An adaptive algorithm for task scheduling for computational grid

IJRET: International Journal of Research in Engineering and Technology eISSN: 2319-1163 | pISSN: 2321-7308
_______________________________________________________________________________________
Volume: 04 Issue: 03 | Mar-2015, Available @ http://www.ijret.org 441
AN ADAPTIVE ALGORITHM FOR TASK SCHEDULING FOR
COMPUTATIONAL GRID
P.Keerthika1
, P.Suresh2
1
Assistant Professor (Senior Grade), Department of CSE, Kongu Engineering College, Tamilnadu, India
2
Assistant Professor (Senior Grade), Department of IT, Kongu Engineering College, Tamilnadu, India
Abstract
Grid Computing is a collection of computing and storage resources that are collected from multiple administrative domains. Grid
resources can be applied to reach a common goal. Since computational grids enable the sharing and aggregation of a wide
variety of geographically distributed computational resources, an effective task scheduling is vital for managing the tasks.
Efficient scheduling algorithms are the need of the hour to achieve efficient utilization of the unused CPU cycles distributed
geographically in various locations. The existing job scheduling algorithms in grid computing are mainly concentrated on the
system’s performance rather than the user satisfaction. This research work presents a new algorithm that mainly focuses on better
meeting the deadlines of the statically available jobs as expected by the users. This algorithm also concentrates on the better
utilization of the available heterogeneous resources.
Keywords: Task Scheduling, Computational Grid, Adaptive Scheduling and User Deadline.
--------------------------------------------------------------------***----------------------------------------------------------------------
1. INTRODUCTION
Even with the emergence of many superfast computers and
the high speed networks, the utilization of the
geographically distributed resources has gained huge
importance. This recognition is mainly because of the low
cost services and the best outcome offered by them. There
are many computing fields that offer high utilization of the
widely available underutilized resources such as grid
computing, distributed computing, parallel computing, etc.
Grid computing has gained more popularity because of it
loosely coupled nature when compared to distributed
computing and parallel computing that mainly deals with
tightly coupled systems. The basic idea of grid Computing is
to utilize the ideal CPU cycles and storage of millions of
computer systems across a worldwide network function as a
flexible, pervasive, and inexpensive accessible pool that
could be harnessed by anyone who needs it, similar to the
way power companies and their users share the electrical
grid [1].
When considering the scheduling of the resources many
factors such as CPU utilization rate, throughput, turnaround
time, waiting time, response time should be focused for all
the processors when assigned with the jobs. Thus the jobs
are assigned to the resources considering the system’s
performance. Thus the scheduling plays an important role in
achieving the best utilization of resources and the better
completion of the submitted jobs. Many scheduling
algorithms have been designed for this purpose. Even then
scheduling is a main focus. There are many algorithms that
are mainly system centric i.e. consider the effective
utilization of resources such as MCT, MET, OLB, Min-min,
Max-min, First Come First Serve (FCFS) Algorithm,
Shortest Job Fastest Resource (SJFR), Longest Job Fastest
Resource (LJFR), etc... But these traditional algorithms
mainly focus on the system performance of the user
expected time for each job. In this paper we have proposed a
new idea that considers the time expected to complete the
job by the user and schedules the job by concentrating on
both the system performance and the user satisfaction.
Then, after scheduling the jobs through our algorithm, we
can guarantee that each job will be assigned to their most
suitable resources, and most of the jobs are completed
within their respective requisition times, thus the satisfaction
of the users is demonstrated. Then in the algorithm of this
paper, we take into consideration as how to satisfy the
demands of the users as much as possible without
decreasing system performance much are just the main
focus. And we compare the efficiency of our algorithm with
the conventional algorithms.
2. RELATED WORKS
Grid scheduling is the process of scheduling jobs over
widely distributed grid resources. This is achieved by using
a grid scheduler. A grid scheduler is different from local
scheduler in that a local scheduler only manages a single site
or cluster and usually owns the resource [2]. There are three
generalized stages in the scheduling procedure of the Grid
computing. They are resource discovering and filtering,
resource allocation and scheduling according to certain
strategies and job submission and job execution
management over multiple administrative domains. As this
paper mainly concentrates on the job scheduling we need to
concentrate on the second stage. Scheduling algorithms can
be divided into two major modes one is static scheduling
and the other one is dynamic scheduling, with each having
their own advantages and disadvantages.

