Information Technology Reference
In-Depth Information
process the largest number of sensor nodes in one hour and its economic cost is
smaller than the instances m1.xlarge and c3.xlarge.
6 Conclusions and Future Works
In this paper we have proposed theoretical models to estimate the performance of
Amazon EC2 instances for processing frost prediction applications based on dif-
ferent number of sensor nodes used. In order to evaluate the proposed method, an
application for frost prediction was used. Next, Amazon EC2 Cloud services were
used to run the application and study the performance of each instance. The per-
formance was evaluated based on two metrics, execution time and economic cost.
In order to conduct the experiments and generate the models we have pre-
sented different test scenarios. Each scenario corresponds to a particular Amazon
EC2 instance. The accuracy of the obtained models was compared with data of
previous executions of the frost prediction application. From the results we con-
clude that the proposed models are suitable to estimate both the execution time
and economic cost. In addition, a typical application case was used to determine
which instance is more suitable for processing frost prediction applications.
Regarding the comparison of the different instances in the typical evaluation
case, it can be concluded that for WSNs formed by few nodes (up to 80) the
t1.micro instance is recommended. Otherwise, for larger number of nodes: 200 -
300, 400 - 700, 800 - 900; it should be used the m1.small, m1.large and m1.xlarge
instances, respectively.
On the other hand, while the c3.xlarge is the EC2 instances with highest per-
formance, we did not observe important differences in the performance compared
to the other tested instances. Moreover, if we also consider its high cost, it is not
recommended for this type of applications. Furthermore, for the case of WSNs
made up of more than 1000 sensor nodes, multiple EC2 instances should be used
in parallel to run the application.
Regarding the frosts prediction method we will test other frost prediction
methods based in machine learning algorithms. It should be mentioned that the
costs of transfer and storage are minimal compared with these ones for using
the on demand instances. For example, there is a pricing of 0.12 U$S per GB of
transfered data for the first 10 TB for month, and if data volume do not exceeds
the GB per month, the transfer is free. However, such costs will be considered
and included in future works.
In this paper we showed that second degree polynomials are a simple and
suitable way for estimating the performance of Amazon EC2 instances. However
we will continue the validation of these polynomials in future works studying
the processing of the frost prediction application using multiple EC2 instances
managed with specific Cloud tools like Star Cluster. The purpose of these fu-
ture experiments is to extend the proposed models in order to estimate how
many machines are needed for optimizing the relationship between the execu-
tion time and economic cost for frost prediction applications and how they must
be managed.
