A Petri net-based Approach to QoS-aware Configuration for Web Services

Transcription

1 A Petri net-based Aroach to QoS-aware Configuration for Web s PengCheng Xiong, YuShun Fan and MengChu Zhou, Fellow, IEEE Abstract With the develoment of enterrise-wide and cross-enterrise alication integration and interoeration towards web service, web service roviders try to not only fulfill the functional requirements of web service users, but also satisfy their non-functional conditions in order to survive in the cometitive market. A hot research toic is how to configure web services to meet their demand under a dynamic heterogeneous environment. This aer builds a web service configuration net based on Petri nets in order to exhibit web service configuration in a formal way. Then, an otimal algorithm is resented to hel choose the best configuration with the otimal quality of services (QoS) to meet users non-functional requirements. Finally, the simulation results and related analysis rove the soundness and correctness of our model and algorithm. I. INTRODUCTION EB service [] framework has evolved to become a W romising technology for the integration of disarate software comonents using Internet rotocols. Web service roviders register web services through an UDDI registry. The web service that they intend to offer is defined by WSDL. Then, web service users/requestors discover the needed web services and send the requests via invocation interfaces. After the resonse from a web service rovider, they invoke those services under SOAP. When any single web service fails to accomlish service requestor s multile function requirements, multile web services need to be dynamically configured together to form a web service comosition. However, configuring web services to meet secial needs is not a trivial task. There are two main considerations. One is functional requirement []. In this resect, the configuration must ensure that all functions required by the requestor are rovided by service elements. comonent architecture [3] rovides a rogramming model through assembly of heterogeneous services. A dynamic configuration can be modeled as a functional decomosition This work was suorted in art by the National High Technology Research and Develoment (863) Program of China under Grant 006AA0Z5 and the China National Science Foundation under Grant The extended version of this aer is submitted to IEEE Transaction on System, Man and Cybernetics: Part A for review. P.C. Xiong and Y.S. Fan are with the Deartment of Automation, Tsinghua University, Beijing 0008, China ( xc03@mails.tsinghua.edu.cn, fanyus@tsinghua.edu.cn) M.C. Zhou is with Deartment of ECE, New Jersey Institute of Technology Newark, NJ USA and School of Electro-Mechanical Engineering, Xidian University, Xi'An, Shanxi 7007, China ( zhou@njit.edu). of the overall requested function. As a formal digrah to resent all of the deendency relationshis, Deendency Grah (SDG) is oularly used to deict deendency relationshi among web services []. Although it rovides a means for web service configuration descritions in order to ensure functional interoerability among collaborating web services, SDG deals with only the functional asect. It is not hard to imagine that service requestors will face with a large number of choices of service configurations that can rovide the similar function. Another consideration is non-functional [-9], such as cost and QoS which are necessary for the evaluation, selection, and configuration of needed service comositions. Nevertheless, although there are many research aers [-5], rojects, such as METEOR-S [7], and middleware, such as GlueQoS [8] and SwinDeW [9] related to QoS aware web service selection, they do not consider that a service configuration is oerating under an environment where the run-time erformance is fluctuating and quality of the individual services are subject to change. In this aer, we address the otimal configuration issues by concentrating on a) modeling and definition of the configuration and b) the otimal QoS searching algorithm under a varying environment. II. MODELING SERVICE CONFIGURATION WITH PETRI NETS The web service rovided may be dynamically reconfigured uon the service comonent udates, resource availability changes, or user requests. We model the configuration roblem as Functional Deendency Configuration (SFDC). A web service rovided may have multile SFDCs. The functional deendency relationshi of a configuration can be divided into two tyes, i.e., function combination and function selection. We describe these two basic tyes by alying AND and OR structures in Petri nets [0] as shown in Fig.. The following definition is used. Definition : A Petri net is a 5-tule PN = ( P, T, I, O, M ) where:. P =,,...,, m > 0, is a finite set of laces. m. T = t, t,..., t, n > 0, is a finite set of transitions, with n P T and P T = ; 3. I : P T 0,, is an inut function that defines the set of directed arcs from P to T;. O: P T 0,, is an outut function that defines the set of directed arcs from T to P;