_______________________________________________________________________________________
In the case of dynamic mode, we don’t do the advanced
estimate of jobs before assigning the jobs to the resource.
Then we monitor the resources available from time to time
and reschedule and reassign the jobs to the best resource
available to improve the system’s performance. We also
have to make some necessary adjustments for reassigning
the jobs to the best resource such as rescheduling and
transferring jobs to most suitable resources. In this paper, we
only take static mode into consideration.
In static mode, a prior estimate of jobs with the available
resources are done and assigned to the suitable resource
based on some strategy. After the submission of the job to a
resource no change will happen to this allocation
dynamically. One of the major advantage of this mode is
this is easy and simple to carry out.
First Come First Serve (FCFS) algorithm schedules the job
to the available resources linearly without considering any
parameters of the job and resource. Shortest Job Fastest
Resource (SJFR) scheduling algorithm mainly aims to
reduce the makespan of all the submitted jobs. The
makespan is the total time taken by all the statically
available jobs to get executed on their scheduled resources.
The shortest job is first scheduled and allocates the jobs to
the fastest resource available [2].
Longest Job Fastest Resource (LJFR) is a scheduling
algorithm, which also tries to reduce the makespan of the
scheduled jobs. The longest job is scheduled to the fastest
resource. Since the longest job is submitted to the fastest
resource the execution time of the longest job is drastically
reduced when compared to its execution time on any other
resource in the grid. As far as execution time is considered it
gives the best results [3].
Opportunistic Load Balancing (OLB) assigns each task, in
arbitrary order, to the next machine that is expected to be
available, regardless of the task's expected execution time
on that machine. The intuition behind OLB is to keep all
machines as busy as possible. The advantage of OLB is its
simplicity. The major drawback of it is, it does not consider
the characteristics of job and resource which results in very
poor makespan [4].
In contrast to OLB, Minimum Execution Time (MET)
assigns each task, in arbitrary order, to the machine with the
best expected execution time for that task, regardless of that
machine's availability. The motivation behind MET is to
give each task to its best machine. It causes severe load
imbalance across machines [4].
Minimum Completion Time (MCT) assigns each task, in
arbitrary order, to the machine with the minimum expected
completion time for that task. This causes some tasks to be
assigned to machines that do not have the minimum
execution time for them. The intuition behind MCT is to
combine the benefits of OLB and MET and hence to
improve the makespan. But all the above mentioned
algorithms deals with only one job at each mapping time
and hence it is not more suitable for the heterogeneous
environment [4].
The Min-min scheduling algorithm deals with the set of all
unscheduled independent tasks. Then the ETC (Expected
Time to Compute) matrix will be computed. ETC matrix
contains the estimated completion time for all the jobs in
every available resources. Then for each job the minimum
completion time will be selected from the ETC matrix. Then
the overall minimum value will be chosen from the
minimum completion time of all unmapped jobs and is
assigned with its best resource. Then the ETC matrix is
recalculated for the remaining unmapped jobs and the above
process is repeated until all the jobs are assigned with the
resources. Min-min maps the tasks in the order that changes
the machine availability status by the least amount that any
assignment could. When compared to the above mentioned
algorithms Min-min has the best makespan. Here, mapping
the task with the shorter execution time to its best machine
allows the tasks to be completed very soon [4].
The Max-min scheduling algorithm is very similar to Min-
min. The Max-min scheduling algorithm also deals with the
set of all unscheduled independent tasks. Then the ETC
(Expected Time to Compute) matrix will be computed for
all the unmapped jobs in every available resource. Then for
each job the minimum completion time will be selected
from the ETC matrix. Then the overall maximum value will
be chosen from the minimum completion time of all
unmapped jobs and is assigned with its best resource. Then
the ETC matrix is recalculated for the remaining unmapped
jobs and the above process is repeated until all the jobs are
assigned with the resources. Here, mapping the task with the
longer execution time to its best machine first allows this
task to be executed concurrently with the remaining tasks.
And hence makespan is improved [5].
As mentioned above, Min-min algorithm is more
advantageous in grid scheduling when compared with other
algorithms. Even though it concentrates on the system
performance the user satisfaction is not taken into
consideration. So we need to modify this aspect and provide
an algorithm that takes both the user satisfaction and system
performance into account. Here we are going to discuss a
new scheduling strategy that guarantees user’s deadline to
be satisfied through our new scheduling strategy without
sacrificing the system’s performance [6].
3. PROPOSED SCHEDULING MODEL
Grid scheduling is the process of scheduling application
tasks over grid resources. There are two main concepts in
this scheduling process namely system’s performance and
user satisfaction. Essentially, all the conventional
algorithms mainly concentrate on the system’s performance.
As these algorithms focus on the proper utilization of the
computational power of the grid resources, these are referred
to us as the system-centric algorithms.
The application-centric algorithms chiefly focus on the
contentment of the demands i.e. deadlines of the
independent tasks provided by the corresponding users.
Here the satisfaction of the users is mainly concentrated
rather than focusing on the performance characteristics of
the grid resources.

