Pre-allocation Strategies of Computational Resources in Cloud Computing using Adaptive Resonance Theory-2

International Journal on Cloud Computing: Services and Architecture(IJCCSA),Vol.1, No.2, August 2011
DOI : 10.5121/ijccsa.2011.1203 31
Pre-allocation Strategies of Computational
Resources in Cloud Computing using Adaptive
Resonance Theory-2
Dr.T. R. Gopalakrishnan Nair1
, P Jayarekha2
1
Director, Research and Industry Incubation Centre(RIIC), DSI, Bangalore.
trgnair@yahoo.com
2
Research Scholar, Dr. MGR University Dept. of ISE, BMSCE, Bangalore .
Member, Multimedia Research Group, RIIC, DSI, Bangalore.
Jayarekha2001@yahoo.co.in
ABSTRACT
One of the major challenges of cloud computing is the management of request-response coupling
and optimal allocation strategies of computational resources for the various types of service requests. In
the normal situations the intelligence required to classify the nature and order of the request using
standard methods is insufficient because the arrival of request is at a random fashion and it is meant for
multiple resources with different priority order and variety. Hence, it becomes absolutely essential that we
identify the trends of different request streams in every category by auto classifications and organize pre-
allocation strategies in a predictive way. It calls for designs of intelligent modes of interaction between
the client request and cloud computing resource manager. This paper discusses about the corresponding
scheme using Adaptive Resonance Theory-2.
KEYWORDS
Adaptive resonance theory-2, cloud computing, pre-allocation strategies.
1. INTRODUCTION
During the past few years, cloud computing has emerged as an enabling technology and it has
been increasingly adopted in many areas including business, science and engineering. A cloud is
an aggregation of resources/services possibly distributed and heterogeneous and operated by an
autonomous administrative body which is run by a company or an organization (e.g., Amazon,
Google or Microsoft). Resources in a cloud are not restricted to hardware, such as processors
and storage devices, but it can also be software services or Web service in various forms of
instances.
A primary driving force of the recent cloud computing paradigm is its inherent cost
effectiveness[12]. The cloud computing environments are charged on the service usage, similar
to many basic utilities like electricity and water. A request-response model is very appealing for
both service providers and consumers and much concentration in research is required to achieve
different models. As the cloud computing expands, few of the major challenges will be in the

32
domain of fluctuating service request volume and unpredictable request arrival patterns. This
unstructured demand supply scenario existing between providers and consumers can hinder the
effective utilization of cloud computing environments [4]. The problem of service request
scheduling in cloud computing systems is addressed in this paper. We consider a three-tier cloud
structure, which consists of infrastructure vendors, service providers and consumers, and the
latter two parties are of particular interest to us. Clouds are primarily driven by economics and
hence the service provider aims to accommodate as many requests as possible with its objective
of maximizing profit. This business interest may conflict with many of the performance
parameter especially the service time. Identification of different request streams for different
category and organizing the resource request pattern in a predictive way to reduce the response
time, are the main objectives of this paper.
Among the multiple forms of cloud utilization, users tend to store more shared data on
the cloud; the data of interest is searched with the help of service providers. One major
difference existing between the cloud and the web is that the cloud data lacks the explicit link
structure present in the web; this link structure plays an important role in improving the
efficiency of web search.
The objects in a cluster might suggest that a particular set of objects are regularly accessed
sequentially and hence it might be beneficial to prefetch them as soon as the first one is
requested [13]. Further, decisions used to perform object prefetching, can in turn drive object
placement decisions. For example, since objects are prefetched together, it might make sense to
ensure that the objects reside on a disjoint set of hardware (nodes, routers, etc.). Furthermore,
since many clients store data on the cloud, the cloud can build more robust models for
clustering, prefetching, and object placement by aggregating similar access patterns for data
across multiple users [13].
Hence it is required to develop a clustering and prefetching technique to effectively
manage the cloud’s storage. Resulting in reduction of the number of jobs being rejected and
increase the profit of the service provider globally.
2. RELATED WORKS
2.1 Related work in clustering
The clustering of users based on their web access pattern is an active area of research in
Web usage mining. R. Cooley et al. [2] have proposed a taxonomy of Web Mining and they
present various research issues. In addition, the research in web mining is centred on the
extraction and applications of clustering and prefetching. Both these issues are clearly discussed
in [5]. It has been proven in this scheme that a 97.78% of prediction hierarchy is achieved.
2.2 Related work in prefetching
Prefetching means fetching the objects much prior to the user request arrival. There are
some existing prefetching techniques, but they possess some deficiency. In this the client
suffers from start-up delay for the first-time access since, prefetching action is only triggered
when a client starts to access that object. However, an inefficient prefetching technique causes
wastage of network resources by increasing the web traffic over the network.[7] J Yuan et al.,
has proposed a scheme, in which proxy servers aggressively prefetch media objects without a
pattern before they are requested. They make use of servers’ knowledge about access patterns to
ensure the accuracy of prefetching, and have tried to minimize the prefetched data size by
prefetching only the initial segments of media objects. [10] KJ Nesbit et al., has proposed a

