OPTIMAL COMPLEX SERVICES COMPOSITION IN SOA SYSTEMS

Transcription

1 Key words SOA, optial, coplex service, coposition, Quality of Service Piotr RYGIELSKI*, Paweł ŚWIĄTEK* OPTIMAL COMPLEX SERVICES COMPOSITION IN SOA SYSTEMS One of the ost iportant tasks in service oriented architecture paradig based systes is the task of coposition of the coplex service. Ai of this paper is to focus on the very last phase of the coposition process, where proper atoic service versions are picked to execute the whole coplex service. An exact algorith is proposed in this paper solving two tasks based on popular quality of service delivery approaches: best-effort and differentiated services. The procedure of graph reduction has been introduced to generate space of possible solutions in such way that decision is ade in ore efficient way. The graph-fold algorith was tested in siulation environent where its perforance was copared to five reference algoriths. Moreover ability to deliver quality at desired level has been tested. 1. INTRODUCTION In systes based on SOA (Service-Oriented Architecture) paradig services delivered to endusers (coplex services) are coposed with use of atoic services (services that have atoic functionality). The functionality of a coplex service is an aggregation of functionalities of atoic services [5]. In general syste is distributed, what eans that applications acting as atoic services can be installed on any achine with counication interface available. An user requesting a service fro the syste forulates a request that is specifying functionality deanded. To deliver requested functionality syste uses service coposition procedure which consists of choosing proper atoic services with an execution order to satisfy user requireent. Moreover, user is able to forulate an additional ters of service delivery which involves non-functional aspect of delivering the service Quality of Service requireents. Such coplex request for service is called SLA Service Level Agreeent and uniquely defines functional and non-functional user needs SERVICE COMPOSITION TASK In general, the task of coplex service coposition consists of finding, for given ordered set of required functionalities (stated in the SLA), an order of atoic services versions execution such that non-functional requireents are et. The task of coplex service coposition can be decoposed into three sequential subtasks (see figure 1): 1. Coplex service structure coposition transforation of the SLA into set of required functionalities and the precedence relations between the. The result of this task is coplex service structure represented as a directed graph (not connected in general) of required functionalities. * Institute of Coputer Science, Wroclaw University of Technology, Wybrzeze Wyspianskiego 27, Wroclaw. piotr.rygielski@pwr.wroc.pl, pawel.swiatek@pwr.wroc.pl

2 2. Coplex service scenario coposition transforation of coplex service structure graph into single and consistent graph of required functionalities with precisely defined order of execution of all atoic functionalities. Since it is possible, that single functionality is delivered by ore than one atoic service version, the scenario graph represents in fact a faily of execution graphs where eber graphs differ in atoic service versions applied to deliver required atoic functionality. 3. Coplex service execution plan coposition choice of particular atoic services in coplex service scenario graph such that non-functional requireents of coplex service are et. In order to deliver coplex service with requested functionality and non-functional properties various optiization tasks need to be solved on consecutive stages of coplex service coposition task (i.e.: stages of coplex service structure, scenario and execution plan coposition) [9]. Fig. 1. Decoposition of the process of coplex service coposition 1.2. COMPLEX SERVICE EXECUTION PLAN Ai of this paper is to focus on the coplex service execution plan deterination. Order constraints and atoic services selected under coplex service scenario deterination process should be processed further to select proper atoic service versions for execution. Atoic service versions can be available at the distributed execution syste in any versions (each version differs fro other in nonfunctional aspect) so choosing proper execution plan can be treated as quality optiization task with popular QoS criteria i.e., execution tie, cost, security, etc. 2. THE COMPLEX SERVICE SCENARIO At the third stage of coplex service coposition process it is assued that a graph called coplex service execution scenario is given. The scenario defines unabiguously the order in which atoic functionalities should be delivered in order to fulfill the user functional requireent. The scenario is odeled as a graph GC = ( VC, EC) where VC is a vertex set where one vertex represents functionality to be delivered and EC is a set of edges which defines the order of delivering functionalities. Exeplary coplex service scenario is presented on the figure 2.

