Research Article QoS-Aware Multiobjective Optimization Algorithm for Web Services Selection with Deadline and Budget Constraints

Transcription

1 Hindawi Publishing Corporation Advances in Mechanical Engineering Volume 214, Article ID , 7 pages Research Article QoS-Aware Multiobjective Optimization Algorithm for Web Services Selection with Deadline and Budget Constraints Xianzhong Han, Yingchun Yuan, Chen Chen, and Kejian Wang College of Information Science & Technology, Agriculture University of Hebei, Baoding 711, China Correspondence should be addressed to Yingchun Yuan; nd hd yyc@163.com Received 2 February 214; Accepted 24 July 214; Published 19 August 214 Academic Editor: Hongxing Wei Copyright 214 Xianzhong Han et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. The problem of QoS-aware multiobjective optimization is an important issue for Web services selection in distributed computing environment. In this paper, a novel algorithm called (multiobjective optimization algorithm for web service selection) is proposed through analyzing the genetic operators such as constraint handling, the initial population generation, fitness assignment, and diversity preservation. Compared with (Yu et al., 27), simulation results show that the feasible objective region can be filled uniformly with the optimal solutions obtained by under different test applications. In the case of higher constraints especially, can obtain more high-quality and evenly distributed nondominated solutions than. 1. Introduction Service-oriented architecture (SOA) has been a major software framework for distributed applications such as e- business and enterprise systems. These distributed applications, that is, composition services, can be dynamically and flexibly composed by integrating existing component services developed independently [1]. Usually, service providers can offer some different service versions with similar functionalityanddifferentqos(qualityofservice)levels.qosforweb services refer to various nonfunctional characteristics such as response time, execution price, availability, and reliability [2]. It is the way to select the best services to compose an efficient andcomplexservicewithqosassurancethatmatters. QoS-oriented service selection is essentially a multiobjective optimization problem, and it is usually an NPhard problem. In the present study, multiple objectives are usually combined into a scalar objective function via a predetermined aggregating function for optimization. Wen et al. [3] presented a QoS objective function with high performance/cost ratio. Zeng et al. [4] transformed the five QoS attributes into a scalar linear weighted sum objective function; the above approach is also found in the literatures [5 7]. Although the method is simple to implement with low time complexity, it can only obtain a particular optimum point on the tradeoff surface. This drawback of such an approach is that the weights are difficult to determine precisely, especially when there is insufficient information or knowledge concerning the optimization problem. Multiobjective optimization problems involve simultaneous optimization of several incommensurable and often competing objectives. Often, there is no single optimal solution, but rather a set of alternative solutions. They are known as Pareto-optimal solutions. Multiobjective evolutionaryalgorithmsseemtobeparticularlysuitedforsolving this kind of optimization problems [8].Basedonmultiobjective evolutionary algorithm, several Web service selection algorithms are developed [9 11]. Yu et al. [12] proposeda workflow execution planning approach using multiobjective evolutionary algorithm (), which developed on basis of NSGAII [13]. It implemented simultaneously to optimize the execution time and service price while meeting time and cost constraints. This class of methods can generate a set of nondominated solutions in single run and it is easy for users to select the most suitable solution for the current issue. Following the guideline of literature [12], we proposed a new Pareto-based multiobjective optimization algorithm

2 2 Advances in Mechanical Engineering for Web service selection (), which implements the simultaneous optimization of execution time and cost while meeting deadline and budget constraints. Experimental results show that the population can converge and uniformly distribute in the Pareto front. The rest of this paper is organized as follows. Section 2 defines the optimization problem for Web service selection, the constraints definition, and the problem assumptions. Section 3 describes the proposed algorithm for Web service selection. Section 4 shows the performance evaluation and comparison of two algorithms. The paper is concluded in Section Problem Description Following the literature [12], a Web service composition application considered in this paper can be described as a workflow application, which is modeled as DAG (directed acyclic graph). Let G = {V,E} denote the DAG graph, wherenodesetv = {1,2,...,n}denotes the set of tasks (or activities), and directed arc (edge) set E represents intertask data or control dependencies. Task 1 and n are dummy singlestart