Вы находитесь на странице: 1из 12

GRID SCHEDULING USING VARIOUS PERFORMANCE MEASURES

– A COMPARATIVE STUDY

1
Dr.K.Vivekanandan, 2D.Ramyachitra
1
Professor, BSMED, Bharathiar University, Coimbatore
E mail: vivekbsmed@gmail.com
2
Assistant Professor, School of Computer Science and Engineering, Bharathiar University, Coimbatore
E mail: jaichitra1@yahoo.co.in

ABSTRACT

Grid Computing, an extension of distributed computing, allows sharing of geographically


distributed resources across multiple administrative domains. As the users can access the
resources transparently without knowing where they are physically located, there are many
challenges that have to be considered. One of the challenges involves scheduling the jobs to
the appropriate resources. This paper gives a survey on scheduling algorithms used in grid
environment. The algorithms have been implemented using gridsim, a simulator used for
creating a grid environment. Various performance measures such as makespan, resource
utilization, cost and profit are used for comparing the algorithms.

Keywords:MET, MCT, Max Min, Min Min, OLB, GA, Tabu Search, SA, Grid, Scheduling.

1 NTRODUCTION computational and storage resources among users


located all over the world [3]. Traditional approaches
maintain a centralized server or hierarchically
organized servers to index the resource information.
A Computational Grid is a collection of Centralized system has all the functional software
heterogeneous resources such as computational components on a single computer but centralized
devices, networks, online instruments, storage servers becomes a problem in the case of highly
archives etc that provides an enormous potential of dynamic environment where many resources join,
capabilities that can be brought to bear on large leave and change characteristics at any time and also
distributed applications and are becoming prevalent it does not scale to large number of grid nodes across
platforms for high performance and resource intensive autonomous organizations. As grid computing is a
applications [1]. In the current networking technology form of distributed computing, where the resources
and based on the availability of the bandwidth, the are geographically distributed, centralized form of
computing resources are aggregated together to form a systems will be a bottleneck for grid computing.
grid computing environment. These computational Hierarchical approaches provide better scalability and
resources which provide free or chargeable services fault tolerance but it takes a long time for resource
may consists of one or more processing elements of information to be updated from the leaf nodes to the
same or different technologies [2]. Data Grids are root node [4]. For peer to peer architecture single
predicted to be the solution to the large computational point of failure does not exist, but in the case of
power and data storage requirements of many research centralized system, if the centralized hub is broken, it
projects. It also enables the sharing of distributed will lead to failure of the entire system. For grid
computing, communication topology is required failure. If one of the resources gets failed, other will
because it deals with different computers in different get over it. It is so scalable and upgrading can be done
geographical locations. Even though there is a on the fly. The split up jobs can also be run in parallel
communication overhead in the case of grid with high speed on many nodes. MPI can also be used
computing, several computers with unutilized for message passing among computer resources.
resources in different geographical locations are used These features make grid computing attractive for the
to solve a single large problem so that overall enterprises also. This enterprise grid computing
computational power is enhanced. reflects the use of grid computing within the context
of a business or enterprise rather than for scientific
One of the characteristic of the grid is users no applications. Stateful nature of business applications,
need to have idea where the required resource for the typical underlying multi-tier architecture where
solving their problem is located. Users also no need to request execution follows a complex path through a
care about the resources they use when their jobs are diverse set of components etc makes these enterprise
running. From the viewpoint of the resources, they do applications harder to deploy on a Grid infrastructure
not need to give a permit to each user to access them [8].
directly but only tasks which have limited access
rights can be allowed to be executed [5]. As grid As grid environments include many servers
involves heterogeneity which arises mainly from the across various administrative domains, managing and
large variety of resources available within each utilizing those resources and keeping configurations
category, the Grid resource broker hides the in synchronization will be challenging in large
complexities of the grid computing environment from environments. Simulation is the only feasible way to
a user. This grid resource broker discovers resources analyze algorithms on large scale distributed systems
that the user can access using information services, as in grid environment. Simulation works well by
negotiates for access costs; maps application jobs to making the analysis system simple by avoiding the
resources, starts the execution process and monitors overhead of co-ordination of real resources compared
the progress of execution [6]. These heterogeneous to using the real system in real time environment.
resources are managed using metadata, whose Gridsim toolkit allows modeling and simulation of
purpose is to provide information about the features of entities in grid computing users , applications,
resources and their effective use. Meta data can resources and resource brokers for design and
provide information regarding which resources are evaluation of scheduling algorithms. Some of the
available, how resources can be accessed, when they features of gridsim includes modeling of
will be available etc. So, metadata can represent a key heterogeneous types of resources, resource capability
element for the discovery and utilization of resources can be defined in the form of MIPS, resources can be
on the grid [7]. Apart from hiding the heterogeneity of located in any time zone, advance reservation can be
the resources, another interesting characteristic is the made for the resources, submission of unlimited
transparent pooling of many kinds of resources such number of applications for the execution, network
as computing power, storage, data and services. This speed between the resources can be specified etc [9].
enables the applications deployed on the grid to
transparently share resources and also the ability to The remaining part of the paper is organized as
allocate and reserve virtualized resources. This follows. Section 2 deals with scheduling in grid
virtualization facilitates sharing of resources and also environment. Section 3 shows the comparison of the
allows the preservation of QoS for time critical scheduling algorithms’ using simulation results and
applications [8]. finally section 4 gives the conclusion.

