Optimization Task Scheduling Algorithm in Cloud Computing
Subject Areas : Cloud, Cluster, Grid and P2P ComputingSomayeh Taherian Dehkordi 1 , Vahid Khatibi Bardsiri 2
1 - Department of Computer Engineering. Kerman Branch. Islamic Azad University. Kerman. Iran
2 - Department of Computer Engineering. Kerman Branch. Islamic Azad University. Kerman. Iran
Keywords:
Abstract :
[1] F. Durao, S.F.J, Carvalho, A. Fonseka, and C.V. Garcia , “A systematic review on cloud computing”,The Journal of Supercomputing, Springer US, Vol. 68, 2014 , pp. 1321-1346.
[2] W. Mingxin, “Research on Improvement of Task Scheduling Algorithm in Cloud Computing” , Applied Mathematics & Information Sciences An International Journal, Vol.9, 2015, pp. 507-516.
[3] T. Buchert, C. Ruiz, L. Nussbaum, and O. Richard, “ A survey of general-purpose experiment management tools for distributed systems”, Future Generation Computer Systems , Vol.45,2015, pp.1-12.
[4] T. Mathew, K.C. Sekaran, and J. Jose,” Study and analysis of various task scheduling algorithms in the cloud computing environment”, Advances in Computing, Communications and Informatics (ICACCI), International Conference, 2014, pp-658-664.
[5] X. Xu, N. Hu, and W.Q. Ying, “Cloud Task and Virtual Machine Allocation Strategy Based on Simulated Annealing-Genetic Algorithm” , Applied Mechanics and Materials, Applied Science, Materials Science and Information Technologies in Industry,2014, pp. 391-394.
[6] F. Ramezani, L. Jie, and J. F. Hussain, “Task Scheduling Optimization in Cloud Computing Applying Multi-Objective Particle Swarm Optimization. Springer-Verlag Berlin Heidelber”,2013, pp.237-251.
[7] N. Rasouli, M.R.Meybodi, and H.Morshedlou, “Virtual machine placement in cloud systems using Learning Automata ”, Fuzzy Systems (IFSC), 2013 13th Iranian Conference on. Qazvin, 2013, pp. 1-5.
5
Journal of Advances in Computer Engineering and Technology
Optimization Task Scheduling Algorithm in Cloud Computing
Received (Day Month Year)
Revised (Day Month Year)
Accepted (Day Month Year)
Abstract— Since software systems play an important role in applications more than ever, the security has become one of the most important indicators of softwares.
Cloud computing refers to services that run in a distributed network and are accessible through common internet protocols. Presenting a proper scheduling method can lead to efficiency of resources by decreasing response time and costs. This research studies the existing approaches of task scheduling and resource allocation in cloud infrastructures and assessment of their advantages and disadvantages.
Afterwards, a compound algorithm is presented in order to allocate tasks to resources properly and decrease runtime. In this paper we proposed a new method for task scheduling by learning automata (LA). This method where has named RAOLA is trained by historical information of task execution on the cloud, then divide task to many classes and evaluate them. Next, manage virtual machine for capture physical resources at any period based on rate of task classes, such that improve efficiency of cloud network.
Index Terms—Resource allocation; cloud environment; learning automata.
I. INTRODUCTION
I
n cloud computing, IaaS approach is to increase the efficiency and utilization of resource, equipment and existing networks [1]. In this approach, it is attempted to manage request executions so that operational costs such as energy consumption in datacenters, and costs of networks would be reduced [2]. Cloud computing by numerous virtual machines on some data centers increase capability of response to requests. Here, managing virtual machines and physical resources, besides scheduling policies of tasks, is a significant issue. An inappropriate scheduling may involve numerous resource for a series of requests while, an optimized scheduling with less resources and better management give the same response. In cloud computing there are many serial request of users at the same times and conditions is causing similar condition for cloud [3]. Iteration of similar events indicates that a learning algorithm is able to provide suitable efficiency in such conditions. Appling intelligent methods based on learning to the cloud, is increasing in the field of optimization of tasks scheduling and resource allocation [4]. In this research, we will attempt to present a method based on scheduling information of the past of cloud so that it would be
able to contribute performance optimization of the cloud scheduler.
A method based on learning automata (LA) will be presented for realize optimum indicate of a behavioral cloud in different time period. The article is organized into five sections. Section we define new formalization for scheduling of resource in cloud environment. In section three, Min-min and Sufferage algorithms are explained. Then, we define learning automata (LA) as an appropriate model from cloud environment. Assessing performance of the proposed algorithm in section four. Section five contains the conclusions.