and single-terminal nodes. Each arc (i, j) E represents the precedence constraint from i to j; thatis,taskj cannot start before i has finished. For any activity i in graph G, there are many candidate services which can fulfill i. Allcandidateservicesfori are called the service pool of activity i, denoted as SP(i). Let l(i) denote the service pool length; that is, l(i) = SP(i). In the same pool, different services have different processing times and costs. Generally, the faster a service executes, the more ausershouldpay[12]. Let s ij be the jth service name for conducting activity i, t ij the processing time on service s ij, and c ij the cost for executing i on service s ij. Thus, 3-tuple (s ij,t ij,c ij,) denotes the jth service that can fulfill activity i. The service pool of activity i can be expressed as SP(i) = {(s ij,t ij,c ij,) 1 j l(i)}.theservicepoolofdummy activities is empty. For any application G = (V, E), oneofitssolutionscan be obtained if each task allocates a service instance using some selection policy. Let S be the solution space of G. For x S, the required execution time is denoted as T(x), and the required total execution cost is denoted as C(x). Let f i (1 i n) be the completion time of task i; thent(x) = f n. Let z ik be -1 variable. It means that its value is 1 when node i chooses service level k to perform it and, otherwise; then C(x) = i V 1 k l(i) z ik c ik. Let D be the given time constraint (deadline), which is also the latest completion time of task n. LetB be the given cost constraint (budget), which is the upper limit of the required total cost. The objective of this problem is to select an appropriate service level for each task so that the duration and the total cost can be simultaneously minimized while meeting the given deadline and budget constraints. Assume that the number of service providers is large enough to accommodate the number of tasks that need to be run simultaneously at any given time, so that the waiting time for an available server is negligible. Thus, each service can provide the expected execution time of any task at any service level. Users can choose properly service levels for these tasks. The constrained multiobjective optimization problem can be formulated as follows: Minimize y=f(x) = (T (x),c(x)), x S, (1) s.t. f i f j l(i) k=1 l(i) k=1 z ik =1, i V, (2) t ik z ik, (i, j) E, (3) T (x) <D, (4) C (x) <B, (5) z ik {, 1}, i V, 1 k l(i). (6) The objective function (1) corresponds to minimizing the duration and the total cost. Equation (2) ensures that exactly only one service should be chosen for each task. Constraint (3) meets precedence constraints. Constraint (4) meets the given deadline constraints. Constraint (5)guaranteesthatthe required total cost is less than or equal to budget B. Constraint (6)guaranteesthatz ik is -1 variable. 3. Algorithm A new approach, called (multiobjective optimization algorithm for web service selection) is proposed in this paper. uses a mixture of established and new techniques in order to find multiple Pareto-optimal solutions in parallel. The flow of the algorithm is as follows. Step (1). Handle constraints based on penalty functions. Step (2). Encode the problem and determine the population size. Step (3). Generate an initial population P and create the empty external nondominated set P. Step (4). Copy nondominated members of P to P. Step (5). Operate crossover and mutation on population P. Step (6). Update population P and preserve its diversity. Step (7). Calculate the fitness of each individual in P as well as in P. Step (8). Select individuals from populations P and P (multiset union), until the mating pool is filled. Step (9). If the maximum number of generations is reached, then stop; otherwise go to Step (5). In algorithm, every chromosome is encoded by a permutation of integer; each integer represents a selected service instance for executing its corresponding task. We

3 Advances in Mechanical Engineering 3 apply two-point crossover and one-point mutation operators as usual. In Step (8), binary tournament selection with replacement is used. In the next four subsections, all technologies applied in are described in detail. f 2 F (n) x 3.1. Handling Constraints. The most common way of handling constraints is by using penalty techniques like exterior penalty method [14]. Two penalty functions are developed to handle the deadline and budget constraints. Let C B (x) be the budget penalty function and let T D (x) be the deadline penalty function; they are defined as follows: 1 min C (x) B if C (x) >B C B (x) ={ otherwise, T (x) D if T (x) >D T D (x) ={ otherwise. Since satisfying deadline and budget requirements is the primary goal of the Web service selection, the penalty is added separately to the optimization objectives. The new objective function is defined as follows: (7) Minimize g (x) = (y 1 (x),y 2 (x)), x S, (8) where y 1 (x) = T(x) + T D (x) and y 2 (x) = C(x) + C B (x) Determining the Population Size. In the MO evolution process, population size has great influence on the exploration efficiency of the solution space. It is difficult to determine an optimal population size. A fixed value of population