Optimized Assignment of Independent Task for Improving Resources Performance in Computational Grid

International Journal of Grid Computing & Applications (IJGCA) Vol.6, No.1/2, June 2015
DOI:10.5121/ijgca.2015.6203 27
Optimized Assignment of Independent Task for
Improving Resources Performance in
Computational Grid
Sarpreet Singh1
, R.K. Bawa2
,
1
Assistant Professor, Department of Computer Science& Engineering,
Sri Guru Granth Sahib World University, Fatehgarh Sahib
2
Professor, Department of Computer Science, Punjabi University Patiala 2
Abstract
Grid computing has emerged from category of distributed and parallel computing where the
heterogeneous resources from different network are used simultaneously to solve a particular problem that
need huge amount of resources. Potential of Grid computing depends on my issues such as security of
resources, heterogeneity of resources, fault tolerance & resource discovery and job scheduling. Scheduling
is one of the core steps to efficiently exploit the capabilities of heterogeneous distributed computing
resources and is an NP-complete problem. To achieve the promising potential of grid computing, an
effective and efficient job scheduling algorithm is proposed, which will optimized two important criteria to
improve the performance of resources i.e. makespan time & resource utilization. With this, we have
classified various tasks scheduling heuristic in grid on the basis of their characteristics.
Keywords
Grid Computing, Scheduler, ETC matrix, Makespan, Resource utilization
1. Introduction
Grids computing is evolved from existing technology like distributed & parallel computing , web
services, Internet, various cryptography technologies having more intelligent security features and
virtualization feature for the next-generation of e-Science and e-business applications [1]. Grid
computing is based on the cooperation of multiple processing elements on multiple distributed &
dynamic machines, to boost the computational power in the scientific problem and other field
which require huge capacity of the CPU cycles & other resources. Due to the dynamic nature of
the resources, a grid is another form of heterogeneous computing (HC) system. Different kind of
resources used in grid like computation resources (i.e. a machine sharing its CPU), storage
resources(i.e. a machine sharing it’s RAM or disk space), communication resources (i.e. sharing
of bandwidth or a communication path), software and licenses and many other special equipment
(i.e. sharing of devices). [2]
The resource management system (RMS) which is central component of a grid computing accepts
requests from user and assigns resource from the overall pool of grid resources which fulfill user

28
requirement. There are number of issues in RMS such as resource discovery, resource scheduling,
resource monitoring, resources inventories, resource provisioning, load balancing, fault isolation,
service level management system [3]. However, grid scheduling and grid load balancing are the
main issues [4].
Scheduling the jobs to the resources in grid computing is also complicated due to the distributed,
dynamic and heterogeneous nature of the resources [5].Scheduling is one of the core steps to
efficiently exploit the capabilities of heterogeneous distributed resource. A scheduling is a
process that maps and manages the execution of inter-dependent tasks/independent task on the
distributed resources. It allocates suitable resources to various tasks to satisfy the agreement
between resource producer & resource consumer. Poor scheduling will reduce performance of the
nodes participating in grid system. Due heterogeneous, dynamic nature of resources, the problem
of mapping tasks on distributed services belongs to a class of problems known as NP-hard
problem [6]. Due to dynamic & heterogeneous nature of resources, neither algorithm can produce
optimal schedule. We can only make our best to find the most suitable solution for users.
Working of grid scheduler passing from four stages i.e. a) Resource discovery b) Resource
selection c) Job selection d) Job execution. Scheduler is responsible for selecting resources from
pool resources and schedule tasks in such a way that the resource usage policy and task
constraints are satisfied, in terms of execution time, deadline & cost of the resources utilized [7].
Scheduler returns the “best” such schedule. “Best” is defined by some performance metric [8]. All
scheduling algorithms which were mostly considering only single parameter i.e.
makespan(completion time of last finished task or job) which is not sufficient to produce idle grid
system. According to Wright [9], a scheduler is designed to satisfy one or more of the following
common objectives: (a) maximizing system throughput; (b) maximizing resource utilizations; (c)
maximizing economic gains; and (d) minimizing the turn-around time for an application. To
minimize the overall turn-around time of the tasks and thus increase the throughput of the system,
it is important that right resources be assigned to every task. Since task may have dependency or
independency between each other’s, so there are two case to assign tasks in grid. First case is for
dependent task. In this, split the task into multiple subtasks. Then these subtasks assigned to a
resource which fulfill the requirements of task. These tasks may have dependency among
themselves. This process is termed matching. After this order of execution of the subtasks is
identified. During this ordering, dependencies may be considered & the process of ordering for
execution is called scheduling. The overall process of matching and scheduling is termed
mapping. Second case is for mapping is for independent tasks. The aim of mapping in both above
cases is to maximize an objective function, which is based on QoS attributes such as execution
time, response time or those requested by the users of the HC system [10]. To achieve such
objective, certain heuristics have been developed for mapping. In the following section, various
mapping heuristics will be discussed and the assumptions made while describing the heuristics.
2. Related Work
Mapping heuristics can be either static or dynamic. In Static mapping information regarding all
participating nodes is collected & matching and scheduling decisions are made before the actual
execution of the task. In dynamic mapping, matching and scheduling decisions are made on the
fly as the application executes. The accuracy of static mapping heuristics depends on the accuracy
of these estimates. The result scheduling in the grid systems are based on the mapping heuristics.

