A Linear Programming Driven Genetic

Algorithm for Meta-Scheduling on Utility Grids

Saurabh Garg, Pramod Konugurthi and Rajkumar Buyya
Grid Computing and Distributed Systems Laboratory, CSSE
Department of Computer Science and Software Engineering
The University of Melbourne, Australia
{sgarg, pramod, raj}@csse.unimelb.edu.au
AbstractThe user-level brokers in grids consider
individual application QoS requirements and minimize
their cost without considering demands from other users.
This results in contention for resources and sub-optimal
schedules. Meta-scheduling in grids aims to address this
scheduling problem, which is NP hard due to its
combinatorial nature. Thus, many heuristic-based
solutions using Genetic Algorithm (GA) have been
proposed, apart from traditional algorithms such as
Greedy and FCFS.
We
propose
a
Linear
Programming/Integer
Programming model (LP/IP) for scheduling these
applications to multiple resources. We also propose a novel
algorithm LPGA (Linear programming driven Genetic
Algorithm) which combines the capabilities of LP and GA.
The aim of this algorithm is to obtain the best metaschedule for utility grids which minimize combined cost of
all users in a coordinated manner. Simulation results show
that our proposed integrated algorithm offers the best
schedule having the minimum processing cost with
negligible time overhead.

I. INTRODUCTION
Grid computing enables harnessing of a wide range
of heterogeneous, distributed resources for solving
compute- and data-intensive applications [23].
Recently, it has been rapidly moving towards a pay-asyou-go model wherein providers expect an economic
compensation for the computational resources or
services offered to grid users. In one side there are users
with applications to execute and, in the other side, there
are providers [7] willing to offer their resources or
computing services in return for regular payments.
Environments with this decoupling of users from
providers are generally termed as utility grids [24].
Scheduling in utility grids is complex due to the
distributed ownership of the resources. Moreover,
consumers and providers are independent from one
another and have different access policies, scheduling
strategies and objectives [16]. Previous work has
proposed grid market infrastructures [25][14] [15] for
utility grids. Although this work provides the basis for
resource markets, application scheduling considering
aspects such as the distributed resource ownership and

cost minimization under these scenarios is still in its

infancy. Furthermore, existing Grid brokers, that
consider cost minimization, are generally single-user
based [17][18][25]. These single-user brokers may
lead to sub-optimal schedules and are certainly not
designed with the aim of minimizing the cost for a
group or community of grid users.
In this work we focus on meta-scheduling of
different applications from a community of users
considering a commodity market. As an example of a
user community, we consider the financial institution
Morgan Stanley that has various branches across the
world. Each branch has computational needs and QoS
constraints that can be satisfied by grid resources. In
this scenario, it is more appealing for the company to
schedule various applications in a coordinated manner.
Furthermore, another goal is to minimize the cost of
using resources to all users across the community (the
company in this case). This meta-scheduling problem
is NP hard due to its combinatorial nature.
This paper proposes a novel LP driven metascheduling algorithm that maps independent
applications with a given deadline and budget to rented
resources in the grid with the aim of minimizing the
combined cost of all the users who share a metascheduler. We have combined the benefit of LP and
GA to obtain an efficient solution for the scheduling
problem in the grid with concurrent users and multiple
heterogeneous resources. Although only three QoS
parameters are considered (i.e. number of processors,
deadline and budget), the proposed algorithm is
general enough to deal with more parameters. A
simulation study shows the effectiveness of the
proposed algorithms compared to other greedy and
genetic algorithms.
The main contributions of this paper are: (a) an
LP/Integer Programming (IP) model for the metascheduling problem; (b) Modified Mincost (MMC)
algorithm for approximating an LP/IP optimal solution
for those applications that can run on only one grid
resource (single-grid node) and (c) LPGA for
scheduling single-grid node applications.

