International Journal of Application or Innovation in Engineering& Management (IJAIEM)

Web Site: www.ijaiem.org Email: [email protected]

Volume 3, Issue 7, July 2014 ISSN 2319 - 4847

Volume 3, Issue 7, July 2014 Page 269



ABSTRACT
Cloud Computing is in its naive stage as a lot of research is going to schedule the problems in best possible method. With
scheduling the cost effectiveness is also looking up for the effectiveness of the system. As a result, in this paper, we have proposed
a machine learning based methods to make the system proactive. This machine learning system may play an important role in
scheduling the requests in an efficient manner. Simulation results show the effectiveness of the work.
Keywords: Proactive, machine learning, scheduling
1. INTRODUCTION
Cloud Computing provides the services to the consumer on the internet. In the course of cloud computing we can access
anything that we want from ubiquitously to any computer without perturbing anything likes about their management, cost
and storage, Infrastructure and so on. Cloud computing is delivering the services to the consumers through the Internet
connection. As a replacement for maintaining data or updating applications on your personal hard drive according to your
requirement and necessities, you can utilize this service via the Internet and access the data from any location and also
use its applications whenever required. The cloud provides the services consent to individuals, enterprise and businesses
utilize the infrastructure, storage, resources, hardware and software that are managed by providers from different
geographical locations. Examples of cloud services like social sites and online file storage, Email clients like Yahoo,
Gmail, Wikipedia, YouTube, Skype or Bit Torrent and many more. Cloud computing provides resources from a pool like
servers, storage, services, processing power, network bandwidth and particular many user applications [1]. The cloud
computing provides access to application, information, data and resources at any time from any geographical location,
merely necessitate a connection of internet. This is implemented with virtualization technology where one or more servers
can be configured and partitioned into many self-sufficient virtual servers and perform functionality independently and
appearing to consumers as a single physical device and consumers are rented virtual machines (VMs) dynamically as
they require. Cloud Computing helps production organizations, organizations, company, government institutions,
academic organizations in cutting down operational costs. The following definition of cloud computing has been
developed by the U.S. National Institute of Standards and Technology (NIST) [2] Cloud computing is a model for
enabling convenient, on-demand network access to a shared pool of configurable computing resources (e.g., networks,
servers, storage, applications, and services) that can be rapidly provisioned and released with minimal management effort
or service provider interaction. This cloud model promotes availability and is composed of five essential characteristics,
three service models, and four deployment modelsCloud computing can be distinct as a distributed and parallel system
with a set of virtualized and dependable computers that are provisioned animatedly [3] and easy to get as one or many
resources accessed by conciliation among the provider of cloud and consumers based on service-level agreements. Various
examples of the Cloud computing infrastructures/platforms are Microsoft Azure, Aneka, Amazon EC2 and Google App
Engine and so on. National Institute of Standards and Technology (NIST) have explained cloud computing uniqueness
that illustrate and distinguish the services of cloud from conformist approaches of cloud computing. The characteristics of
cloud computing, consist of on-demand self service, measured service broad network access, rapid elasticity and resource
pooling.On-demand self service means that customers and typically organizations can request the service provider for
resources and also manage resources for computing. Broad network access, provides services to be accessible anywhere
and anytime over the public networks or private networks.Pooled resources mean that consumers access number of
resources from a cloud where number of resources are available, generally in distant data center. Pay per use measured
service Cloud provider has different policies to produce a bill. According to the use of services are measured. Consumers
are billed as per usage basis. Services can be measured as larger or smaller depend upon usage of consumer [4]. If we
move towards scheduling the efficient strategies are still under research. A lot of strategies we have discussed in the next
section. The sequence of the paper is that the next section 2 contains the related work, after related work we will go to our
proposed framework under Section 3 and later we will see the experimental setup and results as section 4, further with a
conclusion and future work as final section i.e. section 5
Machine Learning based decision making by
brokers in cloud computing
Kiran Bala
1
and Sahil Vashist
2


1
CEC Landran(Pb)

2
CEC, Landran(Pb)
International Journal of Application or Innovation in Engineering& Management (IJAIEM)
Web Site: www.ijaiem.org Email: [email protected]
Volume 3, Issue 7, July 2014 ISSN 2319 - 4847

