Efficient Multiple Ant Colony Algorithm for Job Scheduling In Grid Environment

Transcription

1 Efficient Multiple Ant Colony Algorith for Job Scheduling In Grid Environent S.Uarani 1, L.M.Nithya 1, A.Shanuga 2 1 SNS College of Technology, Coibatore 2 Bannari Aan Institute of Technology, Erode. Abstract:-Grid coputing is now being used in any applications that are beyond distribution and sharing of resources. One priary issue associated with the efficient utilization of heterogeneous resources in a grid environent is task scheduling. Task Scheduling is an iportant issue of current ipleentation of grid coputing. The deand for scheduling is to achieve high perforance coputing. If large nubers of tasks are coputed on the geographically distributed resources, a reasonable scheduling algorith ust be adopted in order to get the iniu copletion tie. Typically, it is difficult to find an optial resource allocation for specific job that iniizes the schedule length of jobs. So the scheduling proble is defined as NP-coplete proble and it is not trivial. Heuristic algoriths are used to solve the task scheduling proble in the grid environent and ay provide high perforance or high throughput coputing or both. In this paper, a new iproved ultiple ant colony optiization scheduling algorith is proposed. The proposed scheduler allocates the best suitable resources to each task with inial akespan and execution tie. The experiental results are copared which shows that the algorith produces better results when copared with the ultiple ant colony algorith. Keywords: grid coputing; task scheduling; ultiple ant colony optiization; ant colony algorith; iproved ultiple ant colony algorith 1. INTRODUCTION: Grid coputing has eerged as an iportant new field, distinguished fro conventional distributed coputing by its focus on large-scale resource sharing, innovative applications, and in soe cases, high-perforance orientation. Grid is a wide-scale, distributed coputing infrastructure that proises to support resource sharing and coordinated proble solving in dynaic, ulti institutional Virtual Organization [7]. Grid coputing is applying the resources of any coputers in a network for a single proble at the sae tie - usually to a scientific or technical proble that requires a great nuber of coputer processing cycles or access to large aounts of data. Scheduling in a grid coputing syste is not as siple as scheduling on a any achine because of several factors. These factors include the fact that grid resources are soeties used by paying custoers who have interest in how their jobs are being scheduled. Also, grid coputing systes usually operate in reote locations so scheduling tasks for the clusters ay be occurring over a network. It is because of these reasons that looking at scheduling in grid coputing is an interesting and iportant proble to exaine [1]. There are two types of scheduling naely static scheduling and dynaic scheduling in grid coputing syste. For the static scheduling, jobs are assigned to suitable resources before their execution begin. However, for the dynaic scheduling, reevaluation is allowed of already taken assignent decisions during job execution [2]. Figure 1: Grid Coputing Environent It can trigger job igration or interruption based on dynaic inforation about the status of the syste and the workload. This job scheduling is the apping of jobs to specific physical resources, trying to iniize soe cost function specified by the user. This is a NP-coplete proble and different heuristics ay be used to reach an optial or near optial solution. Effective coputation and job scheduling is rapidly becoing one of the ain challenges in grid coputing and is seen as being vital for its success. In this paper, we propose a new iproved ultiple ant colonies optiization approach for achieving optial task scheduling in grid coputing. This paper is organized as follows. Section II describes the use of ant and ultiple ant colony algoriths in grid coputing. In Section III, It covers proble forulation and ethodology used in existing syste. In Section IV, our new iproved MACO approach is proposed for the scheduling proble. In Section V, coputational experients and coparison studies are reported. Soe concluding rearks are ade in Section VI. 3388

2 2. RELATED WORKS ON ACO AND MACO IN GRID COMPUTING ENVIRONMENT: One otivation of grid coputing is to aggregate the power of widely distributed resources, and provide nontrivial services to users. To achieve this goal, an efficient grid scheduling syste is an essential part of the grid. Rather than covering the whole grid scheduling area, survey provides a review of the subject ainly fro the perspective of scheduling algoriths. The state of current research on scheduling algoriths for the new generation of coputational environents will be reviewed in this section. Here, a collection of soe coonly used heuristics algoriths literature has been selected and discussed in the following section. Ant Colony Optiization (ACO) [3] has been used as an effective algorith in solving the scheduling proble in