JOURNAL OF TELECOMMUNICATIONS, VOLUME 3, ISSUE 1, JUNE 2010

55

Rule-based HCH Scheduling and Resource Management in Computational Grid L.Mohammad Khanli, F.Davardoost Abstract- Grid system aggregates the heterogeneous resources from every part of the world to form a complicated computing. Resource Management and Scheduling have an important role in enhancing the performance of system. Moreover, well-allocated mechanism could satisfy user’s demand. In this paper we use an Active Grid Information Server (AGIS) for optimal resource selection which supports EvenCondition-Action (ECA) rules. ECA rule-based system can support ad-hoc, adaptive, flexible and dynamic scheduler which allows the system designer to modify scheduler policy. Since there are many resources and users expects to get appropriate Quality of Service (QOS) according to the tasks requirements and cost in the minimum time, therefore in this paper, we propose new rules for classifying tasks and resources and also trading-rule for calculating the cost of resources according to the reliability and credibility during the time. Furthermore for minimizing the task execution time we propose new scheduling called HCH (Hierarchical model based on Classification by Hybrid strategy). The result shows that the new proposed approach has a shorter Makespan in comparison of other approaches. Index Terms—Classification, ECA rule, Fault tolerance, Hierarchical, Hybrid, Grid Computing, Resource Manager, Quality of Service, Trading

——————————  —————————— users according to task requirements and cost. As we 1 INTRODUCTION all know, cost and time are both important QoS attribute. Hence, most of current algorithms can not rid computing emerged in the mid 1990s as a next provide comprehensive QoS assurance [3]. We propose generation of cluster and distributed system [7]. new rules for classifying task and resource in order to Grid computing utilizes the distributed find the best match between them. Furthermore for heterogeneous resources in order to support supporting the QoS According to the cost that user complicated computer problem [2,10]. Heterogeneity, a could pay we propose new Trading-Rule which basic characteristic of grid system, can be occurred in enables the resource adjusting its price according to its many aspects such as task heterogeneity, resource credibility and reliability during the time. The cost of heterogeneity and interconnection heterogeneity. resource indicates the level of QoS. It means that high Resource heterogeneity may be classified in to two cost’s resource supports high Quality of Service. types: temporal heterogeneity, which effectiveness of Therefore the user who pays much will get more QoS. node may vary with time and spatial which depend on In order to reduce the time of task’s executing we the variation among computers [6]. proposed new scheduling called HCH (Hierarchical Resource allocation and scheduling in the grid model based on Classification by Hybrid strategy) environment are confronted with a great challenge [3]. approach. Hierarchical architecture with three levels As the number of grid system component increases, (Central Scheduler, Local Scheduler and Resources) is the probability of a failure in the grid computing considered. In our system resources with almost same becomes higher than traditional parallel computing specification are placed in a certain class and under the furthermore, users not only expect to commitment to control of Local Scheduler. In our work hybrid strategy perform a task in minimum time but also expect which is a combination of static and dynamic strategy commitments to the level of QoS[15]. Well-allocated is presented. Therefore system has an ability of mechanism could enhance the performance of system switching between the static and dynamic according and satisfy the user demand [13]. the state of system. Active Grid Information Server (AGIS) is a resource The rest of this paper is organized as follows, manager for optimal resource selection and fault section 2 of this study focus on related work in the tolerant service that support Even-Condition-Action field of scheduling and resource management. In (ECA) rules. An ECA rule-based system can support section 3 the structure of AGIS with elements are ad-hoc, adaptive, flexible and dynamic scheduler that detailed and according to the events raised inn grid is modified at runtime. Active database system environment we propose new additional rules and the support mechanism that enables them to respond interactions of rules are presented in the activation automatically to the events that are taking place either graph of ECA-rules. Finally we describe our propose inside or outside the database system itself [15]. HCH scheduling based on rules and then the measure In this paper, we propose new rules for satisfying used for performance evaluation are defined. In section 4 we present and analyze the simulation ———————————————— results. Finally section 5 provides concluding remarks.

