THE EFFECT OF THE RESOURCE CONSUMPTION CHARACTERISTICS OF CLOUD APPLICATIONS ON THE EFFICIENCY OF LOW-METRIC AUTO SCALING SOLUTIONS

International Journal on Cloud Computing: Services and Architecture (IJCCSA) Vol. 8, No.1/2/3/4/5, October 2018
DOI: 10.5121/ijccsa.2018.8501 1
THE EFFECT OF THE RESOURCE CONSUMPTION
CHARACTERISTICS OF CLOUD APPLICATIONS ON
THE EFFICIENCY OF LOW-METRIC AUTO SCALING
SOLUTIONS
Ha Van Cu1
, Nguyen Hong Son2
1
R&D Lab, Faculty of Information Technology, Post and Telecommunication Institute of
Technology, Ho Chi Minh City, Vietnam
2
Faculty of Information Technology, Post and Telecommunication Institute of
Technology, Ho Chi Minh City, Vietnam
ABSTRACT
Auto scaling is a service provided by the cloud service provider that allows provision of temporary
resources to the subscriber’s systems to prevent overloading. So far, many methods of auto scaling have
been proposed and applied. Among them, solutions based on low-level metrics are commonly used in
industry systems. Resource statistics are the basis for detecting overloading situation and making
additional resources in a timely manner. However, the effectiveness of these methods depends very much
on the accuracy of the overload calculation from low-level metrics. Overloading is mentioned in solutions
that usually favor a shortage of CPU resources. However, the demand for resources comes from the
application running on that each application has the characteristics of demanding different resource types,
with different CPU, memory, I/O ratios so it can not just be statistically on CPU consumption. The point of
view here is that even though based on low level resources, the source for calculation and forecasting is the
characteristic of the resource needs of the application. In this paper, we will develop an empirical model to
assess the effect of the application's resource consumption characteristics on the efficiency of the low-
metricauto scaling solutions and propose an auto scaling solution that is calculated based on statistics of
different types of resources. The results of the simulations show that the proposed solution based on
multiple resources is more positive.
1. INTRODUCTION
Over decades of development, cloud computing has become a key component of modern
information systems. Many of the advances in technology have contributed to the development of
the cloud. Everyone recognizes that thanks to the explosive growth of broadband technology, the
community has been able to use cloud computing smoothly. Cloud computing services are also
abundant and easily accessible thanks to the development of Internet technology. Sales from
cloud computing is on the rise. In a world trend that is entering Industry 4.0, cloud computing
continues to affirm its position with many intelligent applications that use the cloud platform.
Today, the deployment of the entire enterprise IT infrastructure in the cloud is no longer a rarity
for many reasons. First of all, the need to reduce the cost of infrastructure and operating costs
because sharing resources and energy in the cloud is much cheaper than investing in private
infrastructure. Another reason is that many modern applications that businesses want to use,
especially those that target mobile end-users, IoT, and ecommerce, are trending toward using
services provided from the cloud such as web services, no SQL database. Implementing,
changing, and upgrading applications in the cloud is also quicker and more convenient.
Information security services are also well supported.In addition to the ability to protect
infrastructure and services from cloud service providers, subscribers can rapidly deploy their own

