Professional Documents
Culture Documents
Abstract-Resource provIsioning in Cloud computing is an There might be another possibility that users are not able to
important problem as the resources are scarce. Better utilization fully utilize the resources allocated to them. If provider is able
of the existing unutilized resources may solve the future resource to know the percentage of underutilized resources, provider
scarcity problem. To decide, how much percentage of unutilized may use the concept of redistribution implying the concept of
allocated resources the Cloud manager can take back to solve the distributive justice in order to satisfy its customer and to keep
problem of scarcity of resource in future, an econometric model its reputation intact. Provider may take some part (portion) of
of distributive j ustice is identified and mapped to resource unutilized resources back from the users and allocate them to
redistribution in Cloud computing. The proposed model uses a
the requesting client. Suppose a provider does not have enough
Probit model to get the value of coefficient which helps to
resources to accept a request, provider does a survey of its
determine the percentage of unutilized allocated resources to be
clients to identify those clients that are utilizing the allocated
taken back. Data is increasing at the enormous rate warranting
mass storage. Thus, Cloud storage has been considered an
resources below a defined threshold. Provider may give some
important scarce resource. The proposed model focuses on the lucrative offer to such clients in terms of cost slashing etc. if
Cloud storage as a resource. It is assumed that service provider they agree to leave some of their resources. That allows the
must reserve certain amount of resource to accept sharply provider in getting some resources back to be allocated to the
increased resource request. The proposed model also identifies requesting clients. In Cloud computing, resource reservation
this constraint and proposes a model of distributive j ustice that and on-demand plans are two provisioning plan which are
addresses this. Experimental results reveal the efficacy of the offered by Cloud providers for computing resources [2] .
proposed model. Reservation plan is a static Cloud resource provisioning policy.
In this plan, user rent computing resources for a long time and
Keywords-Cloud, econometric model, probit model, Cloud providers charge price one time (e.g. one year). Quick scale up
storage. or quick scale down is not possible in this plan. In on-demand
plan, providers charge by pay-per-use basis (e.g. 1 day).
I. INTRODUCTION
Therefore, with on-demand plan, when consumers need
Most of the internet connected devices use the Cloud resources for the fluctuated and unpredictable demands,
technology knowingly or unknowingly now days. Cloud consumers can dynamically provision resources. If provider
computing provides a platform for using computing resources gets some resources from reservation-plan-users and allocates
like CPU, bandwidth, memory etc. via internet without owning these resources to the on-demand-plan-requesting clients, it
these resources on a pay per-use basis. Thus, there is no need to will increase the provider' s profit and will meet the users
buy the required computing resources to run any application. demand also. It is to note that with the on-demand plan, the
Imperative is to rent the computing resources that too only for price for resources is costlier than that of the reservation plan
the time period in which the resources are needed. Cloud [2] .
resources are offered in form of a service on a payment basis.
This paper has identified the problem of resource scarcity
Many services of Yahoo, Microsoft and Google etc. have been
and proposed a model to solve this problem based on
designed to act as Cloud from time to time. Cloud users feel of
econometric model. Considering the importance of the Data
infmite resources in the Cloud, which is just an illusion. As
centre management, Cloud storage has been considered an
both the number of users and data are increasing at the rapid
important and scarce resource and the proposed model focuses
pace, scarcity of resources is bound to happen in future due to
on the Cloud storage as a resource.
limited resources. Service provider' s reputation may be
affected in the Cloud service market, if the provider declines a The paper is organized as follows. Section II mentions the
request due to limitation of resources. How to schedule these related work. Section III highlights the motivation behind this
scarce resources amongst the growing number of Cloud users work. Section IV details the proposed model based which is
are an important and a big challenge for the Cloud service based on Econometric model. Section V mentions the
providers. In order to satisfy the requests, provider may opt for experimental evaluation of the proposed model by inferring the
Cloud federation i.e. renting resources from some other results obtained. Section VI concludes the work.
providers [ 1 ] . Provider may have to periodically rent the
resources as and when the resources fall short of.