3 Fig. 2. Exeplary scenario of coplex serviced which goal is to ake a cell phone call. Each of distinguished functionalities can be delivered by any versions of atoic services. An exeplary coplex service - which goal is to ake a cell phone call - consists of four operations that have to be executed. Assuing that user does not have oney to execute this service, first user has to get the oney, for exaple by getting a loan. A loan can be given by any institutions where each can be treated as another atoic service version. Siilar situation is with buying a cell phone user can buy sae phone in any stores and with si card any cellular operators can offer user a si card. Finally user can ake a call using a newly bought cell phone and si card using various versions of calling service e.g. VoIP call, gs call, video call etc. An user forulating its requireent for service defines a set of quality paraeters that should be delivered to satisfy the user. Of course those requireents concern the whole process of delivering a service and should be taken into consideration during the coposition especially in the execution plan deterination stage. 3. PROBLEM FORMULATION A task of execution plan deterination is the last task that needs to be solved in the coplex service coposition procedure. In general the task of deterination of optial execution plan of coplex service can be described as follows. The systes task is to pick one version of each atoic service defined uniquely by the scenario in such way, that quality requireents are satisfied. Noteworthy is fact that picking two atoic service versions defines also a counication path that is going to be used to counicate between these services assuing that there is always chosen the best possible path if there is ore than once. The task is siilar to the task of finding an optial path in graph considered in earlier work [4] but the difference is that in this approach the solution is represented as a graph - not as a single path - and the forer algoriths can not be used. Forally the task of finding optial a coplex service execution plan can be forulated as follows: For given: Scenario graph GC = ( VC, EC) A l-th service request represented by the SLA l Coplex service nonfunctional requireent Ψ l Quality criterion Q( G, Ψl ) Find: Such set of atoic services versions for scenario GC that quality criterion is iniized as *,..., as * = arg in Q( G( { as1 j,..., as kj }, E), Ψ l ) where: 1 j k kj k as 1 jk,..., as as denotes 1 j k jk -th version of k-th atoic service kjk k k

4 The proble stated above can be easily transfored into ulti-choice knapsack proble in the following way. The task is to pick exactly one ite (atoic service version) fro each group (atoic service) in such way that quality of represented by picked ites (coplex service) is optial in sense of choosen quality criterion Q. The solution of the forulated proble is known and soe algoriths solving such proble has been presented i.e. in [7]. But the proble forulated at the beginning of this chapter can not be directly transforated into ultichoice knapsack proble because of scenario of execution of coplex service is given as a graph in general. In order to transforate proble of deterination of optial execution plan into ulti-choice knapsack proble, one has to reduce scenario graph into connected directed graph having all vertices with input and output degree no higher than one. In order to solve stated probles the procedure of graph reduction is proposed in next chapter of this paper. 4. SCENARIO GRAPH REDUCTION PROCEDURE The goal of scenario graph reduction procedure is to transfor the scenario graph into connected graph where all vertices have input and output degree no higher than one. Transforation of input graph scenario consists of deterination of sub graph - called a pattern - and defining a result of reduction for this defined pattern. Moreover all versions of atoic services present within this pattern should be aggregated in the new resulting vertex in such way that latter distinguishing of the original versions in possible. The ost iportant part of pattern reduction is to find an aggregating function that aggregates values of nonfunctional paraeters into new ones in newly created vertex of the reduction process. The whole procedure of pattern reduction has been described below REDUCTION PROCEDURE In order to reduce the scenario graph into for that stated optiization probles can be solved, the set of patterns with aggregation function has to be defined. On the figure 3 there has been presented two scenario subgraph patterns (subfigures a and c) with corresponding resulting vertex (subfigures b and d respectively) obtained using aggregation function. The original versions of the atoic services are being erged into new ones in the newly created vertex. The aggregation functions for the exaples given on the figure 3 considering execution tie as a nonfunctional paraeter - are su and su-ax - respectively for serial reduction (subfigures a,b) and split reduction (subfigures c,d). Therefore the resulting versions of k-th atoic service obtained in process of serial reduction has the following values of execution tie: d ( k1) = d( i 1,1 ) + d( i,1); d ( k2 ) = d( i 1,1 ) + d( i,2); d ( k3 ) = d( i 1,2 ) + d( i,1); d ( k4 ) = d( i 1,2 ) + d( i,2); where d ( i,1) denotes execution tie of first version of i-th atoic service. Respectively the execution ties for split reduction procedure can by calculated by following the following pattern: d ( k1 ) = d( i 1,1 ) + ax{ d( i,1 ),..., d( i + n,1) }; d ( k2 ) = d( i 1,1 ) + ax{ d( i,1 ),..., d( i + n,2) }; and for the last -th version of newly created service: d( k ) = d( i 1,2) + ax{ d( i,2),..., d( i + n,2) };. The nuber of versions in the newly created service can be calculated using the equation 1