2 5. M : P N, is an m column vector whose ith comonent reresents the number of tokens in the ith lace, where N = 0,. An initial marking is denoted by M t t t 3 Figure. AND/OR structure in Petri nets Post set of t is the set of outut laces of t, denoted by t. Preset of t is the set of inut laces of t, denoted by t. Post (Pre) set of is the set of outut (inut) transitions of, denoted by and resectively. Definition : set S = WS s dummy where = is a finite set of web services, and s dummy is WS s z dummy service. Definition 3: Functional Deendency Configuration Net (S-net) : an acyclic Petri net PN is an S-net if:. s S, P, s is maed to a web service lace. If s is a dummy service, s is maed to. t T, if t > and relationshi between and dummy 8 t P., t, there is an AND. t T, if t =, t = and >, there is an OR relationshi among ( ). 3. Each transition has only one inut arc and at least one outut arc.. There is a lace denoted by ' with no inut arc, which corresonds to the service the user requests. M ( ') = and 0 M ( ) = 0, ' A marking in an S-net is changed according to the following transition rule: ) A transition t T is said to be enabled if and only if M ( ) I(, t), P and we denote M [ t >. We denote E(M) as an enabled transition set under M. ) Firing t at M leads to M ', denoted as M [ t > M ', where M '( ) = 0 if ( t, ) I; M '( ) = if ( t, ) Oand M '( ) = M ( ) otherwise. Definition : An SFDC, at timeζ, denoted as C( ', ζ ) is a subset of P such that (a) ' C( ', ζ ), and (b) C( ', ζ ), if, t and '' t, '' C( ', ζ ). III. AUTOMATIC WEB SERVICE CONFIGURATION WITH OPTIMAL QOS Each functionality of a service may have several QoS arameters and can be evaluated by measurements. Ran sorts the QoS arameters into runtime, transaction suort, configuration management and cost, and security related ones []. Each is made u of several metrics and sub-metrics. For instance, higher values of throughut, reliability, robustness and flexibility are desired. Lower values of resonse time, latency and cost are also referred. We suose that a certain QoS arameter of a non-dummy web service lace at time ζ is ψ (, ζ ). In general, higher value of ψ (, ζ ) means higher quality. For examle, ψ (, ζ ) can be throughut and denotes the number of comleted service requests over a time eriod. To reflect the cost, ψ (, ζ ) is defined as the value of zero minus the rice involved in requesting the service. If the web service reresented by is not available, ψ (, ζ ) =. Because the whole functionality of a service deends on its own functionality and that of its deendent sub-services, we assume that a QoS arameter for a C( ', ζ ) is a function of the QoS arameter of all the non-dummy deendent services, i.e., QoS( C( ', ζ)) = f ( ψ(, ζ) C( ', ζ) \ ). dummy For examle, the function to calculate the configuration cost can be simly summing u of all the non-dummy services ψ (, ζ ), i.e., QoS( C( ', ζ )) = ψ(, ζ ) C( ', ζ )\ dummy Given ' and ζ, the algorithm to search for the best C( ', ζ ) of S-net PN, i.e., an SFDC with the highest QoS( C( ', ζ )) is as follows OtimalConfiguration( ', ζ ) Initialization: Get the ψ (, ζ ) for all the non-dummy web service laces Set all the transitions as unmarked find=false a marking stack S = M 0 OtimalQoS= //The otimal QoS OtimalS= //The otimal marking stack CStack= //The configured web service lace stack While ( S ) Begin Get the marking at the to of the stack M = STo. If t where M[ t > and t has never been marked with M Mark t with M; For every t' E( M), if t' Push t into CStack; M ' = Mt [ > ; t, mark t ' with M;