Volume 3, Issue 7, July 2014 Page 270

2. RELATED WORK
Talebi [5] proposed a consistent scheduling algorithm in a cloud computing environment. In this explains the PPDD and
RSDC algorithm.In this scheduling algorithm the job is alienated in many sub jobs. Now the request time and
acknowledge time of jobs is calculated independently in order to establish stability in the jobs. For each job the scheduling
is completed by calculating the acknowledge time and request time from shared job. So the system efficiency is improved.
Entezari-Maleki [6] proposed a RASA algorithm for scheduling the task. It is a combination of traditional two scheduling
algorithms; Min-min and Max-min. RASA takes the benefits from both Min-min and Max-min algorithms and also help
to remove their disadvantages. To attain this, Firstly RASA calculates the time of completion of the tasks according to
available grid resources. After that, apply the both algorithms. For small task execution RASA uses the Min-min
algorithm and for the tasks that is large, apply the Max-min algorithm for execution the large tasks so that avoid the
delay of large tasks. RASA maintain the concurrency in the executions of both types of tasks large or small tasks. While
the limit of each task, tasks arriving rate, communication cost, task execution cost of each of the resources that are not
measured. The results defined that RASA is performing better in large level than existing scheduling algorithms.
Guruprasad [7] has proposed a scheduling algorithm depend upon on Quality of service and priority system. In this
scheduling algorithm serving the consumer requests according to priority that is assigned to each admitted queue.
Calculating the tolerable delay and service cost of the requests and then added into the admission queue. This policy
achieves high service completion time rate and guaranteed QoS with the planned architecture of cloud. Also improve the
throughput and service response time of the request when it is in queue. This policy gives the uppermost preference for
the users of extremely paid, in general the cost of other servicing in the also increases. Al-rahamawy [8] has proposed an
algorithm depend upon the RASA algorithm. In this Max-min algorithm improved the execution time of the expected
task based on selective basis instead of complete time of request. Petri nets describe the simultaneous performance of
dispersed systems. Max-min algorithm gaining schedules better than original Max-min and RASA algorithm with similar
lower makespan. Bansiwal [9] has proposed a scheduling differentiated algorithm in a cloud environment. The
scheduling algorithm is used for reducing the load and improve the utilization of resources. The scheduling algorithms
are static and dynamic. This algorithm is applied on static load balancing and uses non-preemptive priority queuing
model for behavior estimation. In this algorithm one application is created. This application is a web application to do a
number of actions. These actions are mainly file storing, uploading and downloading. For storing a file, if in one folder
there is not sufficient space, then find out the folder which has less space and stored the file. By clicking the button,
downloading any of file, text, audio and video. Uploading can also be done by clicking the button of uploading. For this
there is requiring of the efficient scheduling algorithm. The QoS requirements and maximum profits provide by cloud
providers are gained by users with this algorithm. Samir Youcef [10] has proposed complementary three bi-criteria
approaches. These are cost-based, time-based and cost-time based. In cloud uses the two workflows, business and
scientific workflow. In this algorithm use scientific workflow that explains with the help of the directed acyclic graph. In
DAG, the nodes define the tasks and edges define the dependencies. The tasks mainly have NP-complete problems mean
the problem of allocation and scheduling of tasks. So it is beneficial to use a heuristic algorithm than deterministic
algorithm. The three approaches enhancing the overall the cost and execution time. This approach provides the
consumers to choose the desired schedule and flexibility to estimate their preference for execution of tasks. HaijunCao
[11] has proposed two major algorithms, dynamic-tuning algorithm based and evaluation-scheduling Algorithm. These
algorithms are implemented in Hadoop. Hadoop is used by many companies like Microsoft, Facebook, etc. In the
dynamic-tuning algorithm, the task is divided into map and reduce phase and map phase further divided into two phase
map and sort. Reduce phase divided into three phases shuffle, sort and reduce. These phases record the values at every
phase. Based on historical values of the map and reduce phase it map and reduce the tasks, estimate the progress of a task
properly and also used for calculation of time for execution of tasks. Some time tasks may not get the resources due to
load or other reasons and launch straggler task. In the evaluation-scheduling algorithm, before initiation of a straggler
task on the node, it finds the idle node and assigns this task to the idle node. Late scheduler is used to find out the idle
node and when, choose the task, then uses the largest left over time approach. So it reduces the resource wastage by
finding the free slot. With this algorithm reduce the resource wastage and calculate the execution time. in-Guang Sang
[12] has proposed algorithm UFPS (Urgency First Parallel Strategy) uses the local and cloud preferred strategy. In local
strategy when the task arrives, then the task scheduler checks the resources at local whether resources are free or not. If
resources are free at local then allocate the task at local server, otherwise check resources on cloud servers. My resources
are free at the cloud then allocated the tasks at cloud. If resources are not found at both then put it on a tail of the queue.
In cloud preferred strategy is opposite to local. Firstly check the cloud server and then locally. If idle resources are not
found, then putting it in the queue. In UFPS algorithm uses both strategies. UFPA define the emergency of tasks.
According to emergency the tasks put in front of the queue. The resources are pretreated and tasks are divided into a
number of tasks after the tasks put into the buffer. Then these tasks are divided into many classes. The classes are divided
into special kind require large computation and less computation. According to this special task put in one group and
allocate to a special server, tasks that require larger computation put into another group and allocate for the server that
have the ability of larger computation and similar smaller computation tasks put in one group and allocate to a smaller
International Journal of Application or Innovation in Engineering& Management (IJAIEM)
Web Site: www.ijaiem.org Email: [email protected]
Volume 3, Issue 7, July 2014 ISSN 2319 - 4847