5 = J j= 0 size( as j ), (1) where size as ) denotes nuber of versions of j-th atoic service and J denotes count of atoic ( j services in the pattern Fig. 3. Two subgraph patterns (serial - a and split - c) and corresponding results of reduction (b and d respectively). Atoic service versions before reduction (a,c) are transfored into new ones (b,d) which nonfunctional paraeters values were also recalculated using proper aggregation functions. In this exaple count of atoic service versions in serial reduction equals to 4 and in the split reduction can be calculated using equation (1). Except of vertices patterns reduction one can distinguish an edge reduction pattern which leads to decrease of graph coplexity. The redundant edge reduction pattern has been presented on figure 4. Fig. 4. Possible edge reduction non violating precedence relations between vertices of scenario graph; a) before edge reoval, b) after reoving redundant edge. Without loosing any inforation one can reove edge ( as i 1, asi+ 1) because of transitivity of precedence relation: asi 1 p asi asi p asi+ 1 asi 1 p asi+ 1. Of course one can distinguish other patterns of edge reoval if and only if one does not cause loss of inforation in case of such reduction. The reductions defined forerly are basic ones and one can define ore of such reductions to cope with ore coplicated scenario graphs in ore efficient way. The coplete reduction operations set should contain at least two entioned vertex reduction operations (serial and split) and redundant edge reduction. This gives guarantee to reduce any graph into desired for.

6 4.2. THE GRAPH-FOLD REDUCTION ALGORITHM In order to reduce the scenario graph to the for that is appropriate for the ulti-choice knapsack proble one has to perfor steps concerning all vertex reduction patterns and all edge reduction patterns. Assuing that each reduction procedure reduces exactly one edge (in case of edge reduction) or exactly one pattern (in case of pattern reduction), the whole reduction process akes axiu VC + EC reductions. Each reduction has also its coputational coplexity which is influenced by: the coplexity of procedure of finding a pattern, coplexity of calculating the values of non-functional paraeters of newly created node, disconnecting reoved vertices and connecting newly created one. All of entioned procedures can have different ipleentations so coplexities ay vary. Trying to estiate coplexities of forer operations the results has been presented in table 1. Operation Coplexity Denotation Finding an pattern O ( kn) n = VC M Calculating new values of nonfunctional paraeters J k largest output degree of O vertex in graph = 1 M - nuber of vertices in the Disconnection and connection O ( n) reduction pattern of graph vertices Finding and reducing J nuber of versions of -th 3 O n log n atoic service redundant edge ( ) Table. 1. Estiation of calculation coplexities of operations used in reduction procedure. Noteworthy is fact that vertices that are disconnected can be reoved and not considered further, so value of n ( n = VC ) is decreasing during the process of reduction. Decreasing speed depends on nuber of vertices that are being disconnected during single reduction operation. The general procedure consists of looped execution of all reduction operations until the vertices count n = 1 (in case of reduction of all vertices of the graph) or input and output degrees of all vertices is no greater than 1. Fig. 5. Exeplary reduction process for scenario with six functionalities (atoic services), each having two versions. After two split reductions and one serial reduction the result is single atoic service with 64 versions. Every newly created scenario gets new nuber for ease of distinguishing.