3 For every, if CStack Delete the token in and set M '( ) = 0 ; Push M ' into S; If (find==true) and QoS(S)<OtimalQoS Po(S) and Po (CStack); Elseif t t where M[ t > but t t has been marked with M Po(S) and Po(CStack); Else If QoS(S)>OtimalQoS //At this time the first or a better configuration is found. find= true; OtimalQoS= QoS(S); OtimalS= Configuration(S); Po(S) and Po(CStack); End Begin End While The QoS for a marking stack S can be calculated as follows: QoS(Stack S) Set temorary web service lace set P = T For every CStack Add to P ; T Get the marking at the to of the stack M = STo. Find all P, M( ) > 0 Add to P ; T Return f ( ψ(, ζ) P \ ) T dummy The configuration for a marking stack S can be calculated as follows: Configuration(Stack S) Set temorary SFDC C( ', ζ ) = For every CStack Add to C( ', ζ ) Get the marking at the to of the stack M = STo. Find all P, M( ) > 0 Add to C( ', ζ ) Return C( ', ζ ) Generally seaking, the QoS arameter of the configuration is deteriorating when more deendent web services are added to the configuration, e.g., comared with an individual web service, the integral of this web service and a deendent web service on it has worse reliability, longer latency and higher cost. Based on this regulation, as the algorithm roceeds, when the QoS arameter of the current configuration is already worse than the otimal one, we sto searching along this configuration ath and retrieve to another one through o and ush oerations. Hence the QoS arameter of C( ', ζ ) found through the algorithm is increasing monotonically, and we can conclude that the algorithm is able to find the best C( ', ζ ) with the highest QoS( C( ', ζ )). Only in the worst case, the algorithm has to search for the whole reachability tree. Theorem [6]: If P\ ', = η, the worst-case dummy comlexity of the above algorithm is aroximately Oe ( η ), where e.788 is the base of the natural logarithm. Though the algorithm is roved to be of exonential time in the worst case, the comlexity grows exonentially only with the number of laces in OR structures (not all laces) in the S-net and also deends on actual circumstance. On the other hand, the configuration ste in our algorithm can be carried out concurrently, meaning that each sibling transition of t, i.e., ( t) \ t can fire with t concurrently. If multirocessors are used, this roerty can seed u the algorithm. Moreover, the algorithm derives the best configuration result whose erformance may well exceed what the user really exects. Hence, if considering users actual exectation, the algorithm can terminate in a shorter time once it finds the first configuration meeting their exectation. The above oints rove the alicability of our algorithm. IV. CASE STUDY AND PERFORMANCE ANALYSIS To evaluate algorithm OtimalConfiguration, we take an examle and measure its configuration ability in user satisfaction rate through simulation. During the execution, at a given time eriod ϒ, we measure the number of the invoking times Invoking( ϒ ) and the number of times the user s request is satisfied Satisfied( ϒ ). Then the user satisfaction rate can be calculated as Cr( ϒ ) = Satisfied( ϒ)/ Invoking( ϒ ). We comare the erformance of the roosed algorithm with two other tyical aroaches. The first method is Fixed, which selects the first feasible configuration and never changes. The other one is Random, which randomly chooses a configuration regardless of its erformance. In this section we take the examle of manufacture execution service configuration as shown in Fig.. The manufacture execution service deends on the order management service to collect orders, worksho management service to schedule and monitor job rogress, and rocess quality control service to track roduct quality. The job

4 rogress reort service further relies on two kinds of monitoring services, i.e., the human insection and live monitoring aroaches. Both of the aroaches deend on a data collection service. The data collection service is a dummy service and shown in a dotted-line box, because the configuration should choose one and only one system to realize it. Figure. Web service deendency grah The S-net corresonding to the web service deendency grah is shown in Fig. 3. For simlicity, here we take the cost of the web services as the QoS arameter. Suose that in a highly varying environment, the cost of a web service changes conforming to a normal distribution. The web services that ' and 8 denote and their cost distribution arameters are shown in Tab.. We also assume that the user affordable rice fluctuates in a normal distribution with a mean of 00 and a standard deviation of 0. The average user s request arrival rate is 0. er second and we aroximately receive 600 web service requests over 3000 second eriod in the simulation environment. ' Figure 3. Web service deendency grah We comare the erformance of different algorithms in our simulation in Fig. and Fig. 5. In the simulation, we choose the first feasible configuration as ',,,, for 9 7 Fixed. And it has the worst success rate because of its uniformly choosing one configuration. Random has better erformance since it has more chances to obtain a better configuration. OtimalConfiguration algorithm kees the best success rate since it can accommodate its choice to the changing environment. Because the algorithm does not have to generate the whole reachability tree each time, it is also better than the exhausted enumeration method. We then study the algorithm erformance under a wider request range, i.e., the user affordable rice has an increased mean from 350 to 500 and a standard deviation of 0. The simulation result is shown in Fig. 5. The curve for our OtimalConfiguration algorithm is always above the curves for Random and Fixed algorithms, which roves a better erformance over the other two. Moreover, under a stringent user s request, e.g., 350, our OtimalConfiguration algorithm can still satisfy over 30% of the requests while the other two almost fail. When we gradually relax user s request to 500, the curve for OtimalConfiguration reaches 00% satisfaction rate raidly, while the other two at a lower seed. Note that, Random is not always better than Fixed. In the simulation environment, because the sum of normal distribution variables also conforms to a normal distribution, we can denote the cost obtained through the Fixed and Random algorithms as normal distributions Nu (, σ ) and R Nu (, σ ) resectively, where R Nuσ (, ) denotes a normal distribution with a mean of u and a standard deviation of σ. From the cost distribution arameters in Tab., it is easy to get u > u and σ σ aroximately. Suose that the F R F R affordable rice conforms to Nu (, σ ). The success rate U U for the Fixed and Random algorithms can be calculated as Φ(( u u ) / σ + σ ) and U F F U Φ(( u u ) / σ + σ ) resectively, where U R R U x ( x ) e t Φ = / d t π is the Lalace function. Under the fixed affordable rice mean of 00, we can calculate the satisfaction rate for Fixed as Φ( 0.899) = 0.8, which is close to the simulation result in Fig.. Hence, if u > u and F R σ = σ F R, Random has a higher user satisfaction rate. But if the first feasible configuration Fixed chooses the case that u < u, for instance, ',,,,,, it has a F R higher user satisfaction rate than that of Random. In summary, the above simulation results show that our algorithm is the best suitable for the highly varying environment, while delivering web services with high QoS uon a user s request to make the configuration decision. TABLE I THE WEB SERVICES AND THEIR COST DISTRIBUTION PARAMETERS. Places ' Web service name Manufacture Execution Order Management Mean cost of F F Standard Deviation of Cost