size is set in literature [12]. For the optimization problem considered in this paper, let N be the number of nodes in a service composition application and let M = max{l(i) i V} be the maximum length of all service pools; the optimization problem may have N M solutions at most. It is shown that the solution space for each optimization problem is dependent on the composition scale and the length of service pool. Considering specific example of the optimization problem and the effective exploration of the solution space, the population size, denoted as L(P), canbe defined as follows: L (P) =M+N. (9) External population size, denoted as L(P ), can be defined by the user. In this paper, it depends on the test application; that is, L(P ) is equal to N Fitness Assignment. Afitnessfunctionisusedtomeasure the quality of the solutions according to the given optimization objectives. Considering the optimum and diversity of solutions, a method called Hybrid Nondominated Sorting andnichetechnique(hnsnt)isproposedtocalculatethe fitness of individuals. Including the evolutionary population and external population, the procedure is a three-stage process and is described as follows. Step 1 (Pareto-based sorting). First, it selects nondominated solutions in population P P as the first level nondominated F (n) z 2 Dominated individuals Nondominated individuals F (n) y Figure 1: The tradeoff curve of two-objective optimization. front, denoted as h i =1. The nondominated solutions in the rest of population are selected as the next level nondominated front, denoted as h i =2. The procedure is repeated until all individualsinthepopulationareclassified. Step 2 (compute dynamic niche radius). Niche techniques aim at preserving diversity in population, and the key problem is to determine the niche radius. Based on literature [15], an approximate calculation method for dynamic niche radius is given.as pointed outin [15], the tradeoffs for an m-objective optimization problem is in the form of an (m 1)-dimensional hypersurface. For the two-objective optimization problem in this paper especially, its one-dimensional surface is actually a curve. Let bethediameterofthehypersphereatgeneration n; it is the interpolated distance of the tradeoff curve covered by the nondominated individuals as shown in Figure 1. Obviously, computing accurately the distance is a complex procedure. However, the length of curve can be approximated by the average distance between the shortest distance min and the longest distance max =d(n) ;thatis, max ) (d(n) min +d(n) 2 f 1, (1) where min = ( 1 )2 +( 2 )2. Let σ (n) share denote the niche radius; its value at each generation population can be dynamically determined by σ (n) share = (11) ( P P ). Step 3 (compute the similarity of individuals). Let D i be the similarity; the processes are as follows.

4 4 Advances in Mechanical Engineering (1) For each j P P, for each j P P &&j m=; (12) =i, d ij = (y 2 (i) y 2 (j)) 2 +(y 1 (i) y 1 (j)) 2, { 1 ( d ij ) if d sh (d ij )= { σ (n) ij <σ (n) share share { otherwise, (2) D i = m/ P P. m+ = sh (d ij ); (13) Therefore, the fitness function of each individual can be defined as follows: fitness i = 1 (h i D i ). (14) 3.4. Preserving the Diversity of Population. The external population P stores the current nondominated solutions, which should be distributed evenly along the Pareto front. Meantime, it also participates in selection. So pruning the external nondominated set while maintaining its characteristics might be necessary or even mandatory. An approach, called NTSC (niche technique of similarity basedoncrowding),isgiven.letj be a nondominated individual in P, and let i be an individual in external population P. All the individuals in P that are dominated by j are denoted as a set K and then the NTSC procedure is described as follows. Step (1). Get any untreated individual j in P. Step (2).InitializesetK empty. Step (3).If i P dominates j,thengotostep(11). Step (4).SearchallindividualsinP dominated by i and add them to set K. Step (5). Remove those individuals of set K from P and add j to P. Step (6).IfP <L(P ),thengotostep(11). Step (7). ComputethesimilarityD k of individual k P via Step (3) in Section 3.3. Step (8). If there is an infeasible solution in P,thengotoStep (1). Step (9). Delete the feasible individual with the highest value D k,k P. Step (1). Delete the infeasible individual with the highest value D k,k P. Step (11). If all individuals in P have been treated, then stop; otherwise, go to Step (1). In this approach, nondominated infeasible solutions are stored when the number of feasible solutions is less than the upper limit of nondominated set. When updating, the infeasible solutions are preferentially eliminated from external population. 