II. Related Works
Genetic algorithm is based on biological concept of generation of the population, a rapid growing area of Artificial intelligence. GA’s are inspired by Darwin’s theory about Evolution. According to the Darwin “Survival of the fittest”. It also a used as the method of scheduling in which the tasks are assigned resources according schedules in context of scheduling, which tells about which resource is to be assigned to which task. Genetic Algorithm is based on the biological concept of population generation. In [5] the genetic algorithm as a heuristic method for search an optimized solution in a large space of solutions, has proposed. At first step, random initializing of a chromosome population is performed for a certain duty. Each chromosome has a fitness value (makespan). Results of scheduling Task for machines are saved in a chromosome. After first population is produced, all chromosomes in the population evaluate themselves with the case based on the fitness value. In this evaluation having little makespan is a better mapping. In [6], another scheduling algorithm is introduced based on combining methods for scheduling tasks in the cloud. In order to reach utilization and efficiency of maximum results, this algorithm simultaneously concentrates on tasks and resources to reach a general optimization. In this method, independent and dividable tasks which require different computation and different memories are scheduled by a genetic algorithm efficiently. In this method, it is assumed that the cloud system is heterogeneous. In other words, all resources of processes and communications are performed heterogeneously. Therefore, by considering memorial limitations and the most requests for cloud computations efficiency, the proposed method is attempting to provide a scheduling method based on GA.
In [7], another method to schedule resources of virtual machines (VMs) is introduced based on genetic algorithm. In order to schedule the system's variety and data of the past, behavior of the data which cause imbalance in the system load must be considered. Therefore, the object of provided scheduling is load balancing of the selected system. Here, genetic algorithm approach can perform scheduling of VMs located in the system by using data of the past and the current situation so that it solves the problem of load imbalance and travel cost. Generally, in the cloud environments, in order to schedule virtual machines (VMs), only data of current situation are used and previous situations of the system which have caused load imbalance in the system are not considered.
When task load balance occurs, the number of travels of virtual machines increases and it causes head costs for the cloud. Current resources of VM are allocated to each physical node and this choice has the least load on the system, this subject can create a context to introduce a genetic algorithm which provides a suitable scheduling for the resources of the system by using data of the past and current situation of the system.
III. Cloud Computing
In this section we present the architecture of the cloud network based on its components and elements for task scheduling and resource allocation. Assume that a cloud computing system is consisted of several heterogeneous process units including "m" units (m>1). The template of the tasks in this system is explained by figure 1.
Fig. 1. Task scheduling in cloud environment.
If "T" is the task queue T= {T1, T2 …Tm} including "m" numbers of users' requests, then "T" always changes dynamically. Therefore, in the time of "i", the characteristics of the user's request are Ti= {T1i, Ti2 …Tim}. If physical resources of R (processor and space) are obvious in advance R= {R1, R2 …Rn}, virtualization to the required number of Vi, are allocated on physical resources. An appropriate solution of "S" is consisting of allocation Vis to task queue. Other words, by Si as an optimized solution in the time "i", the appropriate response of allocation of VMs is provided. Therefore, the set of S = {S1, S2 …Sr} indicates the series of scheduling resources to VMs and allocation of VMs to tasks.
Figure 1, shows the activity cycle of the cloud. The series of V= {V1, V2… VP} shows the number of virtual machines (VM). Here, presenting an optimized scheduling can be achieved by finding a possible solution Si.
Tasks are dynamic, so ready time of tasks (arrival) to "T" is not obvious in before. Each Ti task has an arrival time (ai), the worst execution (ei), and duration (di). Also, the basis of required space (si) involves the number of processors (ci) besides a part of bandwidth of virtual machine of Vj which will be applied by Ti (BW (i, j)) to execute itself. Some issues related to Ti are considered here. Therefore, data of T are in the form of an m×6 matrix. If T(i,j) refers to the resources of Ti when Vj is allocated to it, then the factor of utilization extent "Vj" is equal to the tasks which are responded by the jth virtual machine in the unit of time:
(1)
Now the total utilization is:
(2)
In this paper we assume that physical resources in the cloud environment are constant during scheduling and only task queue are working dynamically. These resources include cases below:
ü Space: Disc = {Disc1, …, Discz}
ü Host: Host = { Host1, …, Hostu }
ü Processor: CPU = {CPU1, …, CPUv}
ü Datacenters: DC = {DC1, …, DCw}
ü Cost of network: Cost = {Cost1, …, Costn}
ü Bandwidth: Bandwidth = {BW1, …, BWs}
Cloud environment scheduler dynamically faces a queue of tasks in each moment. When resources are provided by a virtual machine for at least one of the existing tasks in the queue, allocation of VMs to the tasks are performed based on the duration di. If Ve is a part of involved virtual machines, VF =V-Ve would be the free part of resources in the cloud. Therefore, a list of all tasks on T queue which their required resources exist in VF, construct T' queue. Now scheduling approach must be able to allocate VMs in a form to maximize resource utilization. One response to scheduling and giving S response is that it sorts tasks based on least execution time or shortest task first (STF). Now, T" is the sorted queue of existing tasks and the scheduler allocates tasks to VMs existing in VF as the time it is possible.