International Journal on Cloud Computing: Services and Architecture (IJCCSA) Vol. 8, No. 1/2/3/4/5, October 2018
2
security measures to protect their applications. Another reason is that the power and capacity
provided by the cloud for the subscriber can vary according to demand. Building and providing
auto scaling services to subscribers of cloud service providers like Amazon EC2, Azure is very
meaningful.With auto scaling, when the enterprise resources are overloaded, the cloud can
dynamically add temporary resources to maintain the normal service of the information system
and also automatically retrieve the resources when it returns to normal state.The auto scaling
service not only benefits the cloud clients but also facilitates the cloud service provider to
maximize revenue as it can increase the number of subscribers and can leverage idle resources in
flexible way.From the meaning of auto scaling that the implementation of this technique has
become a topic of concern.
Typically, cloud systems allow subscribers to choose auto scaling types along with select the load
balancer. How the auto scaling technique implemented within the service system depends on the
cloud providers. The auto scaling technique follows the trend of positive surveillance, based on
the state of the system to make the decision to scale resources. Decision-making can be made
using accurate measurement data or forecasting technologies, but whatever the type, auto-scaling
solutions are set up with parameters that directly or indirectly relate to processing resources.
However, processing resources are not the entire story of meeting service requirements, because
when the type of resources are not scarce and other resources are insufficient, service requests are
still waiting. As in [8], it has been shown that memory intensive applications also heavily
influence the performance of the cloud. We therefore need to approach the problem of auto
scaling from the perspective of resource consumption characteristics of applications run on the
cloud. The applications will feature different resource needs, such as simulation, brute force
applications that use large amounts of processing power, and memory intensive applications like
data analysis applications. There are also many applications that require both processor power
and memory in high for execution, such as graphics and animation applications.
In this study, we will demonstrate that auto scaling solutions based only on processing resources
are not quality-assured when cloud-based applications are altered.First of all, we build a typical
auto scaling algorithm based on processor utilization monitoring and model a subscriber cloud
infrastructure with an autoscaler to derive parameter calculations necessary for evaluating the
auto scaling solution. The evaluation will be through computer simulation results based on the
Cloudsim tool. The rest of the paper is structured as follows: Section 2 shows some researches
that deal with auto scaling and application load on the cloud. Section 3 details the construction of
a subscriber cloud infrastructure model with an autoscaler as the basis for simulating and
evaluating auto scaling solutions. This section also presents a typical auto scaling algorithm based
on processor utilization monitoring used for evaluation purposes in the paper. Section 4 presents
computer simulations along with analysis and recommendations for auto scaling solutions
tailored to the changing of cloud applications. The paper ends with the conclusions set out in
section 5.
2. RELATED WORKS
The auto scaling solution can be grouped by reactive and proactive type in which reactive will
adjust resource when overload occurs while proactive will forecast and actively adjust resource
first to not overload appear. It is also possible to distinguish auto scaling solutions based on the
use of metric measurements as either low-resource measurement or high-level measurement [11].
Low level parameters such as resource utilities and high level parameters such as number of
users, request rate, average response time, session creation rate, throughput, service time, and
request mix. We have explored the real-life auto scaling (industry system) and typical works
related to the construction of auto scaling solutions. We are particularly interested in load-bearing
solutions and applications running on cloud-based systems requiring auto scaling. Law-based

3
solutions are widely used in the industry autoscalers such as Amazon EC2 [5], Microsoft Azure
or OpenStack, the key components are pre-established rules based on resource measurements. For
example, “if CPU utilization Exceeding 0.7 adds an instance and if less than 0.3 the instance is
destroyed." This method is reactive and based on low level resource measurement. This is the
solution that can be implemented by cloud providers only. Also related to low-level
measurements in [2], the authors propose a hybrid method between reactive and proactive applied
to an intermediary enterprise managing a virtual private cloud that provides resources to
subscribers. The authors also mentioned the necessity of selecting the metric and determining the
exact thresholds for auto scaling from which to provide the GoS quantity and the standard for
auto scaling based on GoS. However, in terms of performance, GoS service time is still
calculated on the basis of CPU processing power and workload only distinguishes between the
two types of requirements is OD (on demand) and AR (advance reservation). Other proactive
methods based on high-level metrics such as in [1] have analyzed and pointed to the challenges of
auto scaling. The first challenge is to accurately predict future workload, and then calculate the
amount of resources needed to increase or decrease according to the workload and finally to
overhead the transition when the number of resources changes. Based on the analysis of the
challenges, the authors proposed a method of distributing resources using model predictive
techniques. Resources are then distributed or retrieved based on the application's optimal utility
over a predetermined prediction time. However, the workload used here is the number of current
users in the system, which does not reflect the characteristics of the application. Similarly, in [3]
an adaptive control approach can be used to change the number of VMs allocated to a service
based on load change monitoring and also future load forecasts. However, the basis of calculation
is based only on the speed of the service request. Another auto scaling solution is proposed in [4]
by predicting future service needs using standard time-series analysis techniques. The scaling
depends on the change of throughput in 10 minutes. Similarly, the authors in [6] also used time
series analysis to propose a resource forecasting model based on double exponential smoothing to
improve forecast accuracy, but based on the overall variability of workload.
In general, active monitoring solutions are based on actual metrics of processing resources or
forecasting methods using high-level metrics such as request rate, average response time, session
creation rate, throughput, service time, and request mix. Most solutions do not take into account
the difference in demand for different resource types of applications. With auto scaling solutions
for the web app like [7], if the demand for resources is balanced between types (CPU, Memory,
Bandwidth), forecast accuracy will not be affected much. But when the application is switched to
CPU intensive or memory intensive, be inadequate is inevitable.
In the interest of evaluating solutions, [9] also developed a fuzzy rule-based system to compare
auto-scaling solutions using reinforcement learning techniques, Fuzzy Q-learning (FQL) and
Fuzzy SARSA learning (FSL), the results also show that FSL is effective for periodic workloads.
Similarly, in [10] the authors provide a simulation model for evaluating different auto scaling
policies that use workflows as the input for evaluations. However, both [9] and [10] do not
address the specific needs of different resource types of applications.
3. CLOUD MODEL WITH AN AUTOSCALER
As stated for the purpose of this paper, we need to examine whether the effectiveness of the auto
scaling solutions is affected as the demand for different resource types of application changes. To
do that we will firstly model a subscriber cloud infrastructure that subscribes to horizontal auto
scaling.

