Professional Documents
Culture Documents
Abstract—In recent years, cloud computing has developed study gradually. In cloud computing, users have their own
rapidly under the vigorous promotion of industry and preferences for resources. Some users need real-time
academia. With the expansion of cloud computing, users’ assurance, some need low cost, and the others need strong
special needs for cloud resources have gradually improved. As stability. Therefore, we need different QoS parameters to
a business model, cloud computing must pay more attention to
user demands for services and provide users with high-quality
measure users’ satisfaction according to their different needs.
services. As one of the key technologies in cloud computing, At present, task scheduling algorithm is more concerned
task scheduling is mainly responsible for assigning user tasks about the efficiency rather than the fairness of resource
to the appropriate resources. However, the existing scheduling allocation, often resulting in the excessive use of resources
algorithms do not take full account of users’ different needs. In with strong computing capability [1].
this paper, we consider multidimensional QoS requirements, In this paper, we propose a task scheduling algorithm
and introduce Berger model to judge the fairness of the based on QoS-DPSO. Our main contributions can be
resource allocation results. We also improve the Particle summarized as follows:
Swarm Optimization(PSO) algorithm by adjusting its ЬWe consider multidimensional QoS requirements, and
parameters dynamically and making the position coding introduce the Berger’ s model to judge the fairness of the
discrete. Then, we propose a task scheduling algorithm based system’s resource allocation results.
on QoS-DPSO. The simulation results show that this algorithm
ˊ
can effectively carry out user tasks and reflect more fairnessˊ
Ь We improve the PSO algorithm by adjusting its
parameters dynamically and making the position coding
Keywords-cloud computing; task scheduling; Particle Swarm discrete.
Optimization(PSO);Berger model The rest of this paper is organized as follows: Section 2
presents related work. In section 3, we introduce the problem
and describe our model for cloud task scheduling. In section
I. INTRODUCTION 4, we introduce standard PSO algorithm. Our task scheduling
Cloud computing is of most concern in today's IT algorithm based on QoS-DPSO is presented in section 5.
industry technology. It is the development and extension of Section 6 shows the simulation results and performance
distributed computing, grid computing and parallel analysis. Finally, section 7 gives conclusions and an outlook
computing. Its application form is the commercial realization on possible continuations of our work.
of above technologies, allowing users to access on-demand
II. RELATED WORK
computing power, storage capacity and information service.
In a cloud environment, the task number is huge and Task scheduling of cloud computing refers to dispatch
resource capability is various, so how to schedule tasks the computing tasks to different resources according to
efficiently has become the focus of cloud computing certain rules. This is an NP-complete problem in the general
research.Currently, research on cloud task scheduling is not form [2]. So, past work have proposed many heuristics based
much, which is mainly divided into traditional and intelligent approaches to solve this problem.
scheduling algorithm. However, current research in this area Since task scheduling is a NP-Complete problem,
has just started, and there is not a task scheduling algorithm Genetic Algorithm (GA) has been used for scheduling jobs
that can be applied to all areas and get the best scheduling [3]. However, GA may not be the best approach. Salman et
results at the same time. Each scheduling algorithm has its al. [4] have shown that the performance of PSO algorithm is
own applications and limitations. Therefore, doing some faster than GA in solving static task assignment problem for
useful exploration in this area is of practical significance. homogeneous distributed computing systems based on their
Cloud computing is dynamic, distributed, heterogeneous test cases. Lei et al. [5] have shown that the PSO algorithm is
in general and the resource management strategies of each able to get better schedule than GA based on their simulated
organization are different, so the quality of service (QoS) experiments for Grid computing. In addition, the results
mechanisms of cloud computing has become the focus of the presented by Tasgetiren et al. [6] have provided evidence
7 90
For different user need, we need to establish different
+RVW evaluation criteria to measure their satisfaction.
90
8VHU1 Suppose that T {T1 , T2 ,, Tn } represents the set of
7Q 90Q
independent tasks, and n represents the total number of tasks.
We use a four-tuple to represent the task parameters, namely
Ti {TID , Tclasstype , Tlength , Texp t } .