The rest of the paper is organized as follows. In the

next section, we describe the related work on metaschedulers and heuristic based scheduling. In Section 3,
we present the problem definition, which is given with
the LP model of the scheduling problem. Then, in
Section 4, we present LPGA scheduling algorithm. In
Section 5, we discuss simulation study of various
algorithms and their comparison with the proposed
algorithms. Finally, Section 6 concludes the paper and
described the future work.
II. RELATED WORK
Market based mechanisms can be divided into two
based on market models i.e. auctions and commodity
market model. As our work is relevant for commodity
markets in Grid thus in this section we will compare our
with other resource allocation mechanisms for this
market model. Cost-based resource management and
market-oriented models have proven to be useful for
resource management [18]. The most relevant works to
our research are [9][19][21] and [20]. In our previous
work Gridbus Broker[9] and Nimrod/G[25], a greedy
approach is proposed to schedule a parameter sweep
application with deadline and cost constraints. To be
precise, our previous work on grid scheduling [9][25]
was focused on application level scheduling, i.e. a
personal broker for efficient deployment of an
individual application on utility grids. But this work is
focused on scheduling of many applications from users
having different QoS requirements with the aim of
global optimization across many user applications.
G-commerce [19] is another economic-based study
that applies strategies for pricing grid resources to
facilitate resource trading, and it compares auction and
commodity market models using these pricing models.
Feng [21] proposed a deadline cost optimization model
for scheduling one application with dependent tasks.
These studies has many limitation (1) algorithms
proposed are not designed to accommodate concurrent
users competing for resources (2) also application
model considered are simple i.e. independent task or
parametric sweep application. In this work we have
modeled both independent and parallel applications
submitted by concurrent users. Similarly, Dogan [20]
proposed a metascheduling algorithm considering many
concurrent users but application model was again very
simple. They assumed that each application consists of
one task and also each application is independent. In
this paper, we have considered multiple and concurrent
users competing for resources in a meta-scheduling
environment to minimize combined cost of all users.
Moreover, in this paper, we presented a formal
theoretical mathematical model for optimal scheduling

of jobs with Deadline and Budget constraints which is

essential to develop an effective heuristic that may
provide a near optimal solution.
In the grid environment, where each user has
different QoS constraints, the scheduling problem
becomes a Generalized Assignment Problem (GAP)
which is a well known NP hard problem. GAP can be
solved using heuristic based GA [2][4]. Thus many GA
based heuristics are proposed in the literature
Wiessman et al. [6] proposed a novel GA based
algorithm which schedules a divisible data intensive
application. Martino et al. [10] presented a GA based
scheduling algorithm where the goal of superscheduling was to minimize the release time of jobs.
These GA based heuristic based solutions [13][10][6]
do not consider QoS constraints of concurrent users
such as budget and deadline. The viability of
evolutionary algorithm-based scheduling such as GA
for realistic scientific workloads is demonstrated by
systems like Mars [11]. Mars is a meta-scheduling
framework for scheduling tasks across multiple
resources in a grid. In a recent study [5], five heuristic
based algorithms were compared, which found that GA
produces the best utility for users in comparison to
heuristics such as Min-Min. Due to these reasons we
have used GA seeded with greedy as an algorithm to
compare with.
From scheduling approaches outside the Grid
computing, it is well known that LP/IP can offer
optimal solutions in minimization/ maximization
scheduling problems when constraints and objective
functions are linear. In Feltl et al. [4], an LP/IP based
initial solution and followed by an intelligent
replacement of offspring based on Martello et al. [3] is
proposed. But these models do not consider the
deadline and budgets constraints. As they are
developed based on specific domain knowledge, they
cant be applied directly to grid scheduling problems,
and hence have to be enhanced accordingly.
The main contribution of this paper is to develop a
mathematical model and an LP/IP based GA algorithm
to minimize the combined procession cost of user
application in a concurrent users environment for
utility Grids. To best of our knowledge, the solution
methodology proposed doesnt overlap with existing
techniques for scheduling problems.
III. PROBLEM DEFINITION
A. System Model
Figure 1 shows the interaction of the MetaBroker
with resource providers and users. Grid users submit
their application jobs to the MetaBroker with their

QoS requirements. The QoS requirements consist of

budget, deadline, and number of processing elements
(PEs) i.e. CPU required. Each application job consists
of independent tasks with each task requiring one PE. A
resource comprises of homogeneous PEs. Two kinds of
users application jobs considered are:
Multi-grid node (MGN) jobs can split into
independent tasks, which can run on different
resources. For example, bag of tasks
applications.
Single-grid node (SGN) jobs which require a
complete set of PEs on the same resource and
fail if all PEs are not available. For example,
synchronous parallel applications.
The MetaBroker gathers information about
resources such as PEs available, grid middleware, cost
of PEs per unit time for each user and PEs capacity (e.g.
Millions of Instruction per Seconds (MIPS) rating) from
the resource providers who have agreed to rent their
computational resources to MetaBroker. The
MetaBroker generates a combined schedule based on
algorithms discussed in subsequent sections. As the
demand for resources may exceed the supply in the grid,
some jobs are placed in the queue for latter assignment
when resources are available or freed.
User
Community

R1

R2

MetaBroker

4) Send
mapping
to users

2) MetaBroker will get

Current load from
Resources

5) User
sends
job to
Resource

R3

R4

R5

R6

Resources
Figure 1. Grid model with MetaBroker

B. Problem Formulation
The objectives of meta-scheduling in the above
environment with m resources and n jobs are as
follows:

Maximize the total number of users job

submission.