29
The component called mapper maintains a two dimensional matrix known as the expected time to
compute (ETC) matrix. It contains the expected execution times of a task that are running on all
participating node in the grid. execution time of the tasks on different nodes is represented by row
entries in ETC & time taken by a node to execute tasks is represented by column entries in ETC.
2.1 Heuristic descriptions
This section, discuss five heuristics for scheduling tasks to various machines which are
participating in grid. Then we propose an efficient heuristic called Optimized Assignment of
Independent Task (OAIT) algorithm
Opportunistic Load Balancing: OLB assigns tasks to the next available machine in grid. If
more than one machine is available, any one machine is chosen arbitrarily. It does not consider
expected execution time of the task on that machine. [11]
Fast Greedy or Minimum Completion Time (MCT): This heuristic assigns each task to the
machine which completes in minimum time. MCT heuristic does not consider execution time of a
task on that machine, so some time execution time of task will be increase on the machine where
the task has to be assigned. [12]
Minimum Execution Time (MET): In this heuristic task is assigned in to machine which give
least execution time for that task’s execution. Due to this, it can cause load imbalance on
resources. MET has advantage over MCT, is its simplicity of implementation. [13]
Min-min: The Min-min heuristic begins with the set of all unmapped tasks. Then, minimum
completion time for each task is found on every machine. Next, the task with the overall
minimum completion time is selected and assigned to the corresponding machine (hence the
name Min-min). Last, the newly mapped task is removed from U, and the process repeats until all
tasks are mapped (i.e., U is empty). Min-min & MCT both heuristic produce minimum
completion time. [13][14]
Max-Min: The max-min mapping heuristic is similar to the min-min mapping heuristic. The first
step of this heuristic is identical to the min-min heuristic. In the second step, instead of choosing
the task having the minimum of earliest completion times among all the tasks, max-min chooses
the task having maximum of earliest completion times and assigns it to the corresponding
machine. The machine availability time is updated and the process is repeated for every task in
the metatask. The max-min mapping heuristics would generally outperform min-min mapping
heuristics, when the number of short tasks is greater than that of long tasks. [12][14].
3. Proposed OAIT Algorithm
All discussed heuristics in previous section are known for reducing makspan time based on
different set of inputs. But utilization of resources is not up to the mark. While proposed OAIT
algorithm reduces makspan as well as optimizes the resource utilization & balance the load on all
participating nodes.