Volume 3, Issue 7, July 2014 Page 271

server whose ability is smaller than others. After that, these tasks scheduled both at local and cloud cluster
simultaneously. If the tasks find idle resources at local then allocate the resources at local cluster and will quit scheduling
at cloud cluster. Similarly, if found in the cloud cluster, then will renounce scheduling at local cluster. If not, find idle
resources at both then put it in a tail of the queue. UFPS enhance the efficiency, reduce time and economic increases.
3. PROPOSED FRAMEWORK
In the proposed framework we have used learning techniques to acquire the proactive behavior and we try to find the type
of analysis to the systems which helps in finding the right mapping system (request-to-resource) as shown in figure 1. In
this framework we have adopted three different approaches for decision making (proactive) also shown in:
3.1 One step or single learning
This system is used when a new type of request is generated from users. In the single learning concept, it first randomly
puts the requests to any of the virtual machine for processing and notices the performance and output. In the next time
when the same type of requests has generated, based on the performance of older one it is decided whether to go for the
same or another.
3.2 Use of predefined rules
In order to refine the behavior of our learning system, a set of rules for the decision on the certain requests can be
predefined and our algorithm has the capability to find out the can be used to figure out more specifically in comparable
cases which may go against the rules. In this framework design certain type requests like priority requests, compute
intensive requests, etc. can be granted by special privileges and govern under special rules.
3.3 Use of Multi learning
By feeding multi-learning into the cloud system the broker got the capability to find the intersection of all the single
learning results. In multi-learning the broker tries to learn from all the results from that was previously mishandled
during single learning. In this learning system the accuracy of the right decision is of high probability.


Figure 1: A descriptive view for proactive behavior in Cloud Computing

In our approach
Traditional algorithm for scheduling, all the information regarding the job is already known and static, but now when
faced with dynamic behavior then need to a proactive approach. Proactive approach studies the dynamic behavior of jobs
and makes solutions that can attain a performance that is acceptable at execution time. Proactive behavior involves acting
in advance of a future situation, rather than just reacting. In this consumer behavior is studied before, according to
behavior of consumer algorithm create and execute. The consumer behavior is almost same. Consumer request for similar
International Journal of Application or Innovation in Engineering& Management (IJAIEM)
Web Site: www.ijaiem.org Email: [email protected]
Volume 3, Issue 7, July 2014 ISSN 2319 - 4847

Volume 3, Issue 7, July 2014 Page 272

application and require similar resources. So that proactive algorithm studies the behavior and attains the best
performance. In this, there is a one datacenter, host, virtual machine created. When cloudlets given to a broker, broker
schedule the request according to the algorithm. In the algorithm broker first get the information about virtual machines-
virtual machine status, memory, storage etc. After that check the size of cloudlets. The broker gets the list of last
processed cloudlets from Call_recomendder_agent and Call_recomendder_agent store the information of all last
processed cloudlets and send information to the broker. The broker receives the list and then schedules the cloudless by
schedule_jobrequest. After that, all cloudlets executed and produced the results.
4. EXPERIMENTAL SETUP AND RESULTS
In this section we use cloudsim as our simulating tool, the experimental setup we have used 30 virtual machines (VM1 to
VM30) under 1 data center (DC1). Configuration of virtual machine as shown in Table 1.

Further 20 user bases have been used to launch requests to the virtual machines. Since to cause the heterogeneity, we have
taken the different priority levels of random selection clauses. These results are compared with the [13] as shown in
figure 2 as per their success rate of scheduling. As shown in figure for initial requests PISA is working well as during this
time PSA is learning the requests and putting them into database for future uses.
0
0.2
0.4
0.6
0.8
1
10 12 14 16 18 20 22 24 26 28 30
PISA
PSA
VMS
(
%
)
S
u
c
c
e
s
s

R
a
t
e

Figure 2: Success Rate PISA vs PSA
In the Figure 2 the comparison of both algorithms, from starting PISA algorithm performs better because this algorithm
allocates the cloudlets according to priority and executes the cloudlets. PSA algorithm schedules the cloudlets according
the randomly generated priority and use machine learning algorithm. By this at start point it allocates the cloudlets
randomly to the virtual machine and then learns which virtual machine performs better. So after that allocate the
cloudlets according to performance of virtual machines. So at starting PISA perform better, but the PSA success rate is
higher than PISA.