Let Pi represent the information which MetaBroker

receives from grid resource i at a given instance T.
P : ( ri , ni , ci,j , vi ) where
i
ri is the Resource Id,
ni is number of free PEs available on the resource,
ci,j is the cost of using a single resource per second
for the job/user j,
vi is the MIPS speed of one of the PE,
and i I = {1..m}
Let Qj represent the QoS demand which MetaBroker
receives from the user j at a given time instance T.
Qj: (uj, jj, bj, dj, Mk, mj) where
uj is the user id,
jj is the job id, since one user can submit multiple
jobs, hence combination (uj,jj) is unique,
bj is the budget constraint the user specifies,
dj is the deadline constraint for finishing the job,
Mk size of each task in the job in terms of Millions
of Instructions (MI) [22][9],
mj is the number of PEs the user requires for
executing the job,
and j J = {1..n}, k K = 1..m j .

1) User will
send
QoS req. to
Metabroker

3) Map users
to
Resources

Scheduling based on QoS, i.e., budget and

deadline.
Minimize combined cost for all the users by
meeting the budget and deadline constraints.

The mathematical model for scheduling of SGN

jobs is as follows:
m n
c = cij xijrij Max (M / v )
(1)
k k i
i=1 j =1
In Equation (1), the variable rij denotes the number
of PEs allotted to a job j on a resource i. Note that, in
case of SGN jobs, the value of rij would be either 0 or
mj (the number of PEs required by the job j). The
variable xij denotes that whether job j is mapped to
resource i or not. Equation (1) denotes cost function
which all users have to pay to resource providers. The
constraints are following :
n
rij xij ni , i I
(2)
j =1
m
rij xij = m j , j J
(3)
i =1

xij {0,1}, i I , j J

(4)

xij = 1, j J

i =1

(5)

m
rij cij b j , j J
i =1

(6)

MAX ( M k / vi ) xij d j , i I , j J , k K

(7)

cij , rij , b j and d j 0, i I , j J

(8)
Equation (2) denotes the resources capacity
constraints. Equation (3) denotes the PE requirement of
users job i.e., mj. Equations (4) and (5) denote the
assignment matrix, which indicates that a job can be
assigned to a maximum of one resource. Equation (6) is
added for satisfying the budget constraints and Equation
(7) is added for the time constraints. Equation (1) is
solved for minimization to obtain an optimal cost
efficient schedule. All the quantities, i.e., the resources,
cost, deadline and budget are greater than zero, hence
the Equation (8). Note that Equations (4) and (5) make
the problem NP-hard as these constraints make the
problem a 0-1 assignment problem which is well known
to be NP-hard [3][4][8].
LP/IP model for scheduling MGN jobs is almost
same as for SGN jobs. The only difference is that model
for MGN jobs will not include constraints (4) and (5).
C. Genetic Algorithm Formulation
Genetic Algorithms (GAs) [8] provide robust search
techniques that allow a high-quality solution to be
derived from a large search space in polynomial time,
by applying the principle of evolution. A genetic
algorithm combines the exploitation of best solutions
from past searches with the exploration of new regions
of the solution space. An individual (chromosomes)
represents any solution in the search space of the
problem. A genetic algorithm maintains a population of
individuals that evolves over generations. The quality of
an individual in the population is determined by a
fitness-function. The fitness value indicates how good
the individual is compared to others in the population. A
typical genetic algorithm consists of the following steps:
(1) Initial population generation;
(2) Evaluation;
(3) repeat
(4) Selection;
(5) Crossover;
(6) Mutation;
(7) Evaluation;
(8) until (convergence).
Convergence can be a different criterion for different
problems, but generally a no change in the solution for
n generations is considered as convergence. The most
important aspect in GA is the solution space encoding
and the fitness function. We have taken the same fitness
function and solution space encoding as given in our
previous work [8]. The genetic operations, i.e.,