4. Experimental Results The proposed algorithm () is compared with algorithm [12], which is developed on NSGAII. The strategy of the population size in can also be adopted by. All these algorithms are coded in Java and performed on a Pentium IV with a 2.93 GHz processor and512mbramusingtheoperatingsystemwindowxp Parameter Setting. Since no standard test instances are available to test the presented algorithms, a DAG graph random generator is developed to generate application DAG with different scales and different structures [16]. The number of nodes varies between 3 and 12 with an increment of 3; that is, V = {3, 6, 9, 12}. The maximum out-degree of the DAG ranges from 1 to 3; that is, out degree = {1, 2, 3}. For each task in a DAG application, the length of its service pool length is 2. For those services in the same service pool, time, and cost quality parameters are randomly generated, but their execution costs are inversely proportional to their processing times. The behaviors of algorithms are also observed at three constraint levels, namely, relaxed constraint, medium constraints, and tight constraint. The relaxed constraint level assumes that users require relatively large deadline and budget, while tight constraint level assumes that users require low deadline and budget. The relaxed/tight deadlines and budgets of an application are determined by the maximum and minimum time/cost for the service composition execution. The deadline D and budget B are defined as follows: D=T min +θ (T max T min ), (15) B=C min +θ (C max C min ), (16) where T min is the minimum completion time of a composite instance and T max is the maximum completion time of a composite instance when all tasks selected are the slowest service for execution. The C min and C max have the same meaning of T min and T max.thevalueofθ varied from.3 to.7 to generate the tight, medium, and relaxed deadlines and budgets. On all test applications, 2 generations were simulated per optimization run. The probabilities of crossover (twopoint) and mutation were fixed (.8 and.1, resp.). The sizes

5 Advances in Mechanical Engineering 5 Table 1: The parameters setting and the experimental results of performance measures. N M P P U MOA C(X MOA,X MOE ) C(X MOE,X MOA ) U MOA U MOE I MOA I MOE Tight Middle Relaxed Tight Middle Relaxed Tight Middle Relaxed Tight Middle Relaxed Aver of evolutionary population and external population were determined. These parameters are shown in Table Performance Measures. In this section, three different quantitative interpretations of statistical MO optimization performances are applied to validate the proposed, together with its comparison with [12]. (1) Coverage of Two Sets (C) [15].Thecoverageoftwosets(C) is a measure to compare the domination of two populations in a pairwise manner. Let X,X Xbe two sets of decision vectors, the function C maps the ordered pair (X,X ) to the interval [, 1]. Consider C(X,X )= {a X ; a X :a a }. (17) X The value C(X,X ) = 1 means that all solutions in X are dominated by or equal to solutions X. While C(X,X )=means none of the points in X are covered by the population X.SinceC(X,X ) is not necessarily equal to C(X,X ),bothc(x,x ) and C(X,X ) have to be considered. (2) Uniform Distribution (U) [17]. An important metric in MO optimization concerns the distribution of individuals. The measure formula is referenced in literature [17]. The metricisabletomeasurethespreadofnondominatedindividuals effectively. The smaller its value, the better uniformity for the nondominated set. (3) The Number of Optimal Solutions (I). The metric concerns the numbers of nondominated individuals searched at the last generation. If the metric value is bigger, the algorithm should have better performance Experimental Results. Basedonthethreeperformance measures, the proposed has been compared with [12] for all above test applications. In our proposed method, four members of initial population were produced by MPBL [16], where θ {,.7,.8, 1} in (16). The and were independently executed 4 times for each test application. The arithmetic means of these results using themetricsdiscussedaboveareshownintable 1,whereMOA denotes the new approach and MOE denotes the compared method. For the measure of C, it can be seen from the results in Table 1 that has the higher mean value as compared to in different situations. The average value is up to 9.8%. Also, can almost dominate when the constraint is higher (middle, tight). The measure value, especially, is equal to 1 in case of N=6and higher (middle, tight) constraints. It clearly indicates that has the capability of providing useful nondominated solutions. Withrespecttothenumberofoptimalsolutions(I), it can be seen that can obtain a large number of nondominated solutions when the constraint is tight. The algorithm obtained more solutions than the proposal with medium and relaxed constraints. This indicates that has better performance than in tighter constraint. Although the number of solutions obtained by is more than those of, the solutions can almost be dominated by. In the context of U-measure of the nondominated population, gives the lower value of U than except for two test examples; that is, N=3and N=6with tight constraint. It is shown that has more capability to distribute the nondominated individuals evenly. This is because that does not have any operation to take care of the population distribution, whereas our approach applied the niche technologies to both fitness assignment and preservation of population diversity. So it can discover missing tradeoffs at each generation.