'DWD&HQWHU
906FKHGXOHU 903URYLGLRQHU
%URNHU
&,6 Where TID represents the task ID. Tclasstype represents the
task category. According to the QoS classification, Tclasstype is
Figure 1. Task scheduling model of cloud computing set to: When classtype = 1, it represents the task preferring
execution time. When classtype = 2, it represents the task
1056
preferring bandwidth. When classtype = 3, it represents the both show unfairness. In cloud environment, it can provide
task preferring running cost. Tlength represents the task length, better quality of service when the actual resource allocation
referring to the computation amount required by the task. is higher than the expected resource allocation, so too much
injustice is allowed. This function is used to judge the
Texpt represents the task expectation. According to the QoS
fairness of the resource allocation results. Based on QoS
classification, Texpt is set to: When classtype =1, it represents classification, we describe the fairness evaluation function
expected execution time. When classtype =2, it represents of each type of task as follows:
expected bandwidth. When classtype =3, it represents a) Execution time expectation
expected running cost. J i ln(tVM / texp t )
D. Scheduling target Where texpt indicates the expected execution time of
1) Fairness evaluation function in Berger model users, which is set by users according to their expectation.
In Berger model, the justice of distribution can be tVM shows the actual execution time, which is calculated as:
judged through comparison between local structure and
Tlength
reference structure. Moreover, it is worth to mention that the tVM
definition of distributive justice evaluation function in VM mips VM pesNumber
Berger model. I.e., the execution time of subtask is equal to the ratio of
JEF ln( AR / JR) its computation amount and the operation speed of the
Where denotes constant. AR(Actual Reward)represents virtual machine assigned to it.
the actual distribution, namely social distribution that actors b) Bandwidth expectation
actually obtained. JR(Just Reward)represents justice J i ln(bwVM / bwexp t )
distribution, namely expected distribution generated by
actors. The intuitive meaning of this function is that when Where bwexpt represents the expected bandwidth of users.
the actual and expected distribution of actors are equal, bwVM indicates the task actual allocated bandwidth, which is
namely, when the function value is 0, fairness set up. When
equal to the bandwidth of the virtual machine assigned to it.
the function value is greater than 0, namely, when the actual
distribution is greater than the expected distribution, the c) Cost expectation
result is excessive distribution , i.e., nonjustice. When the J i ln(cos tVM / cos texp t )
function value is less than 0, namely, when the actual Where cos texp t indicates the task expected running cost.
distribution is less than the expected distribution, the result
is too little distribution , i.e., nonjustice. cos tVM indicates the task actual running cost, which is
2) Fairness evaluation function based on QoS calculated as:
In cloud computing, the fairness of resource uses Tlength
reflected in those areas: cloud can provide reasonable cos tVM VM price
VM mips VM pesNumber
available resource for user tasks according to the
characteristics and the preferences of user tasks. It also I.e., the task running cost is equal to the product of its
enables different users to get what they want in satisfactory. execution time and the virtual machine’s cost per unit of
Next, we will define the function of task justice and time.
system justice under the cloud environment. Definition2. System Justice( J ). Suppose that
Definition1. Task Justice( J i ). A task receives a fair T {T1 , T2 ,, Tn } represents the set of system tasks, and
execution, when the actual allocation resources for it are J {J1 , J 2 ,, J n } shows the set of the corresponding
close to the expected resources, to the maximum extent. Set fairness evaluation function, then the fairness evaluation
the fairness evaluation function of task Ti as: function of the system is defined as:
n
J i ln( ARi / ERi ) J | Ji |
Where is the balancing constant, and
(0,1] . ARi is
i 1
1057
IV. STANDARD PSO INTRODUCTION on the number of sub-tasks, and a particle corresponds to a
Particle swarm optimization is a bionic intelligent task scheduling strategy in fact.
Suppose that n = 10, m = 3,then the particle
optimization algorithm that mimics the behavior of animal
(3,1,2,3,3,1,2,2,1,3) could be a feasible scheduling strategy.
swarm. It is proposed by the American psychologist
The particle encoding is shown in Table 1, where the task-
Kennedy and electrical engineer Eberhart in 1995. PSO is a
resource pair (1,2) represents the sub-task 1 is assigned to the
stochastic optimization algorithm based on swarm
resource 2 for execution.
intelligence, using the information sharing mechanism of
bird group individual to obtain the optimal solution. PSO TABLE I. PARTICLE CODING EXAMPLE
algorithm can remember personal best information and
global best information through collaboration between Subtask Number Virtual Machine Number
1 3
individuals. Each particle in the population represents a 2 1
possible solution of the problem to be optimized[14]. 3 2
Set a swarm which is composed of m particles in a D- 4 3
dimensional search space. The position of particle i is 5 3
6 1
xi {xi1 , xi 2 ,, xiD } , and its velocity is vi {vi1 , vi 2 ,, viD } .