30
When a job is received, mostly heuristic try to find node, which executes job in minimum time
without considering current load on that node. As a result some nodes will be overload & load
distribution is imbalance. But OAIT heuristic, assigning bigger jobs before assignment of smaller
jobs by considering current load on nodes as shown in fig 1. N represent node. Length of Bar
represent the execution capacity of node i.e. more length more execution capacity. Let us consider
two jobs one is heavy job J5 (i.e. with maximum number of instruction) assign to fast node N0 &
second light job J1 (i.e. with minimum number of instruction) assign to low execution capacity
node N3.As a result waiting period for longer jobs will be reduced & also till the completion of
longer jobs many smaller jobs will be executed parallel on other light nodes. This will give better
resource utilization & reduce the overall turnaround time.
N0 J5
N1 J0
N2 J4
N3 J1
N4 J3
N5 J2
Fig 1: Job distribution
3.1 Proposed grid scheduling architecture
Grid is basically a collection of resources. Each resource consists of number of machine & each
machine may have more than one number of processing elements (PE). Fig 1 shows the proposed
Grid scheduling architecture. The architecture is composed of Job queue, Grid scheduler (GS),
Machines, Local scheduler (LS), processing elements (PE). User submits job to Job queue. The
GS analyzes the job’s requirements and then select appropriate machine, after gathering
information from various Local Scheduler which are connected with GS. Then GS pass job to
selected Local Scheduler. LS have complete information about their local PE like execution
capacity, current load etc. This information is used by LS, to select the best PE using the
proposed scheduling policy & assign the jobs for execution.
Fig 2: Grid Scheduling Architecture

31
3.2 Terminology used in Proposed OAIT algorithm:
 JOBQUEUE [id][# instruction]. Its 2D array having two parameters i.e. job Id & number
of instructions.
 Job Size: Define by number of instructions.
 Node (Mi,PEj): Define by ith
PE on jth
machine.
 Turnaround time (TAT): It is the numeric value of node. Its interval from the time of
submission of a job till completion of job. If the TAT value of any node is non-zero,
means node is busy to executing task & calculated by:
TAT = Old TAT of node + (number of instruction /Power of node (Hz))
 Power of node: Capacity to executing jobs. Define by MIPS(millions of instruction per
second)
 Expected time to compute (ETC) matrix: its 2D array, where rows represent participating
machines in grid & column entry represents all PE on that machine. Each entry in ETC
represent by TAT of that node. Shown in fig 3.
PE1 PE.. PEn
MC1
MC2
MC..
MCn
Fig 3. ETC Matrix
 Temp_TAT_F[][]: Its representation is similar to ETC matrix & use to store TAT value
of each node (Mi,PEj) for executing job having minimum number of instructions.
 Temp_TAT_L[][]:Its representation is similar to ETC matrix & use to store TAT value of
each node (Mi,PEj) for executing job having maximum number of instructions.
3.3 OAIT Algorithm
1. Create Grid resource by defining number of machines & PE with MIPS rating.
2. Submit jobs in array JOBQUEUE[id][# instruction]
3. Sort the jobs of JOBQUEUE [id][# instruction] in ascending order, based on second
parameter of JOBQUEUE(i.e. # instruction).
4. Initialize entry of ETC [][] matrix to zero.
5. Scheduler component read two jobs at same time, first Job#min i.e. one with minimum
number of instructions & second Job#max i.e. with maximum number of instructions from
JOBQUEUE [][] , until JOBQUEUE will be empty
a) Calculate TAT for Job#min & Job#max on every node separately, & store in array
"Temp_TAT_f [][] ” & "Temp_Tat_L [][]” respectively.
b) Choose the minimum value (TATmin) from array Temp_Tat_L[][] & select its
corresponding Machine(MCl) & processing element (PEl)
c) Choose the maximum value (TATmax)from array Temp_Tat_f[][] but less than
the TATmin i.e. (TATmax<TATmin) & select its corresponding machine MCf & PEf

