Presented in the Second International Conference on Advanced Data and Information Engineering (DaEng-2015), Bali, Indonesi, April 25-26. To be published in LNEE-Springer (IN PRESS). Job Scheduling Technique for Infrastructure as a Service Cloud Using an Enhanced League Championship Algorithm Shafi’i Muhammad Abdulhamid1,2 , Muhammad Shafie Abd Latiff*1 and Mohammed Abdullahi 1,3 1 2 Faculty of Computing, Universiti Teknologi Malaysia, Johor Bahru MALAYSIA. Department of Cyber Security Science, Federal University of Technology Minna, NIGERIA. 3 Department of Mathematics, Ahmadu Bello University, Zaria, Kaduna State, NIGERIA. (E-mail: shafii.abdulhamid@futminna.edu.ng, shafie@utm.my, muham08@gmail.com) Abstract. League Championship Algorithm (LCA) is a sports-inspired population based algorithmic framework for global optimization over a continuous search space first proposed by Ali Husseinzadeh Kashan in the year 2009. A general characteristic between all population based optimization algorithms similar to the LCA is that, it tries to progress a population of achievable solutions to potential areas of the search space when seeking the optimization. In this paper, we proposed a job scheduling algorithm based on an enhanced LCA optimization technique for the infrastructure as a service (IaaS) cloud. Three other established algorithms i.e. First Come First Served (FCFS), Last Job First (LJF) and Best Effort First (BEF) were used to evaluate the performance of the proposed algorithm. All four algorithms assumed to be nonpreemptive. The parameters used for this experiment are the average response time, average completion time and the makespan time. The results obtained shows that, LCA scheduling algorithm perform moderately better than the other algorithms as the number of virtual machines increases. Keywords: League Championship Algorithm; IaaS Cloud; Job Scheduling Algorithm; Cloud Computing; Cloud Scheduling; Optimization Algorithm. 1 Introduction The league championship algorithm (LCA) is a new optimization scheme designed based on the inspiration of soccer competitions in a championship league. The LCA is a population based algorithmic framework for global optimization over a continuous search space first proposed by Kashan [1]. A more detailed LCA can found in [2]. It is a stochastic population based algorithm for continuous global optimization which tries to imitate a championship situation where synthetic football clubs participate in an artificial league for a number of weeks. This algorithm has been tested in many areas and performed creditably well as compared to other known optimization schemes or heuristics algorithms [3, 4]. In infrastructure-as-a-service (IaaS) cloud computing, computational resources such as the virtual machines (VMs) are provided to remote clients in the form of rents (pay per use). For a cloud client, he/she can demand multiple cloud services concurrently [5]. Job scheduling had been studied in high performance computing in cloud systems. However, the autonomous attribute and the resource heterogeneity within the clouds and the VM execution necessitate different schemes for job scheduling in the IaaS cloud computing, especially in the federated heterogeneous multi-cloud system [6]. 1 Presented in the Second International Conference on Advanced Data and Information Engineering (DaEng-2015), Bali, Indonesi, April 25-26. To be published in LNEE-Springer (IN PRESS). The aim of this paper is to propose a job optimization scheme in a heterogeneous IaaS cloud computing system based on t he LCA scheme, which enables a non-preemptable task scheduling. Job scheduling in the IaaS cloud is a nondeterministic polynomial (NP-hard) problem. Section two reviewes some related works in LCA and also in jobs scheduling in IaaS cloud. Section three puts forward a proposed scheme from jobs scheduling in IaaS cloud by enhancing the LCA.Section four presents the simulation and results, while section five presents the conclusion and future works. 2 Related Works Abdulhamid and Abd Latiff [7] presents a paper that proposed a League Championship Algorithm (LCA) based scheme for global optimization tasks schedule in cloud environment, which mimics the sport league championships. It is a new algorithm for numerical function optimization. Kashan and Karimi [8] tests the effectiveness of the proposed optimization algorithm by measuring the test functions from a recognized yardstick, usually adopted to authenticate new constraint-handling algorithms strategy. While Sebastián and Isabel [9] presents an implementation of the LCA in a Job Shop scheduling in an industrial situation. Diangang, Zhiya [10] presents an efficient VM scheduling scheme in IaaS cloud computing system. Shen, Deng [11] and [12, 13] presents a category of cloudbased, online, hybrid scheduling procedures that reduces cost by using both ondemand and reserved case in point. Sun, Ji [14] puts forward a VM scheduling technique, motion and disaster recovery scheme for IaaS cloud environment based on a runtime and average usage of the three layers of IaaS cloud. A double combinatorial resource allocation scheme was also proposed for distributed cloud computing in [15]. 