7 2
Each particle’s best position is pi { pi1 , pi 2 ,, piD } , and the 8 2
9 1
best position of all particles is pg ( pg1 , pg 2 ,, pgD ) . For
10 3
the particle i in the k-th iteration, its d-dimensional We obtain the resource’s subtasks distribution by
velocity vidk 1 and position xidk 1 is updated by the following decoding the particle. The particle decoding example is
equation: shown in Table 2, where the subtasks performed by resource
1 are {3,8,10}. Subtasks performed by resource 2
vidk 1 vidk c1r1 ( pidk xidk ) c2 r2 ( pgd
k
xidk )
are{1,4,7,9}. Subtasks performed by resource 3 are {2,5,6}.
k 1 k 1
x id x vk
id id TABLE II. PARTICLES DECODING EXAMPLE
Where is the inertia weight, using to balance the Virtual Machine Number Subtask Number
exploration and development capabilities of PSO. Larger 1 2,6,9
makes the particle have a better exploration ability, and 2 3,7,
smaller makes the particle have a better development 3 1,4,5,10
1058
max min
i max i Initialization
I max parameters
E. Updating and legalization processing of vector and Random initial
position population
The vector is updated according to equation . The
position is updated as follows: Firstly, by formula, we Calculate the fitness value
obtain a new coding sequence containing illegal coding.
Then, we make the illegal coding legalized. The main
processing methods include taking absolute value, taking Determine Pi and Pg
integer upwardly, taking remainder, etc.. Specific methods
are described as follows :
k=k+1 Update the velocity and
xidk 1
xidk 1
(0, N ] position
xidk 1 = 1 xidk 1 0
mod( xidk 1 , N ) x k 1
id ! (0, N ]
No Whether the
maximum
Where, xidk 1 represents taking the absolute value of
Iterations
illegal coding and taking the integer upwardly then. Yes
mod( xidk 1 , N ) represents x k 1 -to-N modulo operation. By
id Output for optimal
this legalization processing method, we can obtain legal and results
valid position vectors.
F. QoS-DPSO algorithm flow
Figure 2. QoS-PSO algorithm flowchart
Specific steps of QoS-DPSO algorithm are described as
follows:
Step1 Initialize parameters: Set the iteration’s maximum VI. SIMULATION AND EVALUATION
number I max , the population size NP, inertia weight
In this paper, we use the platform CloudSim for
range min , max , learning factor c1 , c2 and other simulation. In order to achieve our own task scheduling
parameters. policy, we need to extend DatacenterBroker class in
Step2 Initialize population randomly: Set the initial Cloudsim platform, and overload its bindCloudletTbVM ()
positions and velocities of the population randomly. method[16].
Step3 Calculate the fitness value of all the particles A. Simulation environment and parameter settings
according to equation (8).
Step4 For each particle xi , we compare the fitness 1) Parameter settings of PSO algorithm
Parameters of PSO algorithm is set as follows:
function value of it and its best position pi ever experienced.
If the value of xi is smaller, xi will be recorded as the best TABLE III. ALGORITHM’S MAIN PARAMETERS
1059
TABLE IV. VIRTUAL MACHINE PARAMETERS The comparison of the task J value is shown in Figure 4 .
Variable Name Variable Value Or Range The smaller the absolute value is, the better fairness the task
Virtual machine ID {0,1,2,3} receives. Obviously, the resource allocation of QoS-DPSO
CPU number (pesNumber) {3,2,2,1} g
algorithm p
can meet the user’s expectations better.
Computing speed(mips) {300,400,200,350}
Bandwidth(bw) {1200,2000,1000,1500}
Cost per unit of time(price) {2,1.8,1.5,1.1}
1060
REFERENCES
[1] M.A. Arfeen, K. Pawlikowski, A. Willig. A Framework for Resource
Allocation Strategies in Cloud Computing Environment. Computer
Software and Applications Conference Workshops ( COMPSACW ) ,
IEEE 35th Annual, 2011, pp. 261 -266.
[2] J. D. Ullman. Np-complete scheduling problems. J. Comput.Syst. Sci.,
10(3), 1975.