32
d) if( MCf ==MCl && PEf ==PEl )
i. Choose the second largest value (TATsec_max) from array
Temp_Tat_f[][] but less than the TATmin i.e. (TATsec_max<TATmin) &
select its corresponding MCf & PEf
ii. ETC[MCf][PEf]=TATsec_max;
ETC [MCl][PEl]= TATmin;
e) Else update array ETC[][]
i. ETC [MCf][PEf]= TATmax;
ii. ETC [MCl][PEl]= TATmin;
f) Go to step 5.
6. After every second, another thread decrement every TAT value of ETC matrix
7. If all entry in ETC matrix become zero then exit
Initialize all entry of ETC matrix to zero, represent nodes are free. JOBQUEUE receive jobs from
user & rearrange in ascending order according to priority. Priority may be defined by any criteria,
but in proposed OAIT priority is number of instructions. Proposed OAIT heuristic read two jobs
from JOBQUEUE at same time i.e. one with minimum instruction (Job#min) & second with
maximum instruction (Job#max). Find the node NL, which execute Job#max in minimum time &
calculate TATmin. For better resources utilization & reduce the makespan for all pending Jobs ,
OAIT finds slowest node from pool of nodes which executes Job#min with maximum time &
calculate TATmax , but less than TATL i.e. ( TATmax < TATmin). Update the ETC matrix by final
Tmax & TATmin corresponding to the node. After assignment of all jobs, decrement thread called,
which decrement the TAT value of busy nodes by one after every second, till TAT value of nodes
becomes zero.
4. Experimental Detail:
OAIT, MET, Min-Min, Max-Min, MCT heuristics are implemented on Grid simulator (GridSim).
Table 1 shows the scenario of Grid, in which four machines are used with different number of PE
to execute the Jobs. Each PE has specific MIPS rating. Table 2 shows set of 40 jobs with number
of instruction (MI) that are executed on Grid simulator using different scheduling heuristic.
Machine PE0 (mips) PE1 (mips) PE2 (mips) PE3 (mips)
Mc0 90 150 600
Mc1 80 350
Mc2 700
Mc3 450 60 200 120
Table 1
Job id # instruction (MI) Job id # instruction(MI)
Job0 20000 job20 2000
job1 16000 job21 10000
job2 12000 job22 3000

33
job3 13000 job23 70000
job4 19000 job24 650000
job5 15000 job25 30000
job6 60000 job26 30000
job7 7000 job27 89900
job8 2000 job28 32900
job9 45000 job29 12000
job10 3000 job30 4000
job11 65500 job31 2000
job12 18000 job32 34000
job13 5500 job33 130000
job14 30000 job34 99900
job15 14500 job35 14500
job16 300000 job36 18000
job17 120000 job37 3000
job18 320000 job38 184000
job19 315000 job39 150000
Table2
5. Result & Conclusion
The proposed OAIT algorithm is evaluated for Makespan time, Resource utilization and
compared with the other existing heuristics. Table 3 compare the makespan time of the proposed
OAIT algorithm and the existing MET, Min-Min, Max-Min, MCT heuristics. From fig 4, it is
found that proposed OAIT algorithm reduces the Makespan time i.e. executes all the jobs in
minimum time, comparatively to other existing heuristics.
Scheduling Heuristics Makespan time(sec)
OAIT 1082
MET 4222
Min-Min 3413
Max-Min 4899
MCT 1496
Table 3.Compartition of Makespan time

34
Fig 4 Makespan Time
Table 4 compares Resource utilization time (how much time resources are used to executing the
assigned jobs) by proposed OAIT algorithm and other existing MET, Min-Min, Max-Min, MCT
heuristics. From fig 5, it is found that proposed OAIT algorithm gives better result as
comparatively to other existing heuristics. OAIT algorithm use resources in optimized way & for
constant time till the execution of last job. This will also help to distribute computational load
among all participating nodes according to capacity of node. While MET, Max-Min, Min-Min
shows unbalanced execution load, which lead to poor utilization of resources. MCT shows better
result for utilizing resources but not so well as comparatively than proposed OAIT algorithm.
Nodes
OAIT MET Max-Min Min-Min MCT
Time(sec) Time(sec) Time(sec) Time(sec) Time(sec)
Mc-0 /PE-0 954 0 9604 0 355
Mc-0/PE-1 1066 0 8016 3065 915
Mc-0 /PE-2 1043 0 5566 5743 1191
Mc-1 /PE-0 999 0 8849 0 243
Mc-1 /PE-1 1079 0 7345 2279 1000
Mc-2 /PE-0 1072 4222 7842 3640 1496
Mc-3 /PE-0 1050 0 6759 2004 1170
Mc-3 /PE-1 1007 0 9765 0 116
Mc-3 /PE-2 1010 0 7107 3449 427
Mc-3 /PE-3 1082 0 7921 1399 382
Table 4: Comparison of Resource utilization
1082
4222
3413
4899
1496
0
1000
2000
3000
4000
5000
6000
Makespan