_______________________________________________________________________________________
This proposed algorithm mainly deals with the statically
available jobs and hence it is of static scheduling mode. In
particular this algorithm deals with a list of jobs at a time
and has two phases in scheduling such as task prioritizing
and resource selection. In the task prioritizing phase sets the
priority of each task with the user deadline as the parameter
and generates a scheduling list by sorting the tasks
according to their rank values. The resource selection phase
selects tasks in the order of their priorities and maps each
selected task on its optimal resource. So our algorithm falls
into list scheduling algorithms. This list scheduling is further
classified into batch mode, dependency mode and
dependency-batch mode.
Notation Definition
CTi,j
Completion time of the job or task
Ji in the resource rj
RTj Ready time of the resource rj
ETi,j
Execution time of the job or task Ji
in the resource rj
DCTi,j
Difference in time between the
deadline given by the user and the
calculated completion time for the
job in available resources
MinDCTi,j
The minimum value from the
difference values DCTi,j for the
given job
UTi
User requisition time or the
deadline given by the user for the
jobs in U
Batch mode scheduling algorithms are initially designed for
scheduling parallel independent tasks, such as bag of tasks
and parameter tasks, on a pool of resources. Since the
number of resources is much less than the number of tasks,
the tasks need to be scheduled on the resources in a certain
order. A batch mode algorithm intends to provide a strategy
to order and map these parallel tasks on the resources, in
order to complete the execution of these parallel tasks at
earliest time [7]. Even though batch mode scheduling
algorithms aim at the scheduling problem of independent
tasks; they can also be applied to optimize the execution
time of the submitted jobs which consists of a lot of
independent parallel tasks with a limited number of
resources.
Dependency mode scheduling algorithms are derived from
the algorithms for scheduling a task graph with
interdependent tasks on distributed computing environments
[8]. It intends to provide a strategy to order and map the
submitted tasks on heterogeneous resources based on
analyzing the dependencies of the entire task graph, in order
to complete these interdependent tasks at earliest time.
Unlike batch mode algorithms, it ranks the priorities of all
tasks in the submitted jobs at one time.
Dependency-batch mode scheduling algorithms combine
dependency mode and batch mode. It first assigns the rank
to the jobs like that of the Batch mode scheduling algorithm
and then it adapts dependency mode scheduling algorithm to
schedule the independent tasks within the submitted jobs. As
this algorithm deals with the independent tasks, the first step
involved in it is the assignment of rank to the tasks on
considering the deadline of the user as the parameter. The
next step involved in it is the resource selection and hence
our algorithm falls into the batch mode scheduling
algorithm.
Let us consider the mathematical representations to denote
the relationships between the resources and jobs. And also
to introduce the parameters the parameters involved in our
algorithm such as execution time, completion time, ready
time, etc. that have been used in our algorithm. The resource
set is represented as P= { r1, r2, r3,………., rm}. As the
grid environment deals with the heterogeneously distributed
grid resources the number of resources available may be
huge. As we consider the static environment both the jobs
submitted and the resource available are taken as fixed and
they do not change over time.
The jobs submitted can be enclosed within the job set which
is represented as U= {J1, J2, J3,…………,, J4}. The jobs
submitted are considered as the independent tasks that can
be executed in parallel with other available tasks. Also the
jobs are considered as static i.e. they number of tasks
submitted are fixed and they do not change with time. The
users submitting their jobs for execution are represented as
C= {C1, C2, C3, ……………., Cp}. The users submit the
jobs with the requisition time i.e. within which the job needs
to be completed which can also be called the demanded
deadline of the user for the submitted jobs.
4. PROPOSED ALGORITHM
ETC matrix is constructed for all the jobs with every
available resource. Secondly, the job with the minimum
deadline is considered. The deadline of the selected job
given by the user is compared with different ETC values.
The job is allocated to the resource that has the minimum
difference value. Then, the job is removed from the job set.
Next, the waiting time of the resource is changed and the
ETC matrix is recalculated for the remaining unmapped
jobs. Above steps are repeated until all the jobs are
scheduled.
4.1 Calculation of ETC Matrix
ETC matrix is basically the execution time matrix. A prior
estimation of the execution time of the submitted jobs in
every available resource is computed by considering the
characteristics of the tasks and resources. In the proposed
system, 512 tasks and 16 resources are considered. Based on
the characteristics, the tasks can be classified into High Task
and Low Task. Similarly the resources can be classified into
High Machine and Low Machine
By considering the various characteristics of the job and
resource into account, ETC matrix can be designed and it
can be classified into Consistent, Inconsistent and Partially-
consistent. An ETC matrix is said to be consistent, if a