II. RELATED WORK using redistribution technique, proposed model gets some
percentage of unused allocated resources back and then Cloud
The demand-&-supply model of resources in distributed
provider may assign these resources to the new users.
system is analogous to demand-&-supply model of commodity
in economics. Resources provider in former is same as the III. MOTIVATION
commodity supplier that supplies variety of resources such as
CPU, memory, disks etc. for its consumer. Consumers, of Many countries devote large budget for redistribution of
aforementioned resources, in distributed system resemble the income though governmental redistribution is costly. There
commodity buyers in economics. Consumers have to pay a fee should be some knowledge about how much resources society
to acquire resources in order to meet its demand of resources is ready to pay for distributive justice. Giacmo Corneo and
[3 ] . Some resource allocation strategy, based on economics Christina M. Fong in [9], proposed an econometric model
model, has been proposed in the literatures that are listed as which helps to estimate the willingness of individual to pay for
follows. distributive justice. There may be different attitude for
redistribution within groups of individual with similar market
To investigate the effectiveness of resource allocation income. This work [9] assumes that people give preference to
strategy based on the economics model for the benefits to the equity, i.e. for living in a society where "one gets what one
providers, consumers and society, a grid economy is simulated deserves, and deserves what one gets". Measures of
in [4] , in which providers and consumers pursue their own individual' s support or individual' s opposition for
interests and resulting effects on system welfare is measured. It redistribution, fairness of market outcomes, and pre-fiscal
is found that economics based strategies fares well. Since in a incomes are only required for estimation of this model. It
grid system number of consumers, their resource demand and derives a formula
resource supply can be highly dynamic, a model is suggested in
[5] that consider an environment in which both providers and
consumers act as selfish agents. A distributed negotiation
mechanism is introduced where agents negotiate for both that defines maximum amount an individual is willing to
contract price and de-commitment penalty, allowing de pay for distributive justice from the fiscal data and estimated
2
user according to provider, and the redistribution Re of
unutilized resource. Let fairness f be 1 if ratio of unutilized
resource to total allocated resource to a particular user is less
than a threshold t and 0 otherwise, where 0 < t S; 1 . For
example, assume that a user has 1 00 units of resources, 40
units of unutilized resources and t = 0. 3 . Here, ratio (40/1 00)
is greater than t and according to provider, allocation of
resources to this user is unfair; i.e. f = O .
Similarly, let Re b e 1 i f the provider redistributes unutilized
resource and 0 otherwise. It is asswned that justice can be done
if either allocation of resources is fair and provider does not
URe = alogRRe + {J C 1 - 2 f)
Utility with redistribution
decreases a user's utility. Here, it is assumed that utility (U) is
U = alogR + {J} .
cardinally measurable and unit comparable and takes the form (6)
UN = alogRN + {J (2f - 1)
Utility without redistribution
U = alogR - {J (2) (7)
U = a 1 0 g (R - W) + {J
written as {J} = -{J . The three remaining cases including first one are
(3) checked in table 2
Since utility is constant (i.e. does not change), from above two TABLE 2: RELATIONSHIP BETWEEN FAIRNES S , REDISTRIBUTION, PERCEPTION
a logCR - W) + {J
equations (2) and (3 ) we obtain, OF JUST DISTRIBUTION AND UTILITY
UN -fJ
Solving the equation (4), we get
W = 1 - exp
0 0 -I
3
In this theoretical model, it is assumed that if redistribution of Therefore,
unutilized resource provides higher utility to the users then
only they support redistribution. 2f3 ( 1 - r ) Bf
a 2 AR u
((r -l)Bt)
Now for mathematical framework, it can be expressed in terms
� - exp
(8) j =
R = 1 2AR u
(13)
=
T2 30 and at any instance provider has 28 unit unallocated
Here r > 0 and governmental redistribution of income IS
storage (PUAS 28), coordinator will go with the proposed
more progressive for smaller value of r. model to get some unused allocated storage back. If at any
Inserting the value of RN and RRe ' from ( 1 1 ) in (8) and (9), and instance provider has 1 5 units of unallocated storage, which is
r r
less than Tl and as well as T2, coordinator will go with the
P [ d = 1 I R u, [, x] = P [ a log(s) + (r - l) a 10g(R u) +
substituting the resulting equations into ( 1 0), one obtains
proposed model as well as for Cloud federation.
4
provider has all customer related information like unused 5) Storage in Tier Context (tier) : Generally, three types of
allocated space of a particular user. tier storage are considered in Cloud; Tier l , Tier2 and Tier3 .
Different storage providers offer storage as a service with Tierl storage is most scalable, most reliable and highest
different characteristics. For example, as a service user of performing storage, generally used for business critical
AT&T Synaptic Storage, one can replicate two copies remotely applications. Tier2 storage is used for unstructured data,
but as a client of Peerl CloudeOne Storage, it is possible to backup, and multimedia etc. Tier3 provides low performance
have one ' s data replicated at any or all of the 1 7 data centers. with high scalable capacity, so is used for long term storage for
Nirvanix Cloud Storage offers one node, two nodes and three archival and offline storage. For three types of tiered storage,
nodes replication. Thus, characteristics of storage vary from one integer number between 1 and 3 inclusively has been
provider to provider. For experimental purposes here, we have
randomly assigned to each user.
considered types of user, replication of node, tiered memory,
6) Number of child account (ace) and supported
number of child account and number of supported application.