crossover, mutation and inversion are standard

operations in GA.
IV. PROPOSED ALGORITHMS
A. Linear Programming based Algorithm for
Scheduling MGN Jobs
As discussed in the previous section, the
mathematical model for MGN jobs scheduling will be
converted to an LP/IP model after removing
constraints (4) and (5). LP/IP can be then easily solved
using standard integer programming algorithms. For
simulation purposes, we have used the Coin-OR library
[12].
B. Linear Programming based Algorithm for
Scheduling SGN Jobs (LPGA)
Due to the absence of constraints (4) and (5), LP/IP
solutions obtained in section 4.1 are infeasible for the
SGN jobs scheduling problem. We need to
approximate LP/IP to a feasible solution. For this
problem, we have designed a novel algorithm MMC to
approximate an infeasible LP/IP solution to a feasible
solution for the SGN jobs scheduling problem.
1) Modified MinCost (MMC) Algorithm: Our
MMC algorithm is inspired from Minimum Cost
Algorithm which is a well-known method for finding
the feasible solution near to optimal in the transport
problems. In the Minimum Cost algorithm, we find the
minimum cost resource, assign maximum number of
jobs to that and reduce the demand and supply. The
process is repeated until all jobs are allocated.
In MMC, we take an optimal solution from the LP
based algorithm (discussed in the previous section) and
then approximate it to obtain a feasible solution for
SGN scheduling problem. This algorithm gives a
solution, which is between the optimal, and the greedy
solutions, i.e. it will give a better solution than greedy
but not an optimal as shown in the evaluation section.
The solution obtained from this algorithm acts as an
initial solution to GA.
The pseudo code for MMC is given in Algorithm 1,
which shows how MMC approximates an LP/IP
solution. It takes as input a list of jobs which is a
structure containing information regarding how
different tasks of a job are mapped to resource
providers and how many PEs are required by the job.
This job list is generated from algorithm for scheduling
MGN jobs. First, MMC algorithm makes a copy of the
input, and then sorts job list on the basis of number of
resources to which the job is mapped (line 1-2). This is
done so that we get a solution as closer as possible to
the optimal LP/IP solution.

We handle differently the jobs which are mapped

only to (a) one resource and (b) more than one resource.
For jobs which are mapped only to one resource, we
freeze their mappings. The mappings will not change
during the algorithm, and thus we reduce the total
capacity of resource (line 4, 5, 6). Then, these jobs are
added later to the schedule list (line 7). The jobs, which
were mapped to more than one resource, will be mapped
to a resource pi which has the capacity to allocate the
full job j1 and is most economical (line 9-11). Other
jobs which are allocated to this resource will be
remapped to resource where job j1 was allocated other
than pi, if deadline and budget constraints are satisfied;
otherwise they are allocated to dummy resources (line
15). Other jobs, which are not allocated in the above
process, will be mapped to the dummy resource (line
19). These jobs are remapped to real resources using a
greedy approach.
Algorithm 1: ModifiedMinCost (Listjobs)
1
2
3
4
5
6
7
8
9
10
11
12
13
15
16
17
18
19
20
21
22
23

ListJListjobs.clone();
sort(ListJ);
for i0 to ListJ.size()-1 do
j1ListJ(i);
if (numofProviders(j1)==1) then
changeAvailableCapacity(Provider(j1),j1);
Goto step 20
else
PJ1.providerlist.clone();
sort(P); //number of PE allocated on
Provider
for p0 to P.size() do
if(remainingCapacity(p)>PERequired(j1))
AllocateFull(j1,p);
InterchangeCapacity(p,p.joblist[]);
Goto step 20
endfor
endif
addjob(j1,DummyResProv);
endfor
ScheduleDummyResJob();
MakeScheduleList(ListJ);
return ListJ;

Complexity: Complexity of MMC algorithm for m

resources and n jobs is O(nlogn + mn +
O(interchangeCapacity() function)*n). We have taken a
simple implementation of interchangeCapacity function
using arrays, which has worst-case complexity as
O(mn). Thus, the worst-case complexity of the
algorithm becomes O(mn2), which is very high upper
bound over real complexity as average number of jobs
per resource will be less than total number of jobs.
2) LPGA Algorith: LPGA seeds the approximate LP
solution with Genetic algorithm for generating a
solution near to optimal solution. In LPGA, first we
solve an LP/IP problem without constraints (4) and (5)

and then MMC approximates the solution obtained in

the previous step. Finally, using the feasible solution
from MMC, we iterate various genetic operations to
get the best solution.
Algorithm 2. LPGA()
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35