grid coputing. ACO is inspired by a colony of ants that work together to find the shortest path between their nest and food source. Every ant will deposit a cheical substance called pheroone on the ground after they ove fro the nest to food sources and vice versa. Therefore, they will choose the shortest or optial path based on the pheroone value. The path with high pheroone value is shorter than the path with low pheroone value. This behavior is the basis for a cooperative counication. Balanced job assignent based on ant algorith for coputing grids called BACO was proposed by [4]. The research ais to iniize the coputation tie of job executing in grid environent which focused on load balancing factors of each resource. By considering the resource status and the size of the given job, BACO algorith chooses optial resources to process the subitted jobs by applying the local and global pheroone update technique to balance the syste load. Local pheroone update function updates the status of the selected resource after job has been assigned and the job scheduler depends on the newest inforation of the selected resource for the next job subission. Global pheroone update function updates the status of each resource for all jobs after the copletion of the jobs. Fro the experiental result, BACO is capable of balancing the entire syste load regardless of the size of the jobs. It is based on the iproved ACO that proposed by [5]. The pheroone update function in this research is perfored by adding encourageent, punishent coefficient and load balancing factor. The initial pheroone value of each resource is based on its status where job is assigned to the resource with the axiu pheroone value. The strength of pheroone of each resource will be updated after copletion of the job. If a resource copleted a job successfully, ore pheroone will be added by the encourageent coefficient in order to be selected for the next job execution. If a resource failed to coplete a job, it will be punished by adding less pheroone value. The load of each resource is taken into account and the balancing factor is also applied to change the pheroone value of each resource. An Enhanced Ant Colony Algorith proposed in [6] is allocating an application to a host fro a pool of available hosts and applications by selecting the best atch. The proposed algorith uses two types of ET (Execution Tie) atrices and finds the list of available resources in the grid environent, for the ET atrix and start the scheduling. Based on the ET atrix the algorith calculates the probabilistic akespan. The proposed scheduler allocates adopt the syste environent freely at runtie. This resource optially and adaptively in the scalable, dynaic and distribute controlled environent. Multiple Ant Colony Optiization (MACO) approach [7] presented for load balancing in circuit-switched networks. MACO uses ultiple ant colonies to search for alternatives to an optial path. One of the ipetuses of MACO is to optiize the perforance of a congested network by routing calls via several alternatives paths to prevent possible congestion along an optial path. In MACO, each group of obile agents corresponds to a colony of ants, and the routing table of each group corresponds to a pheroone table of each colony [8]. By adopting the MACO approach, it ay be possible to reduce the likelihood that all obile agents establish connections using only the optial path [8]. The advantage of using MACO in circuit-switched routing is that it is ore likely to establish connections through ultiple paths to help balance the load but does not increase the routing overhead. IMACO fraework is proposed in [9]. In this fraework there are two levels of interaction the first one is the colony level and the second one is the population level. The colony level interaction can be achieved through the pheroone depositing process within the sae colony; the pheroone updating echanis is responsible for the ipleentation of this kind of interaction. The population level interaction is achieved by evaluating the pheroones of different colonies using soe evaluation function; the responsibility here is of the pheroone evaluating echanis [10]. IMACO-AVG ethod [11] proposed exploration and exploitation process. Exploration and exploitation is controlled by the paraeter q 0 whose value is in [0, 1]. It is usually used in ant s probabilistic decision as trade-off between exploitation and exploration. Setting q 0 to zero eans that the algorith uses a pure exploration while pure exploitation is reached by setting q 0 to one. This technique enables the utilized ant colonies to work with different levels of exploration. Soe will prefer high exploration of new areas of search space while other colonies will prefer high exploitation search history. Inforation exchange in ultiple ant colony algoriths [12] is proposed that parallelization where an inforation exchange between several colonies of ants is done every k generations for soe fixed k. They show by using siulations how uch the running tie of the algorith 3389