G

L. Mohammad Khanli, assistance professor, computer science Dept., university of Tabriz. F.davardoost is with the Computer Engineering Department, University of Islamic Azad of Shabester. © 2010 JOT http://sites.google.com/site/journaloftelecommunications/

JOURNAL OF TELECOMMUNICATIONS, VOLUME 3, ISSUE 1, JUNE 2010 56

2 RELATED WORKS Many research has be done in the area of scheduling and resource management in grid environment. Some approaches are listed as follows. In [15] an Active Grid Information Server (AGIS) is proposed. AGIS is a resource manager which supports ECA-Rules. These ECA-Rules are used for optimal resource selection and also fault tolerant service that satisfies QoS requirement. In this paper we propose new additional rules for better scheduling and resource management. Complexity of scheduling problem increases with the size of grid and variation of resources. In order to provide optimal or near optimal solution some heuristic approaches, such as Ant Colony Optimization algorithm, Annealing method, genetic algorithms are used [7,5]. Objective of some of these methods is to minimize the overall execution time without concentration on overhead of large scale system like grid, so mapping the tasks to resources will be NP-Hard problem [12]. Authors of [1], proposed hybrid strategies in order to combine static and dynamic strategies Combination of first-come-first-served algorithm with genetic algorithm is an example of this approach. The static phase can help make instantaneous decisions thus reduce the system response time. The main objectives of GA are to maximum node utilization and well balanced load across all nods [1]. They proposed hybrid strategy to assist in the selection of effective node. First the value of each node is estimated with the value of function then the nodes with high value are selected for execution of the task. Since the effectiveness of nodes may vary with time, thus their system could be able to adjust in accordance with variation of node status. If the nodes become ineffective, dispatcher will select the node with highest value and send unscheduled task toward it [8,11]. In this paper we also will use hybrid strategy in our proposed system. Yagoubi and Medebber, have proposed dynamic Tree-based model to represent Grid Architecture. Their model is independent from any Grid physical Architecture. They defined hierarchical strategy in order to achieve two main objectives (i) Reduction of the mean response time of tasks submitted to Grid computing and (ii) Reduction of the communication costs during task transferring [12]. We also will use Hierarchical model in this paper. Hierarchical structure model for resource management and scheduling is proposed by Zikos and Karatza. They have been suggested site allocation scheduling of no clairvoyant tasks in 2-level heterogeneous grid architecture. The system is heterogeneous regarding the number of resources in site not the variation of resource specification. The aim was the reduction of site load information traffic, while at the same time means response time of tasks [6].

3 THE ARCHITECTURE ACTIVE GRID INFORMATION SERVER Figure 1 shows the hierarchical architecture for AGIS rule processing. The Active Grid Information Server cooperates with a Grid Portal, a Fault Detect, GR (Grid Resource) and lower level of AGIS. In Fig. 1 6(a), (b) and (c) represent the run-time procedures carried out for processing the registered rules. The event detector waits for an event that can be either internal or external and checks if the condition of triggered rule is satisfied. When the condition is satisfied, the corresponding action is executed immediately. According to the following steps the procedure is taken placed. 1. User submits task from Grid Portal. A Grid Portal provides an interface for user to utilize the resource and services provided by grid. 2. Task is sent to appropriate lower level AGIS according to the proposed HCH scheduling. 3. Task is sent to Grid Resource for creating the process and execution. 4. In order to detecting the fault the process of execution is monitoring. 5. The result of task’s execution is sent to the Grid Portal.

Fig. 2. Architecture for AGIS ECA rule processing

3.1 RULES According to the event and the points raised in grid environment different rules are already applied. In this paper, we also propose new rules in order to improve the previous system. New rules are R6, R7, R13, R14, R15, R16, R17 ,R18 ,R22 AND R23 which are identified with blue-gray nodes in Activation Graph of AGIS ECA rule shown in Figure 2.The rules descriptions are as follows.

3.1.1 TASK RULES R1: Task-Arrival-rule: Activated when task arrives to the system. R2: Task-Finish-Rule: Indicating that the resource finishes the execution of the remote task and returns the result.