ScheduleListnull;
while (current_time < next_schedule_time) do
RecvResourcePublish(Pj); //from providers
RecvJobQos(Qj); //users
endwhile
if (numJobs>numResource) AddDummyNode();
ListPListProviders.clone();
Sort(ListP) //cost of resource
for each job j in Q do
compute QoS index
endfor
ListJ=sort(Q); //descending order of QoS index
LPschedule=GetLPsolution();//from Coin Library
ScheduleList= ModifiedMinCost(LPschedule);
AddinPopulation(POPU_LIST, ScheduleList);
generateNextGenerationPopulation(){
for each i in (population_size-POPU_LIST.size())
cromos=GenerateRandomCromosome(i);
POPU_LIST.add(cromos);
endfor
ComputeFitnessfunction();
bestCromos=SearchBestCromosome(POPU_LIST);
if (termination) return bestCromos
doSelection(){
SelectBestCrossoverRateCromosome() // based
onRoulette wheel selection policy
DoCrossover()
DoMutation()
AddtoPopulationList(POPU_LIST);
Goto step 16
}
}
Schd_List =GenerateSchedulelist(bestCromos);
for each element in Schd_List do
notifyuser();
endfor

Algorithm 2 gives the pseudo code. As the

MetaBroker at a particular time instance does
scheduling, it waits for jobs from users (taken as QoS
requirement for job in line 4) and resource load from
provider (line 3). Then we have to add dummy job
queue with high cost and number of PEs to make
scheduling problem balanced, i.e. to balance supply of
resource service and demand by users for computing
services (line 6). After that, we sort resource list on the
basis of their cost and job list on the basis of QoS
index (line 7-12). After sorting resource list and job
list, we get the LP/IP solution which is obtained
without considering constraints (3) and (4) discussed in
section 3.2 (line 13). This solution is approximated
using MMC algorithm (line 14) which is used to
generate initial population space for the genetic
algorithm (line 17-20). From initial population space,
we select the best chromosome values based on the

fitness function. These best chromosomes are then

mutated to generate the next generation population (line
21-30). This operation is repeated until the convergence
criterion is reached or maximum specified limit of
iteration is reached (line 23). Based on the chromosome,
schedule list is generated and users are notified about
mappings (line 32-34). One limitation to the problem
formulation is that if the number of resources requested
is more than the number of resources available for the
broker, then the problem results in an infeasible
solution. Therefore, we propose addition of dummy
resources with infinite capacity and having more cost
than any of the resources available on the grid. This
enables the algorithm to converge and assign some of
the jobs to the dummy resources. Dummy resource jobs
are rolled over to the next schedule period, as they
cannot be executed.
V. PERFORMANCE EVALUATION
A. Simulation Methodology
We use the GridSim toolkit [1] to simulate a grid
environment and the MetaBroker for scheduling users
application. We simulated the grid with heterogeneous
resources having different MIPS [22] ratings and each
resource having different number of PEs in the range of
4 to 12 with mean number of PEs as 8. A Grid having
resources in the range of 25 to 200 is generated. The
usage cost per second of each resource is varied (using
Gaussian distribution) between 4G$-5G$ (G$ means
Grid dollar) with mean cost of 4.5G$. Execution time in
the simulator setup for the several hundred runs restricts
us to model 50 jobs per run. Thus, the MetaBroker
schedules the jobs (which consist of several tasks, each
require one PE to run) submitted by 50 concurrent users
with simulation interval of 50 seconds. The jobs are also
simulated with varying QoS requirements. Jobs with
average number of 5 tasks and having 10-50%
variations in the number of tasks (using Gaussian
distribution) are considered and all the jobs are
submitted within 20 seconds of simulation start time.
Average estimated run time for jobs is also taken to
be 400 seconds and varied by 20% using Gaussian
distribution. As far as QoS parameters are concerned,
i.e.., budget and deadline, three simulations are done for
different deadline scenarios as listed below:
Experiment-1: Tight deadline (estimated time+(50
seconds with 20% variation)).
Experiment-2: Medium deadline (estimated time+(250
seconds with 20% variation)).
Experiment-3: Relaxed deadline (estimated time+(500
seconds with 20% variation)).
As far as the budget is concerned, all the jobs are
given relaxed budget constraints (i.e., twice the cost of

average job execution time). Our performance

evaluation examines the relative performance (users
spending) of LPGA with respect to 3 other following
meta-scheduling algorithms for the above type of jobs
by varying the number of resources:
1. Greedy based meta-scheduling algorithm.
2. Heuristic GA algorithm (HGA) [8].
3. Modified Mincost algorithm (MMC).
The results for the three different simulation
scenarios for both MGN LP based algorithm and SGN
LP driven Genetic Algorithm (LPGA) are discussed in
the next section.
B.