4
Subscribers are scaled automatically by the number of virtual machines added or reduced once is
N machine, we have:
numOfVMScale(t)=N x scaleEvent(t) (1)
with numOfVMScale(t) is number of virtual machines (VM) added or reduced at time t
scaleEvent (t) = 1 or 0 or -1 depending on the decision function, called D(). In case the auto
scaling algorithm is implemented in an active metric monitoring style, the autoscaler keeps track
of the remaining resources,denoted as remainResource (t), which introduced to the function D()
along with a threshold.
scaleEvent(t)=D(remainResource(t),Threshold) (2)
Considering the load generated by the request i to VM as I instrutions, C memory in Megabytes,
and bandwidth B in Mbps, denoted as request(Ii,Ci,Bi) and load of VM is total of load of requests
at concerning time t, called requests(t). The configuration of the virtual machine has MIPS
processing resources, the executable memory M, and BW bandwidth, written as VMresources
(MIPS, M, BW). The remaining resources will be computed using the f() method based on
requests(t) and VMresources (MIPS, M, BW).
remainResource(t)= f(requests(t),VMresources(MIPS,M,BW)) (3)
With n requests at concerning time t, requests(t) as following:
requests t = request I , C , B 4
The method f() can be a way of subtraction of processing power in MIPS or memory capacity in
Megabytes.
To evaluate the effectiveness of the use of hiring resources including additional virtual machine
resources under the auto scaling service here define a metric called VMxTime, if the time hourly
called a VM hour and a VM hour means opening a virtual machine for an hour. The cost of
consuming resources is attributed to the VM hour by taking the virtual machine number
multiplied by the corresponding open time in hour. Let VM Usage be the virtual machine
consumption during some time T, it is expressed as follow:
= !"# $ %$
&
'
5
NumOfVMs(t) is numbers of VMs opening at time t. The item is expressed by formulation:
numOfVMs(t)=currentNumofVMs(t) + numOfVMsScale(t) (6)
The larger the VM Usage, the greater the cost to the subscriber
.
Determine the average completion time of a request (Ii, Ci, Bi):
Let I be the total number of instructions for all requests processed during time T and C is the total
memory capacity of all memory needs of all requests processed during time T. The T/I ratio is the
average time that an instruction is completed and T/C ratio is the average time required to
response one Mbyte. If the CPU biased request, use the ratio T/I and if the request for the RAM
intensive, use the ratio T/C. We can refer to the 1000 units, the request has 1000 instructions, or
100MB memory requirements, multiply 1000T/I or 100T/C.To evaluate the quality of the auto
scaling mechanism uses the average virtual machine consumption per request and the average
time spent completing a request. Compare cases based on these two parameters.