Grid computing can be used for applications that


can be split and sent to different resources for
execution. After completion of the execution, the 2 SCHEDULING IN GRID ENVIRONMENT
results can be concatenated and analyzed. There are
many policies for sending the jobs based upon their
characteristics. This will result in efficient utilization Scheduling is the process of allocation of tasks to
of the idle resources. There is no single point of the appropriate resources. Before scheduling the tasks
to the resources, the sequence of tasks in which the information regarding tasks and resources should be
execution will be performed and the time duration of known before the execution starts. In the case of
each activity should also be determined. Job shop dynamic scheduling, task is allocated on the fly as the
scheduling is one of the most well known online application executes. Both static and dynamic
problems and the basic models were developed with scheduling is adopted in grid computing depending
makespan as the objective. Makespan is the time upon the application and the available information.
taken by a given set of resources to execute a Some of the heuristics such as OLB, MET, MCT, Min
sequence of tasks. Scheduling algorithms at the min, Max min, Duplex, Genetic Algorithm, SA, Tabu
beginning were developed with an objective of Search, A* [13], ACO , PSO etc are widely adopted
minimization of the makespan. for scheduling in grid environment and are discussed
here.
During 1960’s mathematical programming such
as integer and dynamic programming has been applied Opportunistic Load Balancing (OLB) assigns
to job shop scheduling problems. But as job shop each to the machine that is available in an arbitrary
scheduling problems are NP Complete problems, it is order and it does not consider the expected execution
not possible to find the exact schedule. So many time on that machine. In Minimum Execution Time
researchers proposed new techniques and heuristics (MET), the task is assigned in an arbitrary manner to
that give optimal schedule which was better compared the machine with the best expected execution time for
to scheduling using mathematical programming. the task, regardless of the availability of the machine.
Dispatching rules or heuristics have been applied to Minimum Completion Time (MCT) assigns each task
scheduling problems which is classified into different in arbitrary order to the machine with the minimum
classes according to the performance criteria for expected completion time for that task. Min min is
which they have been applied [10]. Class 1 contains based on the MCT but Min min considers all
simple priority rules based on processing times, unmapped tasks during each mapping decision and
deadlines, slack and arrival time. Class 2 contains MCT only considers one task at a time. In Max min,
combinations of rules from class 1. Class 3 contains the task with maximum completion time is allocated
rules that are referred to as Weight Priority Indexes. to the machine that has minimum completion time.
Even though a considerable number of works studied Duplex performs both Min min and Max min and uses
on the above said rules, most of the work concentrated the better solution [13]. Genetic algorithm is an
on processing time i.e., makespan. Performance evolutionary technique for large space search. It
measures other than makespan that were optimized operates on a population of solutions rather than on a
were deadline, tardiness, throughput and utilization. single solution. Much work has been done using GA
Some research works developed optimization for grid scheduling [14]. A set of chromosomes are
measures on the resource provider’s side also. Apart initialized and the fitness is calculated for the
from these performance measures, other Quality of chromosomes. Good chromosomes are selected,
Service includes cost for resource consumer, crossover is performed and mutation is done. This
reliability of the resource, income, profit for resource process is repeated until a good mapping is
provider etc. This paper has compared several performed. Simulated annealing is an iterative
algorithms using the QoS such as makespan, cost, technique like genetic algorithm but only considers
income and profit. one possible solution for each task at a time. It
probabilistically allows poorer solutions to be
In general scheduling algorithms can be classified accepted and this probability is based on a system
into local and global scheduling [11]. Processes are temperature that decreases for each iteration. Initial
assigned to the time slices of a single processor in the temperature is the makespan of the initial mapping.
case of local scheduling. Decision has to be taken Then the mapping is mutated and the temperature is
where to execute the process in the case of global reduced by certain percentage and the makespan is
scheduling. Global scheduling uses information about found out. This process is iterated until an optimum
the system to allocate processes to multiple processors solution is got. Tabu search(TS) explores the search
and obviously grid scheduling falls into the global space of all feasible solutions by a sequence of moves.
scheduling branch [12]. Global scheduling can be It performs a number of iterations and at each
either static or dynamic. In static scheduling, all the iteration, TS moves to the best solution that is not
forbidden and thus independent of local optima. A* is is not good compared to other algorithms with respect
a search technique based on a µ-ary tree, beginning at to all the performance measures. Makespan is high,
a root node that is a null solution. The nodes in the resource utilization is poor, the cost of using it is also