Performance Results
1) LP/IP based Meta-Scheduling Algorithm for
MGN Jobs: The results are compiled in Table 1. In
tight deadline out of 50 jobs (246 tasks) only 34 jobs
(170 tasks) were scheduled where as in medium and
relaxed deadlines all jobs i.e. 50 (246 tasks) were
scheduled. The reason behind this is the tight deadline,
for which many jobs missed the deadline constraints
due to simulation interval. Thus, even though the
number of resources increased in number, the number
of scheduled jobs remained same.
TABLE 1.
TOTAL COST SPENT BY USERS FOR MGN JOBS
Number
Of
Providers
25
50
100
150
200

Total
Cost in
Medium
Deadline
259832.6
261059.7
260689.1
260653.5
260664

Total
Cost in
Relaxed
Deadline
194668.9
198771.5
198462.5
198360.5
198455.3

Total
Cost in
Tight
Deadline
245280.8
247655
247251.2
247245.4
247015.7

Moreover, in case of relaxed deadline, user

spending is the minimum in comparison to other
deadline types because more jobs are scheduled to
resources with least cost. Also, even though it seems
that this trend is not followed when we compare tight
deadline and medium deadline scenarios, it results in
more number of jobs scheduled in the case of medium
deadline. It is also interesting to note that in the tight
deadline scenario, with the increase in number of
resources; total cost spent by users is also decreasing.
2) SGN Job Scheduling Algorithms: The
performance of four meta-scheduling algorithms has
been compared by varying the number of jobs
scheduled and users spending. The results are
compiled and presented in Fig. 2 to Fig.4. In the first
set of results, i.e., Fig. 2a, 3a and 4a, aggregated

revenue earned for the completed jobs are plotted on Yaxis. In the first set of results, i.e. Fig. 2b and 3b, X-axis
presents the grid resources while Y-axis presents the
jobs completed
(out of the 50 concurrent jobs
submitted). Due to space constraint we have not
presented the number of jobs completed with number of
resources and also trend in medium deadline case is
very similar to trend in case of relaxed deadline.

the results demonstrate the user spending for each job

decreases when LPGA is used by Meta-Broker for
combined scheduling of jobs. This is because MMC
algorithm is giving better schedule than greedy
algorithm.
1600
1400

Greedy
MMC
HGA
LPGA

1200

User Spending

1600

1000

1400

Greedy
MMC
HGA
LPGA

Users Spending

1200
1000

800
600
400

800

200

600

0
25

50

400

100

150

200

Number of Resources

(a)

200

Figure 3. Medium Deadline

0
25

50

100

150

200

Number of Resources

(a)

1400

40

Greedy
MMC
HGA
LPGA

1200

35
1000

Users Spending

Number of Jobs

30
25
20

800

600

400

15
200

10

LPGA
HGA
Greedy
MMC

0
25

50

100

150

200

Number of Resources

0
2

10

25

50

100

150

200

(a)

Number of Resourses
60

(b)
50

The figures 2a, 3a and 4a for relaxed, medium and

tight deadlines show that LPGA has outperformed all
other meta-scheduling algorithms, thus minimizing the
users spending. In case of all deadline, even though
initially, user spending in schedule given by HGA and
LPGA seems to be same, as number of resources
increases, LPGA decreases the user spending if
compared to other algorithms. Since the cost variations
within the grid resources are not significant (i.e. 4.5 G$
with 0.5G$) only up to 7-8 % cost benefit was noticed.
However, more benefit can be anticipated if the
variations are higher.
More interesting results can be observed when we
compare performance of MMC algorithm and greedy
approach. It is clear from the graphs that cost gain for
users in case of MMC algorithm can be as much as 30%
more beneficial for users than greedy approach. Thus,