5
Auto scaling algorithm:
The service is done by horizontal scaling and assuming that the load balancer is well-balanced,
ensuring that the deviation and the load balancer are matched as described in [12].
1.Set the threshold value Vth;
2.Set idle time T1; //before destroying a VM
3.Set sample time T2;// statistical time
4.Parallel:
Each subscribed VM {
Collect and calculating the average CPU utility during T2;
If (the average CPU utility > Vth) {
create VM;
register the VM to Loadbalancer;
}
5.Parallel:
Each opened VM {
Count the idle time;
If (idle time>T1) {
Delete VM from register list of Loadbalancer;
destroy the VM;
}
4. COMPUTER SIMULATIONS
The simulation programs in this section are made with Cloudsim. We build a data center with
configuration information as follows:
Data center has 50 hosts, each host has 32 Pe, 1000MIPS/Pe, 16GB RAM, 10Gbps, 1000GB
storage. Virtual machine scheduling according to time-shared policy.
Virtual machines have 1000MIPS with 1Pe; 4096MB RAM, 1000Mbps, and the scheduling
policy for cloudlets is time-shared. Suppose the enterprise was granted 4 VMs initially.
The request (cloudlet) corresponds to length = 5000MI, filesize = 1913 and outputsize = 1000,
pe=1.Cloudlets are also customized according to different resource consumption characteristics
for testing.
4.1 AUTO SCALING USING THE CPU THRESHOLD
In this simulation,we use the CPU utilization threshold of 0.95 (Vth). The provisioning is a series
of requests, with each request being an application executed on a virtual machine with a very high
CPU consumption characteristic and a relatively small memory requirement of less than 0.25. We
calculate VMxTime consumption statistics, service completion statistics. Figure 1 shows the
number of virtual machines opened and closed over time (NumOfVMs (t)) for this experiment.

6
Figure 1. Virtual machines opened and closed over time with low RAM requests
With statistical results, the average VMxTime cost for a request is 11.22 VM seconds and the
average completion time for a request is 34 seconds.
Next, experiments performed with applications still have very high CPU resource requirements
and average RAM needs are below 0.45. The resulting virtual machine number is opened and
closed as shown in Figure 2.
Figure 2. Virtual machines opened and closed over time with moderate RAM requests
The average virtual machine consumption per request is 16.3 VM seconds and the average
completion time for a request is 47 seconds. The results show that compared to the small RAM
demand, the virtual machine consumption per request has increased by 45% and the average
request completion time has increased by about 38%.
Continuing to test the impact of memory needs on auto scaling, we have applications with low
CPU utilization, maximum 0.4, and memory requirements are high, over 0.5. The result of the
number of virtual machines opened and closed over time is depicted in Figure 3.
4
5
6
5
4
3
2
1
0
0
1
2
3
4
5
6
7
0 10.1 10.9 121.2 122.41 124.5 125.5 127 130
4
5
4
3
2
1
0
0
1
2
3
4
5
6
0 20.1 103 106 114 127 130

7
Figure 3. The number of virtual machines opened and closed over time, corresponding to the needs of small
CPU and high RAM
As can see in Figure 3, almost no virtual machines have been added for a long time, but the cost
of virtual machine time has increased significantly, with an average of 20.7 VM seconds per
request. Possibly due to lack of resources, the average time to complete a request also jumps to 63
seconds. In general, when the application has a large CPU and small RAM, the parameters are
doubled.
4.2 AUTO SCALING USING BOTH CPU AND MEMORY THRESHOLDS
From the simulation results above, we propose performing auto scaling considering both the CPU
threshold and Memory threshold. To ensure SLA in real-world situations, applications that have
different resource-consuming characteristics need to perform auto scaling across multiple
resource thresholds with the rule that whenever any threshold is violated, resources must be
added immediately. We validated the solution with a CPU threshold of 0.95 and a RAM threshold
of 0.9. The applications used for the simulation have high CPU requirement, above 0.71 and
memory requirement are also high, over 0.5.
Figure 4. Number of virtual machines opened and closed over time with two types of threshold solution
The result of the number of virtual machines opened and closed over time with solutions using
two types of resource thresholds is shown in Figure 4. Although the number of VMs added is
greater but the average virtual machine cost per request is only 12.7 VM seconds, has dropped a
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
0
1
2
3
4
5
6
7
8
9
0 5.1 5.69 11.14 26.24 27.69 27.8 28 29