tree represent partial mapping and these partial high and the profit obtained out of it also is very less.
mappings has one more task mapped than the parent The results for various resource matrices 128 tasks x 4
node. Each parent node generates children and resources, 256 tasks x 8 resources, 512 tasks x 16
becomes inactive. The tree is pruned to make the resources, 1024 tasks x 32 resources shows that MET
execution time traceable. When a node is added, the does not perform as well as other algorithms. This
tree is pruned by deactivating the leaf node with the performance can be seen from the results of the table
largest makespan of the partial solution. This process 7 and table 8, where only one resource that has the
continues until a leaf node representing a complete highest speed is utilized whereas other resources are
mapping is reached [13]. not utilized. OLB performs slightly better compared
to MET in terms of makespan and resource utilization.
3 COMPARISON USING SIMULATION The cost incurred is also less for OLB compared to
MET and profit obtained for OLB is better compared
The scheduling algorithms have been tested using to MET. This is because, in OLB, the machines that
Gridsim, a simulator for the grid environment. are available are considered for allocation of tasks to
Consistent and inconsistent resources as well as the machines. Max min, Min min and MCT
heterogeneity between the tasks and resources are performance with respect to makespan and resource
used for testing the algorithms. Heterogeneity in tasks utilization are more or less similar. GA , Tabu and SA
is variation in size between the tasks and in resources; performs better and time taken by these algorithms is
it is variation in speed between them. Table 1 shows less than 25% compared to all the above algorithms
the comparison of various scheduling algorithms for for various resource matrices and resource utilization
low heterogeneity tasks and low heterogeneity is also more than 50%. The reason for less time taken
consistent resources, table 2 shows the results for low by the algorithms GA, Tabu and SA is all the
heterogeneity tasks and high heterogeneity resources, resources are more or less utilized in an average
table 3 shows the results for high heterogeneity tasks manner. This performance can be seen from the
and low heterogeneity resources and finally table 4 results of tables 2 also. Table 3 shows the results of
shows the results for high heterogeneity tasks and scheduling algorithms using consistent resources for
high heterogeneity resources. In table 1, the processor high heterogeneity in tasks and low heterogeneity in
speed ranges from 0.10MIPS to 0.29MIPS for 1024 x resources. Here, OLB takes more time compared to
32 matrix, 0.1MIPS to 0.224MIPS for 512 x 16 other algorithms and this may be due to high
matrix, 0.1MIPS to 0.25MIPS for 256 x 8 matrix and heterogeneity in size of the tasks. Tables 5 and 6
0.1MIPS to 0.24MIPS for 128 x 4 matrix. Here, the shows the results of scheduling algorithms using
speed of the processors differs from 0.01MIPS to inconsistent resources. Here also, OLB takes more
0.19MIPS for 1024 x 32 matrix, 0.001MIPS to time for execution of all the tasks compared to other
0.15MIPS for 512 x 16 matrix, 0.01MIPS to algorithms. In the case of inconsistent resources, the
0.15MIPS for 256 x 8 matrix and 0.04MIPS to speed of a resource is not constant for all the tasks it
0.14MIPS for 128 x 4 matrix. Various performance executes and it changes randomly between the tasks.
measures such as makespan, resource utilization, cost Tables 7 show the individual resource utilization for
and profit has been used for comparing the consistent resources. It is seen from the table that for
algorithms. low heterogeneity tasks, some of the resources for
Makespan is found out using Max Min, Min Min and MCT algorithms are not
utilized, for MET only single resource that has the
highest speed is utilized, but for GA, Tabu and SA, all
the resources are utilized. As the speed of a resource
varies between the tasks, all the resources are utilized
for n number of jobs and m number of for all the algorithms. Figures 1 to 11 shows the
machines, where C[i] is the completion time of a job pictorial comparison of all the algorithms.
i.From table 1, it is seen that the performance of MET
Scheduling Resource Makespan Resource
Cost (Rs.) Profit (Rs.)
algorithm Matrix (ms) Utilization(%)
MET 214155 9.02 5655 13195
OLB 172288 18.11 3892.59 14957
Max – Min 136604 16.03 5185.15 13664.85
Min –Min 137146 15.72 2072.94 16777.059
128 x 4
MCT 134933 15.86 5187.95 13662.05
GA 43760 68.67 4048.57 14796.63
TABU 44040 68.18 440.40 3963.60
SA 46160 66.86 4138.65 15159.74
MET 422477 4.63 8107.89 30501.11
OLB 298776 10.33 8509.73 30099.27
Max – Min 263171 7.46 9843 28766
Min –Min 263544 7.39 4729.475 33879.525
256 x 8
MCT 263716 7.37 9873.51 28735.49
GA 44240 70.42 8541.34 30056.55
TABU 46120 67.55 461.20 4150.80
SA 46476 67.57 8515.73 30081.96
MET 835115 2.32 16035.18 60322.819
OLB 554619 5.26 17927.92 58430.075
Max – Min 520090 3.86 19469.039 56888.96
Min –Min 521897 3.87 9048.098 67309.902
512 x 16
MCT 521998 3.89 19440.33 56917.67
GA 45160 64.54 17712.32 58615.56
TABU 46890 59.25 984.69 3704.31
SA 48670 60.25 17705.06 58641.44
MET 1574416 1.08 23103.89 130922.1
OLB 1068390 2.51 33685.08 120340.92
Max – Min 1030180 1.63 27030.594 126995.405
Min –Min 1030251 1.63 22333.769 131692.23
1024 x 32
MCT 1030579 1.63 27051.40 126974.598
GA 45509 59.25 33514.54 120427.57
TABU 44130 60.79 1323.90 3089.10
SA 45674 58.85 33426.08 120376.31
Table. 1 Performance comparison of scheduling algorithms using consistent resources for low
heterogeneity tasks and low heterogeneity resources