© 2010 JOT http://sites.google.com/site/journaloftelecommunications/

JOURNAL OF TELECOMMUNICATIONS, VOLUME 3, ISSUE 1, JUNE 2010 57

R3: Task-Delayed-Rule: The failed task should be inserted in to Ready-Task in duplicated by the TaskDelayed-Rule. R4: Insert-Ready-Task-Rule: Arrival or failed task should be inserted in to Ready-Task. R5: Delete-Running-Task-Rule: finishing of task is activated this rule which delete the running task. R6: Task Classification-Rule: Classifying of arrival task in to A , B , C , D according to the following rules: If size of task=small & task type=more computing then the task is belong to class A . If size of task=large & task type=more computing then the task is belong to class B . If size of task=small & task type=less computing then the task is belong to class C . If size of task=large & task type=less computing then the task is belong to class D . R7: Running-Task-Rule: After task is assigned to the resource and start to be executed this rule will be activated which shows the state of task.

Here, FT means the number of task which has been finished correctly and returns the results and NFT means the total number of unfinished task. Resource Availability is the probability that a resource will be correctly operational and be able to

3.1.2 RESOURCES RULES

R15: Running-Resource-Rule: Indicate that resource is in running state.

R8: Resource-Advertise-Rule: Indicating that there is new resource arrival. R9: Resource-Black-List-Rule: Generated when the resource is improper for selection. R10: Idle-Resource-Rule: Generated when resource goes from busy to idle. R11: Insert-Idle-Resource-Rule: this rule is activated after classification of new resource or after finishing of task. R12: Delete-Running-Resource-Rule: this rule will be activated after fault detecting and delete the running resource. R13: Resource-Classification-Rule: Classifying of resources in to A, B, C, D according to the following rules: If Bandwidth= Low and Size of Memory =Small and CPU Speed= Quick then the resource is belong to class A. If Bandwidth= High and Size of Memory =Large and CPU Speed= Quick then the resource is belong to class B. If Bandwidth= Low and Size of Memory =Small and CPU Speed= Slow then the resource is belong to class C. If Bandwidth= High and Size of Memory =Large and CPU Speed= Slow then the resource is belong to class D. R14: Trading-Rule: cost of resource is calculated with the following Eq.(1) Cost Credibility Availability

(1)

Resources credibility represents the probability of tasks could be finished and returns results which is calculated according to the Eq.(2) Credibility

FT

FT NFT

[11]

(2)

deliver resource services during the time formulated according to Eq.(3) Availability 1

and will be

MTF

(3)

Here, MTF indicate the mean time to resource failure during the time of . After calculating of resource’s cost, following rules is used for classifying resource according to their cost which each of classes supports different QoS. 1 then it supports less QoS. 3 2 Cost then it supports medium Qos. 3

IF 0 Cost 1 3 2 IF 3

IF

Cost 1 then it supports high Qos.

3.1.3 AGIS-RULE R16: Check-Blocking-Rule: R17: Dispatch-Toward-Class-Rule: Global Scheduler (GS) dispatch task toward appropriate class according to proposed HCH scheduling. R18: Dispatch-To-Machine -Rule: Local Scheduler (LS) dispatch task toward appropriate machine according to the HCH scheduling.

3-1-4 MAINTAINING-RULE R19: Inc-Threshold-Rule: Maintaining High-Threshold. R20: Dec-Threshold-Rule: Maintaining Low-Threshold. R21: Fault-Detect-Rule: Generated when the fault detector detects failure on resources. R22: Block-Class-Rule: Activated when the class become overloaded and pass the High-Threshold. If High Threshold then block highest load class. Where indicate the mean square deviation of classes and could be calculated according to Eq.(4)[8,14] ( wci

wc) 2

(4)

R23: Unblock-Class-Rule: Activated when the class become under loaded and its load becomes less than Low- Threshold. If Low Thresholdand class was blocked already then it unblock the class.

3-2 ACTIVATION GRAPH RULE We create the activation of graph of ECA rules which is shown in Fig. 2 .

JOURNAL OF TELECOMMUNICATIONS, VOLUME 3, ISSUE 1, JUNE 2010 58

Fig. 2. Activation graph of AGIS ECA rule

If there is no cycle in the activation graph, then termination is assured [15]. New rules are identified with blue-gray nodes in Activation Graph.