8
lot although the application has high memory requirements. Meanwhile, the average completion
time of a request has dropped to 11.7 seconds. This demonstrates that the amount of resources has
been added in time, even for applications with high demand for resources
5. CONCLUSIONS
Actually the cloud-based applications are varied and erratic, difficult to forecast. The results in
the paper show that when the application changes, the resource consumption characteristic will
change, which makes the determination of the overload situation inaccurate. If the overload is
incorrect, the auto scaling process does not guarantee SLA and resource wastage. Auto scaling
solutions that use low-level metrics if only using the threshold of CPU resources will have the
risk of over-forecasting the wrong load. Therefore, the combination of resources to predict
overload in auto scaling solutions is necessary. The combination of both the CPU and Memory
thresholds proposed in this paper has reduced the cost of resource consumption and shortened the
response time to a request. This achievement is due to the proposed solution that mitigates the
effects of the change in resource consumption characteristics of the application.
REFERENCES
[1] Nilabja Roy, Abhishek Dubey, Aniruddha Gokhale (2011), Efﬁcient Autoscaling in the Cloud using
Predictive Models for Workload Forecasting, The IEEE 4th International Conference on Cloud
Computing.
[2] Anshuman Biswas, Shikharesh Majumdar, Biswajit Nandy, Ali El-Haraki (2017), A hybrid auto-
scaling technique for clouds processing applications with service level agreements, Journal of Cloud
Computing: Advances, Systems and Applications.
[3] A. Ali-Eldin, J. Tordsson, and E. Elmroth (2012), An Adaptive Hybrid Elasticity Controller for Cloud
Infrastructures. In IEEE NOMS
[4] H. Fernandez, G. Pierre, and T. Kielmann (2014), Autoscaling Web Applications in Heterogeneous
Cloud Infrastructures. In IEEE IC2E, 2014.
[5] Amazon.(2018).Amazon Auto Scaling Service. Retrieved from http://aws.amazon.com/autoscaling/
[6] J. Huang, C. Li, and J. Yu. (2012) Resource prediction based on double exponential
smoothing in cloud computing. In Intl Conf on Consumer Electronics, Communications and
Networks, pages 2056–2060
[7] Mohammad Sadegh Aslanpour, Mostafa Ghobaei-Arani, Adel Nadjaran Toosi (2017),Auto-scaling
Web Applications in Clouds: A Cost-Aware Approach, Journal of Network and Computer
Applications, Vol. 95, Pages 26-41.
[8] Pooja, Asmita Pandey (2014), Impact of Memory Intensive Applications on Performance of Cloud
Virtual Machine, Proceedings of 2014 RAECS UIET Panjab University Chandigarh
[9] Hamid Arabnejad, Claus Pahl, Pooyan Jamshidi and Giovani Estrada (2017), A Comparison of
Reinforcement Learning Techniques for Fuzzy Cloud Auto-Scaling, The 17th IEEE/ACM
International Symposium on Cluster, Cloud and Grid Computing (CCGRID)
[10] Alexey Ilyushkin, Ahmed Ali-Eldin, Nikolas Herbst, Alessandro V. Papadopoulos, Bogdan Ghit,
Dick Epema, Alexandru Iosup (2017), An Experimental Performance Evaluation of Autoscaling
Policies for Complex Workﬂows, Proceedings of the 8th ACM/SPEC on International Conference on
Performance Engineering (ICPE )

9
[11] Chenhao Qu, Rodrigo N. Calheiros, Rajkumar Buyya (2018), Auto-Scaling Web Applications in
Clouds: A Taxonomy and Survey, ACM Computing Surveys, Vol. 51, No. 4, Article 73.
[12] Nguyen Hong Son, Nguyen Khac Chien (2017), Load Balancing in Auto Scaling-Enabled Cloud
Environments, International Journal on Cloud Computing: Services and Architecture (IJCCSA), Vol.
7, No. 5, pp 15-22.
AUTHORS
Ha Van Cu,received his master degree in Computer Science from Hong Bang International University in
HCM City in 2014. Heis doing a PhD candidate in Computer Science. His research interests include cloud
computing,information security and data science.
Nguyen Hong Son, received his B.Sc. in Computer Engineering from the University of Technology in
HCM city, his M.Sc. and PhD in Communication Engineering from the Post and Telecommunication
Institute of Technology Hanoi. His current research interests include communication engineering, network
security, computer engineering and cloud computing.

THE EFFECT OF THE RESOURCE CONSUMPTION CHARACTERISTICS OF CLOUD APPLICATIONS ON THE EFFICIENCY OF LOW-METRIC AUTO SCALING SOLUTIONS

More Related Content

What's hot (17)

Similar to THE EFFECT OF THE RESOURCE CONSUMPTION CHARACTERISTICS OF CLOUD APPLICATIONS ON THE EFFICIENCY OF LOW-METRIC AUTO SCALING SOLUTIONS (20)

Recently uploaded (20)

THE EFFECT OF THE RESOURCE CONSUMPTION CHARACTERISTICS OF CLOUD APPLICATIONS ON THE EFFICIENCY OF LOW-METRIC AUTO SCALING SOLUTIONS