3 Proposed League Championship Algorithm The proposed LCA based job scheduling scheme was designed by enhancing the LCA metaheuristic algorithm inspired based on the metaphor of sporting contests in a round-robin sport leagues. A detailed LCA steps can found in [2, 7] including the seven idealized rules guiding its implementation. 3.1 Parameters Matching In order to achive optimization with the proposed algorithm (LCA) in scheduling cloud jobs, we must first have to match the corresponding variables or parameters of the two systems. To achive this, a simple comparison was made with the variables of a known evolutionary algorithm (EA) and the following matching was achived; • League L = population • week t = iteration • team i = ith member in the population • formation = solution • playing strength ( ) = fitness value 2 Presented in the Second International Conference on Advanced Data and Information Engineering (DaEng-2015), Bali, Indonesi, April 25-26. To be published in LNEE-Springer (IN PRESS). • 3.2 number of seasons S = maximum iterations Winner/Loser Determination One of the most important feature of the LCA is the winner/loser determination scheme. In this resaerch work, we to utilized this feature in determining which job is scheduled on which VM in the IaaS cloud. Considering a normal league system, teams play each other weekly and their game result is evaluated on the basis of win/loss/tie for each of the teams. For instance, in football league, each club is to get three points for win, zero for loss and one for draw/tie. By ignoring, the irregular abnormalities which may ensure even outstanding clubs in a variety of unsuccessful outcomes, it is probable that a more dominant club having a superior playing pattern defeats the lesser team. In an ideal league situation that is free from uncertainty effects, an assumption can be easily made for a linear correlation between the playing pattern of a club and the result of its matches. Utilizing the playing power condition, the winner/loser in LCA is determined in a stochastic approach with criteria that the probability of winning for a club is relative to its degree of fit. Given teams i and j playing a league match at week t, with their formations and and playing powers ( ) and ( ), correspondingly. Let represents the probability of team i to defeat team j at week t ( is defined respectively). Given also ideal value (e.g., a lower limit on the best value). = be an (1) From the idealized 3 rule we can also write: + =1 (2) From equations (1) and (2) above we solve for = (3) In order to find the winner or loser, a random number in between 0 to1 is generated; if , team i wins and team j loses; the generated number is less than or equal to otherwise j wins and i loses. This method of finding the win or lose is in line with the idealized rules. If ( ) be arbitrarily closed to ( ), then can be arbitrarily closed to 1/2. Moreover, if ( )becomes far greater than ( ( ), then approaches to one. Then, the value of feature, we use from the best function (i.e., = min $%,….,) { (+ )}. ), namely ( )» may be unavailable in the value found so far 3 Presented in the Second International Conference on Advanced Data and Information Engineering (DaEng-2015), Bali, Indonesi, April 25-26. To be published in LNEE-Springer (IN PRESS). 3.3 The LCA Enhancement In the original LCA, two teams i and jare to contest for a football match in every week t, the winner is to maintain its formation and its playing strenght in order to play the next match against the next team j. This scenario is similar to the contest between jobs (represented as Cloudlets in CloudSim) in cloud environment in order to secure access to resources for execution. If we consider the jobs (cloudlets) to be the teams I and j, one of the modification we introduced here is that, the winner of the contest will get access to the cloud resources and be executed. Therefore, new cloudlets i-1 and j-1 with new formations Xti-1 and Xtj-1 and also new playing strenghts f(xti-1) and f(xtj-1) will be generated to contest for the next match. This partern will be maintained untill all the cloudlets are executed. Fig. 1. shows the hierarchical execution of jobs using the enhanced LCA scheme. team i f(xti) 1st winner (execute) team j f(xtj) team i-1 f(xti-1) new formation team j-1 f(xtj-1) new formation Σ 2nd winner (execute) team i-2 f(xti-2) new formation Σ i-th winner (execute) solution team j-2 f(xtj-2) new formation Fig. 1. Enhanced LCA Algorithm The total sum of the executed cloudlets can be obtainted by summing-up Σ all the executed jobs, starting from the 1st winner of the contest to the i-th winner. While the makespan time is the maximum completion time of a cloudlets. It is also described as the peroid from the start of the first winner execution to the end of the last cloudlet execution in the schedule. It assumes that the cloudlets are ready at time zero and resources are continuously available during the whole scheduling. Mathematically, makespan can be expressed as; (4) Makespan Cmax = max{Ci’} = max{C1’, C2’, ..., Cn’} where, Ci’ is the completion of task i. The lesser the makespan the better the efficiency 4 Presented in the Second International Conference on Advanced Data and Information Engineering (DaEng-2015), Bali, Indonesi, April 25-26. To be published in LNEE-Springer (IN PRESS). of the algorithm, meaning less time is taken to execute the algorithm. Algorithm 1. League Championship Algorithm [2] 1. Set the league size L and the amount of seasons S and set t=1 2. Create a league timetable 3. Set team formations and establish the playing strengths ( ) along with them. Let the initialization be the teams’ new formation; 4. While [ t ≤ S (L-1]) ] 5. Using the league timetable at week t, find the winner/loser among each pair of teams by means of their playing strength; 6. t=t+1 7. For i = 1 to L 8. Formulate a new formation Xti-1 for team i1 for the next coming fixture, while taking into consideration the team’s new formation f(Xti-1) and previous week events. Determine the playing strength of the new system; 9. If the new formation is the fittest one, regard the new formation as the team’s current finest formation; 0. End For 1. If mod (t, L-1) = 0 2. Generate a league timetable; 3. End if 4. End While 4 Simulation and Results Three other well established algorithms i.e. First Come First Served (FCFS), Last Job First (LJF) and Best Effort First (BEF) were used to evaluate the performance of the proposed LCA based scheduling technique. The parameters used for measuring the scheduling algorithms in this experiment are based on three factors - Average Response Time, the Average Completion Time and Makespan Time. The data set was formed by using the Delft University of Technology workload traces in CloudSim toolkit. The experiment was performed by varying the number of VMs in the IaaS cloud from 10 to 130. All four algorithms assumed to be non-preemptive. Fig. 2 shows a prototype of the simulation setup environment for the experiment. 5 Presented in the Second International Conference on Advanced Data and Information Engineering (DaEng-2015), Bali, Indonesi, April 25-26. To be published in LNEE-Springer (IN PRESS). C lou d U ser C lo ud U ser C lo ud U se r V IR TU A L IN F R A S T R U C T U R E L C A -B ase d S che duling Algorithm VM VM VM VM VM VM VM H yp ervisors H os t H yperv iso rs H ost VM VM H yp ervisors H o st Fig. 2: IaaS Cloud Setup The experiment was repeated five times and the average total completion time for each of the algorithms was captured and tabulated. The total completion time is the total execution time plus the total waiting time of the job. Average Completion Time (Sec) 80000 70000 60000 50000 LJF 40000 FCFS 30000 BEF 20000 LCA 10000 0 10 30 50 No. of70VM 90 110 130 Fig. 3. Average Completion Time Fig. 3. shows that the average total completion times as calculated by the four scheduling schemes. The average total completion time as processed by the LCA scheduling algorithm is shorter than the other three algorithms, i.e. FCFS, LJF and BEF, especially as the VMs increases. The LJF has the longest completion time amongst the algorithms under consideration. This results obtained from the IaaS cloud environment also shows that, LCA scheduling algorithm perform moderately better 6 Presented in the Second International Conference on Advanced Data and Information Engineering (DaEng-2015), Bali, Indonesi, April 25-26. To be published in LNEE-Springer (IN PRESS). than the FCFS and the LJF algorithms throughout the experiment, but only outperformed the BEF as we continue to increase the number of the VMs. Average Response Time (Sec) 25 20 15 FCFS BEF 10 LJF 5 LCA 0 10 30 50 70 90 110 130 No. of VM Fig. 4. Average Response Time Makespan Time (Sec) Fig. 4. shows that the average response time for all the four algorithms have declined by adding more VMs to execute the jobs. The LJF and FCFS scheduling algorithms result in poor response times as compared to LCA at the beginning of the scheduling process. But the BEF outperformed all at the beginning of the experiment. As the number of VMs increased to execute the jobs, the LCA begins to perform better than BEF, FCFS and the LJF. 90000 80000 70000 60000 50000 40000 30000 20000 10000 0 FCFS LJF BEF LCA 10 30 50 70 90 110 130 No. of CloudLets Fig. 5. Makespan Time 7 Presented in the Second International Conference on Advanced Data and Information Engineering (DaEng-2015), Bali, Indonesi, April 25-26. To be published in LNEE-Springer (IN PRESS). Fig. 5. shows the makespan times as calculated by the four scheduling schemes. The makespan time as processed by the LCA scheduling algorithm is lesser than the other three algorithms, i.e. FCFS, LJF and BEF, especially as the number of tasks increases. The FCFS has the highest makespan time amongst the algorithms under consideration. This results obtained from the IaaS cloud environment also shows that, LCA scheduling algorithm perform moderately better than the FCFS and the BEF algorithms throughout the experiment, but only outperformed the LJF as the number of tasks increases. The implication of this result is that, the proposed LCA scheduling scheme will help the cloud customers to save more money while using the cloud. This is because the algorithm helps to reduce the makespan time which is the maximum completion time of tasks, making the customers to spend lesser time in the pay per use IaaS cloud. 