3 decreases with an increasing interval between the inforation exchanges. But it is not discussed how this influences the quality of the solutions. Through our literature survey on current scheduling algoriths are working in the grid coputing scenario, we can find that heterogeneity, dynais, coputation and data separation are the priary challenges concerned by current research on this topic. 3. MULTIPLE ANT COLONY ALGORITHM AND PROBLEM DESCRIPTION: Recently, ant colony optiization (ACO) has been suggested to solve the task scheduling proble. But ACO approaches using a single colony syste ay suffer fro specific local optia because of its tendency to use the positive feedback echanis of pheroone, ultiple ant colonies optiization (MACO) is eployed to avoid this by using several ant colonies to solve cobinatorial optiization probles cooperatively. To iprove the perforance of ACO approaches, MACO considers both positive and negative feedbacks in searching solutions, sharing the search inforation, and exploring a large area of the search space with utual cooperation of ant colonies. So MACO approaches have been explored for several optiization probles [14]. MACO sees to be appropriate approach to iprove the perforance of ACO algorith. This algorith offers good opportunity to explore a large area of the search space and find optial solution. All colonies construct their solutions in parallel [13] and interaction echaniss are designed for sharing experiences aong colonies. Assue that M ant colonies would be used to tackle the scheduling proble, and each colony contains N ants for the search procedure. The Fraework of the MACO Approach: First the initialization of the algorith is presented, Second Local and global phenoenon updating is perfored, Third construct the solution for deterine the akespan tie and then terinal test of the algorith is take place. Initially several colonies of ant syste are created, and then they perfor iterating and updating their pheroone arrays respectively until one ant colony syste reaches its local optiu solution. Every ant colony syste owns its pheroone array and paraeters and records its local optiu solution. Furtherore, once an ant colony syste arrives at its local optiu solution, it updates its local optiu solution and sends this solution to global best-found center. In general, the approach can be briefly sketched as follows. Initialization of Algorith: Assue that M ant colonies would be used to tackle the scheduling proble, and each colony contains N ants for the search procedure. We denote by ant (, n) the nth ant in the th colony. At the beginning, ants are distributed on coputing nodes randoly, and the pheroone value τ j of the th colony on node Cj is initialized as a sall value. Interaction of Multiple Ant Colony Syste: Pheroone is used as the interaction echanis not only between the ants of the sae colony but also aong ant colonies. As traditional ACO approaches, each colony has its own pheroone to interact between the ants of the sae colony. Furtherore, pheroone inforation is also used for the interaction of ant colonies. The colony level interaction is achieved by evaluating the pheroones of different colonies. More specifically, the evaluated pheroone τj on node Cj in ters of pheroone values of all colonies is defined as follows, τ j = Σ ε [1, M] τ j / M (1) Where M-no of ant colonies in syste, τ j is coputed in ters of pheroone values of all colonies The pheroone evaluation echanis averages the pheroone values of all colonies, which represents inforation of all colonies. Based on the average of the available experiences of ants of all colonies, an ant will decide how to choose an edge. Pheroone Updating: As ant (, n) copletes its tour, local pheroone updating is applied on the visited nodes. More specifically, if ant (, n) assigns task Ti to node Cj, the local pheroone update is given by: τ n j = (1- ) τ j + τ j (2) Where - Evaporation rate In global pheroone updating phrase, after all ants of all colonies construct their solutions, the ant achieving the best-so-far solution in its colony will deposit an aount of pheroone on the edges of its path according to the following rule τ n j = (1-λ) τ j + λ τ j (3) Where λ- Evaporation Thus, the ant finding the solution with the iniu akespan can lay a larger intensity of the pheroone on its tour. Solution construction: In this phase, ant (, n) oves through coputing nodes and assigns task Ti to node Cj probabilistically in ters of pheroone and heuristic inforation until all tasks have been allocated. The probability P n ij for ant (,n) to assign the task to node is defined as; P n ij = τ j Ŋ n ij / τ j Ŋ n ij (4) Where Ŋ n ij the heuristic inforation value for evaluating the assignent of task Ti to node Cj for ant (, n). τ j - is coputed in ters of pheroone values of all colonies. This approach terinates when the global pheroone updating is not iproved in successive predefined no of solutions. 3390