7 Exaple of the reduction process has been presented on the figure 5. Scenario graph used in this exaple contains six vertices (functionalities) with precedence relations given as on subfigure 1 of figure 5. Each functionality in considered scenario graph has two versions of atoic service. In the first step there is a split-reduction pattern found and the result of reduction is presented on subfigure 2. In the second step situation repeats split reduction pattern is found and reduced. Notice that nuber of versions in the resulting vertex increase. On subfigure 3 the serial reduction pattern is found and reduced. The algorith stops in step four where only one vertex is left (subfigure 4). Resulting vertex has 64 versions. 5. PROBLEM SOLUTION The proble of ulti-choice knapsack proble forulated in chapter 3 can be solved using heuristic algoriths [2,7], but uch ore interesting solution can be obtained when scenario graph has been reduced to single vertex. In case of such reduction one gets single atoic service with versions. Moreover if there are used data structures like AVL trees [3] as a collection of QoS value for resulting versions, the optial solution of deterination of coplex service execution plan proble can be found in O( log ) tie. Depending on optiization goal given by quality criterionq, one can achieve different graphfold algorith behavior. According to popular quality of service delivery approaches - best-effort approach and differentiated services approach - one can forulate such quality criterion that desired algorith behavior can be obtained COMPLEX SERVICE EXECUTION TIME MINIMIZATION This solution bases on the best-effort approach and finds iniu tie of execution of coplex service. A coplex service execution tie can be iniized using graph-fold algorith when execution tie is considered as a quality paraeter. In this case the following optiization task should be solved: * k = arg in d k, (2) where ( ) k k,..., k 1 d denotes execution tie of -th version of k-th atoic service obtained after graph reduction procedure. According to assuption that resulting collection of atoic service versions is ordered by the d ( k ) value, the solution can be found in O ( 1) and consist on picking the version with the lowest execution tie value. ( ) 5.2. COMPLEX SERVICE EXECUTION TIME GUARANTIES This solution bases on the differentiated services approach and finds optial execution tie value with respect to given execution tie requireent. In this case the following optiization task should be solved: with respect to the following constraint: k * k,..., k 1 * ( d( k ) d ) 2 = arg in, (3) * * ( k ) d d, (4)

8 * where d denotes execution tie requireent given by user. According to assuption that resulting d value, the solution can be found in collection of atoic service versions is ordered by the ( k ) ( ) O log and consist on picking the version with the highest execution tie value but not higher than the user requireent. In case when there is no solution of this task, the request can be handled in one of the following ways: the SLA contract should be renegotiated, the request should be rejected and not executed in the syste or can be served as in the best-effort approach. 6. SIMULATION STUDY In order to evaluate quality delivered by presented graph-fold algorith the siulation environent has been developed in OMNeT++ siulation fraework [8]. The siulator consists of three distinguishable odules: generator odule, abstract Enterprise Service Bus and set of web services representing atoic service versions. The siulation run consisted of generating strea of coplex service requests with interarrival tie given by exponential distribution with ean 0.1 second and rando coplex service execution scenarios. The generated request was processed in abstract ESB odule where the task of finding the optial execution plan was solved. There were six algoriths for finding optial execution plan present in the syste five of those were reference algoriths and sixth one was the graph-fold algorith. Each atoic service version present in the syste had different perforance index generated fro exponential distribution with ean 10 easured in kilobytes per second that can be processed. Moreover request data size that was going to be processed was also generated fro exponential distribution with ean 10KB. The nuber of atoic services was set to 8 (also axiu 8 functionalities were present in generated scenario) each having 10 versions. Using such configured siulation environent there were two experients perfored. First consisted of solving the coplex service execution tie iniization task and second of solving the coplex service execution tie optiization task, where additionally the execution tie requireent has been set. The results of the first experient have been presented on figure 6. The experient lasted for 1000 siulation seconds and consisted on deterination of coplex service execution plan at the oent of request was entering the syste. The first observation shows that graph-fold algorith obtained the lowest execution tie fro all tested algoriths. Moreover the execution tie line on the chart is very close to the constant value of about 3 seconds while rest of algoriths resulted in higher variation of execution tie. Second best was best-avg algorith which consisted of choosing such version of atoic service that average execution tie was lowest. Other algoriths resulted in worse coplex service execution tie. The results of the second experient have been presented on figure 7. The task of the graphfold algorith was to keep coplex service execution tie as close as possible to the requireent which was set to 5 seconds. On the chart presented on figure 7 one can notice, that the tie of execution calculated by the graph-fold algorith was always lower or equal to the requireent. In case when syste was not able to provide required execution tie (in the seconds of siulation) lower value was assured. Analyzing the real coplex service execution tie curve on the figure 7, soe deviations ight be noticed. The real execution tie approaches required value after 150 siulation seconds and was kept within 5% error range till the end of siulation. The deviation of real coplex service execution tie with respect to the calculated tie is caused by the