[3] J. Yu, R. Buyya, and K. Ramamohanarao. Workflow Scheduling
Algorithms for Grid Computing, volume 146, Springer Heidelberg,
2008, pp. 173–214.
Figure 7. Experimental comparison of the second type of task [4] A. Salman. Particle swarm optimization for task assignment problem.
Microprocessors and Microsystems, 26(8):363– 371, November 2002.
Task 8 ~ 11 are the third type of tasks, preferring less [5] L. Zhang, Y. Chen, R. Sun, S. Jing, and B. Yang. A task scheduling
running cost. The comparison of the third type task’s algorithm based on pso for grid computing. International Journal of
running cost is shown in Figure 8. We see that the allocation Computational Intelligence Research, 4(1), 2008.
results of QoS-DPSO algorithm can better meet the user [6] M. F. Tasgetiren, Y.-C. Liang, M. Sevkli, and G. Gencyilmaz. A
particle swarm optimization algorithm for makespan and total
preference and the system’s fairness. flowtime minimization in the permutation flowshop sequencing
problem. European Journal of Operational Research, 177(3):1930–
1947, March 2007.
[7] Umale J S, Chaudhari M P A. Survey on Job Scheduling Algorithms
of Cloud Computing[J], 2014.
[8] Sutherland IE. A futures market in computer time. Commun ACM
1968, 11(6):449–51.
[9] Ferguson D, Yemini Y, Nikolaou C. Microeconomic algorithms for
load balancing in distributed computer systems. In: Proceedings of
the eighth international conference on distributed systems. San Jose:
IEEE Press, 1988. p. 491–9.
[10] Regev O, Nisan N. The popcorn market – An online market for
computational resources. In: Proceedings of the first international
Figure 8. Experimental comparison of the third type of task
conference on information and computation economies. Charleston:
ACM Press, 1998.
Overall, by comparing the two algorithms’ simulation [11] Buyya R. Economic-Based distributed resource management and
results, we see that the algorithm proposed in this paper can scheduling for grid computing, Ph.D. Thesis. Melbourne: Monash
perform user tasks effectively and show better fairness. University, 2002.
[12] Rajkumar Buyya, Chee Shin Yeo, Srikumar Venugopal. Market-
VII. SUMMARY oriented cloud computing: vision, hype, and reality for delivering IT
services as computing utilities. In: Proceedings of the tenth IEEE
In this paper, for the cloud task scheduling problem, we international conference on high performance computing and
consider multidimensional QoS requirements, and introduce communications (HPCC 2008), Dalian, China. Los Alamitos, CA,
Berger model to judge the fairness of the system resource USA: IEEE CS Press, September 25–27, 2008.
allocation results. We also improve PSO by adjusting the [13] Baomin Xu, Chunyan Zhao, Enzhao Hu, Bin Hu, Job scheduling
inertia weight and making the position coding discrete. algorithm based on Berger model in cloud environment, Advances in
Engineering Software 42, PP. 419–425,2011.
Finally, we propose a task scheduling algorithm based on
[14] D. Bratton, Kennedy J.Defining a standard for particle swarm
QoS-DPSO. The results of the simulation indicate that our optimization//Proceedings of the IEEE Swarm Intelligence
algorithm can effectively solve the problem of task Symposium, Honolulu, HI, 2007, pp. 120-127.
scheduling in cloud environment and improve the fairness of [15] A. Ratnaweera 㸪 S.K. Halgamuge 㸪 H.C. Watson. Self-organizing
the system allocation results. hierarchical particle swarm optimizer with time-varying acceleration
coefficients[J]. Evolutionary Computation, IEEE Transactions on,
In the current approach, response time and transmission 2004, 8(3): 240-255.
costs are not considered. In our future work, we would like [16] R.N. Calheiros, R. Ranjan, A. Beloglazov, et al. CloudSim: a toolkit
to improve the parameters of PSO such as the learning factor, for modeling and simulation of cloud computing environments and
and consider more scheduling targets such as reliability and evaluation of resource provisioning algorithms[J]. Software: Practice
load balancing to improve the performance of the algorithm. and Experience, 2011, 41(1): 23-50.
This will make our approach more practical and effective.
ACKNOWLEDGMENT
This work was supported by the National Science
Foundation of China [Grant No. 61272432, 61370092 and
61472033], Hubei Provincial Department of Education
Outstanding Youth Scientific Innovation Team Support
Foundation (T201410) and Fundamental Research Funds for
the Central Universities [TW201502].
1061