33
prefetching algorithm which is based on a global history buffer that holds the most recent
missing addresses in FIFO order. S K Rangarajan, et al., [14] have proposed ART1 algorithm in
which clustering and prediction has resulted in an accuracy of 97%. In this work we have
proposed ART2 NN clustering algorithm for clustering user request arrival pattern. This cluster
helps the server’s agent in prefetching and making the access ready for memory, processors and
most importantly storage, prior to the user request[3][7][11].
2.3 Related work in cloud computing
A service provider rents resources from cloud infrastructure vendors and prepares a set
of services in the form of virtual machine (VM) images; the provider is then able to dynamically
create instances from these VM images [4]. The underlying cloud computing infrastructure
service is responsible for dispatching these instances to run on physical resources as shown in
figure 1. A running instance is charged by the time it runs at a flat rate per time unit. It is in the
service provider's interests to minimize the cost of using the resources offered by the cloud
infrastructure vendor (i.e., resource rental costs) and maximize the revenue (specifically, net
profit) generated through serving consumers’ applications. From the service consumer’s
viewpoint, a service request for an application consisting of one or more services is sent to a
provider specifying two main constraints, time and cost [12].
Although the processing (response) time of a service request can be assumed to be
accurately estimated, it is most likely that its actual processing time is longer than its original
estimate due primarily to due delays (e.g., queuing and/or processing) occurring on the
provider’s side. This time discrepancy issue is typically dealt with using service level
agreements (SLAs). Scheduling strategies in this cloud computing scenario should satisfy the
objectives of both parties. The specific problem addressed in this paper is the scheduling of
consumers’ service requests (or applications) on service instances made available by providers
taking into account costs—incurred by both consumers and providers—as the most important
factor.
Here we try to present, a management server and method for providing a cloud computing
service at high speed and reasonable cost, are provided. The management server provides a
virtual machine to a client as a computing resource. The virtual machine is multiplexed by
operating multiple virtual devices on a single virtual machine. Accordingly, the demand pattern
for computing resources may be predicted in advance and may be provided to a user more
efficiently.

34
Figure 1 Three Tier architecture of Cloud computing
2.4 Related work in ART2.
ART2 is an unsupervised learning and predicting algorithm derived from the resonance
theory[9]. The ART network can be used for identifying patterns. This design allows the user to
control the similarity between the patterns accepted by the same cluster[1]. ART2 can learn
about significant new classes, yet remain stable in response to previously learned classes. Thus,
it is able to meet the challenges in clustering the request arrival pattern where numerous
variations are common. ART networks are configured to recognize invariant properties of a
given problem domain; when presented with data pertinent to the domain, the network can
categorize it on the basis of these features[6][2] This process also categorizes when distinctly
different data are presented and it includes the ability to create new clusters. ART networks
accommodate these requirements through interactions between different subsystems, designed to
process previously encountered and unfamiliar events, respectively. We choose the ART2 neural
network rather than other classifiers because it is capable of incrementally improve the numbers
of clusters if needed.
ART2 networks were designed to process continuous input pattern data. A special
characteristic of such networks is the plasticity that allows the system to learn new concepts and
at the same time retain the stability that prevents destruction of previously learned information
[6].
4. METHODOLOGY
4.1 Preprocessing the request logs
The, the log data of request arrivals are represented as <client_Id,date,
requested_objects,readytime,deadline> format. We have selected a sample format of 50 clients
requesting for 200 different videos.
Consumer
Provider
Infrastructure
Vendor