Scheduling Resource Makespan Resource Profit


Cost (Rs.)
algorithm Matrix (ms) Utilization(%) (Rs.)
MET 162539 3.81 7539.99 11310
OLB 144258 9.57 4840.4 14009.6
Max – Min 134909 4.09 6600.3 12249.7
Min –Min 128 x 4 135194 4.20 7256.41 11593.59
MCT 135119 3.96 6605.9 12244.1
GA 29590 49.08 5005.59 13818.45
TABU 31257 47.91 937.71 3750.84
SA 28841 50.107 5000.05 13826.42
MET 256 x 8 316146 1.82 3860.9 34748.1
OLB 289870 8.40 8352.38 30256.62
Max – Min 263977 2.07 8295.069 30313.93
Min –Min 263636 2.30 15153.394 23455.605
MCT 263886 2.04 8373.04 30235.96
GA 88140 28.73 8500.34 30066.37
TABU 96280 27.30 962.80 3851.20
SA 93440 27.54 8524.95 30059.36
MET 613820 0.84 19089.5 57268.5
OLB 559846 3.87 14871.6889 61486.311
Max – Min 520982 0.96 17830.418 58527.5819
Min –Min 512 x 16 522346 0.86 34838.15 4151.8599
MCT 525856 0.93 17840.945 58517.055
GA 113025 19.09 5041.09 61403.78
TABU 118025 18.35 141.63 4579.37
SA 112025 19.20 14787.43 61472.4
MET 1459203 0.88 46207.79 107818.20
OLB 1051922 1.24 30216.24 123809.758
Max – Min 1279819 0.95 38513.40 115512.599
Min –Min 1024 x 32 1288136 2.48 22330.879 131695.120
MCT 1284228 0.96 38467.85 115558.15
GA 48285 28.18 29490.48 39922.44
TABU 51208 26.72 46.09 4562.63
SA 52107 26.45 29358.95 124628.81
Table 2 Performance comparison of scheduling algorithms using consistent resources for low
heterogeneity tasks and high heterogeneity resources
Scheduling Resource Makespan Resource
Cost (Rs.) Profit (Rs.)
algorithm Matrix (ms) Utilization(%)
MET 3880766 24.07 268352.09 626154.90
OLB 5113812 25.09 204296.55 690210.45
Max – Min 1898531 63.53 224420.3 670086.7
Min –Min 2304391 51.06 92482.245 802024.755
128 x 4
MCT 3262297 42.88 194271.599 700235.4
GA 1938757 64.43 219516.81 674984.92
TABU 2305957 48.24 166028.9 387400.78
SA 3129511 45.39 194109.51 700393.58
MET 7484688 12.02 377329.88 1419479.11
OLB 8664125 15.27 351807.69 1445001.31
Max – Min 2225282 55.67 420123.31 1376685.68
Min –Min 5541282 26.90 151009.95 1645799.05
MCT 256 x 8 2907328 42.99 452693.55 1344115.45
GA 2266728 56.02 418039.74 1378759.28
TABU 6128712 12.44 198570.27 536875.17
SA 2519668 51.27 287936.65 1345575.56
MET 7970094 5.83 389804.73 1466408.26
OLB 9250594 8.11 424681.035 1431531.965
Max – Min 2226797 27.35 431989.34 1424223.652
512 x 16
Min –Min 4418719 15.99 162751.122 1693461.87
MCT 2738828 21.57 427949.85 1428263.15
GA 2354960 27.54 437389.14 1418792.89
TABU 3212140 19.53 `128838.94 484679.8
SA 2333864 25.27 434205.31 1357046.95
MET 12647516 2.87 504720.44 2860082
OLB 14406703 3.98 675408.861 2689394.13
Max – Min 2303109 18.76 711705.40 2653097.598
Min –Min 4978312 9.82 379995.305 2984807.695
1024 x 32
MCT 2537031 15.63 705748.263 2659054.73
GA 2318479 20.46 729954.02 2634764.97
TABU 3332178 15.92 439423.70 1025321.98
SA 2696481 20.27 756135.12 2670361.54
Table 3 Performance comparison of scheduling algorithms using consistent resources for high
heterogeneity tasks and low heterogeneity resources