_______________________________________________________________________________________
resource Ri executes a task Ti faster than the resource Rk
and Ri executes all other jobs faster than Rk. For a matrix to
be inconsistent, a resource Ri executes some jobs faster than
Rj and some jobs are slower than Rj. Partially-consistent
ETC matrices are also called as semi-consistent matrices. A
semi consistent matrix is a sub matrix of inconsistent matrix
with predefined size.
The proposed algorithm is mainly based on user satisfaction
and system performance. It takes user’s deadlines into
account and makes the job to be executed within the
expected deadline by assigning it to the suitable resource. It
also concentrates on the system performance by reducing
the idle time of the resources and assigning the tasks equally
among the available resources. It considers the ETC matrix
and concentrates on the completion time and hence the
system’s performance is also the major consideration in
addition to user’s satisfaction. The proposed scheduling
process is performed as two major steps. In the first step, we
concentrate on the user satisfaction and in the second step
we consider system performance. Firstly the ETC matrix is
constructed for the available resources with every available
resource. Secondly we consider the job with the minimum
deadline i.e. the job that needs to be completed quickly.
Then the deadline of the selected job given by the user is
compared with that of different ETC values. Then allocate
the job to the resource that has the minimum difference
value. Then remove the job from the job set. Then the
waiting time of the resource is changed and the ETC matrix
is recalculated for the remaining unmapped jobs. Then
continue the above steps until all the jobs are scheduled.
Thus both the user satisfaction and system performance can
be taken into consideration effectively with this algorithm.
In this algorithm, ETC matrix is constructed. Completion
matrix is calculated by adding the ETC values with the
ready time of the resources. With every job submitted the
deadline is acquired as input within which the user expects
the job to be completed. Priority is assigned to the jobs
based on the deadline provided by the user. Initially, the job
with the highest priority is chosen for scheduling.
The difference matrix is computed for the job that has been
chosen. The minimum value is found from the difference
matrix calculated. The job is allocated to the corresponding
resource. Next step is the computation of the ready time and
the execution time of that resource. The completion matrix
is re-computed by adding the ready time of the resource.
The steps from prioritizing the jobs are repeated until all the
submitted jobs are scheduled to their most suitable
resources.
5. RESULTS AND DISCUSSION
Adaptive job scheduling algorithm is compared with Min-
min algorithm and application demand aware scheduling
algorithm. The parameters considered for the comparison
are
 Makespan- the total time taken to complete all the
submitted jobs
 Hit - a task is said to be a hit if it is completed
within the user requisition time.
 Miss - a task is said to be a miss if it is not
completed within the user deadline.
Adaptive job scheduling algorithm has a better makespan
when compared with the application demand aware
scheduling algorithm. Also, the hit is high when compared
with the Min-min algorithm. Even though the application
demand aware scheduling algorithm has a high hit when
compared to the Min-min algorithm, the makespan is high
for the application demand aware scheduling algorithm. The
adaptive job scheduling algorithm has an improved hit when
compared to Min-min and better makespan when compared
to application demand aware scheduling algorithm.
The various characteristics of ETC matrices (i.e. Consistent,
Inconsistent, Partial), diverse tasks (such as high and low
tasks) and variety of machines (such as high and low
machine) are considered for the construction of twelve
different ETC matrices. The application demand aware
algorithm and the adaptive job scheduling algorithm are
compared in terms of makespan, hit and miss.
The Application demand aware algorithm and the proposed
algorithm is compared based on makespan and the values
are given in table 1. The performance analysis is given in
figure 1.
Table 1. Makespan Values
Tasks and Resources
Application
Demand
Aware
Adaptive
Job
Scheduling
High-High Partial (p-hh) 5774893 4504474
High-Low Partial (p-hl) 1470189 546691
Low-High Partial (p-lh) 106230 60084
Low-Low Partial (p-ll) 15524 13854
High-High Inconsistent (i-hh) 6804441 5403760
High-Low Inconsistent (i-hl) 1061204 855583
Low-High Inconsistent (i-lh) 145245 122041
Low-Low Inconsistent (i-ll) 10591 5429
High-High Consistent (c-hh) 9618108 7308179
High-Low Consistent (c-hl) 735913 469402
Low-High Consistent (c-lh) 84511 58645
Low-Low Consistent (c-ll) 11583 4947