35
4.2 Getting the Popularity Value
The popularity value is the most important parameter to get the effective prefetch
operation. This popularity value may consider the long term measurement of request-
frequency, which is neglected in the other algorithms. In this work the reference count
value is used to get the popularity value. The reference count value is highly variable
over short time scales, but this is much smoother over long time scales. This property
makes the popularity value to deal with the long term measurement of request
frequency.
Since the maximum and minimum value of request frequency is known during the
submission of input to the ART2 system, the normalized value of the popularity can be
obtained using the following transformation.
min
d
max
d
min
d
d
−
−
=
δ (1)
The transformed into a range between [0 ,1].
4.3 Extraction of feature vectors
For clustering, we need to extract the popularity of each resource that represents the
frequency of number of times that resource is requested. The pattern vector maps the
access frequency of each base vector element to real values. It is of the form P
={P1,P2,….Pn} where each Pi varies between 0 to 1.
0.8 0.2 0.2 0.1 0.3 0.4 0.5 0.1 0.3
Figure 2 Sample Pattern Vector
Figure 2 is a sample of pattern vector generated during a session.
Each pattern vector has a real value pattern of length 200.For each session we input 50
such pattern to an ART2, since we have 50 clients.
5 PROPOSED ARCHITECTURE AND ALGORITHM
Architecture of ART2 is shown in Figure 3. It is designed for processing analog as well
as binary input patterns. ART2 network module includes two main parts: attentional
subsystem and orienting subsystem. Attentional subsystem preprocess analog input
pattern, and then choose the best matching pattern under competitive selection rule from
the input pattern prototypes[9]. Orienting subsystem carry out similarity vigilance-
testing of the selective pattern prototype and trigger resonance learning and adjusting
weight vectors when vigilance-testing passed, otherwise get rid of the current active
node and search the other new ones. If there is no pattern prototype matching the input
pattern, create a new output node to represent it. Its memory capacity can increase with
the increase of learning patterns. The network allows not only off-line learning but also

36
in an on-line learning and applying way simultaneously, that is, the learning and
applying states are inseparable.
Experimental settings
The performance of the algorithm was thoroughly evaluated using our discrete-event
cloud simulator developed in C/C++. Simulations were carried out with a diverse set of
applications (i.e., various application characteristics) and settings. Owing to the large
scale constitution of cloud computing environments, the number of service instances
(that can be created) is assumed to be unbounded.
The total number of experiments carried out is 100 Specifically, we have generated 5
different numbers of services per application randomly selected from a uniform
distribution, i.e.,U(10, 80), and 7 different simulation durations (2,000, 4,000,
8,000,12,000, 16,000, 20,000 and 30,000).
Figure 3 The ART2 Neural networks
The ART2 neural network algorithm used in this work is summarized below[9].
Neural network configuration:
The parameters required for ART2 formulation has been initially chosen :
Noise inhibition threshold:
0≤θ≤1 (2)
Surveillance Parameter:
0≤ρ≤1 (3)
Error tolerance parameter ETP in the
F1layer: 0≤ETP≤1 (4)
Weight initialization of the neural
network: Top-down: Zji(0)=0 (5)
Bottom-up: Zij(0) ≤ (6)
Operation steps:

37
1. Initialize the sub-layer and layer outputs with zero value and set cycle counter to
one.
2. Apply an input vector I to the sub-layer W of the F1 layer. The output of this layer is:
Wi=Ii+aUi (7)
3. Propagate to the X sub-layer:
W
e
i
w
xi
+
= (8)
4. Calculate the V sub-layer output:
Vi = f(xi) + bf(qi) (9)
In the first cycle the second term of (9) is zero once the value of qi is zero.
The function f(x) is given by:










θ
≥
θ
≤
≤
θ
+
θ
=
x
x
x
0
)
x
(
2
)
x
(
f 2
2
2
(10)
5. Compute the U sub-layer output:
v
e
i
v
i
u
+
=
(11)
6. Propagate the previous output to the P sub-layer: iJ
i
i dZ
u
p +
=
(12)
The J node of the F2 layer is the winner node. If F2 is inactive or if the network is in
its initial configuration
i
u
i
p = (13)
7. Calculate the Q sub-layer output:
p
e
p
q i
i
+
= (14)
8. Repeat steps (2) to (8) until stabilizing the values in F1 layer according to Error (i) =
U(i) - U*(i).
If Error (i) ≤ ETP, the F1 layer is stable.

38
9. Calculate the R sub-layer output:
cP
u
e
cp
u
r
i
i
i
+
+
+
=
(15)
10. Determine if a reset condition is indicated. If ρ (e  ) then, send a reset signal to F2,
mark any active F2 node as not enable for competition, reduce to zero the cycle
counter and return to the step (2). If there is no reset signal and the counter is one,
the cycle counter is increased and passes to step (11). If there is no reset and the
cycle counter is larger than one, then control passes to step (14), once the resonance
was established.
11. Calculate the F2 layer input:
Ji
M
i
i
j Z
P
T ∑
=
=
1
(16)
12. Only the F2 winner node has non-zero output. Any node marked as non capable
by a previous reset signal doesn't participate in the competition.









 =
=
otherwise
)
T
max(
T
)
T
(
g
0
k
j
d
j
(17)
13. Repeat the steps (6) to (10).
14. Update the bottom-up weights of the F2 layer winner node:
d
1
u
Z i
Ji
−
= (18)
15. Update the top-down weights of the F2 layer winner node:
d
1
u
Z i
iJ
−
= (19)
16. Remove input vector, restore inactive F2 nodes and return to the step (1) with a
new input vector.
The initial values chosen for ART2 is as follows
a=10, b=10,d=0.9,c=0.1,e=0.0 θ=0.2 M = 5
Threshold value selection and methodology used
The use of several feature vectors require a special neural network, a supervised
ART2 NN is used. The performance of a supervised or unsupervised ART2 NN depends
on the appropriate selection of the vigilance threshold [10]. If the value of vigilance

39
threshold is near to zero, a lot of clusters will be generated, but if it is greater, then
number clusters will be generated.
5 RESULTS AND DISCUSSION
Each request consists of the type of resources required and configuration details. A
request r is a tuple containing at least (n, rt, d), where n specifies the number of virtual
machines required; rt is the ready time, before which the request is not ready for
execution; and d is the deadline for request completion. A wide range of requests to the
resources can be represented by these parameters.
The users of the infrastructure run different applications with different computing
requirements. Some applications need resources at particular times to meet application
deadlines are called deadline-constrained applications. Whereas other applications are
not strict about the time when they are given resources to execute as long as they are
granted the resources required called best-effort.
Graph 1 Comparison of job rejected
The graph 1 is a comparison for the number of jobs being rejected when no cloud and
with cloud computing.ART2 model is applied for clustering and prefetching the
resources. As shown in the graph the number of jobs rejected is negligible when
compared with that no cloud is applied.
The service provider's interest is to minimize the cost of using the resources offered by
the cloud infrastructure either provided by the vendor or in-house facility and maximize
the revenue (specifically, net profit) generated through serving consumers’ applications.
As shown in the graph 2 cost per task completion remains constant throughout when
clustering and prefetching is not applied. Along with the simulation time cost per task
reduces when ART2 model is used.

