du must foow a norma distribution 4 whose deviation σ du is: σdu 2 = σiz 2 i (21) i A i S So in order to consider the deviation of the tota execution duration in the LP probem, we shoud adopt the foowing obective function: 4 A detaied proof can be found in [26] Max ( 2 =0 ( Q max Q i, W ) + 5 =3 ( Qi, W )) where Q 0 = σ 2 du and W 0 [0, 1] is the weight assigned to the deviation of the tota execution duration Given the above inputs for the LP probem, the output of the LP sover wi be a set of vaues for variabes y i, which indicate the seection or excusion of Web services s i The seected Web services compose an optima execution pan 5 VALIDATION In this section, we present the impementation of the QoSdriven seection of services and some experimenta resuts to evauate the proposed approach 51 Impementation The proposed QoS-driven seection technique is impemented in the SELF-SERV prototype Detaied description of SELF-SERV can be found in [24] In this section, we briefy overview the prototype architecture and discuss its extension to support the service seection approach The prototype architecture (Figure 4) features an interface, a service manager and a poo of services Services communicate via Simpe Obect Access Protoco (SOAP) messages The service manager consists of four modues, namey the service discovery engine, the service editor, the composite service orchestrater and the goba panner The service discovery engine faciitates the advertisement and ocation of services It is impemented using the Universa Description, Discovery and Integration (UDDI), the Web Service Description Language (WSDL), and SOAP In the impementation, we make extensive use of the IBM Web Services Tookit 24 (WSTK24) [14], which is a showcase package for Web services emerging technoogies The service editor provides faciities for defining new services and editing existing ones A service is edited through a visua interface, and transated into an XML document for subsequent anaysis and processing by the service orchestrater The orchestrater is responsibe for scheduing, initiating, and monitoring the invocations to the tasks of a composite service during its execution, and for routing events and data items between these components Finay, goba panner is the modue that pans the execution of a composite service using the goba panning based approach The goba panner is impemented as a inear programming sover based on IBM s Optimization Soutions and Library (OSL) [13] It shoud be noted that QoS information is retrieved via pre-defined operations of services (eg, getexecutiontime()) 52 Experimentation We conducted experiments using the impemented prototype system to evauate our approach In the experiments, a custer of PCs were used to run the prototype system A PCs have the same configuration of Pentium III 933MHz with 512M RAM Each PC runs Windows 2000, Java 2 Edition V130, and Orace XML Deveoper Kit (Orace XDK, for XML parsing) They are connected to a LAN through 100Mbits/sec Ethernet cards This section presents two experimenta resuts The first experiment compares the QoS (22) 418

9 UDDI Service Manager Service Discovery Engine User Web based Interface requests/resuts Service Buider Composite Service Orchestrator Service Advertisements Service Discovery Registry Goba Panner Communication Bus Composite Services CS1 CS2 CS3 is composed of Communities C1 C2 C3 is member of Eementary Services ES1 ES2 ES3 ES4 Figure 4: Architecture of the prototype metrics of the execution of composite services using the goba panning and oca seection approaches The second experiment shows the system costs (ie, computation cost and bandwidth cost) of these two approaches 521 QoS Metrics The purpose of this experiment is to compare the QoS vaues of the goba panning approach with that of the oca seection The comparison was done by measuring price and execution time of the composite services using both approaches In the experiment, we created severa composite services with different number of basic states The services were created by randomy adding states to the composite service shown in Figure 1 The number of states ranges over the vaues 10, 20, 30, 40, 50, 60, 70, and 80 For each composite service, we executed the service 12 times and recorded the price and execution time Since we obtained simiar experimenta resuts for a created composite services, ony the resut of one composite service (with 20 states) is shown (see Tabe 2) for carity reasons From the tabe, we can see that for every instance of the composite service, the goba panning approach gives better QoS than oca seection approach For exampe, for the instance 7 of Tabe 2, the time required to execute the composite service is: (i) 6451 seconds in the goba panning approach, (ii) 9480 seconds in the oca seection approach Simiary, the execution price spends for executing this composite service is: (i) 1231 doars in the goba panning approach, (ii) 1789 doars in the oca seection approach Overa, the average execution time (resp, execution price) is: (i) seconds (resp, 1191 doars) in the goba panning approach, (ii) seconds (resp, 1753 doars) in the oca seection approach 522 System Costs The aim of this experiment is to investigate the system costs of executing composite services using the LP-based goba panning and oca seection approaches The experiment was done by measuring: (i) computation cost (ie, time used for seecting a Web service for each task of the composite service); (ii) bandwidth cost (ie, network bandwidth consumed between goba panner and Web services) Instance Q time(w ) (second) Q price(w ) (doar) No Goba Loca Goba Loca Average: Tabe 2: The QoS of a composite service with 20 states In the experiment, we created severa composite services with different number of tasks The services were created by randomy adding states to the composite service shown in Figure 1 The number of tasks ranges over the vaues 10, 20, 30, 