_______________________________________________________________________________________
The comparison based on number of hit is given in table 2
and figure 2. The performance of the proposed algorithm is
better when compared to the application demand aware
algorithm.
Fig.1 Comparison based on Makespan
Table 2. Hit Count Values
Tasks and
Resources
Application
Demand
Aware
Adaptive
Job
Scheduling
High-High Partial 316 490
High-Low Partial 145 226
Low-High Partial 174 251
Low-Low Partial 208 330
High-High
Inconsistent
230 481
High-Low
Inconsistent
120 220
Low-High
Inconsistent
160 242
Low-Low
Inconsistent 214 311
High-High
Consistent
360 480
High-Low
Consistent
116 217
Low-High
Consistent
176 209
Low-Low
Consistent
211 350
Fig.2 Performance based on Hit Count
6. CONCLUSION
Even though many scheduling algorithms have emerged to
effectively utilize the globally available grid resources, they
mainly concentrate on resource performance. They failed to
contribute to the user’s satisfaction. So, we focused to
overcome the existing drawback with the help of our newly
proposed algorithm. For each job, the user gives user
requisition time and with the help of this we make sure that
most jobs are completed within the deadline requested by
the user. The user satisfaction is the major consideration of
our idea, still the system performance is also preserved to a
greater extent. This algorithm is more beneficial with
respect to the user satisfaction when compared with Min-
min algorithm and has good system performance when
compared with the application demand aware scheduling
algorithm. In the proposed system, only the independent
tasks are considered. So this work can be extended to suite
for dependent tasks. The next task is to study the function
and feasibility of dynamic scheduling, to make some
possible improvements on the current scheduling algorithm,
hoping to make it more flexible and efficient in actual
application.
REFERENCES
[1]. Aysan Rasooli, Mohammad Mirza-Aghatabar and
Siavash Khorsandi (2008) ‘Introduction of Novel Rule
Based Algorithms for Scheduling in Grid Computing
Systems’, Second Asia International Conference on
Modelling & Simulation, pp.138-143.
[2]. G.Sumathi and N.P. Gopalan, “Priority based
scheduling for heterogeneous grid environments”,1-4244-
0411-8/06/$20.00 C 2006 IEEE.
[3]. Zhan Gao, Siwei Luo and Ding Ding, “A Scheduling
Mechanism Considering Simultaneous Running of Grid
Tasks and Local Tasks in the Computational Grid”, 2007
International Conference on Multimedia and Ubiquitous
Engineering(MUE'07) 0-7695-2777-9/07 $20.00 © 2007.

_______________________________________________________________________________________
[4]. Tracy D. Braun, Howard Jay Siegel,2 and Noah Beck,”
A Comparison of Eleven Static Heuristics for Mapping a
Class of Independent Tasks onto Heterogeneous Distributed
Computing Systems”, 2007 International Conference on
Multimedia and Ubiquitous Engineering (MUE'07) 0-7695-
2777-9/07 $20.00 © 2007.
[5]. Y. Zhu, "A Survey on Grid Scheduling Systems",
Department of Computer Science, Hong Kong University of
science and Technology, 2003.
[6]. Sumathi. G., Gopalan. N.P, "Grid Scheduling
Algorithms for Heterogeneous Environment", Proceedings
of the IEEE Int. Conf. on Signal & Image Technology and
Internet Based Systems (SITIS 2005), Cameroon, West
Africa.
[7]. Hao Y, Liu G. and Wenc N, An enhanced load
balancing mechanism based on deadline control on
GridSim, Future Generation Computer Systems, Vol.28,
2012, pp. 657-665.
[8]. Homer Wu, Chong-Yen Lee , Wuu-Yee Chen, Tsang-
Yean Lee, "A Job Schedule Model Based on Grid
Environment ", Proceedings of the First International
Conference on Complex, Intelligent and Software Intensive
Systems (CISIS'7695-2823-6/07 $20.00 © 2007

An adaptive algorithm for task scheduling for computational grid

Recommended

More Related Content

What's hot (19)

Viewers also liked (20)

Similar to An adaptive algorithm for task scheduling for computational grid (20)

More from eSAT Journals (20)

Recently uploaded (20)

An adaptive algorithm for task scheduling for computational grid