4 4. PROPOSED ENHANCED MULTIPLE ANT COLONY ALGORITHMS FOR JOB SCHEDULING IN GRID COMPUTING: One of the ain challenges is to find the best or optial resources to process a particular job in ter of iniizing the job coputational tie. The scheduling proble ais to iniize the total execution tie of tasks as well as to achieve a well-balanced load across all nodes in grid syste. Therefore in this paper one of the ain objectives is to iniize the akaspan, which represents the latest copletion tie aong all the tasks as well as utilizing the resources by efficient way. The proposed enhanced ultiple ant colony algorith is used to solve the large coplex probles. It requires grid scheduling to achieve high perforance. Before starting the grid scheduling, the expected execution tie for each task on each achine ust be estiated and represented by an ET atrix [15]. The ET atrix will have N x M entries, where N is the nuber of independent jobs to be scheduled and M is the nuber of resources which is currently available. Each job s workload is easured by illions of instructions and the capacity of each resource is easured by MIPS. Each row of ET atrix consists of the estiated execution tie for a job on each resource and every colun of the ET atrix is the estiated execution tie for a particular resource of list of all jobs in the job pool. The copletion tie of ith job on the jth achine is CT ij = Ready i + EET ij Where EETij is the expected execution tie of task Ti on node Cj when task Ti is assigned to node Cj, otherwise EETij is 0. Ready (i) is the achine availability tie, i.e. the tie at which achine j copletes any previously assigned tasks. The proposed efficient MACO has changed the basic pheroone updating of ultiple ant colony algorith.in ultiple any colonies; all the ants aintain a separate list in each colony. Whenever they select next task and resource, they are added into the list. At each iteration the ants calculate the new pheroone level of the eleents of the solutions is changed by applying following local updating rule τ j = 1 / EET ij This local pheroone updating rule is ipleented after each ant copletes its tour in each ant colony. Based on local pheroone updating we calculate the global pheroone value. Moreover, global pheroone updating is done by the best ant of each colony after all ants of all colonies coplete their tours, τ j = (1-λ) τ j + λ τ j n The colony level interaction is achieved by evaluating the pheroones of different colonies. More specifically, the (5) (6) (7) evaluated pheroone τj on node Cj in ters of pheroone values of all colonies is defined as follows, τ j = Σ ε [1, M] τ j / M (8) Where M-no of ant colonies in syste, τ j - is coputed in ters of pheroone values of all colonies The probability P n ij for ant (,n) to assign the task to node is defined as; P n ij = τ j Ŋ n ij / τ j Ŋ n ij (9) Where Ŋ n ij the heuristic inforation value for evaluating the assignent of task Ti to node Cj for ant (, n). τ j - is coputed in ters of pheroone values of all colonies. The pheroone evaluation echanis averages the pheroone values of all colonies, which represents inforation of all colonies. Based on the average of the available experiences of ants of all colonies, an ant will decide how to choose an edge. The inclusion of expected execution tie atrix in local pheroone updating is iproving the perforance of ultiple ant colony algoriths. This iproveent is in ters of the decrease in akespan tie. 5. EXPERIMENTAL RESULTS: The proposed ultiple ant colonies construct their solutions in parallel and interaction echaniss are designed for sharing experiences aong colonies. The perforance evaluation of our proposed EMACO algorith and the coparison study with other algoriths for task scheduling have been done on GridSi, which is developed to support siulation of heterogeneous grid resources and application odels. Coparison and discussion: Fro coparison purposes, ACS was run using 10, 20, 30, 40, 50,60,70,80 and 80 ants such that the nuber of ants used by ACS is equal to the su of ants of all colonies in MACO and EMACO. Here we use 4 gridlets and 3 resources and the results of the experients are shown in table 1 and 2. Table 1: probabilistic akespan of MACO Probability Makespan [0][0] 818 [0][1] 652 [0][2] 563 [0][3] 506 [1][0] 491 [1][1] 404 [1][2] 348 [1][3] 343 [2][0] 309 [2][1] 280 [2][2] 176 [2][3]

5 The inclusion of EETij execution tie of the i th job by the j th achine (predicted) in the calculation of probability, that the j th achine will be free, has shown a positive result in perforance iproveent. This iproveent is in ters of the decrease in akespan tie. Table 2: Probabilistic akespan of EMACO Probability Makespan [0][0] 799 [0][1] 550 [0][2] 490 [0][3] 476 [1][0] 359 [1][1] 350 [1][2] 332 [1][3] 312 [2][0] 275 [2][1] 230 [2][2] 141 [2][3] In this ethod first find the proble resource-those with total execution ties equal to the akespan of the solution, and attept to ove or swap set of jobs fro the proble processor to another resource that has the iniu akespan as copared with all other resources. The search is perfored on each proble processor and continues until there is no further iproveent in the global pheroone of the solution. These