Scheduling Resource Makespan Resource


Cost (Rs.) Profit (Rs.)
algorithm Matrix (ms) Utilization(%)
MET 695314 20.17 178398.4 267597.6
OLB 152393 10.12 4890.4499 13959.55
Max – Min 332308 71.80 137410.90 308585.1
Min –Min 714637 40.64 133536.659 312459.34
128 x 4
MCT 425172 58.68 131669.50 314326.5
GA 318895 76.325 137384.07 308494.46
TABU 762741 42.34 22882.23 91528.92
SA 694739 44.11 127048.58 318917.42
MET 3942846 11.67 298031.6 2682284.4
OLB 12732685 13.96 746916.669 2233399.33
Max – Min 2347847 31.87 472019.379 2508296.62
Min –Min 4174881 23.81 747866.54 2232449.46
256 x 8
MCT 4071218 22.95 645159.48 2335156.52
GA 2385458 37.35 465104.28 2515170.82
TABU 5692421 24.165 461086.10 41449774.91
SA 9110100 14.44 469720.71 2510525.67
MET 8325021 5.86 1793097 5379291
OLB 8526668 7.05 1498694.768 5673693.232
Max – Min 6181945 8.77 1737850.466 5434537.534
Min –Min 6482455 7.98 1857856.23 5347676.67
512 x 16
MCT 6519620 7.54 1726440.364 5445947.636
GA 6178605 14.40 1649616.19 5522631.89
TABU 7732511 12.58 603909.11 5435181.98
SA 27122721 7.28 2213584.76 4958678.28
MET 4400751 2.39 979892.99 2286417
OLB 3190903 8.51 646476.81 2619833.18
Max – Min 1031158 11.71 686496.57 2579813.42
Min –Min 1397532 10.08 1479460.41 1786849.58
1024 x 32
MCT 1073271 10.19 671482.305 2594827.695
GA 725923 34.538 644928.78 2621080.31
TABU 1690028 19.38 45546.25 140356.83
SA 922670 28.83 619216.77 2627355.22
Table 4 Performance comparison of scheduling algorithms using consistent resources for high
heterogeneity tasks and high heterogeneity resources
Scheduling Resource Makespan Resource Cost (Rs.) Profit (Rs.)
algorithm Matrix (ms) Utilization(%)
MET 147860 3.82 4119.6 14730.4
OLB 188031 14.99 3976.55 14873.45
Max – Min 140235 31.33 4145.34 14704.65
Min –Min 153515 7.68 3877.3 14972.7
128 x 4
MCT 148922 6.48 4049.29 14800.7
GA 10962 94.77 11637.53 18239.73
TABU 11126 89.84 2160.98 3821.93
SA 10446 93.23 11306.73 16083.28
MET 275282 2.14 8640.77 29968.23
OLB 2574657 12.03 8592.38 30016.62
Max – Min 262484 7.80 8784.32 29824.68
Min –Min 263547 2.16 8462.16 30146.83
256 x 8
MCT 262625 1.90 8145.44 30463.56
GA 11342 87.69 30053.6 15053.79
TABU 13917 81.43 2477.78 1635.20
SA 12030 86.47 29564.4 15610.46
MET 523093 0.88 17981.58 58376.41
OLB 840437 5.59 17283.22 59074.77
Max – Min 526516 6.76 17695.15 58662.84
Min –Min 518453 0.99 17901.78 58456.22
512 x 16
MCT 518094 0.88 17576.33 58781.662
GA 12523 86.96 48507.83 34162.31
TABU 13137 85.83 9535.8 3654.64
SA 12487 71.38 48642.87 35811.2
MET 1034484 0.44 33519.42 120506.58
OLB 1165657 2.77 33623.71 120402.28
Max – Min 1030047 3.07 33611.97 120414.02
Min –Min 1030109 0.48 33744.819 120281.181
1024 x 32
MCT 1030266 0.44 33642.17 120383.82
GA 13270 84.66 131978.52 88504.52
TABU 14070 81.28 4099.86 1739.19
SA 14197 83.43 139052.32 96023.78
Table 5 Performance comparison of scheduling algorithms using inconsistent resources for low
heterogeneity tasks