S = T ʘ VF (3)
IV. Learning Automata
Learning automata (LA) is one of the types of graph-based methods for system modeling, which include a finite series of actions. An LA can be applied to define a finite discrete environment. A special feature provided by LA is when our specific system has a probabilistic environment of events. Here, a series of possible actions are evaluated by this probabilistic environment and its result are given to the automata through a response template and automata selects this response in its next action selection. The final object is that Learning Automata learns to choose the best action among its actions [8]. The general scheme learning automata is shown in figure 2.
Fig. 2. Learning automata and cloud environment.
V. RAOLA Algorithm
In this section, a new method for scheduling task and resources in cloud environment based on learning automata (LA) has been proposed. The proposed method manages scheduling with pre-determined resources and dynamic tasks queue. This method has named RAOLA (Resource Allocation Optimization by Learning Automata). Execution results of RAOLA on a data set of a dynamic queue of T, set of task, by LA can optimize tasks resource allocation and scheduling through identifying behavior pattern of cloud performance.
We present this proposed method for extended online scheduling by ORAOLA algorithm. Also here, according to periodic evaluations of the extent of resource utilization, cloud scheduler results in performance optimization of the broker and the difference is that in this case it faces with computational complexities. Generally, in the cloud environment, scheduling virtual machines of V, is performed regardless of previous situations of the system and is merely performed based on current data of the system. Analysis of scheduling behavior shows the effect of previous situation of the system on the load imbalance of the system [6]. We have assumed that existing resources are constant in the physical machines. Therefore, no travel is performed by virtual machines to balance task load and there is no head costs related to Vis on physical machines. Here, scheduling must have the least load on the system by appropriate selection of Vis for the duties. This prepares a context to provide an optimized algorithm using data of the past. Learning automata can model systems with discrete components that their behaviors have known effects on each other. The most important thing is to record experiences of the past and considering them in the future behavior of a system. Therefore, a correlation must be made between two capabilities (A, B) of recording experiences of the past, and optimizing the current behavior, which are provided by LA. When T consists of numerous similar duties, or the current function of T has been iterated for many times in the past, a scheduling template based on scheduling behavior can be resulted by considering tasks loads. If LA would be a learning automata so that β(n) is the tasks added to Tnew queue in the time "n", and Vf would be the free resources of virtual machines, then the function is defined as F : ∅ ʘ {Tnew, VF} →∅.
Scheduling function is defined as G: {T", VF}. In order to use data of the past to optimize scheduling, score function of р is defined in the following section. Assume that |Ti| is the weight of the required resources for the duty of Ti. Then if the threshold TH = max {|Ti| ʘ Ti is a task} is constant, we would define a parameter named H with the length of TH. While G function schedules the tasks based on T we have:
(4)
Here, we have S (i) =Ti, and all elements of the H parameter are 0 at first. Now, an effective method is proposed to optimize S, based on structure of scoring the tasks when the iteration of tasks are high. Now, we present a learning automata scheduling algorithm. This algorithm which we have called RAOLA, by recording advantages resulted from scheduling tasks with different specifications, shows better performance when task load conditions of the cloud encounters similar situations.
If we consider components existing in the cloud such as VMs or cloudlets as our system components, the situation of these components is completely random in different times. If the performance of scheduling unit of existing tasks in T queue, and the unit of resources allocation management would always have the same priorities based on tasks and resources, their optimization function in similar conditions is verified by a same behavior. Therefore, learning automata, by comparing average application execution (AAE) of tasks as the factor of cloud utilization with the spent finance, can decide whether to allocate tasks to virtual machines and allocating center data to virtual machine. This decision can always be optimized based on past experiences. The function of ∅ in LA is the situation data of the system which consists data of matrixes X and Y. Briefly we have; ∅ = T × V and Tnew = {t'1 ….t's}. If T" would be a part of T" which has exited the queue for execution and a VMj is allocated to it, F function creates a new situation by removing scheduled tasks and adding new tasks, while removing resources allocated of free resources and adding new free resources.