Number of Jobs

Figure 2. Tight Deadline

40

30

20
HGA
LPGA
Greedy
MMC

10

0
2

10

25

50

100

150

200

Number of Resources

(b)
Figure 4. Relaxed Deadline

It can be noticed from Fig. 2b and 4b that the curves

become smoother and peak early compared to tighter
deadline. This trend is due to the relaxed deadline
having more jobs completed with the same number of
resources. As when deadline is more relaxed, meta-

brokers try to complete more and more tasks. With

increase in number of resources, all algorithms schedule
the same number of user jobs. In case of tight deadline
(fig. 2b), as many jobs missed the deadlines (very tight);
only 34 jobs are completed even though the number of
resource increased to 200.
Next simulation study presents the advantage of
LPGA algorithm over HGA algorithm. We have
compared the number of iteration taken by both
algorithms in case of tight deadline. The results shown
in Fig. 5 indicate that in case of LPGA, number of
iterations is reduced with a decrease in user spending.
For example, with 50 resources, the number of iterations
performed in HGA is 1000, whereas in LPGA it is less
than 700. Even though the number of iterations in HGA
is less than in LPGA with 150 resources, the revenue
generated is smaller which can be observed from Fig.
3b. These results also indicate that in spite of taking
more iteration, the HGA could not find the global/better
optimum. Hence, to find a better solution (either the one
found by HGA implementation or any other better
solution) the algorithm would have required a higher
number of iterations. This clearly explains the reason
for a higher number of iterations in case of LPGA when
the number of resources is 150.
1200

Number of GA Iterations

1000

800

HGA

600

LPGA

400

200

The results show that LPGA out performs other

meta-scheduling algorithms such as greedy approach
and HGA by giving the minimum processing cost. In
LPGA, not only number of iterations is reduced by 7%
to 25% compared to HGA, but user spending is also
decreased as the number of resources increases. Our
results also show that MMC can reduce the combined
user spending up to 30% over traditional greedy
approach.
The application of approximation algorithms like
MMC in Grid environment is novel. In addition, this
work shows us that many good heuristics can be
designed after some modifications. In future, our goal
is to study and analyze such algorithms. Also we would
like to integrate this algorithm with some available
meta-schedulers and deploy them on real test beds. We
have given priority to jobs that are mapped to
minimum number of resource providers in the LP
solution. For future work, we intend to improve the
performance of MMC and test it with different priority
indexes such as deadline and execution time. We also
want to investigate other approaches used for the
assignment problem which can be modified for utility
grids. We want to experiment our algorithm with a
wide range of costs to study the effect of costs change.
ACKNOWLEDGMENT
We would like to thank our colleagues Marcos,
Marco Netto, Srikumar Venugopal and Chee Shin Yeo
for their comments and suggestions on this paper.
This research is funded by the International Science
Linkage grant provided by the Department of
Innovation, Industry, Science and Research (DIISR)
and Australia Research Council (ARC) of the
Australian Government.

0
25

50

100

150

REFERENCES

200

Number o f Resources

[1]
Figure 5. Comparison of number of iterations in HGA and LPGA

VI. CONCLUSION AND FUTURE WORK

In this paper, we have modeled the scheduling
concurrent user jobs in utility grids as an LP/IP
problem. Two different types of jobs were considered,
namely multi-grid node jobs and single-grid node jobs.
We also presented a novel algorithm called MMC that
approximates the LP solution to get a single-grid node
job scheduling solution near to optimal. We designed
the LPGA algorithm to decrease the combined user
spending and resource utilization by seeding the
solution from MMC algorithm to genetic algorithm. We
have also taken into consideration user QoS constraints,
i.e., deadline, number of PEs and budget.

[2]
[3]
[4]

[5]

R. Buyya and M. Murshed, GridSim: A Toolkit for the

Modeling and Simulation of Distributed Resource
Management and Scheduling for Grid Computing,
Concurrency and Computation: Practice and Experience
(CCPE), Wiley Press, USA, November - December 2002, vol.
14, pp. 1175-1220.
P. C. Chu, J. E. Beasley, A genetic algorithm for the
generalized assignment problem, Computers and Operations
Research, January 1997, vol. 24, pp. 17-23.
S. Martello and P. Toth, An algorithm for Generalized
Assignment Problem, Operational Research, 1981, vol. 81,
pp. 589-603.
H. Feltl and G. R. Raidl, An improved hybrid genetic
algorithm for the generalized assignment problem,
Proceedings of the 2004 ACM symposium on Applied
computing, Nicosia, Cyprus, March 2004.
K.S. Golconda, F. Ozguner, A comparison of static QoSbased scheduling heuristics for a meta-task with multiple
QoS dimensions in heterogeneous computing, Proceedings

[6]

[7]
[8]

[9]

[10]

[11]

[12]
[13]
[14]

[15]
[16]

[17]

[18]

[19]

[20]