Scheduling Resource Makespan Resource


Cost (Rs.) Profit (Rs.)
algorithm Matrix (ms) Utilization(%)
MET 1081344 23.92 229105.6 665401.4
OLB 3126406 24.56 145151.60 749355.39
Max – Min 567141 57.84 195582.3 698924.7
Min –Min 847547 38.48 225574.69 668932.3
128 x 4
MCT 719656 49.30 219846.45 674660.55
GA 454623 70.54 222395.9 466477.65
TABU 1771476 36.05 648961.13 1147758.68
SA 700102 48.56 221234.97 800757.16
MET 1560797 14.15 300485.24 1496323.76
OLB 256 x 8 3373360 14.66 298103.26 1498705.74
Max – Min 589000 66.17 419947.11 1376861.89
Min –Min 3067328 20.86 389669.33 1407139.66
MCT 743078 38.54 449502.08 1347306.92
GA 755001 49.79 1014982.4 750375.83
TABU 1740652 23.96 169084.57 815115.19
SA 1320677 28.01 2314509.38 176405.65
MET 1747219 6.56 398947.56 1457265.43
OLB 4801890 7.05 459938.89 1396274.10
Max – Min 2127859 13.84 419777.74 1436435.25
Min –Min 1632328 10.31 422842.51 1433370.48
512 x 16
MCT 1036578 10.90 358612.471 1497600.52
GA 531628 36.02 878863.56 547907.17
TABU 1303488 16.22 202318.77 19903.12
SA 773794 31.63 1166248.24 601501.78
MET 2864813 3.47 698004.23 2666798.76
OLB 6951562 3.92 732777.02 2632025.97
Max – Min 6509391 5.66 796115.65 2568687.34
Min –Min 1695610 6.01 742257.79 2622545.20
1024 x 32
MCT 1310531 7.69 637659.57 2727143.42
GA 620573 27.62 1924807.93 1611180.62
TABU 3037900 8.12 1295783.10 636772.85
SA 1603407 13.25 2998679.22 1888930.15
Table 6 Performance comparison of scheduling algorithms using inconsistent resources for high
heterogeneity tasks
ResUtil (l-l Max- Min-
MET OLB MCT GA Tabu SA
heterogeneity) Min Min
R1 0.0 31.40 0.0 0.0 0.0 100 40.08 44.25
R2 0.0 19.54 0.0 0.0 0.0 77.22 51.36 54.24
R3 36.09 11.93 29.66 28.76 28.77 48.72 67.75 68.94
R4 0.0 9.56 34.45 34.11 34.66 64.41 100 100
ResUtil (l-h Max- Min-
MET OLB MCT GA Tabu SA
heterogeneity) Min Min
R1 15.25 3.37 7.82 7.98 7.61 20.51 19.03 21.26
R2 0.0 17.60 0.0 0.0 0.0 100.0 100 22.91
R3 0.0 4.25 8.555 8.82 8.24 22.72 19.88 56.26
R4 0.0 13.06 0.0 0.0 0.0 53.09 52.73 100
ResUtil (h-l Max- Min-
MET OLB MCT GA Tabu SA
heterogeneity) Min Min
R1 0.0 11.25 51.11 40.22 26.72 62.84 10.52 3.66
R2 0.0 30.31 51.68 33.71 98.34 62.57 27.42 36.14
R3 96.29 1.50 99.67 95.89 19.87 100 55.04 41.79
R4 0.0 57.30 51.68 34.43 26.58 32.34 100 100
ResUtil (h-h Max- Min-
MET OLB MCT GA Tabu SA
heterogeneity) Min Min