40
Graph 2 Comparison of cost per task completion
Graph 3 Comparison of job rejected
In graph 3 a comparison is presented between the Tight and Relaxed-Deadline when
cloud computing provides the services by prefetching the resources. ART2 clustering
technique helps in predicting the resources required hence resulting in reduction in the
number of jobs being rejected even at tight deadline.
6 CONCLUSION
Cloud computing has the potential of becoming a revolutionary technology that changes
the way service computing is performed. The principles introduced in this work are
essential for the cloud computing vision to materialize to its full extent.
One of the most important benefits of Cloud computing is the ability for Cloud clients to
adapt the number of resources used based on their actual use. This has great implications
on cost saving as resources are not paid for when they are not used. It follows that an
accurate prediction method would greatly aid a Cloud client in making its auto-scaling
decisions.
Adaptive Resonance Theory-2 identifies the trends of different request streams in every
category by auto classifications and organizes pre-allocation strategies in a predictive
way. In the proposed design of intelligent modes of interaction between the client
request and cloud computing resource manager has resulted in reduction of number of
jobs being rejected and also reduction in cost per task completion.

41
References
1. CARPENTER G, GROSSBERG, S , 1987, ART2: Self-Organization of Stable Category Recognition Codes
for Analog Input Patterns, Applied Optics, 26: p 4916:4930,
2 .COOLEY R., MOBASHER B., SRIVATSAVA J., 1997 ,Web Mining: Information and Pattern
Discovery on the World Wide Web , ICTAI’97.
3.CHRIS TSENG H, 2007, Internet Applications with Fuzzy Logic and Neural Networks: A Survey, Journal
of engineering, computing and architecture
4. Lee,Y. C.,Wang, C., Zomaya, A. Y., Zhou B.B. (2010). Profit-driven Service Request Scheduling in
Clouds. In Proceedings of the International Symposium on Cluster Computing and the Grid (CCGRID).
Melbourne, Australia.
5. FAUSETT, L.V. 1994, Fundamentals of Neural Networks Architectures, Algorithms, and applications.
New Jersey: Prentice Hall International Inc, New Jersey, p .246-287.
6. ISSAM DAGHER , 2006 Art networks with geometrical distances Journal of Discrete Algorithms archive
Volume 4 Issue 4 ISSN:1570-8667 Pages: 538-553.
7.JUNG J, LEE D, CHON K, 2000 Proactive Web caching with cumulative prefetching for large multimedia
Data Computer Networks – Elsevier.
8. KUMAR N, JOSHI R S 2007 Data Clustering Using Artificial Neural Networks Proceedings of National
Conference on Challenges Opportunities in Information Technology (COIT-2007).
9. NACHEV A, GANCHEV I 2003 Data Mining For Browsing Pattern Weblog Data By Art2 Neural
Networks International Journal Information Theories Applications Vol.10.
10.NESBIT K J, SMITH J E, 2004 Data Cache Prefetching using a global history buffer, High Performance
Computer Architecture.
11.MASSEY L 2002 Determination of clustering tendency with ART neural networks Proceedings of
recent advances in soft-computing (RASC02) .
12.Marcos Dias de Assunção, Alexandre di Costanzo, and Rajkumar Buyya. A cost-benefit analysis of
using Cloud computing to extend the capacity of clusters. Cluster Computing, 13(3):335-347, September
2010.
13.Muniswamy-Reddy,K.K., Holland, D. A., Braun, U.,and Seltzer, M. Provenance-aware storage
systems. In proceedings of the 2006 USENIX Annual Technical Conference.
14.RANGARAJAN S K , PHOHA V V , BALAGANI K, SELMIC R R 2008 Web user caching and its application
to prefetching using ART neural networks IEEE Internet Computing, Data Knowledge Engineering,
Vol.65 No.3, p.512-543

Pre-allocation Strategies of Computational Resources in Cloud Computing using Adaptive Resonance Theory-2

More Related Content

Similar to Pre-allocation Strategies of Computational Resources in Cloud Computing using Adaptive Resonance Theory-2 (20)

Recently uploaded (20)

Pre-allocation Strategies of Computational Resources in Cloud Computing using Adaptive Resonance Theory-2