9 no ideal ethod of prediction of single atoic service version execution tie. The decision for each request was ade at the oent when request was entering the syste so the values describing execution tie ight have changed during the execution process. Fig. 6. Siulation results for the six algoriths solving the proble of coplex service execution tie iniization. Fig. 7. Siulation results for the graph-fold algorith solving the task of coplex service execution tie optiization. The calculated copletion tie curve is one predicted by the algorith at the beginning of processing the request.

10 7. CONCLUSION The graph-fold algorith presented in this paper is a good and precise approach for delivering optial coplex service execution plan in the process of coplex service coposition. The graph reduction procedure proposed in this paper copletes the other approaches proposed in the literature [1,2,6,7,10]. The ost iportant weak side of proposed approach is the coplexity of the stated proble. The graph-fold algorith was significantly slower than the reference algoriths what can cause growth of calculation tie in case of coplex service execution scenarios with large aount of functionalities or large aount of atoic service versions available. The fact that presented ethod gives always the precise and optial solution can be used as a reference point for developed heuristic algoriths. Moreover in specific application of graph-fold algorith there can be other set of reduction operations used which appears ore often in input scenario graphs. Such odification can decrease the calculation coplexity for ajority of considered coplex service scenarios. 8. REFERENCES [1] M. ALRIFAI AND T. RISSE, Cobining global optiization with local selection for efficient QoS-aware service coposition, WWW '09: Proceedings of the 18th international conference on World wide web, pp , ACM [2] BERBNER, R., SPAHN, M., REPP, N., HECKMANN, O., AND STEINMETZ, R. Heuristics for QoS-aware Web Service Coposition. In Proceedings of the IEEE international Conference on Web Services (Septeber 18-22, 2006). ICWS. IEEE Coputer Society, Washington, DC, [3] CORMEN T. ET.AL. Introduction to Algoriths, MIT Press, Cabridge, MA, Second edition, (2001) [4] GRZECH A. RYGIELSKI P. ŚWIĄTEK P., QoS-aware infrastructure resources allocation in systes based on service-oriented architecture paradig. Perforance odelling and evaluation of heterogeneous networks, 6th Working International Conference HET - NETs 2010, Zakopane, January 14th-16th, 2010 s [5] GRZECH A. AND ŚWIĄTEK P., Parallel processing of connection streas in nodes of packet-switched coputer counication systes. Cybernetics and Systes. 2008, vol. 39, nr 2, pp [6] MICHAEL C. JAEGER, GERO MÜHL, SEBASTIAN GOLZE, QoS-Aware Coposition of Web Services: An Evaluation of Selection Algoriths in On the Move to Meaningful Internet Systes 2005: CoopIS, DOA, and ODBASE (2005), pp [7] SHAHADAT KHAN AND KIN F. LI AND ERIC G. MANNING, The Utility Model For Adaptive Multiedia Systes In International Conference on Multiedia Modeling, 1997 pp [8] OMNeT++ Web page: [9] RYGIELSKI P., ŚWIĄTEK P., QoS-aware Coplex Service Coposition in SOA-based Systes, in SOA Infrastructure Tools: Concepts and Methods, Springer-Verlang, Berlin [10] YU TAO, ZHANG YUE AND KWEI-JAY LIN, Efficient algoriths for Web services selection with end-to-end QoS constraints, ACM Trans. Web 2007, Vol. 1, No. 1, Article. 6.