Fig.3. proposed HCH Scheduling Algorithm

3.3 PROPOSED HCH SCHEDULING

3-4

In this paper we propose HCH algorithm (Hierarchical model based on Classification by Hybrid strategy). Our propose HCH algorithm is shown in Fig.3. Hierarchical model is kind of scheduling which occurs at multiple levels [4,14]. It is a distributed strategy with local decision-making which support scalability [12]. In this paper we considered 3-level scheduling. Global scheduler, Local Scheduler and third level which is consists of resources. They are different type off resource such as: Computing resource Storage resource Communication resource By the new proposed rule, tasks and resources could be classified and then resources with the same class will be under the control of certain Local Scheduler. We use hybrid strategy which is a combination of static and dynamic. Therefore our system has an ability of switching between static and dynamic according to the state of system. Static, scheduling decision is carried out according to pre-determined approach and does not require collecting system status; therefore, no extra overhead will be created. As in grid environment effectiveness of nodes may vary with time thus, the assignment of tasks has to be dynamically adjusted in accordance with the variation of node status. Dynamic, the status of the system is taken in to account. However, real-time monitoring will cause system overhead. Hybrid combines static and dynamic policy, in order to have the benefit of both advantages [6, 9, 8].

Response_ Time of task is the time period from the arrival to the CS to the time service completion of task. The Response_ Time for each task is calculated by adding the estimated Transmission_ Time, Wait_ Time and Service_ Time of given task when assigned to machine. Response_ Time is defined as Eq. (5)

PERFORMANCE METRIC

Re sponse _ Time Transmission _ Time Wait _ Time Service _ Time

(5)

The estimated transmission time from CS to each LS can be easily determined by Eq. (6) where Mj is the size of given taskj and bandwidthi is bandwidth between CS and LS. Transmission _ Time

Mj bandwidthi

(6)

As all of the machines of LS are connected via highspeed Local network therefore, transmission time of Local Scheduler could be ignored. Wait_ Time of given taski could be estimated by the sum of all service time of tasks in queue of certain machine which have been assigned to that machine before arriving of taski and is defined as in Eq.(7) length of queue

Wait _ Time

Service_ Time

j

(7)

j 1

Service_ Time is defined as the number of instruction of each task by the cpu_speed. Therefore, formulate the calculation as in Eq. (8) Service _ Time

number _ of _ instruction cpu _ speed

(8)

JOURNAL OF TELECOMMUNICATIONS, VOLUME 3, ISSUE 1, JUNE 2010 59

4. SIMULATION RESULTS AND COMPARISON In this paper, the experiments were carried out by a simulation application that we implemented in Matlab. In the first, different type of task and resources with various specifications are generated randomly according to table 1. Then, the proposed approach is compared with Random-Shortest Queue and Hybrid (Random-Deferred) scheduling. In HCH approach computing nodes are classified then they are placed in a certain Local Scheduler. However, in two other scheduling, computing nodes are placed in Local Scheduler without reorganization of their specification since there is no classification method. Random-Shortest Queue: According to this policy, when there is task arrival at Central Scheduler, one of the Local Scheduler is selected with the same probability and the task is routed to Local Scheduler then LS uses information about number of tasks in each queue and selects the node with the least number of tasks waiting in queue [4]. Hybrid (Random- Deferred) Hybrid policy is integrated Random and Deferred policy. In deferred, for each tasks, the Local Scheduler with the minimum number of tasks is selected. There exists the concept of Allocation Interval. A parameter (A_I %) have been introduced which shows the percentage of allocation interval in which the random policy is used. If a task arrives beyond the threshold that (A_I %) defines, then the CS will operates according to deferred policy [4]. Table 1: Rating with fuzzy approach

Fig. 4. Comparison of Makespan

5-Conclusion In this paper we use Active Grid Information Server (AGIS) which consist of Even-Condition-Action (ECA) Rules. ECA-Rule based system supports ad-hoc, flexible system which allows the designer to modify scheduler policy according to the requirements and situation. In order to find the best match between task and resource we propose new Classification-Rule and also for supporting the QoS according to the user’s budge we propose new Trading-Rule which enables the resource adjusting its price according to the credibility and reliability and finally we propose new rule-based HCH scheduling which has an ability of switching between static and dynamic approach according the state of system and minimizing the response time of tasks.