R1 80.66 3.7 66.25 33.72 33.98 71.05 16.95 16.75
R2 0.0 21.11 88.52 83.33 85.86 100 19.62 41.07
R3 0.0 4.81 66.72 22.15 70.9 64.62 33.59 18.63
R4 0.0 11.24 65.71 23.36 43.97 69.63 100 100
Table 7 Performance comparison of scheduling algorithms using individual resource utilization for consistent
resources for 128 x 4 resource matrixes.
ResUtil (low
Max- Min-
heterogeneity MET OLB MCT GA Tabu SA
Min Min
tasks)
R1 3.02 16.69 32.16 5.80 3.91 89.02 67.88 80.12
R2 3.40 16.01 15.48 4.12 3.20 100 97.27 96.89
R3 3.34 19.26 18.44 3.66 3.86 98.72 100 95.94
R4 5.52 8.00 59.25 17.13 14.95 91.34 94.24 100
ResUtil (high
Max- Min-
heterogeneity MET OLB MCT GA Tabu SA
Min Min
tasks)
R1 19.91 54.95 43.22 19.11 57.97 35.41 5.08 30.11
R2 9.71 8.71 43.03 24.53 20.34 91.46 6.67 100
R3 63.69 28.87 45.92 8.34 91.21 55.30 32.45 52.05
R4 2.35 5.71 99.17 23.94 27.67 100 100 12.10
Table 8 Performance comparison of scheduling algorithms using individual resource utilization for
inconsistent resources for 128 x 4 resource matrixes.
and low heterogeneity resources

Fig. 3 : Comparison of scheduling algorithms using consistent


resources and cost for low heterogeneity tasks and low
heterogeneity resources
Fig. 1 : Comparison of scheduling algorithms using consistent
resources and makespan for low heterogeneity tasks and low
heterogeneity resources

Fig. 4 : Comparison of scheduling algorithms using consistent


resources and profit for low heterogeneity tasks and low
heterogeneity resources

Fig. 2 : Comparison of scheduling algorithms using consistent


resources and resource utilization for low heterogeneity tasks
Fig. 7 : Comparison of scheduling algorithms using consistent
resources and cost for low heterogeneity tasks and high
heterogeneity resources

Fig. 8 : Comparison of scheduling algorithms using consistent


Fig. 5 : Comparison of scheduling algorithms using consistent resources and profit for low heterogeneity tasks and high
resources and makespan for low heterogeneity tasks and high heterogeneity resources
heterogeneity
resources

Fig. 6 : Comparison of scheduling algorithms using consistent Fig. 9 : Comparison of scheduling algorithms using consistent
resources and resource utilization for low heterogeneity tasks resources and makespan for high heterogeneity tasks and low
and high heterogeneity resources heterogeneity resources