T = ((T ʘ T"') ʘ Tnew) (5)
VF = ((VF ʘ V (T"')) ʘ Vnew) (6)
Operator is row summing of two matrixes with the same number of columns, and operator ʘ indicates their subtraction. X ʘ Y means that it is equal to a Z matrix having columns equal to X matrix, and its rows consist of that part of X matrix that does not consist components of rows of Y matrix. For instance, for these two matrixes below we have:
Management of raw data scheduling behavior of cloud environment through learning automata presents an optimum way for resource allocation in the cloud based on learning system behavior in different task points. In cloud processing, similar events in the same time and conditions are iterating due to specific request of the users. Therefore, identification of behavior pattern of a cloud by the proposed LA, provides high efficiency in increasing system utilization.
As we introduced the non-linear method for simultaneous online scheduling of tasks and resources, in this part we attempt to present an extension of a proposed online scheduling algorithm its online status and present ORAOLA algorithm. The ORAOLA scheduling algorithm has considered as a developed version of RAOLA for online scheduling. If resource allocation is managed by broker in the cloud, proposed scheduler behavior in RAOLA algorithm can be optimized for online scheduling by help of data transmitted to the broker dynamically from tasks existing in queue in the cloud in each moment.
VI. Algorithm Implemention
A simulator named Cloudsim is used to execute this algorithm in Java language environment. We executed resource allocation management of the algorithm in its simplest case which is responding the memorial request. At first, we conduct scheduling merely based on the smallest duty. A threshold is defined for the requests and we evaluate the extent of utilizations based on this threshold. Step by step we update scheduling behavior and the effect of tasks with different extents in the utilization extent of the H parameter of the cloud and we set the selection priority based on utilization extent of tasks in cloud behavior. Learning automata gets trained step by step by comparing the results. Output results of algorithm execution are indicated in the figure below. As indicated in this figure, normal shape is ascending as utilization makespan increases.
Our result is for the returning times close to the predicted extent, but its growth is more intense as period increases.
Fig. 3. Comparison in resource utilization (resource the time).
Among important indicators existing for scheduling, makespan factor which is equal to subtraction of end time from start time, is very effective in evaluating the method. Here, we compare makespan resulted from existing methods. As indicated explicitly, makespans of non-linear methods increase as the number of cloudlets increase, while proposed methods based on LA are completely declining.
Fig. 4. Returns time on number of cloudlet.
The debt is reduced in this method, since utilization increase is in interaction with debt and exists as a negative index in |Ti|. Therefore, utilization increase is accompanied by debts of responses to the tasks in the cloud.
Fig. 5. Cost per number of cloudlet.
VII. Conclusion
In this research we have proposed an optimized algorithm for scheduling and allocation of virtual machines by using a scheduler learning approach in cloud environment when we don't consider effects due to dynamic change of cloud resources. This algorithm attempts to identify an optimum behavior for scheduling based on learning automata and its optimization capability to investigate the effect of similar tasks in utilization and especially in times with different task loads. The proposed algorithm called RAOLA increases resource utilization explicitly.
Since online scheduling in cloud is of high importance specially in providing existing services in the cloud for users of SLA, we extended the proposed method for online scheduling. The resulted outputs verify the efficiency of the proposed method. Appendixes, if needed, appear before the acknowledgment.
References
[1] F. Durao, S.F.J, Carvalho, A. Fonseka, and C.V. Garcia , “A systematic review on cloud computing”,The Journal of Supercomputing, Springer US, Vol. 68, 2014 , pp. 1321-1346.
[2] W. Mingxin, “Research on Improvement of Task Scheduling Algorithm in Cloud Computing” , Applied Mathematics & Information Sciences An International Journal, Vol.9, 2015, pp. 507-516.
[3] T. Buchert, C. Ruiz, L. Nussbaum, and O. Richard, “ A survey of general-purpose experiment management tools for distributed systems”, Future Generation Computer Systems , Vol.45,2015, pp.1-12.
[4] T. Mathew, K.C. Sekaran, and J. Jose,” Study and analysis of various task scheduling algorithms in the cloud computing environment”, Advances in Computing, Communications and Informatics (ICACCI), International Conference, 2014, pp-658-664.
[5] X. Xu, N. Hu, and W.Q. Ying, “Cloud Task and Virtual Machine Allocation Strategy Based on Simulated Annealing-Genetic Algorithm” , Applied Mechanics and Materials, Applied Science, Materials Science and Information Technologies in Industry,2014, pp. 391-394.
[6] F. Ramezani, L. Jie, and J. F. Hussain, “Task Scheduling Optimization in Cloud Computing Applying Multi-Objective Particle Swarm Optimization. Springer-Verlag Berlin Heidelber”,2013, pp.237-251.
[7] N. Rasouli, M.R.Meybodi, and H.Morshedlou, “Virtual machine placement in cloud systems using Learning Automata ”, Fuzzy Systems (IFSC), 2013 13th Iranian Conference on. Qazvin, 2013, pp. 1-5.