[21]

of the 18th International Parallel and Distributed Processing

Symposium, 2004.
Seonho Kim, Jon B. Weissman, A Genetic Algorithm Based
Approach for Scheduling Decomposable Data Grid
Applications, Proceedings of the 2004 International
Conference on Parallel Processing (ICPP'04), August 2004,
pp. 406-413.
Amazon Elastic Compute Cloud (Amazon EC2),
http://www.amazon.com/gp/browse.html
Pramod Kumar Konugurthi, Krishnan Ramakrishnan, and
Rajkumar Buyya, A Heuristic Genetic Algorithm based
Scheduler for Clearing House Grid Broker, Technical Report,
GRIDS-TR-2007-22, Grid Computing and Distributed Systems
Laboratory, The University of Melbourne, Australia,
November 2007.
Rajkumar Buyya, Manzur Murshed, David Abramson, and
Srikumar Venugopal, Scheduling Parameter Sweep
Applications on Global Grids: A Deadline and Budget
Constrained Cost-Time Optimisation Algorithm, Software:
Practice and Experience (SPE), Volume 35, Issue 5, Pages:
491 - 512, Wiley Press, New York, USA, April 25, 2005.
V. Di Martino, M. Mililotti, Scheduling in a grid computing
environment using genetic algorithms, Proceedings of the
International Parallel and Distributed Processing Symposium,
IPDPS 2002.
A. Bose, B. Wickman, and C. Wood. MARS: A
Metascheduler for Distributed Resources in Campus Grids,
5th International Workshop on Grid Computing, IEEE
Computer Society, 2004.
Coin-Or Project , http://www.coin-or.org/projects/.
J. Cao, D.P. Spooner, S.A. Jarvis and G.R. Nudd, Grid load
balancing using intelligent agents, Future Generation
Computer Systems, 2005.
Neumann, D.; Stoesser, J.; Anandasivam, A.; Borissov, N.:
SORMA Building an Open Grid Market for Grid Resource
Allocation, Proceedings of the 4th International Workshop on
Grid Economics and Business Models, GECON 2007
Jrn Altmann, Costas Courboubetis, John Darlington, Jeremy
Cohen "GridEcon - The Economic-Enhanced Next-Generation
Internet", GECON 2007, LNCS 4685, Springer 2007
B. N. Chun and D. E. Culler, User-centric performance
analysis of market-based cluster batch schedulers, 2nd IEEE
International Symposium on Cluster Computing and the Grid,
May 2002.
Jia Yu, Rajkumar Buyya, and Chen Khong Tham, Cost-based
Scheduling of Workflow Applications on Utility Grids,
Proceedings of the 1st IEEE International Conference on eScience and Grid Computing (e-Science 2005, IEEE CS Press,
Los Alamitos, CA, USA), Melbourne, Australia.
Cheliotis, Kenyon, and Buyya, Grid Economics: 10 Lessons
from Finance. Joint Technical Report, GRIDS-TR-2003-3,
IBM Research Zurich and Grid Computing and Distributed
Systems Laboratory and University of Melbourne, 2003
Wolski, R., Plank, J. S., Brevik, J. and Bryan, T. (2001).
Gcommerce: Market formulations controlling resource
allocation on the computational Grid. International Parallel
and Distributed Processing Symposium, San Francisco, USA,
April 2327.
A. Dogan and F. zgner: Scheduling Independent Tasks with
QoS Requirements in Grid Computing with Time-Varying
Resource Prices, Proceeding of Grid Computing-GRID 2002,
58-69, 2002
Feng H L, A deadline and budget constrained cost-time
optimization algorithm for scheduling dependent tasks in grid
computing. Grid and Cooperative Computing. Second

[22]
[23]
[24]
[25]

International Workshop (GCC 2003), Germany: SpringerVerlag, 2003.113-120.

SPEC 2007, Standard Performance Evaluation Corporation.
http://www.spec.org/.
I. Foster and C. Kesselman, Eds., The Grid: Blueprint for a
Future Computing Infrastructure. San Mateo, CA: Morgan
Kaufmann, 1999.
I. Foster, C. Kesselman, J. Nick, and S. Tuecke, Grid
services for distributed system integration, Computer, vol.
35, no. 6, pp. 37-46, Jun. 2002.
D. Abramson, R. Buyya, and J. Giddy. A Computational
Economy for Grid Computing and its Implementation in the
Nimrod-G Resource Broker. Future Generation Computer
Systems (FGCS) Journal, 18(8):10611074, October 2002