70 MET
R 60
e
50 OLB
40
s 30 Max - Min
U 20
10 Min - Min
t 0
l MCT
512,16
128,4
256,8

1024,32

% GA
TABU
Resource Matrix SA
Fig. 10 : Comparison of scheduling algorithms using consistent Using OptorSim, Journal of Grid Computing., 2:
resources and resource utilization for high heterogeneity tasks 57-69, (2004).
and low heterogeneity resources
[4] Min Cai et al: MAAN: A Multi-Attribute
Addressable Network for Grid Information
MET
Services, Journal of Grid Computing., 2: 3-14,
2500000 (2004).
C 2000000 [5] Seung-Hyun Kim et al: Workflow – Based
OLB
Authorization Services in the Grid, Journal of
o 1500000 Grid Computing., 2: 43-55, (2004).
1000000 [6] Anthony Sulistio, Chee Shin Yeo, and Rajkumar
s Max
Buyya:- Visual Modeler for Grid Modeling and
500000
t 0 Min
Simulation (GridSim) Toolkit., ICCS 2003, LNCS
2659, pp. 1123 – 1132, (2003).
[7]
Min -
Carlo Mastroianni, Domenico Talia and Paolo
Min
Trunfio: Metadata for Managing Grid Resources
in Data Mining Applications, Journal of Grid
MCT
Fig. 11: Comparison of scheduling algorithms using consistent Computing., 2: 85-102, (2004).
resources and cost for high heterogeneity tasks and low [8] R.Jimenez-Peris, M.Patino-Martinez, B.Kemme:
heterogeneity resources
Resource Matrix GA
Enterprise Grids: Challenges Ahead, Journal of
Grid Computing 5: 283-294, (2007).
4 CONCLUSION [9] Rajkumar Buyya, and Manzur Murshed:
Scheduling is one of the important challenges in GridSim: a toolkit for the modeling and
grid computing environment as resources are simulation of distributed resource management
geographically distributed over multiple administrative and scheduling for Grid computing, Concurrency
domains. So, there is a need to find a good schedule, and Computation: Practice and Experience.,
automate the scheduling process and to build a flexible 14:1175-1220, (2002).
and scalable scheduling mechanism. This paper has [10] Albert Jones, Luis C. Rabelo: Survey of Job Shop
given a survey on grid scheduling algorithms Scheduling Techniques, Technical Report,
implemented using gridsim. Performance measures NISTIR, National Institute of Standards and
such as makespan, resource utilization, cost and profit Technology., Gaithersburg, (1998).
are used for comparing the algorithms. The limitations [11] T.Casavant and J.Kuhl: A Taxonomy of
of scheduling algorithms has been clearly depicted by scheduling in General –purpose Distributed
individual resource utilization. From the results, it is Computing Systems, IEEE Trans. On Software
seen that the resources are not utilized in a fair manner. Engineerin.,, vol. 14, no. 2, pp. 141 – 154, (Feb
Some resources are not at all utilized, whereas some 1988).
resources are utilized to the maximum. Rescheduling [12] Fangpeng Dong, Selim G. Akl: Scheduling
of the tasks to the resources may result in better Algorithms for Grid Computing: State of the Art
makespan. and Open PrOLBems., Technical Report No.
2006-504, (Jan 2006).
REFERENCES [13] R.Braun et al: A Comparison of Eleven Static
Heuristics for Mapping a Class of Independent
[1] Francine Berman et al: Adaptive Computing on Tasks onto Heterogeneous Distributed
the Grid Using AppLeS, IEEE Trans. On Parallel Computing Systems, J. of Parallel and
and Distributed Systems, vol. 14, no. 4, (April Distributed Computing., vol. 61, no. 6, pp. 810-
2003). 837, (2001).
[2] Ravin Ahuja, G.Gabrani and Asok De: A [14] Vincenzo Di Martino: SubOptimal Scheduling in
Dynamic Priority Scheduler for Advance a Grid using Genetic Algorithms, Parallel and
Reservation in Grid Computing, International Nature Inspired Computational Paradigms and
Journal of Soft Computing., 4(2): 60-67, (2009). Applications., Elsevier Science Publishers pp:
[3] D.G. Cameron et al: Analysis of Scheduling and 553-565, (2004)
Replica Optimisation Strategies for Data Grids

Вам также может понравиться