5 Conclusion and Future Works The LCA based scheduling technique for jobs optimization in the IaaS cloud has not been adapted in this environment before. After a comprehensive review of the proposed algorithm, the LCA-based scheduling scheme shows great prospects of performing well in this area as it had performed in solving other NP-complete problems in other areas of research. The results obtained from this experiment shows that, LCA-based scheduling algorithm performed better than the FCFS, LJF and BEF algorithms. Especially in reducing the average response time and the average completion time of the jobs. Cloud being a pay-per-user, it implies that, the LCA saves cost for the cloud users than the LJF, BEF or the FCFS scheduling algorithms, as it take lesser time for the response and the completion of job processing. The LCA is a new sport-based optimization technique that has the potential of adaptation in various fields of research. Further reseaches are also required to minimize the makespan time for the scheduling jobs and improve resource allocation within the VMs in the IaaS cloud. This proposed algorithm can as well be extented in other areas such as search techniques in big data, chaotic sequences in some engineering design, assignment problem in graph coloring and known NP-hard problems. Acknowledgments The authors would like to express their appreciation for the support of Universiti Teknologi Malaysia (UTM) Research University Grant Q. J130000.2528.05H87 sponsorship for this research and and the Nigerian Tertiary Education Trust Fund (TetFund) for their support. We also wish to appreciate Dr. Kashan A. H. for his assistance on research materials. References 1. 2. Kashan, A.H.: League championship algorithm: a new algorithm for numerical function optimization: Soft Computing and Pattern Recognition, 2009. SOCPAR'09. IEEE (2009). Kashan, H.A.,: League Championship Algorithm (LCA): An algorithm for global optimization inspired by sport championships: Applied Soft Computing, 16, 171-200 (2014). 8 Presented in the Second International Conference on Advanced Data and Information Engineering (DaEng-2015), Bali, Indonesi, April 25-26. To be published in LNEE-Springer (IN PRESS). 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14. 15. Kahledan, S.,: A League Championship Algorithm for Travelling Salesman Problem: Azad University, Najaf Abad branch, Iran (in Persian), (2014). Edraki, S.,: A new approach for engineering design optimization of centrifuge pumps based on league championship algorithm: Azad University, Science and Research branch, Iran (in Persian), (2014). Li, J., Qiu, M., Ming, Z., Quan, G., Qin, X., and Gu, Z.: Online optimization for scheduling preemptable tasks on IaaS cloud systems: Journal of Parallel and Distributed Computing, 72(5), 666-677, (2012). Mohammed B. B., Abd Latiff. M. S., Abdulhamid S. M. and Cheah T. L.: Grid-based Search Technique for Massive Academic Publications: The 2014 Third ICT International Student Project Conference (ICT-ISPC2014). 2014, IEEE Thailand: Thailand. 175 - 178, (2014). Abdulhamid, S. M. and Abd Latiff M. S.: League Championship Algorithm Based Job Scheduling Scheme for Infrastructure as a Service Cloud: 5th International Graduate Conference on Engineering, Science and Humanities (IGCESH2014) (2014). Kashan, H.A. and B. Karimi: A new algorithm for constrained optimization inspired by the sport league championships. in Evolutionary Computation (CEC), 2010 IEEE Congress on. (2010). Sebastián, A.R. and L.R. Isabel, Scheduling To Job Shop Configuration Minimizing The Makespan Using Champions League Algorithm. Fray Ismael Leonardo Ballesteros Guerrero, OP–Decano de División de Arquitectura e Ingenierías, Universidad Santo Tomás Seccional Tunja, (2014). Diangang, W., et al., Efficient VM scheduling method in IaaS cloud environment. Advanced Computer Control, 59, 447 (2014). Shen, S., Deng, K., Iosup, A, and Epema, D: Scheduling jobs in the cloud using ondemand and reserved instances: Euro-Par 2013 Parallel Processing. Springer. 242-254 (2013). Abdulhamid, S.M., M.S.A. Latiff, and M.B. Bashir: On-Demand Grid Provisioning Using Cloud Infrastructures and Related Virtualization Tools: A Survey and Taxonomy. International Journal of Advanced Studies in Computer Science and Engineering IJASCSE, 3(1), 49 - 59 (2014). Abdulhamid, S.M. Abd Latiff, M.S. and M.B. Bashir: Scheduling Techniques in OnDemand Grid as a Service Cloud: A Review. Journal of Theoretical and Applied Information Technology, 63, (2014). Sun, A., Ji, T., Yue, Q. and Yang, S.: Virtual machine scheduling, motion and disaster recovery model for IaaS cloud computing platform. in Conference Anthology, IEEE. (2013). Sun, J., Wang, X., Li, K., Wu, C., Huang, M., and Wang, X.: An Auction and League Championship Algorithm Based Resource Allocation Mechanism for Distributed Cloud, in Advanced Parallel Processing Technologies, Springer. 334-346, (2013). 9