results are shown in figure 2. Figure 2: MACO and EMACO perforance coparison based on probabilistic akespan Figure 3: ACO, MACO and EMACO perforance coparison Fro the experiental results of figure 3, we can see that all the optial akespan obtained by EMACO is better than the MACO and ACO. Here we use ten ant colonies and the akespan of ACO decreased roughly as the no of ants increased. Due to the influence of stagnation situation, ACO cannot benefit fro the increase of the nuber of ants. In contrast, the solution quality of MACO was iproved through the cooperation of ultiple colonies. But EMACO can be iproved using soe for of local searching process. Figure 4: MACO and EMACO perforance coparison based on siulation tie Figure 4 shows that the siulation tie for enhanced ultiple ant colony algorith and ultiple ant colony algorith. Enhanced ultiple ant colony algorith produces iniu siulation tie when copared with traditional process. This is because the groups of all ant colonies are constructing their local pheroone with the help of expected execution tie atrix in parallelization. Therefore the proposed ultiple ant colonies algorith is a best suited ethod for tracking proble with large data sets. Fro the results it is clearly evident that the proposed efficient ultiple ant colonies algorith offers better optiization a very fast rate and produced better akespan. In other words, these results prove the effectiveness of the EMACO algorith. 6. CONCLUSION: Grid coputing ais to assign tasks to coputing nodes and iniize the execution tie of tasks as well as workload across all nodes. Despite of the intractability, the scheduling proble is of particular concern to both users and grid systes. The proposed EMACO approach achieves optial schedule than the traditional ultiple colony algorith. In EMACO, experiences are shared fro the cooperation of ultiple ant colonies. At the sae tie, the positive and negative feedback are applied to avoid stagnation situation encountered in searching. Here the odified pheroone updating is allocating the resources and obtain the shortest path optially and adaptively in scalable, dynaic and distributed environent. This enhanceent process has been achieved optial scheduling by copleting the tasks with iniu execution tie as well as utilizing the resource in an efficient way. 3392

6 REFERENCES: [1] Foster and Kesselan.C, The Grid:Blueprint for a Future Coputing Infrastructure Morgan Kaufan Publishers, USA, [2] M. Chtepen, Dynaic scheduling in grids syste, Sixth Firw PhD Syposiu, Faculty of Engineering, Ghent University, pp. 110, [3] Stefka Fidanova and Mariya Durchova Ant Algorith for Grid Scheduling Proble, Large Scale Coputing, Lecture Notes in Coputer Science No. 3743, Springer, 2006, pp [4] Chang.R, Chang.J, and Lin.P "Balanced Job Assignent Based on Ant Algorith for Grid Coputing," presented at Proceedings of the 2nd IEEE Asia-Pacific Service Coputing Conference, 2007, pp [5] Yan.H, Shen.X, Li.X, and Wu.M,(2005) "An iproved ant algorith for job scheduling in grid coputing," presented at Proceedings of the Fourth International Conference on Machine Learning and Cybernetics, vol. 5, 2005, pp [6] Kousalya.K and Balasubraanie.P, An Enhanced Ant Colony Algorith for Grid Scheduling Proble, IJCSNS International Journal of Coputer Science and Network Security, Vol.8, No.4,2008 [7] Kwang Mong Si, and Weng Hong Sun, " Multiple Ant Colony Optiization for Load Balancing" journal of coputer applications,2008,pp [8] Kwang Mong Si and Weng Hong Sun," Ant Colony Optiization for Routing and Load - Balancing: Survey and New Directions", Ieee Transactions On Systes, Man, And Cybernetics Part A: Systes And Huans, Vol. 33, No. 5, Septeber [9] H. Kawaura, M. Yaaoto, K. Suzuki, and A. Ohuchi, "Multiple ant colonies algorith based on colony level interactions", IEICE Trans. Fundaentals, Vol. E83-A, No.2, 2000, pp [10] J. Jong, and M. Wiering, "Multiple ant colony syste for the bus-stop allocation proble", Proc of the Thirteenth Belgiu- Netherlands Conference on Artificial Intelligence BNAIC 01, Asterda, Netherlands, 2001, pp [11] Alaa Aljanaby, Ku Ruhana Ku-Mahaud, Norita Md. Norwawi, Interacted Multiple Ant Colonies Optiization Fraework: an Experiental Study of the Evaluation and the Exploration Techniques to Control the Search Stagnation, International Journal of Advanceents in Coputing Technology Volue 2, Nuber 1, March [12] Martin Middendorf, Frank Reischle, Hartut Scheck Inforation Exchange in Multi Colony Ant, Institute four Angewandte Inforatikund Forale Beschreibungsverfahren, Universityat Karlsruhe, D Karlsruhe, Gerany. [13] Kruger.F, Middendorf.M, and Merkle.D, Studies on a Parallel Ant Syste for the BSP Model Unpub. Manuscript, [14] Liang Bai, Yan-Li Hu, Song-Yang Lao, Wei-Ming Zhang, Task Scheduling with Load Balancing using Multiple Ant Colonies Optiization in Grid Coputing International Conference on Natural Coputation (ICNC), [15] K. Kousalya and P. Balasubraanie To Iprove Ant Algorith s Grid Scheduling Using Local Search, International Journal Of Coputational Cognition Vol. 7, No. 4, Deceber