REFERENCES [1]

[2]

[3]

As it shown in Fig.4. In a case, all of task is created randomly, HCH (no blocking) comparing with the other scheduling reduces the system Response_Time, thereby it has the shortest Makespan. However, Hybrid Scheduling performs better with less number of tasks. With comparing Random-Shortest Queue and Hybrid we realize that without depending on the number of tasks Random-Shortest Queue sometimes performs better than Hybrid.

[4]

[5]

[6]

[7]

Y. Li, Y.Yang, M. Ma and L. Zhou, “A hybrid load balancing strategy of sequential tasks for grid computing environments,” Future Generation Computer Systems 25 (2009) 819_828. R. Chang, J. Chang and P. Lin, “An ant algorithm for balanced job scheduling in grids,” Future Generation Computer Systems 25 (2009) 20–27. B. Tang, Z. Zhou, Q. Liu and F. Li, “Market-driven Based Resource Scheduling Algorithm in Computational Grid,” IEEE. International Conference on Computer Science and Software Engineering, DOI 10.1109/CSSE.2008. S. Zikos and H. Karatza, “Resource Allocation Strategies in a 2-level Hierarchical Grid System,” IEEE. 41st Annual Simulation Symposium, DOI 10.1109/ANSS-41.2008. Y. Li, Y. Yang and R. Zhu, “A Hybrid Load balancing Strategy of Sequential Tasks for Computational Grids,” IEEE. International Conference on Networking and Digital Society, DOI 10.1109/ICNDS.2009. S. Zikos and H. Karatza, “Communication cost effective scheduling policies of non-clairvoyant jobs with load balancing in a grid,” Elsevier Inc. The Journal of Systems and Software, doi:10.1016/j.jss.2009. M. Paletta and P. Herrero, “A Simulated Annealing Method to Cover Dynamic Load Balancing in Grid Environment,” Springer-Verlag Berlin Heidelberg, DCAI 2008, ASC 50, pp. 1–10, 2009.

JOURNAL OF TELECOMMUNICATIONS, VOLUME 3, ISSUE 1, JUNE 2010 60

[8]

[9]

[10]

[11]

[12]

[13]

[14]

[15]

K.Q. Yan, S.C. Wang, C.P. Chang and J.S. Lin, “A hybrid load balancing policy underlying grid computing environment,” Elsevier. Ltd. Computer Standards & Interfaces 29 (2007) 161–173. K. Yan, S. Wang, S. Wang and C. Chang, “Towards a hybrid load balancing policy in grid computing system,” Elsevier Ltd. Expert Systems with Applications. doi:10.1016/j.eswa.2009.03.001. Q. Zheng, C.Tham and B.Veeravalli, “Dynamic Load Balancing and Pricing in Grid Computing with Communication Delay,” Springer Science. J Grid Computing, 10.1007/s10723-007-9093-5.2008. S.Choi and R.Buyya “Group-based adaptive result certification mechanism in Desktop Grids, Future Generation Computer Systems 26 (2010) 776_786” B. Yagoubi and M. Medebber, “A Load Balancing Model for Grid Environment,” IEEE 1-4244-1364-8/07/$25.00 ©2007 J. Chen and B. Lu, “Load Balancing Oriented Economic Grid Resource Scheduling,” IEEE. Pacific-Asia Workshop on Computational Intelligence and Industrial Application, DOI 10.1109/PACIIA.2008.324. H. Baghban and A.M. Rahmani “A Heuristic on Job Scheduling in Grid Computing Environment,” IEEE, Seventh International Conference on Grid and Cooperative Computing, DOI 10.1109/GCC.2008.22 L.Mohammad Khanli and M.Analoui. “An approach to grid resource selection and fault management based on ECA rules, Future Generation Computer Systems 24 (2008) 296–316.

Leyli Mohammad Khanli received her B.S. (1995) from Shahid Beheshti University Tehran, Iran, M.S. (2000) from IUST (Iran University of Science and Technology) University and a Ph.D. degree (2007) from IUST (Iran University of Science and Technology) University. All are in computer engineering. She is currently assistant professor in the Department of Computer Science at University of Tabriz. Her research interests include grid computing and Quality of Service management. F.davardoost. received her B.S. degree in software engineering (2006) in software engineering from Islamic Azad University of Shabestar. Now she is a student of M.S. in software engineering. Her research interests include grid computing and quality of service management.