But these characteristics are not fixed for all service providers; application (app) : It is possible for more than one account to
provider can change characteristic parameter according to their access the same Cloud storage. For example, in a university, an
need easily in the proposed model. administrator can provide child account to each student to
access online library. Similarly, Cloud storage can be accessed
A. Parameters Used
by more than one application. For this, a number between 1 0
The proposed model uses various parameters that have and 1 000 has been randomly assigned to each user for number
been deliberated as follows : of supported child account and also between 5 and 1 00 to each
r=
storage as a service to the users. It is possible that users will not discussed in section IV.A). Substituting the estimate
be utilizing all resources assigned to them by the provider. This of Bf /AR u as well as 0.85 into equation ( 1 3), an estimated
work assumes that storage is provided in GB . Different types value of j for each instance is obtained.
of users may have different units of unused allocated storage. B. Analysis ofExperimental Results
In the experiment, a number has been randomly assigned
Results have been analyzed by carrying out the experiment
between 1 0 and 1 000 to type 1 and type2 users, between 2 to
as below.
1 00 to type3 and type4 users. Suppose a user gets a number
Experiment 1 :
288, it means this user has 288GB unused allocated storage. Experiment 1 observes the posItIve response for varying
3) Fairness ofAllocation ofResource (f) : The term (t) has number of users. In fig. 4, it is shown that how many positive
already been discussed in section IV.A. Value of f can be 0 or responses U > 0) out of 1 00 instances, for different number of
1 . Here, without loss of generosity, we have assigned randomly users varying from 1 00 users to 1 000 users, are obtained. It is
o or 1 to each user for fairness of allocation of resources. If a found that number of positive response out of 1 00 instances
user has f= 1 , it means allocation of storage for this user is fair for fixed numbers of users vary irregularly.
N u m b e r of p o s i t i ve response Vs N u mber of u sers
otherwise unfair.
4) Replication of Nodes (RoN) : "Performance is highly w ,-----
50 +---�----------__��____--_.__
affected by the speed-of-light latency, TCP latency, (both of
which are directly correlated to circuit distance between user
40
and files), as well as packet loss. By placing the files closer to N u mber of
positive 30
the user, both speed-of-light latency and TCP latency are response
10
minimized. Packet loss is also minimized because the
probability of packet loss increases as distance increases" [ 1 2].
In this work, it is assumed that provider is providing three o ...."--
... ----"---�---'-------"------"-
1 00 2 00 300 400 ,00 GOO 1 00 BOO 900 1 00 U
types of node replication to place the data closer to the user. N u m ber of users
For three types of node replication, one integer number
between 1 and 3 inclusively has been randomly assigned to Figure 4: Number of positive response out of 1 00 instances
each user.
5
Experiment 2: VI. CONCLUSION
I n this, for the goodness of fit (i.e. measurement o f In this paper, problem of future resource scarcity in Cloud
discrepancy between expected value and observed value) o f is identified and a mathematical model is proposed that is
the proposed model, percentage of correctly predicted measure based on an econometric model to solve the identified problem.
(by comparing the actual value with predicted value) for each For the validity of the model, characteristics of cloud storage
instance has been calculated. Average of percent correctly have been observed from different available storage service
predicted measure of 1 00 instances of 1 00 users has been providers and on a randomly generated data set. It is possible
taken which has been repeated for instances of 200 users and that provider can change the parameters easily according to his
so on (fig. 5). It is observed that as nwnbers of users is need. By observing different results it is concluded that number
increasing, average of percent correctly predicted measure is of positive response and sum of positive response depends on a
decreasing. Notable is that this value is always greater than particular instance and do not depend on the number of users.
50%. This paper considers only the Data storage. Cloud resources
Ave rage goo d n e s s of f i t measu re N u m b e r of
,------
Vs are not limited to storage as other important resources also
users exist in Cloud such as CPU, memory, bandwidth etc. Future
62 work may address other Cloud resources and it will be possible
bO to apply this model on other Cloud resources.
58
• • •
Averag@
good n ess o f f i t 56 REFERENCES
• • •
m e a s u re
of 100 ' 4
• • •
[I] B. Rochwerger, D. Breitgand, E. Levy, A. Galis, K. Nagin, I. M.
Llorente, R. Montero, Y. Wolfsthal, E. Elmroth, 1. Caceres, M. Ben
Instances
52
' 0
Yehuda, W. Emmerich, F. Galan, 'The reservoir model and architecture
48