35
Fig. 5 Resources utilization
REFERENCES
[1] I. Foster and C. Kesselman (Eds.). The Grid: Blueprint for a Future Computing Infrastructure,
Morgan Kaufmann Publishers, USA, 1999.
[2] Jarek Nabrzyski, Jennifer M. Schopf & Jan Weglarz, Grid Resource Management– State of the art
and Future trends, Kluwer Academic Publisher
[3] Sharma, A., & Bawa, S. Comparative analysis of resource discovery approaches in grid
computing.
Journal of Computers, 3(5), 60-64, 2008.
[4] Moallem, A., & Ludwig, S. Using artificial life techniques for distributed grid job scheduling.
Proceedings of the 2009 ACM Symposium on Applied Computing, 1091-1097, (2009).
[5] Li, Y. A bio-inspired adaptive job scheduling mechanism on a computational grid. International
Journal of Computer Science and Network Security, 6(3B), 1-7., 2006
[6] J. D. Ullman, NP-complete Scheduling Problems, Journal of Computer and System Sciences,
10:384-393, 1975
[7] Maozhen Li, Mark Baker, The Grid Core Technologies, A John Wiley & Sons, Inc., 2005.
[8] H. Dail, H. Casanova, and F. Berman, A Decoupled Scheduling Approach for the GrADS
Environm ent, in Proc. 2002 ACM/IEEE conference on Supercomputing,pp.1-14, Baltimore,
Maryland USA, November 2002
[9] D Wright, Cheap Cycles from the Desktop to the Dedicated Cluster: Combining Opportunistic and
Dedicated Scheduling with Condor, Proceeding of HPC Revolution 01, Illinois, 2001.
[10] Howard Jay Siegel and Shoukat Ali. Techniques for mapping tasks to machines in heterogeneous
computing systems. Journal of Systems Architecture, 46(8):627–639, 2000. Available online at:
citeseer.ist.psu.edu/siegel00techniques.html (accessed January 1st, 2009).
[11] R. Braun, H. Siegel, N. Beck, L. Boloni, M. Maheswaran, A. Reuther, J. Robertson,M. Theys, B.
Yao, D. Hensgen and R. Freund, A Comparison of Eleven Static Heuristics for Mapping a Class
of Independent Tasks onto Heterogeneous Distributed Computing Systems, in J. of Parallel and
Distributed Computing, vol.61, No. 6, pp. 810-837, 2001
[12] Tracy D. Braun, Howard Jay Siegel, Noah Beck, Ladislau Bni, Muthucumaru Maheswaran, Albert
I. Reuther, James P. Robertson, Mitchell D. Theys, Bin Yao, Debra A. Hensgen, and Richard F.
Freund. A comparison study of static mapping heuristics for a class of metatasks on heterogeneous
0
2000
4000
6000
8000
10000
12000
Executiontime
Processing Nodes
OAIT MET Max-Min Min-Min MCT

36
computing systems. In Heterogeneous ComputingWorkshop, volume 8, pages 15–29. IEEE
Computer Society, 1999.
[13] Technical Report No. 2006-504 “Scheduling Algorithms for Grid Computing: State of the Art and
Open Problems ’’, Fangpeng Dong and Selim G. Akl ,School of Computing, Queen’s University
Kingston, Ontario ,January 2006
[14] Sameer Singh Chauhan, and R. C. Joshi,” A Weighted Mean Time Min-Min Max-Min Selective
Scheduling Strategy for Independent Tasks on Grid”, In Advance Computing Conference (IACC),
IEEE, pp. 4-9 ISBN: 978-1-4244-4790-9 , 2010
Authors
Er. Sarpreet Singh has done Master in the area of Grid Computing from Department of
Computer science, Punjabi University, Patiala, INDIA. Presently he is working as
Assistant professor at Sri Guru Granth sahib World University, Fatehgarh sahib & doing
his Ph.D from Punjabi University, Patiala, INDIA. His area of interest is Grid & distributed
computing
Dr. Rajesh K. Bawa has done Master’s and Ph.D degree in the area of Numerical
Computing from IIT Kanpur INDIA. Presently he is working as Head & Professor in
Department of Computer science, Punjabi University, Patiala, INDIA. His present area of
interest is Parallel and Scientific Computation. He has published many research papers in
international & national journals.

Optimized Assignment of Independent Task for Improving Resources Performance in Computational Grid

Recommended

More Related Content

What's hot (16)

Similar to Optimized Assignment of Independent Task for Improving Resources Performance in Computational Grid (20)

Recently uploaded (20)

Optimized Assignment of Independent Task for Improving Resources Performance in Computational Grid