6 6 Advances in Mechanical Engineering (a) Tight constraint (θ =.3) 5 4 (a) Tight constraint (θ =.3) (b) Medium constraint (θ =.5) (b) Medium constraint (θ =.5) (c) Relaxed constraint (θ =.7) Figure 2: The Pareto-front with N=6and different constraints (c) Relaxed constraint (θ =.7) Figure 3: The Pareto-front with N= 12 and different constraints. The running results of the second group and the fourth group of test problems in Table 1 are randomly sampled. The population distribution of Pareto optimal solutions in the objectivedomainisshowninfigures2 and 3 (where denotes execution cost C(x) and denotes execution time T(x)). In general, the purpose of producing these figures is to visually inspect the performance of various algorithms in terms of thepopulation distribution, since quantitative

7 Advances in Mechanical Engineering 7 performance alone may provide insufficient information for full observation and understanding. As shown in Figures 2 and 3, is more capable of filling up the gap along the discovered Pareto front in all test application, especially those test examples with higher constraints. Although, hasagoodperformanceundertherelaxedconstraint,itisstill slightly inferior to the. 5. Conclusion A novel multiobjective optimization algorithm (named ) is proposed for solving constrained multiobjective Web service selection problem. In handing constraints, two time and cost penalty functions are added separately to the two objective functions without liner summation. The algorithm generates initial population by existing heuristic algorithm to achieve broader explorations in boundary regions. The population size depended on the complexity of test applications. The algorithm also applies the niche technology to both fitness assignment and population diversity, which aims to discover missing tradeoffs at each generation. Furthermore, extensive quantitative comparisons between and MOEWP on given test applications have been performed. Experimental results show that performed better than in searching and maintaining the uniform distribution of the nondominated solutions on all the test applications. Under tight and medium constraints especially, algorithm can discover more optimal solutions distributed uniformly along the tradeoff surface. Conflict of Interests The authors declare that there is no conflict of interests regarding the publication of this paper. Acknowledgments This work was supported by National Natural Science Foundation of China under Grant (no ) and Natural Science Foundation of Hebei Province in China under Grants (no. F29653 and no. F212489). In addition, our research is also supported by the Youth Science Research Foundation of Hebei Province in China under Grant (no ). References [5] L. Wei, Z. Lanfei, and D. Jianqiang, Research on an adaptive algorithm of service selection in pervasive computing, Journal of Digital Information Management, vol.11,no.6,pp , 213. [6] B. Chen, X. Peng, Y. Yu, and W. Zhao, Requirements-driven self-optimization of composite services using feedback control, IEEE Transactions on Services Computing,214. [7] Y. Xia, B. Cheng, J. Chen, X. Meng, and D. Liu, Optimizing services composition based on improved ant colony algorithm, Chinese Journal of Computers,vol.35,no.2,pp ,212. [8] X. Tao and C. Huowang, Chen Huowang, Evolutionary Algorithms for multi-objective optimization and decision-making problem, Engineering Science,vol.4,no.2, pp.59 68,22. [9] Z. H. Xin, S. Li Wei, P. Xin, Z. WenYun, and P. Moea, A multiobjective decision making aid EA for servicecomposition QoS optimization, Science China: Information Science,vol.43,no.1, pp , 213. [1] Q. Yu and A. Bouguettaya, Efficient service skyline computation for composite service selection, IEEE Transactions on Knowledge and Data Engineering, vol.25,no.4,pp , 213. [11] M. Alrifai, D. Skoutas, and T. Risse, Selecting skyline services for QoS-based web service composition, in Proceedings of the 19th International World Wide Web Conference (WWW '1),pp. 11 2, Raleigh, NC, USA, April 21. [12] J. Yu, M. Kirley, and R. Buyya, Multi-objective planning for workflow execution on Grids, in Proceedings of the 8th IEEE/ACM International Conference on Grid Computing (GRID 7), pp. 1 17, Austin, Tex, USA, September 27. [13] K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan, A fast and elitist multiobjective genetic algorithm: NSGA-II, IEEE Transactions on Evolutionary Computation,vol.6,no.2,pp , 22. [14] A. Homaifar, C. X. Qi, and S. H. Lai, Constrained optimization via genetic algorithms, Simulation,vol.62,no.4,pp , [15] K.C.Tan,T.H.Lee,andE.F.Khor, Evolutionaryalgorithms with dynamic population size and local exploration for multiobjective optimization, IEEE Transactions on Evolutionary Computation,vol.5,no.6,pp ,21. [16] Y. Yuan, X. Li, Q. Wang, and K. Wang, Time optimization heuristics for scheduling budget-constrained grid workflows, Journal of Computer Research and Development, vol.46,no.2, pp ,29. [17] Y.-W. Leung and Y.-P. Wang, U-measure: a quality measure for multi-objective programming, IEEransactiononSystems, Man and Cybernetics A: Systems and Humans,vol.33,no.2,pp , 23. [1] S. Dustdar and W. Schreiner, A survey on web services composition, International Journal of Web and Grid Services, vol.1,no.1,pp.1 3,25. [2] D. A. Menascé, Composing web services: a QoS view, IEEE Internet Computing, vol. 8, no. 6, pp. 88 9, 24. [3] T.Wen,G.Sheng,Q.Guo,andY.Li, Webservicecomposition based on modified particle swarm optimization, Chinese Journal of Computers,vol.36,no.5,pp ,213. [4] L.Zeng,B.Benatallah,A.H.H.Ngu,M.Dumas,J.Kalagnanam, and H. Chang, QoS-aware middleware for Web services composition, IEEE Transactions on Software Engineering, vol. 3, no. 5, pp , 24.