0
10
20
30
40
For 10
Requests
For 20
Requests
For 30
Requests
For 40
Requests
PSA
PISA
%

S
L
A

v
i
o
l
a
t
i
o
n
s

Figure 3: Percentage comparisons of SLA Violations
In the Figure 3 we can observe that there is an improvement of 20% when comparing both techniques and if we look at
SLA violations comparisons for a sequence of runs we can see that the random priority does not make any impact on the
working of PSA over PISA.
5. FUTURE WORK AND COMPARISONS
This paper has explored the machine learning technique for proactive behavior. In this work three types were used to
exploit the past performance of the system. This designed is used to explore the predefined rules impact on the behavior,
but in order to move further as future work we shall work on the higher the uncertainties of the users which need to be
investigated.
International Journal of Application or Innovation in Engineering& Management (IJAIEM)
Web Site: www.ijaiem.org Email: [email protected]
Volume 3, Issue 7, July 2014 ISSN 2319 - 4847

Volume 3, Issue 7, July 2014 Page 273


REFERENCES
[1.] Mohamed Abu Sharkh, Manar Jammal, Abdallah Shami, and Abdelkader Ouda Resource Allocation in a Network-
Based Cloud Computing Environment: Design Challenges in IEEE Communications Magazine November 2013
[2.] 2.http://thecloudtutorial.com/related.html
[3.] Linan Zhu1, Qingshui Li2 and Lingna He3 Study on Cloud Computing Resource Scheduling Strategy Based on the
Ant Colony Optimization Algorithm in IJCSI International Journal of Computer Science Issues, Vol. 9, Issue 5, No
2, September 2012 ISSN (Online): 1694-0814
[4.] 4.Ronnie D. Caytiles and Byungjoo Park A Study on Analysis and Implementation of a Cloud Computing
Framework for Multimedia Convergence Services in International Journal of Software Engineering and Its
Applications Vol. 7, No. 2, March, 2013.
[5.] Arash Ghorbannia Delavar,Mahdi Javanmard , Mehrdad Barzegar Shabestari and Marjan Khosravi Talebi RSDC
(RELIABLE SCHEDULING DISTRIBUTED IN CLOUD COMPUTING) in International Journal of Computer
Science, Engineering and Applications (IJCSEA) Vol.2, No.3, June 2012
[6.] Saeed Parsa and Reza Entezari-Maleki, RASA: A New Task Scheduling Algorithm in Grid Environment in World
Applied Sciences Journal 7 (Special Issue of Computer & IT): 152-160, 2009.Berry M. W., Dumais S. T., OBrien G.
W. Using linear algebra for intelligent information retrieval, SIAM Review, 1995, 37, pp. 573-595.
[7.] Dr. M. Dakshayini, Dr. H. S. Guruprasad An Optimal Model for Priority based Service Scheduling Policy for Cloud
Computing Environment International Journal of Computer Applications (0975 8887) Volume 32 No.9, October
2011Shamsollah Ghanbari, Mohamed Othman A Priority based Job Scheduling Algorithm in Cloud Computing
International Conference on Advances Science and Contemporary Engineering 2012 (ICASCE 2012)
[8.] El-Sayed T. El-kenawy, Ali Ibraheem El-Desoky, Mohamed F. Al-rahamawy Extended Max-Min Scheduling Using
Petri Net and Load Balancing International Journal of Soft Computing and Engineering (IJSCE) ISSN: 2231-2307,
Volume-2, Issue-4, September 2012
[9.] Shalmali Ambike, Dipti Bhansali, Jaee Kshirsagar, Juhi Bansiwal An Optimistic Differentiated Job Scheduling
System for Cloud Computing International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-
9622 www.ijera.com Vol. 2, Issue
[10.] 2, Mar-Apr 2012, pp.1212-1214
[11.] Kahina Bessai, Samir Youcef, Ammar Oulamara, Claude Godart and Selmin Nurcan Bi-criteria workflow tasks
allocation and scheduling in Cloud computing environments , Fifth International Conference on Cloud Computing,
2012, IEEE.
[12.] Xu Zhao, Xiaoshe Dong, Haijun Cao, Yuanquan Fan, Huo Zhu A parameter dynamic-tuning scheduling
algorithm based on history in heterogeneous environments Seventh ChinaGrid Annual Conference, 2012, IEEE.
[13.] Zheng-Wu Yuan ,Xin-Guang Sang A Study on Resource Scheduling Strategy in the
[14.] Enterprise Service Cloud International Conference on Systems and Informatics (ICSAI 2012), 2012
[15.] 13.Hu Wu, Zhuo Tang, Renfa Li A Priority Constrained Scheduling Strategy of Multiple Workflows for Cloud
Computing in School of Computer and Communication of Hunan University, Embedded Systems & Networking
Laboratory, China, ICACT2012.