5 3 5 Worksho Management Process Quality Control Job Scheduling Alarm and Warning SPC Product Tracking 7 8 Rolling Scheduling PLC PLC 7 0 PLC 3 6 Human Insection 3 Live Monitoring System 58 5 System System Bar Code 55 7 RFID 3 8 Otimal Random Fixed Otimal Random Fixed Figure 5. Simulation result under variable affordable rice mean from 350 to 500 V. CONCLUSION Since web service framework brings in a new revolution in traditional comuting, web service configuration issues have been receiving more and more attentions. The imortance of web service configuration is also incarnated in the research on other hot toics such as service fault management, accounting management as well as service comonent selection olicy. This aer rooses a systematic method to manage the existent web services to form a high erformance configuration under a diverse and changeable environment. It resents a service functional deendency configuration net based on Petri nets for the web service resentation and automatic configuration. Based on the configuration net, this aer resents an otimal algorithm able to return the otimal configuration with the best QoS arameter. Comared with a brute force exhaust algorithm, this algorithm often does not have to search for all the ossible configurations in order to obtain the otimal one. Moreover, theoretical evidence and simulation results demonstrate that the configuration results our algorithm returns can achieve the highest user satisfaction rate. Figure. Simulation result under fixed affordable rice mean of 00 REFERENCES [] F. Curbera, M. Duftler, R. Khalaf, W. Nagy, N. Mukhi and S. Weerawarana, Unraveling the Web s Web: An Introduction to SOAP, WSDL, and UDDI, IEEE Internet Comuting, Vol. 6, No., , Mar./Ar. 00. [] P. Hasselmayer, Managing Dynamic Deendencies, Proc. of the th Int l Worksho on Distributed Systems: Oerations and Management, Nancy, France, Oct. 00,.-50. [3] IBM, SAP, Oracle, BEA, Sybase, Siebel Systems, etc. SCA Assembly Model Secification. htt://www-8.ibm.com/develoerworks/library/secification/ws-sca [] S. Ran, A model for web services discovery with QoS, SIGecom Exchanges, Vol., Iss.,. -0, Mar [5] D. Ardagna and B. Pernici, Global and local QoS constraints guarantee in Web service selection, Proc. of 005 IEEE International Conference on Web s, Orlando, FL, July 005,

6 [6] P.C. Xiong, Y. S. Fan and M.C. Zhou, QoS-aware Web Configuration, Submitted to IEEE Transaction on System, Man and Cybernetics, Part A. [7] A. Sheth, J. Cardoso, J. Miller and K. Kochut, QoS for -oriented Middleware, Proc. of the Conference on Systemics, Cybernetics and Informatics, Orlando, FL, July 00, [8] E. Wohlstadter, S. Tai, T. Mikalsen, I. Rouvellou and P. Devanbu, GlueQoS: middleware to sweeten quality-of-service olicy interactions, Proc. of the 6th International Conference on Software Engineering, Edinburgh, United Kingdom, May 00, [9] J. Yan, Y. Yang and G. K. Raikundalia, SwinDeW-a -based decentralized workflow management system, IEEE Transactions on Systems, Man and Cybernetics, Part A, Vol. 36, Iss. 5, , Set [0] M.C. Zhou and K. Venkatesh, Modeling, Simulation and Control of Flexible Manufacturing Systems: A Petri Net Aroach, World Scientific, Singaore, 998.