40, 50, 60, 70, and 80 In addition, we created a set of Web services which were assigned to tasks of composite services as the candidate Web services For each task, the number of candidate Web services we used varies as foows: 5, 10, 20, and 40 services We executed composite services with different number of states and candidate Web services The computation and bandwidth costs for seecting Web services were recorded The resuts for composite services with 40 candidate Web services of each state are shown in figures 5 and 6 respectivey Simiar resuts were obtained for other cases From Figure 5, we can see that for both goba and oca seection approaches, the computation cost increases when the number of tasks and the number of candidate Web services increases As expected, the computation cost of goba panning approach is a itte bit higher than that of oca seection approach For exampe, a composite service with 40 tasks spends: (i) 16 seconds for seecting Web services in goba panning approach, (ii) 07 seconds in oca seection approach 419

10 Computation Cost (in seconds) Statechart has one execution path Number of Tasks in Statecharts Goba Panning Loca Seection Figure 5: The computation cost of seecting services for composite services (each state has 40 candidate Web services) Bandwidth Cost (in KBytes) Statechart has one execution path Number of Tasks in Statecharts Goba Panning Loca Seection Figure 6: The bandwidth cost of seecting service for composite services (each state has 40 candidate Web services) Simiar observations are found regarding bandwidth cost More specificay, for both approaches, the inear increase of the number of tasks and the number of candidate Web services eads amost a inear increase of bandwidth cost (see Figure 6) The bandwidth cost in goba panning is sighty higher than that of oca seection approach For exampe, a composite service with 40 tasks consumes about 2080 KBytes of network bandwidth for seecting Web services in goba panning approach, whie it consumes 1910 KBytes in oca seection approach 6 RELATED WORK In this section, we briefy discuss the reationships between our work and existing Web service standards, Web service composition approaches, and QoS-driven workfow management Severa standardisation proposas aiming at providing infrastructure to support Web services composition have recenty emerged incuding SOAP, WSDL, UDDI, and BPEL- 4WS SOAP defines an XML messaging protoco for communication among services WSDL is an XML-based anguage for describing web service interfaces UDDI provides the directory and a SOAP-based API to pubish and discover services Finay, BPEL4WS provides a anguage for process-based service composition Other proposed notations for service description and composition incude ebxml and DAML-S The above proposas however are compementary to ours Indeed, our proposa aims at everaging the above standards (eg, SOAP, UDDI) to provide a quaitydriven and dynamic service composition mode Web service composition is a very active area of research and deveopment [1, 7, 8] Previous efforts in this area such as CMI [11] and efow [6] have investigated dynamic service seection based on user requirements In particuar, CMI s service definition mode features the concept of a pacehoder activity to cater for dynamic composition of services A pacehoder is an abstract activity repaced at runtime with a concrete activity type A seection poicy is specified to indicate the activity that shoud be executed in ieu of the pacehoder In efow, the definition of a service node contains a search recipe represented in a query anguage When a service node is invoked, a search recipe is executed in order to seect a reference to a specific service Both CMI and efow focus on optimizing service seection at singe task eve (ie oca seection) In addition, no QoS mode is expicity supported In contrast, our approach focuses on optimizing service seection at a composite service eve, based on a generic QoS mode, and using estabished inear programming techniques Reated work on QoS has been conducted in the area of workfow In particuar, there are a number of research proposas addressing the specification and verification of tempora constraints in workfows [9, 3] Other proposas such as METEOR [5] and CrossFow [18, 17] have considered QoS modes with other parameters than time Specificay, [5] considers four quaity dimensions, namey time, cost, reiabiity and fideity However, it does not consider the dynamic composition of services Instead, it focuses on anayzing, predicting, and monitoring QoS of workfow processes Simiary, [18] proposes the use of continuous-time Markov chain to estimate execution time and cost of a workfow Coser to our proposa is the one reported in [17], which expores the issue of dynamicay seecting severa aternative tasks within a workfow, based on quaity parameters, in a simiar way as efow does using search recipes As stated before, this oca seection strategy contrasts with the goba panning approaches that we advocate Other reated research proposas incude [20, 21], which focus on data quaity management in cooperative information systems They investigate techniques to seect best avaiabe data from different service providers based on a set of data quaity dimensions such as accuracy, competeness, and consistency 7 CONCLUSION Dynamic seection of component services is an important issue in Web services composition In this paper, we have presented a genera and extensibe mode to evauate QoS of both eementary and composite services Based on the QoS mode, a goba service seection approach that uses inear programming techniques to compute optima execution pans for composite services has been described We have conducted experiments to compare the proposed technique with the oca seection approach The resuts show that the proposed approach effectivey seects high quaity execution pans (ie, pans which have higher overa 420